diff --git a/.gitattributes b/.gitattributes index c225f8480af7f18798fb3c10302fd31c6e7f0b83..331af9aa009de0a80ed836cace4e52f791e43014 100644 --- a/.gitattributes +++ b/.gitattributes @@ -1428,3 +1428,19 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text 2024/Modeling[[:space:]]Label[[:space:]]Correlations[[:space:]]with[[:space:]]Latent[[:space:]]Context[[:space:]]for[[:space:]]Multi-Label[[:space:]]Recognition/db360f7c-5171-4a2a-8147-499c0fbf9014_origin.pdf filter=lfs diff=lfs merge=lfs -text 2024/Modeling[[:space:]]and[[:space:]]Driving[[:space:]]Human[[:space:]]Body[[:space:]]Soundfields[[:space:]]through[[:space:]]Acoustic[[:space:]]Primitives/3e3fdfc7-9a44-402d-ae28-13ceca875c26_origin.pdf filter=lfs diff=lfs merge=lfs -text 2024/Modelling[[:space:]]Competitive[[:space:]]Behaviors[[:space:]]in[[:space:]]Autonomous[[:space:]]Driving[[:space:]]Under[[:space:]]Generative[[:space:]]World[[:space:]]Model/ed2fa938-68fa-4d43-85a5-df3d24fff9a3_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Momentum[[:space:]]Auxiliary[[:space:]]Network[[:space:]]for[[:space:]]Supervised[[:space:]]Local[[:space:]]Learning/eb53acf2-10ec-44f2-923d-fc539ed31362_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Mono-ViFI_[[:space:]]A[[:space:]]Unified[[:space:]]Learning[[:space:]]Framework[[:space:]]for[[:space:]]Self-supervised[[:space:]]Single-[[:space:]]and[[:space:]]Multi-frame[[:space:]]Monocular[[:space:]]Depth[[:space:]]Estimation/47dcf543-78f7-47b3-955c-778904eede62_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/MonoTTA_[[:space:]]Fully[[:space:]]Test-Time[[:space:]]Adaptation[[:space:]]for[[:space:]]Monocular[[:space:]]3D[[:space:]]Object[[:space:]]Detection/cfa6dde5-bde3-4186-b95b-8bbf6c77c339_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/MonoWAD_[[:space:]]Weather-Adaptive[[:space:]]Diffusion[[:space:]]Model[[:space:]]for[[:space:]]Robust[[:space:]]Monocular[[:space:]]3D[[:space:]]Object[[:space:]]Detection/4d73977a-6d19-47be-84e6-b3aec99a372e_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Monocular[[:space:]]Occupancy[[:space:]]Prediction[[:space:]]for[[:space:]]Scalable[[:space:]]Indoor[[:space:]]Scenes/325d434b-3564-4d20-a82e-4ec4e345eb58_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Motion[[:space:]]Aware[[:space:]]Event[[:space:]]Representation-driven[[:space:]]Image[[:space:]]Deblurring/a9105a02-b537-456a-9778-80b745c423ab_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Motion[[:space:]]Keyframe[[:space:]]Interpolation[[:space:]]for[[:space:]]Any[[:space:]]Human[[:space:]]Skeleton[[:space:]]using[[:space:]]Point[[:space:]]Cloud-based[[:space:]]Human[[:space:]]Motion[[:space:]]Data[[:space:]]Homogenisation/391ac41a-5023-4e3a-b209-fdaecfa62b6a_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Motion[[:space:]]Mamba_[[:space:]]Efficient[[:space:]]and[[:space:]]Long[[:space:]]Sequence[[:space:]]Motion[[:space:]]Generation/1b586349-9221-4bad-a56d-e274ec79c852_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Motion[[:space:]]and[[:space:]]Structure[[:space:]]from[[:space:]]Event-based[[:space:]]Normal[[:space:]]Flow/2dddd958-d905-447a-8afe-98428ea7b185_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Motion-Guided[[:space:]]Latent[[:space:]]Diffusion[[:space:]]for[[:space:]]Temporally[[:space:]]Consistent[[:space:]]Real-world[[:space:]]Video[[:space:]]Super-resolution/dbea7f31-cd1b-433d-b163-70053b8d34db_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Motion-Oriented[[:space:]]Compositional[[:space:]]Neural[[:space:]]Radiance[[:space:]]Fields[[:space:]]for[[:space:]]Monocular[[:space:]]Dynamic[[:space:]]Human[[:space:]]Modeling/6e1b3c80-b2f1-4690-8b23-807237bf60da_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Motion-prior[[:space:]]Contrast[[:space:]]Maximization[[:space:]]for[[:space:]]Dense[[:space:]]Continuous-Time[[:space:]]Motion[[:space:]]Estimation/34d0d1c3-d92d-4348-bbb4-52308320fede_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/MotionChain_[[:space:]]Conversational[[:space:]]Motion[[:space:]]Controllers[[:space:]]via[[:space:]]Multimodal[[:space:]]Prompts/24d4fe30-81b0-462c-a698-38fed56b9fc4_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/MotionDirector_[[:space:]]Motion[[:space:]]Customization[[:space:]]of[[:space:]]Text-to-Video[[:space:]]Diffusion[[:space:]]Models/8325ffaa-f0c0-4f28-929f-63aa2f3d7c76_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/MotionLCM_[[:space:]]Real-time[[:space:]]Controllable[[:space:]]Motion[[:space:]]Generation[[:space:]]via[[:space:]]Latent[[:space:]]Consistency[[:space:]]Model/69320bc9-c9fe-4fc1-8e78-3d61fd4c6087_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Multi-Granularity[[:space:]]Sparse[[:space:]]Relationship[[:space:]]Matrix[[:space:]]Prediction[[:space:]]Network[[:space:]]for[[:space:]]End-to-End[[:space:]]Scene[[:space:]]Graph[[:space:]]Generation/6b85535a-3ad0-4a54-882a-f4d4244fbf81_origin.pdf filter=lfs diff=lfs merge=lfs -text diff --git a/2024/Momentum Auxiliary Network for Supervised Local Learning/eb53acf2-10ec-44f2-923d-fc539ed31362_content_list.json b/2024/Momentum Auxiliary Network for Supervised Local Learning/eb53acf2-10ec-44f2-923d-fc539ed31362_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..60c7c6803c000675ab102d50c0b39a41887eaf20 --- /dev/null +++ b/2024/Momentum Auxiliary Network for Supervised Local Learning/eb53acf2-10ec-44f2-923d-fc539ed31362_content_list.json @@ -0,0 +1,1733 @@ +[ + { + "type": "text", + "text": "Momentum Auxiliary Network for Supervised Local Learning", + "text_level": 1, + "bbox": [ + 238, + 140, + 763, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Junhao Su $^{1*}$ , Changpeng Cai $^{1*}$ , Feiyu Zhu $^{2*}$ , Chenghao He $^{4}$", + "bbox": [ + 282, + 215, + 720, + 229 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xiaojie $\\mathrm{Xu}^5$ , Dongzhi Guan $^{1\\boxtimes}$ , and Chenyang Si $^{3\\boxtimes}$", + "bbox": [ + 316, + 229, + 684, + 247 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Southeast University", + "bbox": [ + 423, + 257, + 580, + 272 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 University of Shanghai for Science and Technology", + "bbox": [ + 323, + 272, + 676, + 287 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 Nanyang Technological University", + "bbox": [ + 377, + 286, + 622, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{4}$ East China University of Science and Technology", + "bbox": [ + 328, + 300, + 671, + 314 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "5 The Hong Kong University of Science and Technology", + "bbox": [ + 313, + 314, + 689, + 328 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Deep neural networks conventionally employ end-to-end backpropagation for their training process, which lacks biological credibility and triggers a locking dilemma during network parameter updates, leading to significant GPU memory use. Supervised local learning, which segments the network into multiple local blocks updated by independent auxiliary networks. However, these methods cannot replace end-to-end training due to lower accuracy, as gradients only propagate within their local block, creating a lack of information exchange between blocks. To address this issue and establish information transfer across blocks, we propose a Momentum Auxiliary Network (MAN) that establishes a dynamic interaction mechanism. The MAN leverages an exponential moving average (EMA) of the parameters from adjacent local blocks to enhance information flow. This auxiliary network, updated through EMA, helps bridge the informational gap between blocks. Nevertheless, we observe that directly applying EMA parameters has certain limitations due to feature discrepancies among local blocks. To overcome this, we introduce learnable biases, further boosting performance. We have validated our method on four image classification datasets (CIFAR-10, STL-10, SVHN, ImageNet), attaining superior performance and substantial memory savings. Notably, our method can reduce GPU memory usage by more than $45\\%$ on the ImageNet dataset compared to end-to-end training, while achieving higher performance. The Momentum Auxiliary Network thus offers a new perspective for supervised local learning. Our code is available at: https://github.com/JunhaoSuO/MAN.", + "bbox": [ + 261, + 397, + 743, + 731 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Local Learning $\\cdot$ Image Classification $\\cdot$ Momentum Auxiliary Network", + "bbox": [ + 259, + 746, + 740, + 776 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal Contribution.", + "bbox": [ + 230, + 797, + 375, + 811 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "Corresponding Authors: Chengyang Si (chenyang.si.mail@gmail.com) and Dongzhi", + "bbox": [ + 230, + 811, + 785, + 825 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "Guan (guandongzhi@seu.edu.cn).", + "bbox": [ + 230, + 825, + 455, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/59658765abdb05b2d54ec752161d340c1c89205054a19306ed2110965081b8ad.jpg", + "image_caption": [ + "Fig. 1: Comparison between different methods with MAN and the original methods in terms of accuracy. Results are obtained using ResNet-110 (K=55) on the various datasets." + ], + "image_footnote": [], + "bbox": [ + 217, + 145, + 787, + 328 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 217, + 411, + 375, + 428 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In deep learning, the prevalent use of end-to-end backpropagation [35] is essential for training complex neural networks, which requires heavy computational work for loss evaluation and successive gradient descent through network layers to refine parameters [17, 24, 29]. This method starkly contrasts with the local signal processing of biological synapses [6, 7, 40] and imposes a 'locking' constraint [19] that delays parameter updates until the entire forward and backward pass is complete. Such constraints exacerbate challenges like reduced parallelism and increased GPU memory usage [28, 38], compromising training efficiency and scalability. An emerging alternative to traditional end-to-end (E2E) training is Local Learning [2, 3, 5, 18, 19, 29, 32, 37, 41], which promises to mitigate the drawbacks of traditional methods. These methods divide the network into discrete, gradient-isolated blocks, each of which is updated by its dedicated auxiliary network according to distinct local objectives [3, 32, 37]. This training strategy enables immediate parameter updates upon local error reception, circumventing the sequential update bottleneck of E2E training, thereby significantly enhancing the efficiency of parallel training [13, 22]. Moreover, local learning reduces the demand on GPU memory by only retaining gradients for the local and auxiliary networks, thus eliminating the need to store extensive global gradient information and saving computational resources [28, 38]. However, even with these advantages, there is still a significant performance gap between local learning and traditional E2E methods, especially as the network is divided into more local blocks. Current local learning techniques mainly focus on improving the design of auxiliary network structures [3] and making local loss functions better to close this performance gap [32, 37]. Yet, these improvements do not completely address the inherent short-sightedness of local learning: the separation into blocks can make each part of the network only focus on its local objectives, possibly ig-", + "bbox": [ + 212, + 446, + 787, + 843 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Junhao Su et al.", + "bbox": [ + 271, + 114, + 382, + 126 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "noring the overall objectives of the network. This can lead to the discarding of globally beneficial information due to the lack of inter-block communication.", + "bbox": [ + 212, + 146, + 782, + 175 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this paper, we introduce the Momentum Auxiliary Network (MAN), a novel network architecture designed to mitigate the inherent limitations of supervised local learning by enhancing inter-block communication. Specifically, MAN not only accepts the current local block as input but also absorbs the parameters of its next block. Upon completion of a forward pass, parameter updates are performed using local gradients refined through Exponential Moving Average (EMA) techniques [14]. This novel approach enables each local block to integrate information from the following block, thereby extending the operational perspective of each block beyond its immediate objectives and further aligning with the global objective of the network. This addresses the critical limitations in local learning frameworks, ensuring a more cohesive alignment with the overarching network objectives. Furthermore, we identify that directly applying EMA parameters is constrained by the feature incongruities across gradient-isolated blocks. To mitigate this, we implement additional learnable bias terms to each auxiliary block, enhancing their ability to share information effectively. The proposed MAN requires only a minimal increase in memory use but significantly improves performance by enhancing the information sharing ability of local blocks. The efficacy of the MAN approach is validated on a suite of benchmark image classification datasets, including CIFAR-10 [21], STL-10 [9], SVHN [30], and ImageNet [12]. The results demonstrate the effectiveness of our method in surpassing the limitations of traditional supervised local learning.", + "bbox": [ + 212, + 176, + 787, + 493 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The contributions of this paper could be summarized as follows:", + "bbox": [ + 238, + 494, + 700, + 508 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose Momentum Auxiliary Network (MAN), designed to facilitate inter-block communication with Exponential Moving Average (EMA), mitigating the short-sightedness issue in conventional supervised local learning techniques and culminating in an elevated overall performance of the network.", + "- MAN is a versatile, plug-and-play approach that can seamlessly integrate with any supervised local learning technique, markedly amplifying their efficacy while requiring only a negligible increase in memory usage.", + "- MAN has demonstrated its effectiveness through experiments on benchmark image classification datasets, achieving state-of-the-art performance. Notably, on the ImageNet [12] dataset, it outperforms E2E training while using significantly less memory." + ], + "bbox": [ + 225, + 518, + 782, + 700 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 736, + 387, + 752 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Local Learning", + "text_level": 1, + "bbox": [ + 215, + 768, + 385, + 784 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Local learning is first proposed as an innovative deep learning algorithm with the intention of utilizing memory more efficiently and adhering more closely to principles of biological plausibility [10]. This approach emerges as a response to", + "bbox": [ + 212, + 794, + 784, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Momentum Auxiliary Network for Supervised Local Learning", + "bbox": [ + 321, + 114, + 732, + 130 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/21efcd67c2d31a85b36be6d1f25ab9a282f7e5aedfdad7e1deda6dcee912e8b2.jpg", + "image_caption": [ + "Fig. 2: Comparison of (a) end-to-end backpropagation, (b) other supervised local learning methods, and (c) our proposed method. Unlike E2E, supervised local learning separates the network into K gradient-isolated local blocks." + ], + "image_footnote": [], + "bbox": [ + 218, + 142, + 789, + 292 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the limitations presented by global E2E training [16], prompting the development of alternative supervised local learning methodologies. Examples of these methodologies include a differentiable search algorithm, which decouples network blocks for block-level learning and selects a manually assisted network for each local block [33], as well as a self-supervised contrasting loss function that leverages its local learning rules [18, 41]. In supervised local learning, training primarily relies on the design of reasonable supervised local loss functions or the manual construction of auxiliary networks.", + "bbox": [ + 212, + 388, + 787, + 511 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Supervised Local Loss: InfoPro [37] presents a method that involves the reconstruction loss for each hidden layer, coupled with cross-entropy loss. This approach forces the hidden layers to preserve vital input information, thereby preventing the early layers from discarding information relevant to the task. PredSim [32] employs layer-wise loss functions for network training. It makes use of two distinct supervised loss functions to generate local error signals for hidden layers. The optimization of these losses takes place during supervised local learning, thus eliminating the necessity to propagate global errors back to the hidden layers. These methods have achieved advanced performance in the field of supervised local learning, but there is still a significant gap compared to the performance of E2E training.", + "bbox": [ + 212, + 516, + 787, + 683 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Auxiliary Networks: InfoPro [37] incorporates two auxiliary networks into its approach. One network consists of a single convolutional layer followed by two fully-connected layers, which are employed for the cross-entropy loss. The other network utilizes two convolutional layers with upsampling for the reconstruction loss. Alternatively, DGL [3] designs a structure that uses three consecutive convolutional layers connected to a global pooling layer, followed by three consecutive fully connected layers. Moreover, DGL restricts the parameter of the auxiliary network to only $5\\%$ of the main network capacity, offering advantages in terms of speed and memory savings. However, these method still perform poorly when the network is divided into a large number of local blocks.", + "bbox": [ + 212, + 689, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Junhao Su et al.", + "bbox": [ + 271, + 114, + 382, + 126 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Alternative Learning Rules to E2E Training", + "text_level": 1, + "bbox": [ + 214, + 146, + 625, + 162 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Due to certain limitations inherent in E2E training, the pursuit of alternative methods to E2E training has gradually garnered attention in recent years [27]. Several efforts have been dedicated to addressing the biologically implausible aspects of E2E training, such as the training methodology of target propagation [4, 23, 25] attempts to train a specialized backward connection by utilizing local reconstruction targets. Additionally, some recent studies have attempted to completely avoid backpropagation in neural networks through forward gradient learning [11, 34]. Meanwhile, the weight transport problem [10]. This has been tackled either by employing distinct feedback connections [1, 26] or by directly disseminating global errors to each hidden unit [8, 31]. While these methods have somewhat alleviated the biological implausibility of E2E training, they still rely on global objectives, which fundamentally differ from biological neural networks that rely on local synapses for information transmission. Furthermore, these methods currently struggle to be effective on large datasets [12].", + "bbox": [ + 212, + 171, + 787, + 383 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 406, + 330, + 422 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Preliminaries", + "text_level": 1, + "bbox": [ + 214, + 436, + 372, + 452 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To set the stage, we first provide a brief overview of traditional end-to-end supervised learning and backpropagation mechanisms. We denote a data sample as $x$ and its corresponding ground-truth label as $y$ . The entire deep network is fragmented into several local blocks. During the forward propagation process, the output from the j-th block serves as the input for the (j+1)-th block, expressed as $x_{j+1} = f_{\\theta_j}(x_j)$ . Here, $\\theta_j$ symbolizes the parameters of the j-th local block and $f(\\cdot)$ represents the forward calculation of the block. We evaluate the loss function $\\mathcal{L}(\\hat{y}, y)$ between the output of the last block and the ground truth label, and propagate it back iteratively to the preceding blocks.", + "bbox": [ + 212, + 462, + 787, + 598 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Supervised local learning strategies [3, 32, 37] integrate auxiliary networks for local supervision. For each local block, an auxiliary network is affixed. The output from a local block is fed to its corresponding auxiliary network, generating the local supervisory signal as $\\hat{y}_j = g_{\\gamma_j}(x_{j + 1})$ . Here, $\\gamma_{j}$ denotes the parameters of the j-th auxiliary network.", + "bbox": [ + 212, + 598, + 787, + 672 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this setup, we update the parameters of the j-th auxiliary network and local block, $\\gamma_{j},\\theta_{j}$ , as follows:", + "bbox": [ + 212, + 674, + 787, + 705 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\gamma_ {j} \\leftarrow \\gamma_ {j} - \\eta_ {a} \\times \\nabla_ {\\gamma_ {j}} \\mathcal {L} (\\hat {y} _ {j}, y) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 403, + 720, + 784, + 736 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\theta_ {j} \\leftarrow \\theta_ {j} - \\eta_ {l} \\times \\nabla_ {\\theta_ {j}} \\mathcal {L} \\left(y _ {\\hat {j}}, y\\right) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 406, + 757, + 784, + 773 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\eta_{a},\\eta_{l}$ are the learning rates of the auxiliary networks and local blocks, respectively. By attaching auxiliary networks, each local block becomes gradient-isolated and can be updated with local supervision rather than global backpropagation.", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Momentum Auxiliary Network for Supervised Local Learning", + "bbox": [ + 321, + 114, + 732, + 130 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Momentum Auxiliary Network", + "text_level": 1, + "bbox": [ + 215, + 146, + 517, + 162 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Existing techniques incorporate supervision signals into individual local blocks, which allows for parallel parameter updates and reduces memory overhead. However, this approach can lead to a short-sightedness problem, where each local block neglects the information from subsequent blocks, ultimately resulting in suboptimal final accuracy.", + "bbox": [ + 212, + 172, + 787, + 247 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To address this issue, we propose a comprehensive information exchange module—the Momentum Auxiliary Network (MAN). The MAN employs the Exponential Moving Average (EMA) mechanism [14] as a conduit for transferring information from subsequent blocks to the current block. In the context of MAN, we update the parameters of the j-th auxiliary network and local block as follows:", + "bbox": [ + 212, + 248, + 787, + 324 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\gamma_ {j} \\leftarrow \\gamma_ {j} - \\eta_ {a} \\times \\nabla_ {\\gamma_ {j}} \\mathcal {L} \\left(\\hat {y} _ {j}, y\\right) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 403, + 338, + 785, + 356 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\gamma_ {j} \\leftarrow E M A \\left(\\gamma_ {j}, \\theta_ {j + 1}\\right) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 424, + 377, + 785, + 393 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\theta_ {j} \\leftarrow \\theta_ {j} - \\eta_ {l} \\times \\nabla_ {\\theta_ {j}} \\mathcal {L} (\\hat {y} _ {j}, y) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 403, + 415, + 785, + 431 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\gamma_{j}$ represents the parameters of the j-th auxiliary network and $\\theta_{j}$ represents those of the j-th local block. After updating with local gradients, $\\gamma_{j}$ undergoes further refinement by incorporating the parameters of the subsequent local block via the EMA, which is a weighted sum operation.", + "bbox": [ + 212, + 438, + 787, + 500 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "However, through experimental results, it becomes apparent that directly applying EMA parameters to update the auxiliary network has its limitations, and it provides limited improvements to each local block. Upon analyzing, we find that the features learned between each local block vary to a certain degree, which hampers the effectiveness of the EMA update method. As a consequence, a learnable bias is introduced to augment the learning capability of the hidden layers within the local block, while also compensating for the deficiencies of the EMA update method.", + "bbox": [ + 212, + 500, + 486, + 739 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3b716f4cdbd8f877ca15468385def71a2b8632344813761894cbd88dfa245ebb.jpg", + "image_caption": [ + "Fig. 3: Details of the Momentum Auxiliary Network. Local $(\\mathrm{i} + 1)$ represents the $(\\mathrm{i} + 1)$ -th gradient-isolated local block, which contains layers from layer m to layer $(\\mathrm{m} + \\mathrm{n})$ , totaling n+1 layers $(\\mathrm{n} \\geqslant 0)$ . We only use the parameters of the first layer to ensure a balance in GPU memory usage." + ], + "image_footnote": [], + "bbox": [ + 498, + 527, + 784, + 632 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Therefore, the parameter update method for the j-th auxiliary network can be written as:", + "bbox": [ + 212, + 739, + 787, + 770 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\left(\\gamma_ {j}, b _ {j}\\right) \\leftarrow \\left(\\gamma_ {j}, b _ {j}\\right) - \\eta_ {a} \\times \\nabla_ {\\left(\\gamma_ {j}, b _ {j}\\right)} \\mathcal {L} \\left(\\hat {y} _ {j}, y\\right) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 354, + 785, + 785, + 805 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\gamma_ {j} \\leftarrow E M A \\left(\\gamma_ {j}, \\theta_ {j + 1}\\right) \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 423, + 824, + 785, + 842 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Junhao Su et al.", + "bbox": [ + 271, + 114, + 383, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $(\\gamma_j, b_j)$ represent the parameters of the j-th auxiliary network. During this process, we update the parameters $\\gamma_j$ of the j-th auxiliary network jointly through the learnable bias $b_j$ and EMA.", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Essentially, the proposed MAN facilitates communication between subsequent blocks through the EMA mechanism [14], effectively resolving the short-sightedness problem inherent in traditional supervised local learning methods. Concurrently, it introduces an independent learnable bias to mitigate information discrepancies caused by feature variations among different local blocks, thereby aligning the output of each block closer to the global target.", + "bbox": [ + 212, + 191, + 782, + 281 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In addition to the two innovative methods mentioned above, MAN demonstrates strong versatility. It can be seamlessly integrated into existing supervised local learning methods, and it excels across various datasets, reflecting its flexibility and generalizability.", + "bbox": [ + 212, + 282, + 782, + 343 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 214, + 364, + 375, + 382 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1 Experimental Setup", + "text_level": 1, + "bbox": [ + 214, + 393, + 428, + 410 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conduct experiments using four widely adopted datasets: CIFAR-10 [21], SVHN [30], STL-10 [9], and ImageNet [12], with ResNets [15] of varying depths serving as the network architectures.", + "bbox": [ + 212, + 417, + 782, + 462 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Three state-of-the-art supervised local learning methods are selected for comparison: PredSim [32], DGL [3], and InfoPro [37]. We then partition the network into K local blocks, each containing an approximately equal number of layers. Our proposed Momentum Auxiliary Network is incorporated only into the first K-1 local blocks. The K-th local block does not employ an auxiliary network and is directly connected to the output classifier. We compare these configurations against traditional E2E and original supervised local learning methods to maintain consistent training settings and eliminate confounding variables.", + "bbox": [ + 212, + 463, + 784, + 584 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2 Implement Details", + "text_level": 1, + "bbox": [ + 214, + 604, + 415, + 619 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In our experiments on CIFAR-10 [21], SVHN [30], and STL-10 [9] datasets with ResNet-32 [15] and ResNet-110 [15], we utilize the SGD optimizer with Nesterov momentum set at 0.9 and an L2 weight decay factor of 1e-4. We employ batch sizes of 1024 for CIFAR-10 and SVHN and 128 for STL-10. The training duration spans 400 epochs, starting with initial learning rates of 0.8 for CIFAR-10 / SVHN and 0.1 for STL-10, following a cosine annealing scheduler [9].", + "bbox": [ + 212, + 628, + 782, + 719 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In our experiments with ImageNet [12], we adopt different training settings for various architectures. We train VGG13 [36] for 90 epochs with an initial learning rate of 0.025. For ResNet-101 [15], ResNet-152 [15], and ResNeXt-101,32 $\\times$ 8d [39], we train them for 90 epochs as well, with initial learning rates of 0.05, 0.05, and 0.025, respectively. We set the batch sizes to 64 for VGG13, 128 for ResNet-101 and ResNet-152, and 64 for ResNeXt-101,32 $\\times$ 8d. We maintain consistency with other training configurations as previously described for CIFAR-10 [21].", + "bbox": [ + 212, + 719, + 784, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Momentum Auxiliary Network for Supervised Local Learning", + "bbox": [ + 321, + 114, + 732, + 130 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Based on our multiple experimental results, we select the hyperparameter momentum value as 0.995, due to its consistently stable and superior performance. We will provide the experimental results of other momentum hyperparameters in the supplementary materials.", + "bbox": [ + 212, + 146, + 787, + 207 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3 Results on Image Classification Datasets", + "text_level": 1, + "bbox": [ + 215, + 229, + 598, + 244 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Results on Image Classification Benchmarks: We start by assessing the accuracy performance of our approach using the CIFAR-10 [21], SVHN [30], and STL-10 [9] datasets. We employ ResNet-32 [15], partitioned into 8 and 16 local blocks, and ResNet-110 [15], divided into 32 and 55 local blocks. As illustrated in Table 1, our MAN significantly bolsters the accuracy of all methods.", + "bbox": [ + 212, + 253, + 787, + 330 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/4f3d82f689a3e522b8bb8528ab4cf6588240a70062adc67c2439434322184a52.jpg", + "table_caption": [ + "Table 1: Performance of different networks with varying numbers of local blocks. The average test error is obtained by 5 experiments. The * means addition of our MAN." + ], + "table_footnote": [], + "table_body": "
DatasetMethodResNet-32ResNet-110
K = 8 (Test Error)K = 16 (Test Error)K = 32 (Test Error)K = 55 (Test Error)
CIFAR-10(E2E(ResNet-32)=6.37,E2E(ResNet-110)=5.42)PredSim [32]20.6222.7122.0824.74
PredSim*14.29(↓6.33)15.58(↓7.13)17.05(↓5.03)18.02(↓6.72)
DGL [3]11.6314.0812.5114.45
DGL*8.42(↓3.21)9.11(↓4.97)9.65(↓2.86)9.73(↓4.72)
InfoPro [37]11.5112.9312.2613.22
InfoPro*9.32(↓2.19)9.65(↓3.28)9.06(↓3.20)9.77(↓3.45)
STL-10(E2E(ResNet-32)=19.35,E2E(ResNet-110)=19.67)PredSim [32]31.9732.9032.0533.27
PredSim*29.97(↓2.00)29.99(↓2.91)30.48(↓1.57)31.55(↓1.72)
DGL [3]25.0527.1425.6728.16
DGL*20.74(↓4.31)21.37(↓5.77)22.54(↓3.13)22.69(↓5.47)
InfoPro [37]27.3229.2828.5829.20
InfoPro*23.17(↓4.15)23.54(↓5.74)24.08(↓4.50)24.74(↓4.46)
SVHN(E2E(ResNet-32)=2.99,E2E(ResNet-110)=2.92)PredSim [32]6.918.089.1210.47
PredSim*5.54(↓1.37)6.39(↓1.69)7.27(↓1.85)8.24(↓2.23)
DGL [3]4.835.055.125.36
DGL*3.80(↓1.03)4.04(↓1.01)4.08(↓1.04)4.52(↓0.84)
InfoPro [37]5.615.975.896.11
InfoPro*4.49(↓1.12)5.19(↓0.78)4.85(↓1.04)4.99(↓1.12)
", + "bbox": [ + 220, + 398, + 787, + 597 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "On the CIFAR-10 dataset [21], our method exhibits considerable improvements in diminishing test errors across various methods. In the relatively shallower network of ResNet-32 (K=16), where individual layers function as gradient-isolated local blocks, we record a reduction in test errors for PredSim [32], DGL [3], and InfoPro [37], from 22.71, 14.08, and 12.93 to 15.58, 9.11, and 9.65 respectively. This translates to a performance enhancement exceeding $25\\%$ for all methods. Even though the performance across all methods in the comparatively deeper network, ResNet-110 (K=55), is somewhat inferior due to the inherent need for more global information in such networks, our method still delivers exceptional performance. It achieves approximately a $20\\%$ improvement across all methods, underscoring the robust effectiveness of MAN in deeper networks.", + "bbox": [ + 212, + 628, + 787, + 794 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "When applied to other datasets, MAN can also reduce the test error of Pred-Sim [32], DGL [3], and InfoPro [37] by at least $5\\%$ , $12\\%$ , and $16\\%$ on the STL-10 [9] dataset. On the SVHN [30] dataset, our improvements over the three", + "bbox": [ + 212, + 795, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Junhao Su et al.", + "bbox": [ + 271, + 114, + 382, + 127 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "methods also surpass $20\\%$ , $21\\%$ , and $13\\%$ . As can be seen, the improvement our MAN introduces to all methods is quite remarkable—comparable even to the accuracy of E2E training—and it significantly mitigates the overwhelming performance issue that has continually plagued supervised local learning.", + "bbox": [ + 212, + 146, + 782, + 207 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Results on ImageNet: We further validate the effectiveness of our approach on ImageNet [12] using four networks of varying depths (ResNets [15] and VGG13 [36]). As depicted in Table 2, when we employ VGG13 as the backbone and divide the network into 10 blocks, DGL [3] achieves merely a Top1-Error of 35.60 and a Top5-Error of 14.2, representing a substantial gap when compared to the E2E method. However, with the introduction of our MAN, the Top1-Error reduces by 3.61 points, and the Top5-Error decreases by 3.36 points for DGL. This significant enhancement brings the performance closer to the E2E method.", + "bbox": [ + 212, + 209, + 784, + 330 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "As illustrated in Table 2, when we use ResNet-101 [15], ResNet-152 [15], ResNeXt-101, $32 \\times 8\\mathrm{d}$ [39] as backbones and divide the network into four blocks, the performance of InfoPro [37] is already below that of E2E. After incorporating our MAN, the Top-1 Error of these three backbone networks can be reduced by approximately $6\\%$ compared to the original, surpassing the performance of E2E training. These results underscore the effectiveness of our MAN on the large-scale ImageNet [12] dataset, even when using deeper networks.", + "bbox": [ + 212, + 332, + 785, + 439 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/580e507a3e07a6eb64b7d026a5fde080f06d3b55364e2ca96f02c81de23ec32c.jpg", + "table_caption": [ + "Table 2: Results on the validation set of ImageNet" + ], + "table_footnote": [], + "table_body": "
NetworkMethodTop1-ErrorTop5-Error
ResNet-101E2E22.035.93
InfoPro(K=2) [37]21.855.89
InfoPro*(K=2)21.65(↓0.20)5.49(↓0.40)
InfoPro(K=4) [37]22.816.54
InfoPro*(K=4)21.73(↓1.08)5.81(↓0.73)
ResNet-152E2E21.605.92
InfoPro(K=2) [37]21.455.84
InfoPro*(K=2)21.23(↓0.22)5.53(↓0.31)
InfoPro(K=4) [37]22.936.71
InfoPro*(K=4)21.59(↓1.34)5.89(↓0.82)
ResNeXt-101, 32 × 8dE2E20.645.40
InfoPro(K=2) [37]20.355.28
InfoPro*(K=2)20.11(↓0.24)5.18(↓0.10)
InfoPro(K=4) [37]21.696.11
InfoPro*(K=4)20.37(↓1.32)5.34(↓0.77)
VGG13E2E28.419.63
DGL [3]35.6014.20
DGL*(K=10)31.99(↓3.61)10.84(↓3.36)
", + "bbox": [ + 233, + 496, + 767, + 696 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Training-Accuracy Curve Analysis: As depicted in the accuracy-epoch curve in Fig. 4, our Momentum Auxiliary Network consistently outperforms the original method in terms of accuracy throughout the entire training process. This underscores its reliability and stability during the training process of classification tasks. Moreover, our MAN achieves a higher accuracy earlier in the later stages of training and attains stability sooner, indicating a faster convergence rate—a critical attribute in large-scale and complex tasks.", + "bbox": [ + 212, + 734, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Momentum Auxiliary Network for Supervised Local Learning", + "bbox": [ + 321, + 114, + 732, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/b07c01938291e0a448ebb3858e0be9c8c8bb17e68a298b6cdb7c67ccbacb0085.jpg", + "image_caption": [ + "Fig. 4: Training-Accuracy curves, the left uses ResNet-32 (K=16) as the backbone, while the right uses ResNet-110 (K=55). Both are utilizing the CIFAR-10 dataset." + ], + "image_footnote": [], + "bbox": [ + 230, + 161, + 480, + 308 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/df3d2ab998231fd5d3874f9afe2e9b6e29d6c8f6cb317212026fe1b08f1dd61a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 511, + 161, + 759, + 306 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "In summary, our Momentum Auxiliary Network significantly enhances the accuracy of traditional supervised local learning methods by promoting information exchange between local blocks. The beneficial information exchange facilitated by Momentum Auxiliary Network significantly mitigates the pervasive issue of shortsightedness within the supervised local learning domain, thereby offering advantages in terms of training stability and convergence speed. This is crucial for training models more effectively and efficiently.", + "bbox": [ + 212, + 378, + 787, + 486 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/ec04e57e122a5123e26bcd488997ca24aaf363e36720f62d19a5682b3eb8f0ab.jpg", + "table_caption": [ + "Table 3: Comparison of GPU memory usage using InfoPro as baseline with different backbones on the ImageNet dataset." + ], + "table_footnote": [], + "table_body": "
MethodResNet-101\nGPU Memory(GB)ResNet-152\nGPU Memory(GB)ResNeXt-101, 32×8d\nGPU Memory(GB)
E2E19.7126.2919.22
InfoPro(K=2) [37]12.06(↓ 38.8%)15.53(↓ 40.9%)11.55(↓ 39.9%)
InfoPro*(K=2)12.32(↓ 37.5%)15.93(↓ 39.4%)11.73(↓ 38.9%)
InfoPro(K=4) [37]10.37(↓ 47.3%)13.48(↓ 48.7%)10.24(↓ 46.7%)
InfoPro*(K=4)10.69(↓ 45.8%)13.91(↓ 47.1%)10.49(↓ 45.5%)
", + "bbox": [ + 236, + 551, + 761, + 672 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Results on GPU memory requirement: Supervised local learning markedly conserves GPU memory by limiting gradient propagation to within local blocks, thereby ensuring that backpropagation transpires exclusively within each respective local block and its corresponding auxiliary network. This strategy significantly diminishes the storage demands for activation parameters and gradient information, which would typically disseminate across local blocks. Consequently, this facilitates substantial savings in GPU memory utilization. Our analysis, featuring a comprehensive comparison of GPU memory usage on the ImageNet dataset, illustrates that the implementation of a Momentum Auxiliary Network", + "bbox": [ + 212, + 703, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Junhao Su et al.", + "bbox": [ + 271, + 114, + 382, + 126 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/e2eb95355a30cf53f58dc887b70013841e734ecb8adc3080590c32728f2664ff.jpg", + "table_caption": [ + "Table 4: Abalation study of MAN. (a) Using DGL as baseline and ResNet-32 $(\\mathrm{K} = 16)$ as backbone on the CIFAR-10 dataset. (b) Using InfoPro as baseline and ResNet-101 $(\\mathrm{K} = 4)$ as backbone on the ImageNet dataset. LB stands for Learnable Bias." + ], + "table_footnote": [], + "table_body": "
EMALBTest Error
××14.08
✓×11.07(↓3.01)
✓✓9.11(↓4.97)
(a)
", + "bbox": [ + 217, + 196, + 390, + 272 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/cd15f42c2505970c794118f68718ff70719e972a1612ae1cbdaaf7ae424e2159.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
EMALBTop1-ErrorTop5-Error
××22.816.54
✓×22.09(↓ 0.72)6.07(↓ 0.47)
✓✓21.73(↓ 1.08)5.81(↓ 0.73)
(b)
", + "bbox": [ + 500, + 198, + 764, + 272 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "can elicit notable performance enhancements while incurring only a minimal uptick in GPU memory consumption.", + "bbox": [ + 212, + 301, + 782, + 332 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "As shown in Table 3, it indicates the substantial GPU memory reduction on the ImageNet [12]. When applying our method to InfoPro [37] on ResNet-101 [15], ResNet-152 [15], and ResNeXt-101, $32 \\times 8d$ [39], dividing the backbone into two local blocks results in a GPU memory saving of $37.5\\%$ to $39.4\\%$ . When the network is divided into four local blocks, the degree of GPU memory savings is further enhanced, exceeding $45\\%$ . Combined with Table 2, it can be seen that we achieve better performance with almost half the GPU memory required for end-to-end training. Further analysis of the memory usage comparison with the original method shows that our approach only increases the GPU memory by about $1\\%$ over the original method, while achieving a performance improvement of over $5\\%$ . These results highlight the excellent balance that the MAN method achieves in terms of GPU memory usage and performance enhancement.", + "bbox": [ + 212, + 333, + 784, + 513 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.4 Ablation Studies", + "text_level": 1, + "bbox": [ + 215, + 539, + 400, + 553 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We conduct an ablation study on the CIFAR-10 [21] dataset and ImageNet dataset to assess the impact of the EMA method [14] and learnable bias in the MAN on performance. For this analysis, we use ResNet-32 (K=16) [15] as the backbone and the original DGL [3] method as a comparison baseline.", + "bbox": [ + 212, + 566, + 784, + 626 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "As shown in Table 4(a), when we only use the EMA [14] in MAN to promote information exchange with subsequent blocks, without adding a learnable bias, the test error decreases from 14.08 to 11.07. When we further add a learnable bias, the test error decreases from 11.07 to 9.11, it is evident that both the EMA method and learnable bias contribute to performance improvement. We hypothesize that the EMA and learnable bias are somewhat complementary in terms of the features they learn. To validate this, we conduct feature visualization.", + "bbox": [ + 212, + 628, + 784, + 733 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "As depicted in Fig. 5(a), we can observe that the original method without the addition of MAN learns very limited and chaotic features. After adding the EMA method alone, the features it learns are clearly more characteristic and defined, indicating that it indeed receives some global features from the information in the subsequent block. However, there are still many blurry features, which may be due to the EMA parameter update method [14], causing information imbalance due to feature differences between different local blocks. When we", + "bbox": [ + 212, + 734, + 785, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Momentum Auxiliary Network for Supervised Local Learning", + "bbox": [ + 321, + 114, + 732, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "add learnable bias to the original method alone, the features it learns are more concentrated and clear, indicating that the learnable bias significantly enhances the learning ability of the current local block's hidden layer, but it still lacks some details because it does not receive more global information from the subsequent block. When we add the EMA method and learnable bias simultaneously, the learned features are not only clear and specific, but also more comprehensive in detail, proving that their learning abilities are complementary, and greatly improving the performance of the original method. The feature visualization in Fig. 5 verifies our previous thoughts.", + "bbox": [ + 212, + 146, + 787, + 282 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/6a5e04a5245e930abbb02927da9156858b05b05c4cd1fdc37fa7c71ca7236623.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 243, + 316, + 336, + 386 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/e86b2ebc2f6adfda5e94bb461ad8384510581d9974c641ed47a94b62988cc62c.jpg", + "image_caption": [ + "(b)", + "Fig.5: Feature maps comparison of (a) Original Method without MAN, (b) MAN with only EMA, (c) MAN with only learnable bias and (d) MAN with both EMA and learnable bias. The feature map are obtained using ResNet-32 $(\\mathrm{K} = 16)$ on CIFAR-10." + ], + "image_footnote": [], + "bbox": [ + 383, + 316, + 475, + 386 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/cfd441b347f678abb1f0f8ef3af1fa200eff7a0e5621a1276fcbfe90fcc6b4e5.jpg", + "image_caption": [ + "(c)" + ], + "image_footnote": [], + "bbox": [ + 524, + 316, + 616, + 386 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/f99499f3083090cf91f95005d7233b93e4bcdb12d86301673c6056e6733178ae.jpg", + "image_caption": [ + "(d)" + ], + "image_footnote": [], + "bbox": [ + 666, + 316, + 756, + 386 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.5 The Effectiveness of EMA", + "text_level": 1, + "bbox": [ + 214, + 492, + 477, + 507 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "To verify the effectiveness of the EMA update method in MAN, we conduct further ablation experiments. As shown in Table 5, using only the parameters of the next layer in MAN results in very limited improvement. In contrast, employing the EMA method to utilize the parameters of the next layer leads to more significant enhancements. This is because the subsequent local block and the current local block learn different features and have different learning objectives, leading to minimal improvement. However, using the EMA method can gradually promote information exchange between local blocks and reduce parameter fluctuations during training, leading to smoother parameter updates. This beneficial interaction brings significant improvements to MAN.", + "bbox": [ + 212, + 517, + 787, + 670 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/46b62afa0abb81d15ce5a5cd531fdbc3b17ede73bf9922d85d3ace06fdbb81dc.jpg", + "table_caption": [ + "Table 5: Abalation study of EMA. (a) Using DGL as baseline and ResNet-32 $(\\mathrm{K} = 16)$ as backbone on the CIFAR-10 dataset. (b) Using InfoPro as baseline and ResNet-101 $(\\mathrm{K} = 4)$ as backbone on the ImageNet dataset. Parameter signifies whether each local block utilizes the parameters of the adjacent next layer connected to it." + ], + "table_footnote": [], + "table_body": "
ParameterEMATestError
××14.08
✓×12.93
✓✓11.07
", + "bbox": [ + 217, + 762, + 410, + 823 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/da282c873bf314ae3e071496196c78c221ed3675ba80279aa6d185ca4a1fad4d.jpg", + "table_caption": [ + "(a)" + ], + "table_footnote": [], + "table_body": "
ParameterEMATop1-ErrorTop5-Error
××22.816.54
✓×22.646.41
✓✓22.096.07
", + "bbox": [ + 500, + 762, + 772, + 823 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "(b)", + "bbox": [ + 625, + 823, + 645, + 835 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Junhao Su et al.", + "bbox": [ + 271, + 114, + 380, + 126 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.6 Linear Separability Analysis", + "text_level": 1, + "bbox": [ + 215, + 146, + 493, + 162 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "To demonstrate the effectiveness of our Momentum Auxiliary Network, we freeze the parameters of the main network and train a classifier for each local block to obtain the classification accuracy of each local block. As depicted in Fig. 6, we use DGL [3] as a baseline, and after integrating our method, the accuracy of the earlier layers decreases, while the accuracy of the middle and later layers significantly improve. The decrease in accuracy of the earlier layers suggests that they have learned more generalized features. Although these features are not beneficial for optimizing local objectives, they are beneficial from a global perspective. The middle and later layers, receiving these globally beneficial features, have seen a significant increase in accuracy. These results illustrate that our proposed method has facilitated information interaction between gradient-isolated local blocks. This addresses the shortsightedness issue existing in the current supervised local learning field and significantly enhances their performance.", + "bbox": [ + 212, + 171, + 787, + 368 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/6db996f1514a558097af01f9586d67f5704f0d5d7d835416605ab9f89d57a5e7.jpg", + "image_caption": [ + "Fig. 6: Comparison of layer-wise linear separability across different learning rules on ResNet-32 and ResNet-110." + ], + "image_footnote": [], + "bbox": [ + 246, + 395, + 498, + 560 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/759438b2911d841f2adcd073e32c9fb54746344980fa8fe40b9372997a11ed2c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 396, + 754, + 560 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.7 Representation Similarity Analysis", + "text_level": 1, + "bbox": [ + 215, + 633, + 549, + 648 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "To further validate the efficacy of MAN, we conduct a Centered Kernel Alignment (CKA) experiment [20]. CKA serves as a metric to assess the similarity between feature representations. If a method's CKA score is closer to 1 in relation to the E2E training method, it indicates that the method's feature learning process is more aligned with that of E2E training. As depicted in Fig. 7, we use DGL [3] as a baseline and incorporate our method. It can be observed that whether adding the EMA method or learnable bias alone, the CKA scores significantly improve compared to the original method. When both the EMA and learnable bias are used in MAN, the CKA score further improves and performs more stably. Notably, the most significant increase in the CKA score occurs in the early and late layers. In conjunction with our previous analysis of the linear separability experiment, this is because if the early layers' learning method is closer", + "bbox": [ + 212, + 657, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Momentum Auxiliary Network for Supervised Local Learning", + "bbox": [ + 321, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "to E2E training, they will focus more on learning general features to optimize the global objective, rather than focusing narrowly on local objectives. While this may result in poorer classification capabilities in the early layers, it greatly contributes to the overall performance improvement of the network. Through the analysis of images and experimental results, it can be demonstrated that MAN enables information interaction between gradient-isolated local blocks, solving the myopia problem present in current supervised local learning methods.", + "bbox": [ + 212, + 146, + 787, + 252 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/41e853fbfca688fd107444a9f7f8ce88d84afe36a94bfae8e47b9a49985d920f.jpg", + "image_caption": [ + "Fig. 7: Assessment of Similarity in Layer-wise Representations. We use Centered Kernel Alignment (CKA) [20] to quantify the degree of similarity in the layer-wise representations between the E2E backpropagation (BP) and our proposed MAN." + ], + "image_footnote": [], + "bbox": [ + 246, + 280, + 500, + 445 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/8773d9182a59cd09b3ff8c23529e403f422635454fb0f3e22848b615db1e5763.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 280, + 756, + 445 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 534, + 359, + 549 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "This study addresses the performance disparities between traditional supervised local learning and end-to-end (E2E) methods in deep learning. We introduce a versatile Momentum Auxiliary Network to tackle the short-sighted problem in the early optimization work related to supervised local learning, facilitating information exchange between gradient-isolated local blocks. We integrate our Momentum Auxiliary Network into three advanced supervised local learning approaches and evaluate their performance across network architectures with varying depths on four widely adopted datasets. The results demonstrate our method's ability to significantly enhance the ultimate output performance of original supervised local learning methods. Particularly when combined with InfoPro [37], our method significantly reduces GPU memory usage while consistently maintaining performance levels closely aligned with E2E approaches.", + "bbox": [ + 212, + 566, + 787, + 748 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Limitations and future works: Despite the superior performance on large-scale problems like ImageNet [12], our method still performs less accurately than E2E on some conventional image classification datasets. This may be due to the MAN using too few information interaction layers when the network is divided into a larger number of local blocks. In future work, we could explore deepening these information interaction layers to achieve better precision performance.", + "bbox": [ + 212, + 750, + 787, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Junhao Su et al.", + "bbox": [ + 271, + 114, + 380, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 217, + 143, + 392, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work was supported by the National Natural Science Foundation of China (52278154), the Natural Science Foundation of Jiangsu (BK20231429), the Fundamental Research Funds for the Central Universities (2242024RCB0008), and assist from the program of Zhishan Young Scholar of Southeast University.", + "bbox": [ + 215, + 175, + 784, + 234 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 258, + 321, + 273 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Akrout, M., Wilson, C., Humphreys, P., Lillicrap, T., Tweed, D.B.: Deep learning without weight transport. Advances in neural information processing systems 32 (2019)", + "2. Belilovsky, E., Eickenberg, M., Oyallon, E.: Greedy layerwise learning can scale to imagenet. In: International conference on machine learning. pp. 583-593. PMLR (2019)", + "3. Belilovsky, E., Eickenberg, M., Oyallon, E.: Decoupled greedy learning of cnns. In: International Conference on Machine Learning. pp. 736-745. PMLR (2020)", + "4. Bengio, Y.: How auto-encoders could provide credit assignment in deep networks via target propagation. arXiv preprint arXiv:1407.7906 (2014)", + "5. Bengio, Y., Lamblin, P., Popovici, D., Larochelle, H.: Greedy layer-wise training of deep networks. Advances in neural information processing systems 19 (2006)", + "6. Bengio, Y., Lee, D.H., Bornschein, J., Mesnard, T., Lin, Z.: Towards biologically plausible deep learning. arXiv preprint arXiv:1502.04156 (2015)", + "7. Caporale, N., Dan, Y.: Spike timing-dependent plasticity: a hebbian learning rule. Annu. Rev. Neurosci. 31, 25-46 (2008)", + "8. Clark, D., Abbott, L., Chung, S.: Credit assignment through broadcasting a global error vector. Advances in Neural Information Processing Systems 34, 10053-10066 (2021)", + "9. Coates, A., Ng, A., Lee, H.: An analysis of single-layer networks in unsupervised feature learning. In: Proceedings of the fourteenth international conference on artificial intelligence and statistics. pp. 215-223. JMLR Workshop and Conference Proceedings (2011)", + "0. Crick, F.: The recent excitement about neural networks. Nature 337(6203), 129-132 (1989)", + "1. Dellafererra, G., Kreiman, G.: Error-driven input modulation: solving the credit assignment problem without a backward pass. In: International Conference on Machine Learning. pp. 4937-4955. PMLR (2022)", + "2. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 248-255. IEEE (2009)", + "3. Gomez, A.N., Key, O., Perlin, K., Gou, S., Frosst, N., Dean, J., Gal, Y.: Interlocking backpropagation: Improving depthwise model-parallelism. The Journal of Machine Learning Research 23(1), 7714-7741 (2022)", + "4. He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.: Momentum contrast for unsupervised visual representation learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9729-9738 (2020)", + "5. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)" + ], + "bbox": [ + 225, + 287, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Momentum Auxiliary Network for Supervised Local Learning", + "bbox": [ + 321, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "16. Hinton, G.E., Osindero, S., Teh, Y.W.: A fast learning algorithm for deep belief nets. Neural computation 18(7), 1527-1554 (2006)", + "17. Huang, G., Liu, Z., Van Der Maaten, L., Weinberger, K.Q.: Densely connected convolutional networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4700-4708 (2017)", + "18. Illing, B., Ventura, J., Bellec, G., Gerstner, W.: Local plasticity rules can learn deep representations using self-supervised contrastive predictions. Advances in Neural Information Processing Systems 34, 30365-30379 (2021)", + "19. Jaderberg, M., Czarnecki, W.M., Osindero, S., Vinyals, O., Graves, A., Silver, D., Kavukcuoglu, K.: Decoupled neural interfaces using synthetic gradients. In: International conference on machine learning. pp. 1627-1635. PMLR (2017)", + "20. Kornblith, S., Norouzi, M., Lee, H., Hinton, G.: Similarity of neural network representations revisited. In: International conference on machine learning. pp. 3519-3529. PMLR (2019)", + "21. Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images (2009)", + "22. Laskin, M., Metz, L., Nabarro, S., Saroufim, M., Noune, B., Luschi, C., Sohl-Dickstein, J., Abbeel, P.: Parallel training of deep networks with local updates. arXiv preprint arXiv:2012.03837 (2020)", + "23. Le Cun, Y.: Learning process in an asymmetric threshold network. In: Disordered systems and biological organization, pp. 233-240. Springer (1986)", + "24. LeCun, Y., Bengio, Y., Hinton, G.: Deep learning. nature 521(7553), 436-444 (2015)", + "25. Lee, D.H., Zhang, S., Fischer, A., Bengio, Y.: Difference target propagation. In: Machine Learning and Knowledge Discovery in Databases: European Conference, ECML PKDD 2015, Porto, Portugal, September 7-11, 2015, Proceedings, Part I 15. pp. 498-515. Springer (2015)", + "26. Lillicrap, T.P., Cownden, D., Tweed, D.B., Akerman, C.J.: Random synaptic feedback weights support error backpropagation for deep learning. Nature communications 7(1), 13276 (2016)", + "27. Lillicrap, T.P., Santoro, A., Marris, L., Akerman, C.J., Hinton, G.: Backpropagation and the brain. Nature Reviews Neuroscience 21(6), 335-346 (2020)", + "28. Löwe, S., O'Connor, P., Veeling, B.: Putting an end to end-to-end: Gradient-isolated learning of representations. Advances in neural information processing systems 32 (2019)", + "29. Mostafa, H., Ramesh, V., Cauwenberghs, G.: Deep supervised learning using local errors. Frontiers in neuroscience 12, 608 (2018)", + "30. Netzer, Y., Wang, T., Coates, A., Bissacco, A., Wu, B., Ng, A.Y.: Reading digits in natural images with unsupervised feature learning (2011)", + "31. Nokland, A.: Direct feedback alignment provides learning in deep neural networks. Advances in neural information processing systems 29 (2016)", + "32. Nokland, A., Eidnes, L.H.: Training neural networks with local error signals. In: International conference on machine learning. pp. 4839-4850. PMLR (2019)", + "33. Pyeon, M., Moon, J., Hahn, T., Kim, G.: Sedona: Search for decoupled neural networks toward greedy block-wise learning. In: International Conference on Learning Representations (2020)", + "34. Ren, M., Kornblith, S., Liao, R., Hinton, G.: Scaling forward gradient with local losses. arXiv preprint arXiv:2210.03310 (2022)", + "35. Rumelhart, D.E., Hinton, G.E., Williams, R.J.: Learning representations by backpropagating errors. nature 323(6088), 533-536 (1986)" + ], + "bbox": [ + 215, + 146, + 785, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Junhao Su et al.", + "bbox": [ + 271, + 114, + 380, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "36. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556 (2014)", + "37. Wang, Y., Ni, Z., Song, S., Yang, L., Huang, G.: Revisiting locally supervised learning: an alternative to end-to-end training. arXiv preprint arXiv:2101.10832 (2021)", + "38. Wu, B., Nair, S., Martin-Martin, R., Fei-Fei, L., Finn, C.: Greedy hierarchical variational autoencoders for large-scale video prediction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2318-2328 (2021)", + "39. Xie, S., Girshick, R., Dollar, P., Tu, Z., He, K.: Aggregated residual transformations for deep neural networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1492-1500 (2017)", + "40. Xie, X., Seung, H.S.: Equivalence of backpropagation and contrastive hebbian learning in a layered network. Neural computation 15(2), 441-454 (2003)", + "41. Xiong, Y., Ren, M., Urtasun, R.: Loco: Local contrastive representation learning. Advances in neural information processing systems 33, 11142-11153 (2020)" + ], + "bbox": [ + 212, + 146, + 787, + 369 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Momentum Auxiliary Network for Supervised Local Learning", + "bbox": [ + 321, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/Momentum Auxiliary Network for Supervised Local Learning/eb53acf2-10ec-44f2-923d-fc539ed31362_model.json b/2024/Momentum Auxiliary Network for Supervised Local Learning/eb53acf2-10ec-44f2-923d-fc539ed31362_model.json new file mode 100644 index 0000000000000000000000000000000000000000..f8aa69319faa8ee12fdea4eabde96d4ad37c93a0 --- /dev/null +++ b/2024/Momentum Auxiliary Network for Supervised Local Learning/eb53acf2-10ec-44f2-923d-fc539ed31362_model.json @@ -0,0 +1,2280 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.24, + 0.141, + 0.764, + 0.187 + ], + "angle": 0, + "content": "Momentum Auxiliary Network for Supervised Local Learning" + }, + { + "type": "text", + "bbox": [ + 0.283, + 0.216, + 0.721, + 0.231 + ], + "angle": 0, + "content": "Junhao Su\\(^{1*}\\), Changpeng Cai\\(^{1*}\\), Feiyu Zhu\\(^{2*}\\), Chenghao He\\(^{4}\\)" + }, + { + "type": "text", + "bbox": [ + 0.318, + 0.231, + 0.686, + 0.248 + ], + "angle": 0, + "content": "Xiaojie \\(\\mathrm{Xu}^5\\), Dongzhi Guan\\(^{1\\boxtimes}\\), and Chenyang Si\\(^{3\\boxtimes}\\)" + }, + { + "type": "text", + "bbox": [ + 0.424, + 0.258, + 0.581, + 0.273 + ], + "angle": 0, + "content": "\\(^{1}\\) Southeast University" + }, + { + "type": "text", + "bbox": [ + 0.325, + 0.273, + 0.678, + 0.288 + ], + "angle": 0, + "content": "2 University of Shanghai for Science and Technology" + }, + { + "type": "text", + "bbox": [ + 0.379, + 0.287, + 0.623, + 0.302 + ], + "angle": 0, + "content": "3 Nanyang Technological University" + }, + { + "type": "text", + "bbox": [ + 0.329, + 0.301, + 0.673, + 0.315 + ], + "angle": 0, + "content": "\\(^{4}\\) East China University of Science and Technology" + }, + { + "type": "text", + "bbox": [ + 0.315, + 0.315, + 0.69, + 0.329 + ], + "angle": 0, + "content": "5 The Hong Kong University of Science and Technology" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.398, + 0.744, + 0.732 + ], + "angle": 0, + "content": "Abstract. Deep neural networks conventionally employ end-to-end backpropagation for their training process, which lacks biological credibility and triggers a locking dilemma during network parameter updates, leading to significant GPU memory use. Supervised local learning, which segments the network into multiple local blocks updated by independent auxiliary networks. However, these methods cannot replace end-to-end training due to lower accuracy, as gradients only propagate within their local block, creating a lack of information exchange between blocks. To address this issue and establish information transfer across blocks, we propose a Momentum Auxiliary Network (MAN) that establishes a dynamic interaction mechanism. The MAN leverages an exponential moving average (EMA) of the parameters from adjacent local blocks to enhance information flow. This auxiliary network, updated through EMA, helps bridge the informational gap between blocks. Nevertheless, we observe that directly applying EMA parameters has certain limitations due to feature discrepancies among local blocks. To overcome this, we introduce learnable biases, further boosting performance. We have validated our method on four image classification datasets (CIFAR-10, STL-10, SVHN, ImageNet), attaining superior performance and substantial memory savings. Notably, our method can reduce GPU memory usage by more than \\(45\\%\\) on the ImageNet dataset compared to end-to-end training, while achieving higher performance. The Momentum Auxiliary Network thus offers a new perspective for supervised local learning. Our code is available at: https://github.com/JunhaoSuO/MAN." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.747, + 0.741, + 0.777 + ], + "angle": 0, + "content": "Keywords: Local Learning \\(\\cdot\\) Image Classification \\(\\cdot\\) Momentum Auxiliary Network" + }, + { + "type": "page_footnote", + "bbox": [ + 0.232, + 0.798, + 0.376, + 0.812 + ], + "angle": 0, + "content": "*Equal Contribution." + }, + { + "type": "page_footnote", + "bbox": [ + 0.232, + 0.812, + 0.787, + 0.827 + ], + "angle": 0, + "content": "Corresponding Authors: Chengyang Si (chenyang.si.mail@gmail.com) and Dongzhi" + }, + { + "type": "page_footnote", + "bbox": [ + 0.232, + 0.827, + 0.456, + 0.84 + ], + "angle": 0, + "content": "Guan (guandongzhi@seu.edu.cn)." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.383, + 0.127 + ], + "angle": 0, + "content": "Junhao Su et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.146, + 0.788, + 0.329 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.339, + 0.788, + 0.382 + ], + "angle": 0, + "content": "Fig. 1: Comparison between different methods with MAN and the original methods in terms of accuracy. Results are obtained using ResNet-110 (K=55) on the various datasets." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.412, + 0.377, + 0.429 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.448, + 0.789, + 0.844 + ], + "angle": 0, + "content": "In deep learning, the prevalent use of end-to-end backpropagation [35] is essential for training complex neural networks, which requires heavy computational work for loss evaluation and successive gradient descent through network layers to refine parameters [17, 24, 29]. This method starkly contrasts with the local signal processing of biological synapses [6, 7, 40] and imposes a 'locking' constraint [19] that delays parameter updates until the entire forward and backward pass is complete. Such constraints exacerbate challenges like reduced parallelism and increased GPU memory usage [28, 38], compromising training efficiency and scalability. An emerging alternative to traditional end-to-end (E2E) training is Local Learning [2, 3, 5, 18, 19, 29, 32, 37, 41], which promises to mitigate the drawbacks of traditional methods. These methods divide the network into discrete, gradient-isolated blocks, each of which is updated by its dedicated auxiliary network according to distinct local objectives [3, 32, 37]. This training strategy enables immediate parameter updates upon local error reception, circumventing the sequential update bottleneck of E2E training, thereby significantly enhancing the efficiency of parallel training [13, 22]. Moreover, local learning reduces the demand on GPU memory by only retaining gradients for the local and auxiliary networks, thus eliminating the need to store extensive global gradient information and saving computational resources [28, 38]. However, even with these advantages, there is still a significant performance gap between local learning and traditional E2E methods, especially as the network is divided into more local blocks. Current local learning techniques mainly focus on improving the design of auxiliary network structures [3] and making local loss functions better to close this performance gap [32, 37]. Yet, these improvements do not completely address the inherent short-sightedness of local learning: the separation into blocks can make each part of the network only focus on its local objectives, possibly ig-" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.323, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "Momentum Auxiliary Network for Supervised Local Learning" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.176 + ], + "angle": 0, + "content": "noring the overall objectives of the network. This can lead to the discarding of globally beneficial information due to the lack of inter-block communication." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.788, + 0.494 + ], + "angle": 0, + "content": "In this paper, we introduce the Momentum Auxiliary Network (MAN), a novel network architecture designed to mitigate the inherent limitations of supervised local learning by enhancing inter-block communication. Specifically, MAN not only accepts the current local block as input but also absorbs the parameters of its next block. Upon completion of a forward pass, parameter updates are performed using local gradients refined through Exponential Moving Average (EMA) techniques [14]. This novel approach enables each local block to integrate information from the following block, thereby extending the operational perspective of each block beyond its immediate objectives and further aligning with the global objective of the network. This addresses the critical limitations in local learning frameworks, ensuring a more cohesive alignment with the overarching network objectives. Furthermore, we identify that directly applying EMA parameters is constrained by the feature incongruities across gradient-isolated blocks. To mitigate this, we implement additional learnable bias terms to each auxiliary block, enhancing their ability to share information effectively. The proposed MAN requires only a minimal increase in memory use but significantly improves performance by enhancing the information sharing ability of local blocks. The efficacy of the MAN approach is validated on a suite of benchmark image classification datasets, including CIFAR-10 [21], STL-10 [9], SVHN [30], and ImageNet [12]. The results demonstrate the effectiveness of our method in surpassing the limitations of traditional supervised local learning." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.496, + 0.702, + 0.51 + ], + "angle": 0, + "content": "The contributions of this paper could be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.52, + 0.784, + 0.594 + ], + "angle": 0, + "content": "- We propose Momentum Auxiliary Network (MAN), designed to facilitate inter-block communication with Exponential Moving Average (EMA), mitigating the short-sightedness issue in conventional supervised local learning techniques and culminating in an elevated overall performance of the network." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.596, + 0.784, + 0.64 + ], + "angle": 0, + "content": "- MAN is a versatile, plug-and-play approach that can seamlessly integrate with any supervised local learning technique, markedly amplifying their efficacy while requiring only a negligible increase in memory usage." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.642, + 0.784, + 0.702 + ], + "angle": 0, + "content": "- MAN has demonstrated its effectiveness through experiments on benchmark image classification datasets, achieving state-of-the-art performance. Notably, on the ImageNet [12] dataset, it outperforms E2E training while using significantly less memory." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.52, + 0.784, + 0.702 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.737, + 0.388, + 0.753 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.769, + 0.386, + 0.785 + ], + "angle": 0, + "content": "2.1 Local Learning" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.785, + 0.842 + ], + "angle": 0, + "content": "Local learning is first proposed as an innovative deep learning algorithm with the intention of utilizing memory more efficiently and adhering more closely to principles of biological plausibility [10]. This approach emerges as a response to" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.383, + 0.127 + ], + "angle": 0, + "content": "Junhao Su et al." + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.143, + 0.79, + 0.294 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.306, + 0.788, + 0.348 + ], + "angle": 0, + "content": "Fig. 2: Comparison of (a) end-to-end backpropagation, (b) other supervised local learning methods, and (c) our proposed method. Unlike E2E, supervised local learning separates the network into K gradient-isolated local blocks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.39, + 0.788, + 0.512 + ], + "angle": 0, + "content": "the limitations presented by global E2E training [16], prompting the development of alternative supervised local learning methodologies. Examples of these methodologies include a differentiable search algorithm, which decouples network blocks for block-level learning and selects a manually assisted network for each local block [33], as well as a self-supervised contrasting loss function that leverages its local learning rules [18, 41]. In supervised local learning, training primarily relies on the design of reasonable supervised local loss functions or the manual construction of auxiliary networks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.517, + 0.788, + 0.684 + ], + "angle": 0, + "content": "Supervised Local Loss: InfoPro [37] presents a method that involves the reconstruction loss for each hidden layer, coupled with cross-entropy loss. This approach forces the hidden layers to preserve vital input information, thereby preventing the early layers from discarding information relevant to the task. PredSim [32] employs layer-wise loss functions for network training. It makes use of two distinct supervised loss functions to generate local error signals for hidden layers. The optimization of these losses takes place during supervised local learning, thus eliminating the necessity to propagate global errors back to the hidden layers. These methods have achieved advanced performance in the field of supervised local learning, but there is still a significant gap compared to the performance of E2E training." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.69, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Auxiliary Networks: InfoPro [37] incorporates two auxiliary networks into its approach. One network consists of a single convolutional layer followed by two fully-connected layers, which are employed for the cross-entropy loss. The other network utilizes two convolutional layers with upsampling for the reconstruction loss. Alternatively, DGL [3] designs a structure that uses three consecutive convolutional layers connected to a global pooling layer, followed by three consecutive fully connected layers. Moreover, DGL restricts the parameter of the auxiliary network to only \\(5\\%\\) of the main network capacity, offering advantages in terms of speed and memory savings. However, these method still perform poorly when the network is divided into a large number of local blocks." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.322, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "Momentum Auxiliary Network for Supervised Local Learning" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.147, + 0.627, + 0.163 + ], + "angle": 0, + "content": "2.2 Alternative Learning Rules to E2E Training" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.172, + 0.788, + 0.385 + ], + "angle": 0, + "content": "Due to certain limitations inherent in E2E training, the pursuit of alternative methods to E2E training has gradually garnered attention in recent years [27]. Several efforts have been dedicated to addressing the biologically implausible aspects of E2E training, such as the training methodology of target propagation [4, 23, 25] attempts to train a specialized backward connection by utilizing local reconstruction targets. Additionally, some recent studies have attempted to completely avoid backpropagation in neural networks through forward gradient learning [11, 34]. Meanwhile, the weight transport problem [10]. This has been tackled either by employing distinct feedback connections [1, 26] or by directly disseminating global errors to each hidden unit [8, 31]. While these methods have somewhat alleviated the biological implausibility of E2E training, they still rely on global objectives, which fundamentally differ from biological neural networks that rely on local synapses for information transmission. Furthermore, these methods currently struggle to be effective on large datasets [12]." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.407, + 0.331, + 0.423 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.438, + 0.373, + 0.453 + ], + "angle": 0, + "content": "3.1 Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.463, + 0.788, + 0.599 + ], + "angle": 0, + "content": "To set the stage, we first provide a brief overview of traditional end-to-end supervised learning and backpropagation mechanisms. We denote a data sample as \\( x \\) and its corresponding ground-truth label as \\( y \\). The entire deep network is fragmented into several local blocks. During the forward propagation process, the output from the j-th block serves as the input for the (j+1)-th block, expressed as \\( x_{j+1} = f_{\\theta_j}(x_j) \\). Here, \\( \\theta_j \\) symbolizes the parameters of the j-th local block and \\( f(\\cdot) \\) represents the forward calculation of the block. We evaluate the loss function \\( \\mathcal{L}(\\hat{y}, y) \\) between the output of the last block and the ground truth label, and propagate it back iteratively to the preceding blocks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.788, + 0.674 + ], + "angle": 0, + "content": "Supervised local learning strategies [3, 32, 37] integrate auxiliary networks for local supervision. For each local block, an auxiliary network is affixed. The output from a local block is fed to its corresponding auxiliary network, generating the local supervisory signal as \\(\\hat{y}_j = g_{\\gamma_j}(x_{j + 1})\\). Here, \\(\\gamma_{j}\\) denotes the parameters of the j-th auxiliary network." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.675, + 0.788, + 0.707 + ], + "angle": 0, + "content": "In this setup, we update the parameters of the j-th auxiliary network and local block, \\(\\gamma_{j},\\theta_{j}\\), as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.405, + 0.721, + 0.785, + 0.737 + ], + "angle": 0, + "content": "\\[\n\\gamma_ {j} \\leftarrow \\gamma_ {j} - \\eta_ {a} \\times \\nabla_ {\\gamma_ {j}} \\mathcal {L} (\\hat {y} _ {j}, y) \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.407, + 0.758, + 0.785, + 0.774 + ], + "angle": 0, + "content": "\\[\n\\theta_ {j} \\leftarrow \\theta_ {j} - \\eta_ {l} \\times \\nabla_ {\\theta_ {j}} \\mathcal {L} \\left(y _ {\\hat {j}}, y\\right) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "where \\(\\eta_{a},\\eta_{l}\\) are the learning rates of the auxiliary networks and local blocks, respectively. By attaching auxiliary networks, each local block becomes gradient-isolated and can be updated with local supervision rather than global backpropagation." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.384, + 0.128 + ], + "angle": 0, + "content": "Junhao Su et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.519, + 0.163 + ], + "angle": 0, + "content": "3.2 Momentum Auxiliary Network" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.173, + 0.788, + 0.248 + ], + "angle": 0, + "content": "Existing techniques incorporate supervision signals into individual local blocks, which allows for parallel parameter updates and reduces memory overhead. However, this approach can lead to a short-sightedness problem, where each local block neglects the information from subsequent blocks, ultimately resulting in suboptimal final accuracy." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.249, + 0.788, + 0.325 + ], + "angle": 0, + "content": "To address this issue, we propose a comprehensive information exchange module—the Momentum Auxiliary Network (MAN). The MAN employs the Exponential Moving Average (EMA) mechanism [14] as a conduit for transferring information from subsequent blocks to the current block. In the context of MAN, we update the parameters of the j-th auxiliary network and local block as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.405, + 0.339, + 0.786, + 0.357 + ], + "angle": 0, + "content": "\\[\n\\gamma_ {j} \\leftarrow \\gamma_ {j} - \\eta_ {a} \\times \\nabla_ {\\gamma_ {j}} \\mathcal {L} \\left(\\hat {y} _ {j}, y\\right) \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.425, + 0.378, + 0.787, + 0.395 + ], + "angle": 0, + "content": "\\[\n\\gamma_ {j} \\leftarrow E M A \\left(\\gamma_ {j}, \\theta_ {j + 1}\\right) \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.405, + 0.416, + 0.786, + 0.433 + ], + "angle": 0, + "content": "\\[\n\\theta_ {j} \\leftarrow \\theta_ {j} - \\eta_ {l} \\times \\nabla_ {\\theta_ {j}} \\mathcal {L} (\\hat {y} _ {j}, y) \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.439, + 0.788, + 0.5 + ], + "angle": 0, + "content": "where \\(\\gamma_{j}\\) represents the parameters of the j-th auxiliary network and \\(\\theta_{j}\\) represents those of the j-th local block. After updating with local gradients, \\(\\gamma_{j}\\) undergoes further refinement by incorporating the parameters of the subsequent local block via the EMA, which is a weighted sum operation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.5, + 0.487, + 0.741 + ], + "angle": 0, + "content": "However, through experimental results, it becomes apparent that directly applying EMA parameters to update the auxiliary network has its limitations, and it provides limited improvements to each local block. Upon analyzing, we find that the features learned between each local block vary to a certain degree, which hampers the effectiveness of the EMA update method. As a consequence, a learnable bias is introduced to augment the learning capability of the hidden layers within the local block, while also compensating for the deficiencies of the EMA update method." + }, + { + "type": "image", + "bbox": [ + 0.499, + 0.529, + 0.785, + 0.633 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.497, + 0.644, + 0.788, + 0.741 + ], + "angle": 0, + "content": "Fig. 3: Details of the Momentum Auxiliary Network. Local \\((\\mathrm{i} + 1)\\) represents the \\((\\mathrm{i} + 1)\\)-th gradient-isolated local block, which contains layers from layer m to layer \\((\\mathrm{m} + \\mathrm{n})\\), totaling n+1 layers \\((\\mathrm{n} \\geqslant 0)\\). We only use the parameters of the first layer to ensure a balance in GPU memory usage." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.741, + 0.788, + 0.771 + ], + "angle": 0, + "content": "Therefore, the parameter update method for the j-th auxiliary network can be written as:" + }, + { + "type": "equation", + "bbox": [ + 0.356, + 0.786, + 0.787, + 0.806 + ], + "angle": 0, + "content": "\\[\n\\left(\\gamma_ {j}, b _ {j}\\right) \\leftarrow \\left(\\gamma_ {j}, b _ {j}\\right) - \\eta_ {a} \\times \\nabla_ {\\left(\\gamma_ {j}, b _ {j}\\right)} \\mathcal {L} \\left(\\hat {y} _ {j}, y\\right) \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.424, + 0.825, + 0.787, + 0.843 + ], + "angle": 0, + "content": "\\[\n\\gamma_ {j} \\leftarrow E M A \\left(\\gamma_ {j}, \\theta_ {j + 1}\\right) \\tag {7}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.323, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "Momentum Auxiliary Network for Supervised Local Learning" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.192 + ], + "angle": 0, + "content": "where \\((\\gamma_j, b_j)\\) represent the parameters of the j-th auxiliary network. During this process, we update the parameters \\(\\gamma_j\\) of the j-th auxiliary network jointly through the learnable bias \\(b_j\\) and EMA." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.192, + 0.784, + 0.282 + ], + "angle": 0, + "content": "Essentially, the proposed MAN facilitates communication between subsequent blocks through the EMA mechanism [14], effectively resolving the short-sightedness problem inherent in traditional supervised local learning methods. Concurrently, it introduces an independent learnable bias to mitigate information discrepancies caused by feature variations among different local blocks, thereby aligning the output of each block closer to the global target." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.283, + 0.784, + 0.344 + ], + "angle": 0, + "content": "In addition to the two innovative methods mentioned above, MAN demonstrates strong versatility. It can be seamlessly integrated into existing supervised local learning methods, and it excels across various datasets, reflecting its flexibility and generalizability." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.365, + 0.376, + 0.383 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.395, + 0.429, + 0.411 + ], + "angle": 0, + "content": "4.1 Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.418, + 0.784, + 0.463 + ], + "angle": 0, + "content": "We conduct experiments using four widely adopted datasets: CIFAR-10 [21], SVHN [30], STL-10 [9], and ImageNet [12], with ResNets [15] of varying depths serving as the network architectures." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.464, + 0.785, + 0.585 + ], + "angle": 0, + "content": "Three state-of-the-art supervised local learning methods are selected for comparison: PredSim [32], DGL [3], and InfoPro [37]. We then partition the network into K local blocks, each containing an approximately equal number of layers. Our proposed Momentum Auxiliary Network is incorporated only into the first K-1 local blocks. The K-th local block does not employ an auxiliary network and is directly connected to the output classifier. We compare these configurations against traditional E2E and original supervised local learning methods to maintain consistent training settings and eliminate confounding variables." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.605, + 0.416, + 0.621 + ], + "angle": 0, + "content": "4.2 Implement Details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.784, + 0.72 + ], + "angle": 0, + "content": "In our experiments on CIFAR-10 [21], SVHN [30], and STL-10 [9] datasets with ResNet-32 [15] and ResNet-110 [15], we utilize the SGD optimizer with Nesterov momentum set at 0.9 and an L2 weight decay factor of 1e-4. We employ batch sizes of 1024 for CIFAR-10 and SVHN and 128 for STL-10. The training duration spans 400 epochs, starting with initial learning rates of 0.8 for CIFAR-10 / SVHN and 0.1 for STL-10, following a cosine annealing scheduler [9]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.785, + 0.84 + ], + "angle": 0, + "content": "In our experiments with ImageNet [12], we adopt different training settings for various architectures. We train VGG13 [36] for 90 epochs with an initial learning rate of 0.025. For ResNet-101 [15], ResNet-152 [15], and ResNeXt-101,32 \\(\\times\\) 8d [39], we train them for 90 epochs as well, with initial learning rates of 0.05, 0.05, and 0.025, respectively. We set the batch sizes to 64 for VGG13, 128 for ResNet-101 and ResNet-152, and 64 for ResNeXt-101,32 \\(\\times\\) 8d. We maintain consistency with other training configurations as previously described for CIFAR-10 [21]." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.383, + 0.128 + ], + "angle": 0, + "content": "Junhao Su et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.208 + ], + "angle": 0, + "content": "Based on our multiple experimental results, we select the hyperparameter momentum value as 0.995, due to its consistently stable and superior performance. We will provide the experimental results of other momentum hyperparameters in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.23, + 0.599, + 0.246 + ], + "angle": 0, + "content": "4.3 Results on Image Classification Datasets" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.255, + 0.788, + 0.332 + ], + "angle": 0, + "content": "Results on Image Classification Benchmarks: We start by assessing the accuracy performance of our approach using the CIFAR-10 [21], SVHN [30], and STL-10 [9] datasets. We employ ResNet-32 [15], partitioned into 8 and 16 local blocks, and ResNet-110 [15], divided into 32 and 55 local blocks. As illustrated in Table 1, our MAN significantly bolsters the accuracy of all methods." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.358, + 0.787, + 0.389 + ], + "angle": 0, + "content": "Table 1: Performance of different networks with varying numbers of local blocks. The average test error is obtained by 5 experiments. The * means addition of our MAN." + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.399, + 0.789, + 0.598 + ], + "angle": 0, + "content": "
DatasetMethodResNet-32ResNet-110
K = 8 (Test Error)K = 16 (Test Error)K = 32 (Test Error)K = 55 (Test Error)
CIFAR-10(E2E(ResNet-32)=6.37,E2E(ResNet-110)=5.42)PredSim [32]20.6222.7122.0824.74
PredSim*14.29(↓6.33)15.58(↓7.13)17.05(↓5.03)18.02(↓6.72)
DGL [3]11.6314.0812.5114.45
DGL*8.42(↓3.21)9.11(↓4.97)9.65(↓2.86)9.73(↓4.72)
InfoPro [37]11.5112.9312.2613.22
InfoPro*9.32(↓2.19)9.65(↓3.28)9.06(↓3.20)9.77(↓3.45)
STL-10(E2E(ResNet-32)=19.35,E2E(ResNet-110)=19.67)PredSim [32]31.9732.9032.0533.27
PredSim*29.97(↓2.00)29.99(↓2.91)30.48(↓1.57)31.55(↓1.72)
DGL [3]25.0527.1425.6728.16
DGL*20.74(↓4.31)21.37(↓5.77)22.54(↓3.13)22.69(↓5.47)
InfoPro [37]27.3229.2828.5829.20
InfoPro*23.17(↓4.15)23.54(↓5.74)24.08(↓4.50)24.74(↓4.46)
SVHN(E2E(ResNet-32)=2.99,E2E(ResNet-110)=2.92)PredSim [32]6.918.089.1210.47
PredSim*5.54(↓1.37)6.39(↓1.69)7.27(↓1.85)8.24(↓2.23)
DGL [3]4.835.055.125.36
DGL*3.80(↓1.03)4.04(↓1.01)4.08(↓1.04)4.52(↓0.84)
InfoPro [37]5.615.975.896.11
InfoPro*4.49(↓1.12)5.19(↓0.78)4.85(↓1.04)4.99(↓1.12)
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.788, + 0.795 + ], + "angle": 0, + "content": "On the CIFAR-10 dataset [21], our method exhibits considerable improvements in diminishing test errors across various methods. In the relatively shallower network of ResNet-32 (K=16), where individual layers function as gradient-isolated local blocks, we record a reduction in test errors for PredSim [32], DGL [3], and InfoPro [37], from 22.71, 14.08, and 12.93 to 15.58, 9.11, and 9.65 respectively. This translates to a performance enhancement exceeding \\(25\\%\\) for all methods. Even though the performance across all methods in the comparatively deeper network, ResNet-110 (K=55), is somewhat inferior due to the inherent need for more global information in such networks, our method still delivers exceptional performance. It achieves approximately a \\(20\\%\\) improvement across all methods, underscoring the robust effectiveness of MAN in deeper networks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.788, + 0.842 + ], + "angle": 0, + "content": "When applied to other datasets, MAN can also reduce the test error of Pred-Sim [32], DGL [3], and InfoPro [37] by at least \\(5\\%\\), \\(12\\%\\), and \\(16\\%\\) on the STL-10 [9] dataset. On the SVHN [30] dataset, our improvements over the three" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.323, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "Momentum Auxiliary Network for Supervised Local Learning" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.208 + ], + "angle": 0, + "content": "methods also surpass \\(20\\%\\), \\(21\\%\\), and \\(13\\%\\). As can be seen, the improvement our MAN introduces to all methods is quite remarkable—comparable even to the accuracy of E2E training—and it significantly mitigates the overwhelming performance issue that has continually plagued supervised local learning." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.21, + 0.785, + 0.331 + ], + "angle": 0, + "content": "Results on ImageNet: We further validate the effectiveness of our approach on ImageNet [12] using four networks of varying depths (ResNets [15] and VGG13 [36]). As depicted in Table 2, when we employ VGG13 as the backbone and divide the network into 10 blocks, DGL [3] achieves merely a Top1-Error of 35.60 and a Top5-Error of 14.2, representing a substantial gap when compared to the E2E method. However, with the introduction of our MAN, the Top1-Error reduces by 3.61 points, and the Top5-Error decreases by 3.36 points for DGL. This significant enhancement brings the performance closer to the E2E method." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.333, + 0.787, + 0.44 + ], + "angle": 0, + "content": "As illustrated in Table 2, when we use ResNet-101 [15], ResNet-152 [15], ResNeXt-101, \\(32 \\times 8\\mathrm{d}\\) [39] as backbones and divide the network into four blocks, the performance of InfoPro [37] is already below that of E2E. After incorporating our MAN, the Top-1 Error of these three backbone networks can be reduced by approximately \\(6\\%\\) compared to the original, surpassing the performance of E2E training. These results underscore the effectiveness of our MAN on the large-scale ImageNet [12] dataset, even when using deeper networks." + }, + { + "type": "table_caption", + "bbox": [ + 0.329, + 0.472, + 0.674, + 0.487 + ], + "angle": 0, + "content": "Table 2: Results on the validation set of ImageNet" + }, + { + "type": "table", + "bbox": [ + 0.235, + 0.497, + 0.768, + 0.698 + ], + "angle": 0, + "content": "
NetworkMethodTop1-ErrorTop5-Error
ResNet-101E2E22.035.93
InfoPro(K=2) [37]21.855.89
InfoPro*(K=2)21.65(↓0.20)5.49(↓0.40)
InfoPro(K=4) [37]22.816.54
InfoPro*(K=4)21.73(↓1.08)5.81(↓0.73)
ResNet-152E2E21.605.92
InfoPro(K=2) [37]21.455.84
InfoPro*(K=2)21.23(↓0.22)5.53(↓0.31)
InfoPro(K=4) [37]22.936.71
InfoPro*(K=4)21.59(↓1.34)5.89(↓0.82)
ResNeXt-101, 32 × 8dE2E20.645.40
InfoPro(K=2) [37]20.355.28
InfoPro*(K=2)20.11(↓0.24)5.18(↓0.10)
InfoPro(K=4) [37]21.696.11
InfoPro*(K=4)20.37(↓1.32)5.34(↓0.77)
VGG13E2E28.419.63
DGL [3]35.6014.20
DGL*(K=10)31.99(↓3.61)10.84(↓3.36)
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Training-Accuracy Curve Analysis: As depicted in the accuracy-epoch curve in Fig. 4, our Momentum Auxiliary Network consistently outperforms the original method in terms of accuracy throughout the entire training process. This underscores its reliability and stability during the training process of classification tasks. Moreover, our MAN achieves a higher accuracy earlier in the later stages of training and attains stability sooner, indicating a faster convergence rate—a critical attribute in large-scale and complex tasks." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.383, + 0.127 + ], + "angle": 0, + "content": "Junhao Su et al." + }, + { + "type": "image", + "bbox": [ + 0.231, + 0.162, + 0.481, + 0.309 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.162, + 0.761, + 0.308 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.321, + 0.788, + 0.351 + ], + "angle": 0, + "content": "Fig. 4: Training-Accuracy curves, the left uses ResNet-32 (K=16) as the backbone, while the right uses ResNet-110 (K=55). Both are utilizing the CIFAR-10 dataset." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.379, + 0.789, + 0.487 + ], + "angle": 0, + "content": "In summary, our Momentum Auxiliary Network significantly enhances the accuracy of traditional supervised local learning methods by promoting information exchange between local blocks. The beneficial information exchange facilitated by Momentum Auxiliary Network significantly mitigates the pervasive issue of shortsightedness within the supervised local learning domain, thereby offering advantages in terms of training stability and convergence speed. This is crucial for training models more effectively and efficiently." + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.512, + 0.789, + 0.542 + ], + "angle": 0, + "content": "Table 3: Comparison of GPU memory usage using InfoPro as baseline with different backbones on the ImageNet dataset." + }, + { + "type": "table", + "bbox": [ + 0.238, + 0.553, + 0.763, + 0.674 + ], + "angle": 0, + "content": "
MethodResNet-101\nGPU Memory(GB)ResNet-152\nGPU Memory(GB)ResNeXt-101, 32×8d\nGPU Memory(GB)
E2E19.7126.2919.22
InfoPro(K=2) [37]12.06(↓ 38.8%)15.53(↓ 40.9%)11.55(↓ 39.9%)
InfoPro*(K=2)12.32(↓ 37.5%)15.93(↓ 39.4%)11.73(↓ 38.9%)
InfoPro(K=4) [37]10.37(↓ 47.3%)13.48(↓ 48.7%)10.24(↓ 46.7%)
InfoPro*(K=4)10.69(↓ 45.8%)13.91(↓ 47.1%)10.49(↓ 45.5%)
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Results on GPU memory requirement: Supervised local learning markedly conserves GPU memory by limiting gradient propagation to within local blocks, thereby ensuring that backpropagation transpires exclusively within each respective local block and its corresponding auxiliary network. This strategy significantly diminishes the storage demands for activation parameters and gradient information, which would typically disseminate across local blocks. Consequently, this facilitates substantial savings in GPU memory utilization. Our analysis, featuring a comprehensive comparison of GPU memory usage on the ImageNet dataset, illustrates that the implementation of a Momentum Auxiliary Network" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.323, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "Momentum Auxiliary Network for Supervised Local Learning" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.784, + 0.187 + ], + "angle": 0, + "content": "Table 4: Abalation study of MAN. (a) Using DGL as baseline and ResNet-32 \\((\\mathrm{K} = 16)\\) as backbone on the CIFAR-10 dataset. (b) Using InfoPro as baseline and ResNet-101 \\((\\mathrm{K} = 4)\\) as backbone on the ImageNet dataset. LB stands for Learnable Bias." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.198, + 0.391, + 0.273 + ], + "angle": 0, + "content": "
EMALBTest Error
××14.08
✓×11.07(↓3.01)
✓✓9.11(↓4.97)
(a)
" + }, + { + "type": "table", + "bbox": [ + 0.5, + 0.199, + 0.765, + 0.273 + ], + "angle": 0, + "content": "
EMALBTop1-ErrorTop5-Error
××22.816.54
✓×22.09(↓ 0.72)6.07(↓ 0.47)
✓✓21.73(↓ 1.08)5.81(↓ 0.73)
(b)
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.303, + 0.784, + 0.333 + ], + "angle": 0, + "content": "can elicit notable performance enhancements while incurring only a minimal uptick in GPU memory consumption." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.334, + 0.785, + 0.515 + ], + "angle": 0, + "content": "As shown in Table 3, it indicates the substantial GPU memory reduction on the ImageNet [12]. When applying our method to InfoPro [37] on ResNet-101 [15], ResNet-152 [15], and ResNeXt-101, \\(32 \\times 8d\\) [39], dividing the backbone into two local blocks results in a GPU memory saving of \\(37.5\\%\\) to \\(39.4\\%\\). When the network is divided into four local blocks, the degree of GPU memory savings is further enhanced, exceeding \\(45\\%\\). Combined with Table 2, it can be seen that we achieve better performance with almost half the GPU memory required for end-to-end training. Further analysis of the memory usage comparison with the original method shows that our approach only increases the GPU memory by about \\(1\\%\\) over the original method, while achieving a performance improvement of over \\(5\\%\\). These results highlight the excellent balance that the MAN method achieves in terms of GPU memory usage and performance enhancement." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.54, + 0.401, + 0.554 + ], + "angle": 0, + "content": "4.4 Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.567, + 0.785, + 0.627 + ], + "angle": 0, + "content": "We conduct an ablation study on the CIFAR-10 [21] dataset and ImageNet dataset to assess the impact of the EMA method [14] and learnable bias in the MAN on performance. For this analysis, we use ResNet-32 (K=16) [15] as the backbone and the original DGL [3] method as a comparison baseline." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.785, + 0.734 + ], + "angle": 0, + "content": "As shown in Table 4(a), when we only use the EMA [14] in MAN to promote information exchange with subsequent blocks, without adding a learnable bias, the test error decreases from 14.08 to 11.07. When we further add a learnable bias, the test error decreases from 11.07 to 9.11, it is evident that both the EMA method and learnable bias contribute to performance improvement. We hypothesize that the EMA and learnable bias are somewhat complementary in terms of the features they learn. To validate this, we conduct feature visualization." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.787, + 0.84 + ], + "angle": 0, + "content": "As depicted in Fig. 5(a), we can observe that the original method without the addition of MAN learns very limited and chaotic features. After adding the EMA method alone, the features it learns are clearly more characteristic and defined, indicating that it indeed receives some global features from the information in the subsequent block. However, there are still many blurry features, which may be due to the EMA parameter update method [14], causing information imbalance due to feature differences between different local blocks. When we" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.382, + 0.127 + ], + "angle": 0, + "content": "Junhao Su et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.284 + ], + "angle": 0, + "content": "add learnable bias to the original method alone, the features it learns are more concentrated and clear, indicating that the learnable bias significantly enhances the learning ability of the current local block's hidden layer, but it still lacks some details because it does not receive more global information from the subsequent block. When we add the EMA method and learnable bias simultaneously, the learned features are not only clear and specific, but also more comprehensive in detail, proving that their learning abilities are complementary, and greatly improving the performance of the original method. The feature visualization in Fig. 5 verifies our previous thoughts." + }, + { + "type": "image", + "bbox": [ + 0.245, + 0.317, + 0.337, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.281, + 0.394, + 0.302, + 0.408 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.384, + 0.318, + 0.477, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.421, + 0.394, + 0.442, + 0.408 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.525, + 0.318, + 0.617, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.563, + 0.394, + 0.583, + 0.408 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image", + "bbox": [ + 0.668, + 0.318, + 0.757, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.702, + 0.393, + 0.724, + 0.408 + ], + "angle": 0, + "content": "(d)" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.419, + 0.788, + 0.462 + ], + "angle": 0, + "content": "Fig.5: Feature maps comparison of (a) Original Method without MAN, (b) MAN with only EMA, (c) MAN with only learnable bias and (d) MAN with both EMA and learnable bias. The feature map are obtained using ResNet-32 \\((\\mathrm{K} = 16)\\) on CIFAR-10." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.493, + 0.478, + 0.508 + ], + "angle": 0, + "content": "4.5 The Effectiveness of EMA" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.518, + 0.788, + 0.671 + ], + "angle": 0, + "content": "To verify the effectiveness of the EMA update method in MAN, we conduct further ablation experiments. As shown in Table 5, using only the parameters of the next layer in MAN results in very limited improvement. In contrast, employing the EMA method to utilize the parameters of the next layer leads to more significant enhancements. This is because the subsequent local block and the current local block learn different features and have different learning objectives, leading to minimal improvement. However, using the EMA method can gradually promote information exchange between local blocks and reduce parameter fluctuations during training, leading to smoother parameter updates. This beneficial interaction brings significant improvements to MAN." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.696, + 0.788, + 0.753 + ], + "angle": 0, + "content": "Table 5: Abalation study of EMA. (a) Using DGL as baseline and ResNet-32 \\((\\mathrm{K} = 16)\\) as backbone on the CIFAR-10 dataset. (b) Using InfoPro as baseline and ResNet-101 \\((\\mathrm{K} = 4)\\) as backbone on the ImageNet dataset. Parameter signifies whether each local block utilizes the parameters of the adjacent next layer connected to it." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.763, + 0.411, + 0.824 + ], + "angle": 0, + "content": "
ParameterEMATestError
××14.08
✓×12.93
✓✓11.07
" + }, + { + "type": "table_caption", + "bbox": [ + 0.304, + 0.824, + 0.323, + 0.836 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "table", + "bbox": [ + 0.5, + 0.763, + 0.774, + 0.824 + ], + "angle": 0, + "content": "
ParameterEMATop1-ErrorTop5-Error
××22.816.54
✓×22.646.41
✓✓22.096.07
" + }, + { + "type": "table_caption", + "bbox": [ + 0.626, + 0.824, + 0.647, + 0.836 + ], + "angle": 0, + "content": "(b)" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.323, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Momentum Auxiliary Network for Supervised Local Learning" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.147, + 0.495, + 0.163 + ], + "angle": 0, + "content": "4.6 Linear Separability Analysis" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.172, + 0.788, + 0.369 + ], + "angle": 0, + "content": "To demonstrate the effectiveness of our Momentum Auxiliary Network, we freeze the parameters of the main network and train a classifier for each local block to obtain the classification accuracy of each local block. As depicted in Fig. 6, we use DGL [3] as a baseline, and after integrating our method, the accuracy of the earlier layers decreases, while the accuracy of the middle and later layers significantly improve. The decrease in accuracy of the earlier layers suggests that they have learned more generalized features. Although these features are not beneficial for optimizing local objectives, they are beneficial from a global perspective. The middle and later layers, receiving these globally beneficial features, have seen a significant increase in accuracy. These results illustrate that our proposed method has facilitated information interaction between gradient-isolated local blocks. This addresses the shortsightedness issue existing in the current supervised local learning field and significantly enhances their performance." + }, + { + "type": "image", + "bbox": [ + 0.248, + 0.396, + 0.499, + 0.561 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.397, + 0.756, + 0.561 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.573, + 0.788, + 0.601 + ], + "angle": 0, + "content": "Fig. 6: Comparison of layer-wise linear separability across different learning rules on ResNet-32 and ResNet-110." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.634, + 0.55, + 0.65 + ], + "angle": 0, + "content": "4.7 Representation Similarity Analysis" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.841 + ], + "angle": 0, + "content": "To further validate the efficacy of MAN, we conduct a Centered Kernel Alignment (CKA) experiment [20]. CKA serves as a metric to assess the similarity between feature representations. If a method's CKA score is closer to 1 in relation to the E2E training method, it indicates that the method's feature learning process is more aligned with that of E2E training. As depicted in Fig. 7, we use DGL [3] as a baseline and incorporate our method. It can be observed that whether adding the EMA method or learnable bias alone, the CKA scores significantly improve compared to the original method. When both the EMA and learnable bias are used in MAN, the CKA score further improves and performs more stably. Notably, the most significant increase in the CKA score occurs in the early and late layers. In conjunction with our previous analysis of the linear separability experiment, this is because if the early layers' learning method is closer" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.382, + 0.127 + ], + "angle": 0, + "content": "Junhao Su et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.253 + ], + "angle": 0, + "content": "to E2E training, they will focus more on learning general features to optimize the global objective, rather than focusing narrowly on local objectives. While this may result in poorer classification capabilities in the early layers, it greatly contributes to the overall performance improvement of the network. Through the analysis of images and experimental results, it can be demonstrated that MAN enables information interaction between gradient-isolated local blocks, solving the myopia problem present in current supervised local learning methods." + }, + { + "type": "image", + "bbox": [ + 0.248, + 0.281, + 0.501, + 0.446 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.281, + 0.758, + 0.446 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.458, + 0.788, + 0.501 + ], + "angle": 0, + "content": "Fig. 7: Assessment of Similarity in Layer-wise Representations. We use Centered Kernel Alignment (CKA) [20] to quantify the degree of similarity in the layer-wise representations between the E2E backpropagation (BP) and our proposed MAN." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.535, + 0.36, + 0.55 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.568, + 0.788, + 0.749 + ], + "angle": 0, + "content": "This study addresses the performance disparities between traditional supervised local learning and end-to-end (E2E) methods in deep learning. We introduce a versatile Momentum Auxiliary Network to tackle the short-sighted problem in the early optimization work related to supervised local learning, facilitating information exchange between gradient-isolated local blocks. We integrate our Momentum Auxiliary Network into three advanced supervised local learning approaches and evaluate their performance across network architectures with varying depths on four widely adopted datasets. The results demonstrate our method's ability to significantly enhance the ultimate output performance of original supervised local learning methods. Particularly when combined with InfoPro [37], our method significantly reduces GPU memory usage while consistently maintaining performance levels closely aligned with E2E approaches." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Limitations and future works: Despite the superior performance on large-scale problems like ImageNet [12], our method still performs less accurately than E2E on some conventional image classification datasets. This may be due to the MAN using too few information interaction layers when the network is divided into a larger number of local blocks. In future work, we could explore deepening these information interaction layers to achieve better precision performance." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.323, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Momentum Auxiliary Network for Supervised Local Learning" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.393, + 0.163 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.176, + 0.785, + 0.236 + ], + "angle": 0, + "content": "This work was supported by the National Natural Science Foundation of China (52278154), the Natural Science Foundation of Jiangsu (BK20231429), the Fundamental Research Funds for the Central Universities (2242024RCB0008), and assist from the program of Zhishan Young Scholar of Southeast University." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.259, + 0.323, + 0.274 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.289, + 0.786, + 0.331 + ], + "angle": 0, + "content": "1. Akrout, M., Wilson, C., Humphreys, P., Lillicrap, T., Tweed, D.B.: Deep learning without weight transport. Advances in neural information processing systems 32 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.332, + 0.786, + 0.372 + ], + "angle": 0, + "content": "2. Belilovsky, E., Eickenberg, M., Oyallon, E.: Greedy layerwise learning can scale to imagenet. In: International conference on machine learning. pp. 583-593. PMLR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.373, + 0.786, + 0.4 + ], + "angle": 0, + "content": "3. Belilovsky, E., Eickenberg, M., Oyallon, E.: Decoupled greedy learning of cnns. In: International Conference on Machine Learning. pp. 736-745. PMLR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.401, + 0.786, + 0.428 + ], + "angle": 0, + "content": "4. Bengio, Y.: How auto-encoders could provide credit assignment in deep networks via target propagation. arXiv preprint arXiv:1407.7906 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.429, + 0.786, + 0.455 + ], + "angle": 0, + "content": "5. Bengio, Y., Lamblin, P., Popovici, D., Larochelle, H.: Greedy layer-wise training of deep networks. Advances in neural information processing systems 19 (2006)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.457, + 0.786, + 0.482 + ], + "angle": 0, + "content": "6. Bengio, Y., Lee, D.H., Bornschein, J., Mesnard, T., Lin, Z.: Towards biologically plausible deep learning. arXiv preprint arXiv:1502.04156 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.484, + 0.786, + 0.51 + ], + "angle": 0, + "content": "7. Caporale, N., Dan, Y.: Spike timing-dependent plasticity: a hebbian learning rule. Annu. Rev. Neurosci. 31, 25-46 (2008)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.511, + 0.786, + 0.551 + ], + "angle": 0, + "content": "8. Clark, D., Abbott, L., Chung, S.: Credit assignment through broadcasting a global error vector. Advances in Neural Information Processing Systems 34, 10053-10066 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.552, + 0.786, + 0.607 + ], + "angle": 0, + "content": "9. Coates, A., Ng, A., Lee, H.: An analysis of single-layer networks in unsupervised feature learning. In: Proceedings of the fourteenth international conference on artificial intelligence and statistics. pp. 215-223. JMLR Workshop and Conference Proceedings (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.608, + 0.786, + 0.634 + ], + "angle": 0, + "content": "0. Crick, F.: The recent excitement about neural networks. Nature 337(6203), 129-132 (1989)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.635, + 0.786, + 0.675 + ], + "angle": 0, + "content": "1. Dellafererra, G., Kreiman, G.: Error-driven input modulation: solving the credit assignment problem without a backward pass. In: International Conference on Machine Learning. pp. 4937-4955. PMLR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.676, + 0.786, + 0.716 + ], + "angle": 0, + "content": "2. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 248-255. IEEE (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.717, + 0.786, + 0.758 + ], + "angle": 0, + "content": "3. Gomez, A.N., Key, O., Perlin, K., Gou, S., Frosst, N., Dean, J., Gal, Y.: Interlocking backpropagation: Improving depthwise model-parallelism. The Journal of Machine Learning Research 23(1), 7714-7741 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.759, + 0.786, + 0.799 + ], + "angle": 0, + "content": "4. He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.: Momentum contrast for unsupervised visual representation learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9729-9738 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.8, + 0.786, + 0.84 + ], + "angle": 0, + "content": "5. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.289, + 0.786, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.382, + 0.127 + ], + "angle": 0, + "content": "Junhao Su et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.175 + ], + "angle": 0, + "content": "16. Hinton, G.E., Osindero, S., Teh, Y.W.: A fast learning algorithm for deep belief nets. Neural computation 18(7), 1527-1554 (2006)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.176, + 0.786, + 0.217 + ], + "angle": 0, + "content": "17. Huang, G., Liu, Z., Van Der Maaten, L., Weinberger, K.Q.: Densely connected convolutional networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4700-4708 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.218, + 0.786, + 0.259 + ], + "angle": 0, + "content": "18. Illing, B., Ventura, J., Bellec, G., Gerstner, W.: Local plasticity rules can learn deep representations using self-supervised contrastive predictions. Advances in Neural Information Processing Systems 34, 30365-30379 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.26, + 0.787, + 0.3 + ], + "angle": 0, + "content": "19. Jaderberg, M., Czarnecki, W.M., Osindero, S., Vinyals, O., Graves, A., Silver, D., Kavukcuoglu, K.: Decoupled neural interfaces using synthetic gradients. In: International conference on machine learning. pp. 1627-1635. PMLR (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.301, + 0.787, + 0.342 + ], + "angle": 0, + "content": "20. Kornblith, S., Norouzi, M., Lee, H., Hinton, G.: Similarity of neural network representations revisited. In: International conference on machine learning. pp. 3519-3529. PMLR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.343, + 0.787, + 0.37 + ], + "angle": 0, + "content": "21. Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.371, + 0.787, + 0.412 + ], + "angle": 0, + "content": "22. Laskin, M., Metz, L., Nabarro, S., Saroufim, M., Noune, B., Luschi, C., Sohl-Dickstein, J., Abbeel, P.: Parallel training of deep networks with local updates. arXiv preprint arXiv:2012.03837 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.412, + 0.787, + 0.439 + ], + "angle": 0, + "content": "23. Le Cun, Y.: Learning process in an asymmetric threshold network. In: Disordered systems and biological organization, pp. 233-240. Springer (1986)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.44, + 0.787, + 0.466 + ], + "angle": 0, + "content": "24. LeCun, Y., Bengio, Y., Hinton, G.: Deep learning. nature 521(7553), 436-444 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.467, + 0.787, + 0.522 + ], + "angle": 0, + "content": "25. Lee, D.H., Zhang, S., Fischer, A., Bengio, Y.: Difference target propagation. In: Machine Learning and Knowledge Discovery in Databases: European Conference, ECML PKDD 2015, Porto, Portugal, September 7-11, 2015, Proceedings, Part I 15. pp. 498-515. Springer (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.523, + 0.787, + 0.563 + ], + "angle": 0, + "content": "26. Lillicrap, T.P., Cownden, D., Tweed, D.B., Akerman, C.J.: Random synaptic feedback weights support error backpropagation for deep learning. Nature communications 7(1), 13276 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.565, + 0.787, + 0.592 + ], + "angle": 0, + "content": "27. Lillicrap, T.P., Santoro, A., Marris, L., Akerman, C.J., Hinton, G.: Backpropagation and the brain. Nature Reviews Neuroscience 21(6), 335-346 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.592, + 0.787, + 0.633 + ], + "angle": 0, + "content": "28. Löwe, S., O'Connor, P., Veeling, B.: Putting an end to end-to-end: Gradient-isolated learning of representations. Advances in neural information processing systems 32 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.634, + 0.787, + 0.66 + ], + "angle": 0, + "content": "29. Mostafa, H., Ramesh, V., Cauwenberghs, G.: Deep supervised learning using local errors. Frontiers in neuroscience 12, 608 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.661, + 0.787, + 0.688 + ], + "angle": 0, + "content": "30. Netzer, Y., Wang, T., Coates, A., Bissacco, A., Wu, B., Ng, A.Y.: Reading digits in natural images with unsupervised feature learning (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.689, + 0.787, + 0.716 + ], + "angle": 0, + "content": "31. Nokland, A.: Direct feedback alignment provides learning in deep neural networks. Advances in neural information processing systems 29 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.717, + 0.787, + 0.744 + ], + "angle": 0, + "content": "32. Nokland, A., Eidnes, L.H.: Training neural networks with local error signals. In: International conference on machine learning. pp. 4839-4850. PMLR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.745, + 0.787, + 0.785 + ], + "angle": 0, + "content": "33. Pyeon, M., Moon, J., Hahn, T., Kim, G.: Sedona: Search for decoupled neural networks toward greedy block-wise learning. In: International Conference on Learning Representations (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.786, + 0.787, + 0.813 + ], + "angle": 0, + "content": "34. Ren, M., Kornblith, S., Liao, R., Hinton, G.: Scaling forward gradient with local losses. arXiv preprint arXiv:2210.03310 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.814, + 0.787, + 0.841 + ], + "angle": 0, + "content": "35. Rumelhart, D.E., Hinton, G.E., Williams, R.J.: Learning representations by backpropagating errors. nature 323(6088), 533-536 (1986)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.323, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Momentum Auxiliary Network for Supervised Local Learning" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.175 + ], + "angle": 0, + "content": "36. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.175, + 0.788, + 0.217 + ], + "angle": 0, + "content": "37. Wang, Y., Ni, Z., Song, S., Yang, L., Huang, G.: Revisiting locally supervised learning: an alternative to end-to-end training. arXiv preprint arXiv:2101.10832 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.218, + 0.788, + 0.272 + ], + "angle": 0, + "content": "38. Wu, B., Nair, S., Martin-Martin, R., Fei-Fei, L., Finn, C.: Greedy hierarchical variational autoencoders for large-scale video prediction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2318-2328 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.273, + 0.788, + 0.314 + ], + "angle": 0, + "content": "39. Xie, S., Girshick, R., Dollar, P., Tu, Z., He, K.: Aggregated residual transformations for deep neural networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1492-1500 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.315, + 0.788, + 0.342 + ], + "angle": 0, + "content": "40. Xie, X., Seung, H.S.: Equivalence of backpropagation and contrastive hebbian learning in a layered network. Neural computation 15(2), 441-454 (2003)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.343, + 0.788, + 0.371 + ], + "angle": 0, + "content": "41. Xiong, Y., Ren, M., Urtasun, R.: Loco: Local contrastive representation learning. Advances in neural information processing systems 33, 11142-11153 (2020)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.371 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Momentum Auxiliary Network for Supervised Local Learning/eb53acf2-10ec-44f2-923d-fc539ed31362_origin.pdf b/2024/Momentum Auxiliary Network for Supervised Local Learning/eb53acf2-10ec-44f2-923d-fc539ed31362_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..b4cc29da90f539ce5e3c87379a12d5c2c54132cb --- /dev/null +++ b/2024/Momentum Auxiliary Network for Supervised Local Learning/eb53acf2-10ec-44f2-923d-fc539ed31362_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fd662352e3197cb1ed844eeec1fac4224ea191df1ec672dce3ec19afc1d47a03 +size 4177036 diff --git a/2024/Momentum Auxiliary Network for Supervised Local Learning/full.md b/2024/Momentum Auxiliary Network for Supervised Local Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..33b3b73ce7724a59b68cc620213c8488bf4d913c --- /dev/null +++ b/2024/Momentum Auxiliary Network for Supervised Local Learning/full.md @@ -0,0 +1,295 @@ +# Momentum Auxiliary Network for Supervised Local Learning + +Junhao Su $^{1*}$ , Changpeng Cai $^{1*}$ , Feiyu Zhu $^{2*}$ , Chenghao He $^{4}$ + +Xiaojie $\mathrm{Xu}^5$ , Dongzhi Guan $^{1\boxtimes}$ , and Chenyang Si $^{3\boxtimes}$ + +$^{1}$ Southeast University + +2 University of Shanghai for Science and Technology + +3 Nanyang Technological University + +$^{4}$ East China University of Science and Technology + +5 The Hong Kong University of Science and Technology + +Abstract. Deep neural networks conventionally employ end-to-end backpropagation for their training process, which lacks biological credibility and triggers a locking dilemma during network parameter updates, leading to significant GPU memory use. Supervised local learning, which segments the network into multiple local blocks updated by independent auxiliary networks. However, these methods cannot replace end-to-end training due to lower accuracy, as gradients only propagate within their local block, creating a lack of information exchange between blocks. To address this issue and establish information transfer across blocks, we propose a Momentum Auxiliary Network (MAN) that establishes a dynamic interaction mechanism. The MAN leverages an exponential moving average (EMA) of the parameters from adjacent local blocks to enhance information flow. This auxiliary network, updated through EMA, helps bridge the informational gap between blocks. Nevertheless, we observe that directly applying EMA parameters has certain limitations due to feature discrepancies among local blocks. To overcome this, we introduce learnable biases, further boosting performance. We have validated our method on four image classification datasets (CIFAR-10, STL-10, SVHN, ImageNet), attaining superior performance and substantial memory savings. Notably, our method can reduce GPU memory usage by more than $45\%$ on the ImageNet dataset compared to end-to-end training, while achieving higher performance. The Momentum Auxiliary Network thus offers a new perspective for supervised local learning. Our code is available at: https://github.com/JunhaoSuO/MAN. + +Keywords: Local Learning $\cdot$ Image Classification $\cdot$ Momentum Auxiliary Network + +![](images/59658765abdb05b2d54ec752161d340c1c89205054a19306ed2110965081b8ad.jpg) +Fig. 1: Comparison between different methods with MAN and the original methods in terms of accuracy. Results are obtained using ResNet-110 (K=55) on the various datasets. + +# 1 Introduction + +In deep learning, the prevalent use of end-to-end backpropagation [35] is essential for training complex neural networks, which requires heavy computational work for loss evaluation and successive gradient descent through network layers to refine parameters [17, 24, 29]. This method starkly contrasts with the local signal processing of biological synapses [6, 7, 40] and imposes a 'locking' constraint [19] that delays parameter updates until the entire forward and backward pass is complete. Such constraints exacerbate challenges like reduced parallelism and increased GPU memory usage [28, 38], compromising training efficiency and scalability. An emerging alternative to traditional end-to-end (E2E) training is Local Learning [2, 3, 5, 18, 19, 29, 32, 37, 41], which promises to mitigate the drawbacks of traditional methods. These methods divide the network into discrete, gradient-isolated blocks, each of which is updated by its dedicated auxiliary network according to distinct local objectives [3, 32, 37]. This training strategy enables immediate parameter updates upon local error reception, circumventing the sequential update bottleneck of E2E training, thereby significantly enhancing the efficiency of parallel training [13, 22]. Moreover, local learning reduces the demand on GPU memory by only retaining gradients for the local and auxiliary networks, thus eliminating the need to store extensive global gradient information and saving computational resources [28, 38]. However, even with these advantages, there is still a significant performance gap between local learning and traditional E2E methods, especially as the network is divided into more local blocks. Current local learning techniques mainly focus on improving the design of auxiliary network structures [3] and making local loss functions better to close this performance gap [32, 37]. Yet, these improvements do not completely address the inherent short-sightedness of local learning: the separation into blocks can make each part of the network only focus on its local objectives, possibly ig- + +noring the overall objectives of the network. This can lead to the discarding of globally beneficial information due to the lack of inter-block communication. + +In this paper, we introduce the Momentum Auxiliary Network (MAN), a novel network architecture designed to mitigate the inherent limitations of supervised local learning by enhancing inter-block communication. Specifically, MAN not only accepts the current local block as input but also absorbs the parameters of its next block. Upon completion of a forward pass, parameter updates are performed using local gradients refined through Exponential Moving Average (EMA) techniques [14]. This novel approach enables each local block to integrate information from the following block, thereby extending the operational perspective of each block beyond its immediate objectives and further aligning with the global objective of the network. This addresses the critical limitations in local learning frameworks, ensuring a more cohesive alignment with the overarching network objectives. Furthermore, we identify that directly applying EMA parameters is constrained by the feature incongruities across gradient-isolated blocks. To mitigate this, we implement additional learnable bias terms to each auxiliary block, enhancing their ability to share information effectively. The proposed MAN requires only a minimal increase in memory use but significantly improves performance by enhancing the information sharing ability of local blocks. The efficacy of the MAN approach is validated on a suite of benchmark image classification datasets, including CIFAR-10 [21], STL-10 [9], SVHN [30], and ImageNet [12]. The results demonstrate the effectiveness of our method in surpassing the limitations of traditional supervised local learning. + +The contributions of this paper could be summarized as follows: + +- We propose Momentum Auxiliary Network (MAN), designed to facilitate inter-block communication with Exponential Moving Average (EMA), mitigating the short-sightedness issue in conventional supervised local learning techniques and culminating in an elevated overall performance of the network. +- MAN is a versatile, plug-and-play approach that can seamlessly integrate with any supervised local learning technique, markedly amplifying their efficacy while requiring only a negligible increase in memory usage. +- MAN has demonstrated its effectiveness through experiments on benchmark image classification datasets, achieving state-of-the-art performance. Notably, on the ImageNet [12] dataset, it outperforms E2E training while using significantly less memory. + +# 2 Related Work + +# 2.1 Local Learning + +Local learning is first proposed as an innovative deep learning algorithm with the intention of utilizing memory more efficiently and adhering more closely to principles of biological plausibility [10]. This approach emerges as a response to + +![](images/21efcd67c2d31a85b36be6d1f25ab9a282f7e5aedfdad7e1deda6dcee912e8b2.jpg) +Fig. 2: Comparison of (a) end-to-end backpropagation, (b) other supervised local learning methods, and (c) our proposed method. Unlike E2E, supervised local learning separates the network into K gradient-isolated local blocks. + +the limitations presented by global E2E training [16], prompting the development of alternative supervised local learning methodologies. Examples of these methodologies include a differentiable search algorithm, which decouples network blocks for block-level learning and selects a manually assisted network for each local block [33], as well as a self-supervised contrasting loss function that leverages its local learning rules [18, 41]. In supervised local learning, training primarily relies on the design of reasonable supervised local loss functions or the manual construction of auxiliary networks. + +Supervised Local Loss: InfoPro [37] presents a method that involves the reconstruction loss for each hidden layer, coupled with cross-entropy loss. This approach forces the hidden layers to preserve vital input information, thereby preventing the early layers from discarding information relevant to the task. PredSim [32] employs layer-wise loss functions for network training. It makes use of two distinct supervised loss functions to generate local error signals for hidden layers. The optimization of these losses takes place during supervised local learning, thus eliminating the necessity to propagate global errors back to the hidden layers. These methods have achieved advanced performance in the field of supervised local learning, but there is still a significant gap compared to the performance of E2E training. + +Auxiliary Networks: InfoPro [37] incorporates two auxiliary networks into its approach. One network consists of a single convolutional layer followed by two fully-connected layers, which are employed for the cross-entropy loss. The other network utilizes two convolutional layers with upsampling for the reconstruction loss. Alternatively, DGL [3] designs a structure that uses three consecutive convolutional layers connected to a global pooling layer, followed by three consecutive fully connected layers. Moreover, DGL restricts the parameter of the auxiliary network to only $5\%$ of the main network capacity, offering advantages in terms of speed and memory savings. However, these method still perform poorly when the network is divided into a large number of local blocks. + +# 2.2 Alternative Learning Rules to E2E Training + +Due to certain limitations inherent in E2E training, the pursuit of alternative methods to E2E training has gradually garnered attention in recent years [27]. Several efforts have been dedicated to addressing the biologically implausible aspects of E2E training, such as the training methodology of target propagation [4, 23, 25] attempts to train a specialized backward connection by utilizing local reconstruction targets. Additionally, some recent studies have attempted to completely avoid backpropagation in neural networks through forward gradient learning [11, 34]. Meanwhile, the weight transport problem [10]. This has been tackled either by employing distinct feedback connections [1, 26] or by directly disseminating global errors to each hidden unit [8, 31]. While these methods have somewhat alleviated the biological implausibility of E2E training, they still rely on global objectives, which fundamentally differ from biological neural networks that rely on local synapses for information transmission. Furthermore, these methods currently struggle to be effective on large datasets [12]. + +# 3 Method + +# 3.1 Preliminaries + +To set the stage, we first provide a brief overview of traditional end-to-end supervised learning and backpropagation mechanisms. We denote a data sample as $x$ and its corresponding ground-truth label as $y$ . The entire deep network is fragmented into several local blocks. During the forward propagation process, the output from the j-th block serves as the input for the (j+1)-th block, expressed as $x_{j+1} = f_{\theta_j}(x_j)$ . Here, $\theta_j$ symbolizes the parameters of the j-th local block and $f(\cdot)$ represents the forward calculation of the block. We evaluate the loss function $\mathcal{L}(\hat{y}, y)$ between the output of the last block and the ground truth label, and propagate it back iteratively to the preceding blocks. + +Supervised local learning strategies [3, 32, 37] integrate auxiliary networks for local supervision. For each local block, an auxiliary network is affixed. The output from a local block is fed to its corresponding auxiliary network, generating the local supervisory signal as $\hat{y}_j = g_{\gamma_j}(x_{j + 1})$ . Here, $\gamma_{j}$ denotes the parameters of the j-th auxiliary network. + +In this setup, we update the parameters of the j-th auxiliary network and local block, $\gamma_{j},\theta_{j}$ , as follows: + +$$ +\gamma_ {j} \leftarrow \gamma_ {j} - \eta_ {a} \times \nabla_ {\gamma_ {j}} \mathcal {L} (\hat {y} _ {j}, y) \tag {1} +$$ + +$$ +\theta_ {j} \leftarrow \theta_ {j} - \eta_ {l} \times \nabla_ {\theta_ {j}} \mathcal {L} \left(y _ {\hat {j}}, y\right) \tag {2} +$$ + +where $\eta_{a},\eta_{l}$ are the learning rates of the auxiliary networks and local blocks, respectively. By attaching auxiliary networks, each local block becomes gradient-isolated and can be updated with local supervision rather than global backpropagation. + +# 3.2 Momentum Auxiliary Network + +Existing techniques incorporate supervision signals into individual local blocks, which allows for parallel parameter updates and reduces memory overhead. However, this approach can lead to a short-sightedness problem, where each local block neglects the information from subsequent blocks, ultimately resulting in suboptimal final accuracy. + +To address this issue, we propose a comprehensive information exchange module—the Momentum Auxiliary Network (MAN). The MAN employs the Exponential Moving Average (EMA) mechanism [14] as a conduit for transferring information from subsequent blocks to the current block. In the context of MAN, we update the parameters of the j-th auxiliary network and local block as follows: + +$$ +\gamma_ {j} \leftarrow \gamma_ {j} - \eta_ {a} \times \nabla_ {\gamma_ {j}} \mathcal {L} \left(\hat {y} _ {j}, y\right) \tag {3} +$$ + +$$ +\gamma_ {j} \leftarrow E M A \left(\gamma_ {j}, \theta_ {j + 1}\right) \tag {4} +$$ + +$$ +\theta_ {j} \leftarrow \theta_ {j} - \eta_ {l} \times \nabla_ {\theta_ {j}} \mathcal {L} (\hat {y} _ {j}, y) \tag {5} +$$ + +where $\gamma_{j}$ represents the parameters of the j-th auxiliary network and $\theta_{j}$ represents those of the j-th local block. After updating with local gradients, $\gamma_{j}$ undergoes further refinement by incorporating the parameters of the subsequent local block via the EMA, which is a weighted sum operation. + +However, through experimental results, it becomes apparent that directly applying EMA parameters to update the auxiliary network has its limitations, and it provides limited improvements to each local block. Upon analyzing, we find that the features learned between each local block vary to a certain degree, which hampers the effectiveness of the EMA update method. As a consequence, a learnable bias is introduced to augment the learning capability of the hidden layers within the local block, while also compensating for the deficiencies of the EMA update method. + +![](images/3b716f4cdbd8f877ca15468385def71a2b8632344813761894cbd88dfa245ebb.jpg) +Fig. 3: Details of the Momentum Auxiliary Network. Local $(\mathrm{i} + 1)$ represents the $(\mathrm{i} + 1)$ -th gradient-isolated local block, which contains layers from layer m to layer $(\mathrm{m} + \mathrm{n})$ , totaling n+1 layers $(\mathrm{n} \geqslant 0)$ . We only use the parameters of the first layer to ensure a balance in GPU memory usage. + +Therefore, the parameter update method for the j-th auxiliary network can be written as: + +$$ +\left(\gamma_ {j}, b _ {j}\right) \leftarrow \left(\gamma_ {j}, b _ {j}\right) - \eta_ {a} \times \nabla_ {\left(\gamma_ {j}, b _ {j}\right)} \mathcal {L} \left(\hat {y} _ {j}, y\right) \tag {6} +$$ + +$$ +\gamma_ {j} \leftarrow E M A \left(\gamma_ {j}, \theta_ {j + 1}\right) \tag {7} +$$ + +where $(\gamma_j, b_j)$ represent the parameters of the j-th auxiliary network. During this process, we update the parameters $\gamma_j$ of the j-th auxiliary network jointly through the learnable bias $b_j$ and EMA. + +Essentially, the proposed MAN facilitates communication between subsequent blocks through the EMA mechanism [14], effectively resolving the short-sightedness problem inherent in traditional supervised local learning methods. Concurrently, it introduces an independent learnable bias to mitigate information discrepancies caused by feature variations among different local blocks, thereby aligning the output of each block closer to the global target. + +In addition to the two innovative methods mentioned above, MAN demonstrates strong versatility. It can be seamlessly integrated into existing supervised local learning methods, and it excels across various datasets, reflecting its flexibility and generalizability. + +# 4 Experiments + +# 4.1 Experimental Setup + +We conduct experiments using four widely adopted datasets: CIFAR-10 [21], SVHN [30], STL-10 [9], and ImageNet [12], with ResNets [15] of varying depths serving as the network architectures. + +Three state-of-the-art supervised local learning methods are selected for comparison: PredSim [32], DGL [3], and InfoPro [37]. We then partition the network into K local blocks, each containing an approximately equal number of layers. Our proposed Momentum Auxiliary Network is incorporated only into the first K-1 local blocks. The K-th local block does not employ an auxiliary network and is directly connected to the output classifier. We compare these configurations against traditional E2E and original supervised local learning methods to maintain consistent training settings and eliminate confounding variables. + +# 4.2 Implement Details + +In our experiments on CIFAR-10 [21], SVHN [30], and STL-10 [9] datasets with ResNet-32 [15] and ResNet-110 [15], we utilize the SGD optimizer with Nesterov momentum set at 0.9 and an L2 weight decay factor of 1e-4. We employ batch sizes of 1024 for CIFAR-10 and SVHN and 128 for STL-10. The training duration spans 400 epochs, starting with initial learning rates of 0.8 for CIFAR-10 / SVHN and 0.1 for STL-10, following a cosine annealing scheduler [9]. + +In our experiments with ImageNet [12], we adopt different training settings for various architectures. We train VGG13 [36] for 90 epochs with an initial learning rate of 0.025. For ResNet-101 [15], ResNet-152 [15], and ResNeXt-101,32 $\times$ 8d [39], we train them for 90 epochs as well, with initial learning rates of 0.05, 0.05, and 0.025, respectively. We set the batch sizes to 64 for VGG13, 128 for ResNet-101 and ResNet-152, and 64 for ResNeXt-101,32 $\times$ 8d. We maintain consistency with other training configurations as previously described for CIFAR-10 [21]. + +Based on our multiple experimental results, we select the hyperparameter momentum value as 0.995, due to its consistently stable and superior performance. We will provide the experimental results of other momentum hyperparameters in the supplementary materials. + +# 4.3 Results on Image Classification Datasets + +Results on Image Classification Benchmarks: We start by assessing the accuracy performance of our approach using the CIFAR-10 [21], SVHN [30], and STL-10 [9] datasets. We employ ResNet-32 [15], partitioned into 8 and 16 local blocks, and ResNet-110 [15], divided into 32 and 55 local blocks. As illustrated in Table 1, our MAN significantly bolsters the accuracy of all methods. + +Table 1: Performance of different networks with varying numbers of local blocks. The average test error is obtained by 5 experiments. The * means addition of our MAN. + +
DatasetMethodResNet-32ResNet-110
K = 8 (Test Error)K = 16 (Test Error)K = 32 (Test Error)K = 55 (Test Error)
CIFAR-10(E2E(ResNet-32)=6.37,E2E(ResNet-110)=5.42)PredSim [32]20.6222.7122.0824.74
PredSim*14.29(↓6.33)15.58(↓7.13)17.05(↓5.03)18.02(↓6.72)
DGL [3]11.6314.0812.5114.45
DGL*8.42(↓3.21)9.11(↓4.97)9.65(↓2.86)9.73(↓4.72)
InfoPro [37]11.5112.9312.2613.22
InfoPro*9.32(↓2.19)9.65(↓3.28)9.06(↓3.20)9.77(↓3.45)
STL-10(E2E(ResNet-32)=19.35,E2E(ResNet-110)=19.67)PredSim [32]31.9732.9032.0533.27
PredSim*29.97(↓2.00)29.99(↓2.91)30.48(↓1.57)31.55(↓1.72)
DGL [3]25.0527.1425.6728.16
DGL*20.74(↓4.31)21.37(↓5.77)22.54(↓3.13)22.69(↓5.47)
InfoPro [37]27.3229.2828.5829.20
InfoPro*23.17(↓4.15)23.54(↓5.74)24.08(↓4.50)24.74(↓4.46)
SVHN(E2E(ResNet-32)=2.99,E2E(ResNet-110)=2.92)PredSim [32]6.918.089.1210.47
PredSim*5.54(↓1.37)6.39(↓1.69)7.27(↓1.85)8.24(↓2.23)
DGL [3]4.835.055.125.36
DGL*3.80(↓1.03)4.04(↓1.01)4.08(↓1.04)4.52(↓0.84)
InfoPro [37]5.615.975.896.11
InfoPro*4.49(↓1.12)5.19(↓0.78)4.85(↓1.04)4.99(↓1.12)
+ +On the CIFAR-10 dataset [21], our method exhibits considerable improvements in diminishing test errors across various methods. In the relatively shallower network of ResNet-32 (K=16), where individual layers function as gradient-isolated local blocks, we record a reduction in test errors for PredSim [32], DGL [3], and InfoPro [37], from 22.71, 14.08, and 12.93 to 15.58, 9.11, and 9.65 respectively. This translates to a performance enhancement exceeding $25\%$ for all methods. Even though the performance across all methods in the comparatively deeper network, ResNet-110 (K=55), is somewhat inferior due to the inherent need for more global information in such networks, our method still delivers exceptional performance. It achieves approximately a $20\%$ improvement across all methods, underscoring the robust effectiveness of MAN in deeper networks. + +When applied to other datasets, MAN can also reduce the test error of Pred-Sim [32], DGL [3], and InfoPro [37] by at least $5\%$ , $12\%$ , and $16\%$ on the STL-10 [9] dataset. On the SVHN [30] dataset, our improvements over the three + +methods also surpass $20\%$ , $21\%$ , and $13\%$ . As can be seen, the improvement our MAN introduces to all methods is quite remarkable—comparable even to the accuracy of E2E training—and it significantly mitigates the overwhelming performance issue that has continually plagued supervised local learning. + +Results on ImageNet: We further validate the effectiveness of our approach on ImageNet [12] using four networks of varying depths (ResNets [15] and VGG13 [36]). As depicted in Table 2, when we employ VGG13 as the backbone and divide the network into 10 blocks, DGL [3] achieves merely a Top1-Error of 35.60 and a Top5-Error of 14.2, representing a substantial gap when compared to the E2E method. However, with the introduction of our MAN, the Top1-Error reduces by 3.61 points, and the Top5-Error decreases by 3.36 points for DGL. This significant enhancement brings the performance closer to the E2E method. + +As illustrated in Table 2, when we use ResNet-101 [15], ResNet-152 [15], ResNeXt-101, $32 \times 8\mathrm{d}$ [39] as backbones and divide the network into four blocks, the performance of InfoPro [37] is already below that of E2E. After incorporating our MAN, the Top-1 Error of these three backbone networks can be reduced by approximately $6\%$ compared to the original, surpassing the performance of E2E training. These results underscore the effectiveness of our MAN on the large-scale ImageNet [12] dataset, even when using deeper networks. + +Table 2: Results on the validation set of ImageNet + +
NetworkMethodTop1-ErrorTop5-Error
ResNet-101E2E22.035.93
InfoPro(K=2) [37]21.855.89
InfoPro*(K=2)21.65(↓0.20)5.49(↓0.40)
InfoPro(K=4) [37]22.816.54
InfoPro*(K=4)21.73(↓1.08)5.81(↓0.73)
ResNet-152E2E21.605.92
InfoPro(K=2) [37]21.455.84
InfoPro*(K=2)21.23(↓0.22)5.53(↓0.31)
InfoPro(K=4) [37]22.936.71
InfoPro*(K=4)21.59(↓1.34)5.89(↓0.82)
ResNeXt-101, 32 × 8dE2E20.645.40
InfoPro(K=2) [37]20.355.28
InfoPro*(K=2)20.11(↓0.24)5.18(↓0.10)
InfoPro(K=4) [37]21.696.11
InfoPro*(K=4)20.37(↓1.32)5.34(↓0.77)
VGG13E2E28.419.63
DGL [3]35.6014.20
DGL*(K=10)31.99(↓3.61)10.84(↓3.36)
+ +Training-Accuracy Curve Analysis: As depicted in the accuracy-epoch curve in Fig. 4, our Momentum Auxiliary Network consistently outperforms the original method in terms of accuracy throughout the entire training process. This underscores its reliability and stability during the training process of classification tasks. Moreover, our MAN achieves a higher accuracy earlier in the later stages of training and attains stability sooner, indicating a faster convergence rate—a critical attribute in large-scale and complex tasks. + +![](images/b07c01938291e0a448ebb3858e0be9c8c8bb17e68a298b6cdb7c67ccbacb0085.jpg) +Fig. 4: Training-Accuracy curves, the left uses ResNet-32 (K=16) as the backbone, while the right uses ResNet-110 (K=55). Both are utilizing the CIFAR-10 dataset. + +![](images/df3d2ab998231fd5d3874f9afe2e9b6e29d6c8f6cb317212026fe1b08f1dd61a.jpg) + +In summary, our Momentum Auxiliary Network significantly enhances the accuracy of traditional supervised local learning methods by promoting information exchange between local blocks. The beneficial information exchange facilitated by Momentum Auxiliary Network significantly mitigates the pervasive issue of shortsightedness within the supervised local learning domain, thereby offering advantages in terms of training stability and convergence speed. This is crucial for training models more effectively and efficiently. + +Table 3: Comparison of GPU memory usage using InfoPro as baseline with different backbones on the ImageNet dataset. + +
MethodResNet-101 +GPU Memory(GB)ResNet-152 +GPU Memory(GB)ResNeXt-101, 32×8d +GPU Memory(GB)
E2E19.7126.2919.22
InfoPro(K=2) [37]12.06(↓ 38.8%)15.53(↓ 40.9%)11.55(↓ 39.9%)
InfoPro*(K=2)12.32(↓ 37.5%)15.93(↓ 39.4%)11.73(↓ 38.9%)
InfoPro(K=4) [37]10.37(↓ 47.3%)13.48(↓ 48.7%)10.24(↓ 46.7%)
InfoPro*(K=4)10.69(↓ 45.8%)13.91(↓ 47.1%)10.49(↓ 45.5%)
+ +Results on GPU memory requirement: Supervised local learning markedly conserves GPU memory by limiting gradient propagation to within local blocks, thereby ensuring that backpropagation transpires exclusively within each respective local block and its corresponding auxiliary network. This strategy significantly diminishes the storage demands for activation parameters and gradient information, which would typically disseminate across local blocks. Consequently, this facilitates substantial savings in GPU memory utilization. Our analysis, featuring a comprehensive comparison of GPU memory usage on the ImageNet dataset, illustrates that the implementation of a Momentum Auxiliary Network + +Table 4: Abalation study of MAN. (a) Using DGL as baseline and ResNet-32 $(\mathrm{K} = 16)$ as backbone on the CIFAR-10 dataset. (b) Using InfoPro as baseline and ResNet-101 $(\mathrm{K} = 4)$ as backbone on the ImageNet dataset. LB stands for Learnable Bias. + +
EMALBTest Error
××14.08
✓×11.07(↓3.01)
✓✓9.11(↓4.97)
(a)
+ +
EMALBTop1-ErrorTop5-Error
××22.816.54
✓×22.09(↓ 0.72)6.07(↓ 0.47)
✓✓21.73(↓ 1.08)5.81(↓ 0.73)
(b)
+ +can elicit notable performance enhancements while incurring only a minimal uptick in GPU memory consumption. + +As shown in Table 3, it indicates the substantial GPU memory reduction on the ImageNet [12]. When applying our method to InfoPro [37] on ResNet-101 [15], ResNet-152 [15], and ResNeXt-101, $32 \times 8d$ [39], dividing the backbone into two local blocks results in a GPU memory saving of $37.5\%$ to $39.4\%$ . When the network is divided into four local blocks, the degree of GPU memory savings is further enhanced, exceeding $45\%$ . Combined with Table 2, it can be seen that we achieve better performance with almost half the GPU memory required for end-to-end training. Further analysis of the memory usage comparison with the original method shows that our approach only increases the GPU memory by about $1\%$ over the original method, while achieving a performance improvement of over $5\%$ . These results highlight the excellent balance that the MAN method achieves in terms of GPU memory usage and performance enhancement. + +# 4.4 Ablation Studies + +We conduct an ablation study on the CIFAR-10 [21] dataset and ImageNet dataset to assess the impact of the EMA method [14] and learnable bias in the MAN on performance. For this analysis, we use ResNet-32 (K=16) [15] as the backbone and the original DGL [3] method as a comparison baseline. + +As shown in Table 4(a), when we only use the EMA [14] in MAN to promote information exchange with subsequent blocks, without adding a learnable bias, the test error decreases from 14.08 to 11.07. When we further add a learnable bias, the test error decreases from 11.07 to 9.11, it is evident that both the EMA method and learnable bias contribute to performance improvement. We hypothesize that the EMA and learnable bias are somewhat complementary in terms of the features they learn. To validate this, we conduct feature visualization. + +As depicted in Fig. 5(a), we can observe that the original method without the addition of MAN learns very limited and chaotic features. After adding the EMA method alone, the features it learns are clearly more characteristic and defined, indicating that it indeed receives some global features from the information in the subsequent block. However, there are still many blurry features, which may be due to the EMA parameter update method [14], causing information imbalance due to feature differences between different local blocks. When we + +add learnable bias to the original method alone, the features it learns are more concentrated and clear, indicating that the learnable bias significantly enhances the learning ability of the current local block's hidden layer, but it still lacks some details because it does not receive more global information from the subsequent block. When we add the EMA method and learnable bias simultaneously, the learned features are not only clear and specific, but also more comprehensive in detail, proving that their learning abilities are complementary, and greatly improving the performance of the original method. The feature visualization in Fig. 5 verifies our previous thoughts. + +![](images/6a5e04a5245e930abbb02927da9156858b05b05c4cd1fdc37fa7c71ca7236623.jpg) +(a) + +![](images/e86b2ebc2f6adfda5e94bb461ad8384510581d9974c641ed47a94b62988cc62c.jpg) +(b) +Fig.5: Feature maps comparison of (a) Original Method without MAN, (b) MAN with only EMA, (c) MAN with only learnable bias and (d) MAN with both EMA and learnable bias. The feature map are obtained using ResNet-32 $(\mathrm{K} = 16)$ on CIFAR-10. + +![](images/cfd441b347f678abb1f0f8ef3af1fa200eff7a0e5621a1276fcbfe90fcc6b4e5.jpg) +(c) + +![](images/f99499f3083090cf91f95005d7233b93e4bcdb12d86301673c6056e6733178ae.jpg) +(d) + +# 4.5 The Effectiveness of EMA + +To verify the effectiveness of the EMA update method in MAN, we conduct further ablation experiments. As shown in Table 5, using only the parameters of the next layer in MAN results in very limited improvement. In contrast, employing the EMA method to utilize the parameters of the next layer leads to more significant enhancements. This is because the subsequent local block and the current local block learn different features and have different learning objectives, leading to minimal improvement. However, using the EMA method can gradually promote information exchange between local blocks and reduce parameter fluctuations during training, leading to smoother parameter updates. This beneficial interaction brings significant improvements to MAN. + +Table 5: Abalation study of EMA. (a) Using DGL as baseline and ResNet-32 $(\mathrm{K} = 16)$ as backbone on the CIFAR-10 dataset. (b) Using InfoPro as baseline and ResNet-101 $(\mathrm{K} = 4)$ as backbone on the ImageNet dataset. Parameter signifies whether each local block utilizes the parameters of the adjacent next layer connected to it. + +
ParameterEMATestError
××14.08
✓×12.93
✓✓11.07
+ +(a) + +
ParameterEMATop1-ErrorTop5-Error
××22.816.54
✓×22.646.41
✓✓22.096.07
+ +(b) + +# 4.6 Linear Separability Analysis + +To demonstrate the effectiveness of our Momentum Auxiliary Network, we freeze the parameters of the main network and train a classifier for each local block to obtain the classification accuracy of each local block. As depicted in Fig. 6, we use DGL [3] as a baseline, and after integrating our method, the accuracy of the earlier layers decreases, while the accuracy of the middle and later layers significantly improve. The decrease in accuracy of the earlier layers suggests that they have learned more generalized features. Although these features are not beneficial for optimizing local objectives, they are beneficial from a global perspective. The middle and later layers, receiving these globally beneficial features, have seen a significant increase in accuracy. These results illustrate that our proposed method has facilitated information interaction between gradient-isolated local blocks. This addresses the shortsightedness issue existing in the current supervised local learning field and significantly enhances their performance. + +![](images/6db996f1514a558097af01f9586d67f5704f0d5d7d835416605ab9f89d57a5e7.jpg) +Fig. 6: Comparison of layer-wise linear separability across different learning rules on ResNet-32 and ResNet-110. + +![](images/759438b2911d841f2adcd073e32c9fb54746344980fa8fe40b9372997a11ed2c.jpg) + +# 4.7 Representation Similarity Analysis + +To further validate the efficacy of MAN, we conduct a Centered Kernel Alignment (CKA) experiment [20]. CKA serves as a metric to assess the similarity between feature representations. If a method's CKA score is closer to 1 in relation to the E2E training method, it indicates that the method's feature learning process is more aligned with that of E2E training. As depicted in Fig. 7, we use DGL [3] as a baseline and incorporate our method. It can be observed that whether adding the EMA method or learnable bias alone, the CKA scores significantly improve compared to the original method. When both the EMA and learnable bias are used in MAN, the CKA score further improves and performs more stably. Notably, the most significant increase in the CKA score occurs in the early and late layers. In conjunction with our previous analysis of the linear separability experiment, this is because if the early layers' learning method is closer + +to E2E training, they will focus more on learning general features to optimize the global objective, rather than focusing narrowly on local objectives. While this may result in poorer classification capabilities in the early layers, it greatly contributes to the overall performance improvement of the network. Through the analysis of images and experimental results, it can be demonstrated that MAN enables information interaction between gradient-isolated local blocks, solving the myopia problem present in current supervised local learning methods. + +![](images/41e853fbfca688fd107444a9f7f8ce88d84afe36a94bfae8e47b9a49985d920f.jpg) +Fig. 7: Assessment of Similarity in Layer-wise Representations. We use Centered Kernel Alignment (CKA) [20] to quantify the degree of similarity in the layer-wise representations between the E2E backpropagation (BP) and our proposed MAN. + +![](images/8773d9182a59cd09b3ff8c23529e403f422635454fb0f3e22848b615db1e5763.jpg) + +# 5 Conclusion + +This study addresses the performance disparities between traditional supervised local learning and end-to-end (E2E) methods in deep learning. We introduce a versatile Momentum Auxiliary Network to tackle the short-sighted problem in the early optimization work related to supervised local learning, facilitating information exchange between gradient-isolated local blocks. We integrate our Momentum Auxiliary Network into three advanced supervised local learning approaches and evaluate their performance across network architectures with varying depths on four widely adopted datasets. The results demonstrate our method's ability to significantly enhance the ultimate output performance of original supervised local learning methods. Particularly when combined with InfoPro [37], our method significantly reduces GPU memory usage while consistently maintaining performance levels closely aligned with E2E approaches. + +Limitations and future works: Despite the superior performance on large-scale problems like ImageNet [12], our method still performs less accurately than E2E on some conventional image classification datasets. This may be due to the MAN using too few information interaction layers when the network is divided into a larger number of local blocks. In future work, we could explore deepening these information interaction layers to achieve better precision performance. + +# Acknowledgement + +This work was supported by the National Natural Science Foundation of China (52278154), the Natural Science Foundation of Jiangsu (BK20231429), the Fundamental Research Funds for the Central Universities (2242024RCB0008), and assist from the program of Zhishan Young Scholar of Southeast University. + +# References + +1. Akrout, M., Wilson, C., Humphreys, P., Lillicrap, T., Tweed, D.B.: Deep learning without weight transport. Advances in neural information processing systems 32 (2019) +2. Belilovsky, E., Eickenberg, M., Oyallon, E.: Greedy layerwise learning can scale to imagenet. In: International conference on machine learning. pp. 583-593. PMLR (2019) +3. Belilovsky, E., Eickenberg, M., Oyallon, E.: Decoupled greedy learning of cnns. In: International Conference on Machine Learning. pp. 736-745. PMLR (2020) +4. Bengio, Y.: How auto-encoders could provide credit assignment in deep networks via target propagation. arXiv preprint arXiv:1407.7906 (2014) +5. Bengio, Y., Lamblin, P., Popovici, D., Larochelle, H.: Greedy layer-wise training of deep networks. Advances in neural information processing systems 19 (2006) +6. Bengio, Y., Lee, D.H., Bornschein, J., Mesnard, T., Lin, Z.: Towards biologically plausible deep learning. arXiv preprint arXiv:1502.04156 (2015) +7. Caporale, N., Dan, Y.: Spike timing-dependent plasticity: a hebbian learning rule. Annu. Rev. Neurosci. 31, 25-46 (2008) +8. Clark, D., Abbott, L., Chung, S.: Credit assignment through broadcasting a global error vector. Advances in Neural Information Processing Systems 34, 10053-10066 (2021) +9. Coates, A., Ng, A., Lee, H.: An analysis of single-layer networks in unsupervised feature learning. In: Proceedings of the fourteenth international conference on artificial intelligence and statistics. pp. 215-223. JMLR Workshop and Conference Proceedings (2011) +0. Crick, F.: The recent excitement about neural networks. Nature 337(6203), 129-132 (1989) +1. Dellafererra, G., Kreiman, G.: Error-driven input modulation: solving the credit assignment problem without a backward pass. In: International Conference on Machine Learning. pp. 4937-4955. PMLR (2022) +2. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 248-255. IEEE (2009) +3. Gomez, A.N., Key, O., Perlin, K., Gou, S., Frosst, N., Dean, J., Gal, Y.: Interlocking backpropagation: Improving depthwise model-parallelism. The Journal of Machine Learning Research 23(1), 7714-7741 (2022) +4. He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.: Momentum contrast for unsupervised visual representation learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9729-9738 (2020) +5. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016) + +16. Hinton, G.E., Osindero, S., Teh, Y.W.: A fast learning algorithm for deep belief nets. Neural computation 18(7), 1527-1554 (2006) +17. Huang, G., Liu, Z., Van Der Maaten, L., Weinberger, K.Q.: Densely connected convolutional networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4700-4708 (2017) +18. Illing, B., Ventura, J., Bellec, G., Gerstner, W.: Local plasticity rules can learn deep representations using self-supervised contrastive predictions. Advances in Neural Information Processing Systems 34, 30365-30379 (2021) +19. Jaderberg, M., Czarnecki, W.M., Osindero, S., Vinyals, O., Graves, A., Silver, D., Kavukcuoglu, K.: Decoupled neural interfaces using synthetic gradients. In: International conference on machine learning. pp. 1627-1635. PMLR (2017) +20. Kornblith, S., Norouzi, M., Lee, H., Hinton, G.: Similarity of neural network representations revisited. In: International conference on machine learning. pp. 3519-3529. PMLR (2019) +21. Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images (2009) +22. Laskin, M., Metz, L., Nabarro, S., Saroufim, M., Noune, B., Luschi, C., Sohl-Dickstein, J., Abbeel, P.: Parallel training of deep networks with local updates. arXiv preprint arXiv:2012.03837 (2020) +23. Le Cun, Y.: Learning process in an asymmetric threshold network. In: Disordered systems and biological organization, pp. 233-240. Springer (1986) +24. LeCun, Y., Bengio, Y., Hinton, G.: Deep learning. nature 521(7553), 436-444 (2015) +25. Lee, D.H., Zhang, S., Fischer, A., Bengio, Y.: Difference target propagation. In: Machine Learning and Knowledge Discovery in Databases: European Conference, ECML PKDD 2015, Porto, Portugal, September 7-11, 2015, Proceedings, Part I 15. pp. 498-515. Springer (2015) +26. Lillicrap, T.P., Cownden, D., Tweed, D.B., Akerman, C.J.: Random synaptic feedback weights support error backpropagation for deep learning. Nature communications 7(1), 13276 (2016) +27. Lillicrap, T.P., Santoro, A., Marris, L., Akerman, C.J., Hinton, G.: Backpropagation and the brain. Nature Reviews Neuroscience 21(6), 335-346 (2020) +28. Löwe, S., O'Connor, P., Veeling, B.: Putting an end to end-to-end: Gradient-isolated learning of representations. Advances in neural information processing systems 32 (2019) +29. Mostafa, H., Ramesh, V., Cauwenberghs, G.: Deep supervised learning using local errors. Frontiers in neuroscience 12, 608 (2018) +30. Netzer, Y., Wang, T., Coates, A., Bissacco, A., Wu, B., Ng, A.Y.: Reading digits in natural images with unsupervised feature learning (2011) +31. Nokland, A.: Direct feedback alignment provides learning in deep neural networks. Advances in neural information processing systems 29 (2016) +32. Nokland, A., Eidnes, L.H.: Training neural networks with local error signals. In: International conference on machine learning. pp. 4839-4850. PMLR (2019) +33. Pyeon, M., Moon, J., Hahn, T., Kim, G.: Sedona: Search for decoupled neural networks toward greedy block-wise learning. In: International Conference on Learning Representations (2020) +34. Ren, M., Kornblith, S., Liao, R., Hinton, G.: Scaling forward gradient with local losses. arXiv preprint arXiv:2210.03310 (2022) +35. Rumelhart, D.E., Hinton, G.E., Williams, R.J.: Learning representations by backpropagating errors. nature 323(6088), 533-536 (1986) + +36. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556 (2014) +37. Wang, Y., Ni, Z., Song, S., Yang, L., Huang, G.: Revisiting locally supervised learning: an alternative to end-to-end training. arXiv preprint arXiv:2101.10832 (2021) +38. Wu, B., Nair, S., Martin-Martin, R., Fei-Fei, L., Finn, C.: Greedy hierarchical variational autoencoders for large-scale video prediction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2318-2328 (2021) +39. Xie, S., Girshick, R., Dollar, P., Tu, Z., He, K.: Aggregated residual transformations for deep neural networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1492-1500 (2017) +40. Xie, X., Seung, H.S.: Equivalence of backpropagation and contrastive hebbian learning in a layered network. Neural computation 15(2), 441-454 (2003) +41. Xiong, Y., Ren, M., Urtasun, R.: Loco: Local contrastive representation learning. Advances in neural information processing systems 33, 11142-11153 (2020) \ No newline at end of file diff --git a/2024/Momentum Auxiliary Network for Supervised Local Learning/images.zip b/2024/Momentum Auxiliary Network for Supervised Local Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..3aa06be0be9af6c0bc7d5fdbcd24dad04d08e91a --- /dev/null +++ b/2024/Momentum Auxiliary Network for Supervised Local Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:acea2424a4a0d06b5376c20db28bbaecb9311844a0ca13724560ced29ec1fe86 +size 471586 diff --git a/2024/Momentum Auxiliary Network for Supervised Local Learning/layout.json b/2024/Momentum Auxiliary Network for Supervised Local Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..c4dc8a21ae73cf74e74bf4a9590d2db0a7d08404 --- /dev/null +++ b/2024/Momentum Auxiliary Network for Supervised Local Learning/layout.json @@ -0,0 +1,8183 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 146, + 111, + 467, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 111, + 467, + 148 + ], + "spans": [ + { + "bbox": [ + 146, + 111, + 467, + 148 + ], + "type": "text", + "content": "Momentum Auxiliary Network for Supervised Local Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 173, + 171, + 441, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 173, + 171, + 441, + 182 + ], + "spans": [ + { + "bbox": [ + 173, + 171, + 441, + 182 + ], + "type": "text", + "content": "Junhao Su" + }, + { + "bbox": [ + 173, + 171, + 441, + 182 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 173, + 171, + 441, + 182 + ], + "type": "text", + "content": ", Changpeng Cai" + }, + { + "bbox": [ + 173, + 171, + 441, + 182 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 173, + 171, + 441, + 182 + ], + "type": "text", + "content": ", Feiyu Zhu" + }, + { + "bbox": [ + 173, + 171, + 441, + 182 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 173, + 171, + 441, + 182 + ], + "type": "text", + "content": ", Chenghao He" + }, + { + "bbox": [ + 173, + 171, + 441, + 182 + ], + "type": "inline_equation", + "content": "^{4}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 194, + 182, + 419, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 182, + 419, + 196 + ], + "spans": [ + { + "bbox": [ + 194, + 182, + 419, + 196 + ], + "type": "text", + "content": "Xiaojie " + }, + { + "bbox": [ + 194, + 182, + 419, + 196 + ], + "type": "inline_equation", + "content": "\\mathrm{Xu}^5" + }, + { + "bbox": [ + 194, + 182, + 419, + 196 + ], + "type": "text", + "content": ", Dongzhi Guan" + }, + { + "bbox": [ + 194, + 182, + 419, + 196 + ], + "type": "inline_equation", + "content": "^{1\\boxtimes}" + }, + { + "bbox": [ + 194, + 182, + 419, + 196 + ], + "type": "text", + "content": ", and Chenyang Si" + }, + { + "bbox": [ + 194, + 182, + 419, + 196 + ], + "type": "inline_equation", + "content": "^{3\\boxtimes}" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 259, + 204, + 355, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 259, + 204, + 355, + 216 + ], + "spans": [ + { + "bbox": [ + 259, + 204, + 355, + 216 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 259, + 204, + 355, + 216 + ], + "type": "text", + "content": " Southeast University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 198, + 216, + 414, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 216, + 414, + 228 + ], + "spans": [ + { + "bbox": [ + 198, + 216, + 414, + 228 + ], + "type": "text", + "content": "2 University of Shanghai for Science and Technology" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 231, + 227, + 381, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 227, + 381, + 239 + ], + "spans": [ + { + "bbox": [ + 231, + 227, + 381, + 239 + ], + "type": "text", + "content": "3 Nanyang Technological University" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 201, + 238, + 411, + 249 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 238, + 411, + 249 + ], + "spans": [ + { + "bbox": [ + 201, + 238, + 411, + 249 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 201, + 238, + 411, + 249 + ], + "type": "text", + "content": " East China University of Science and Technology" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 192, + 249, + 422, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 192, + 249, + 422, + 260 + ], + "spans": [ + { + "bbox": [ + 192, + 249, + 422, + 260 + ], + "type": "text", + "content": "5 The Hong Kong University of Science and Technology" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 160, + 315, + 455, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 315, + 455, + 579 + ], + "spans": [ + { + "bbox": [ + 160, + 315, + 455, + 579 + ], + "type": "text", + "content": "Abstract. Deep neural networks conventionally employ end-to-end backpropagation for their training process, which lacks biological credibility and triggers a locking dilemma during network parameter updates, leading to significant GPU memory use. Supervised local learning, which segments the network into multiple local blocks updated by independent auxiliary networks. However, these methods cannot replace end-to-end training due to lower accuracy, as gradients only propagate within their local block, creating a lack of information exchange between blocks. To address this issue and establish information transfer across blocks, we propose a Momentum Auxiliary Network (MAN) that establishes a dynamic interaction mechanism. The MAN leverages an exponential moving average (EMA) of the parameters from adjacent local blocks to enhance information flow. This auxiliary network, updated through EMA, helps bridge the informational gap between blocks. Nevertheless, we observe that directly applying EMA parameters has certain limitations due to feature discrepancies among local blocks. To overcome this, we introduce learnable biases, further boosting performance. We have validated our method on four image classification datasets (CIFAR-10, STL-10, SVHN, ImageNet), attaining superior performance and substantial memory savings. Notably, our method can reduce GPU memory usage by more than " + }, + { + "bbox": [ + 160, + 315, + 455, + 579 + ], + "type": "inline_equation", + "content": "45\\%" + }, + { + "bbox": [ + 160, + 315, + 455, + 579 + ], + "type": "text", + "content": " on the ImageNet dataset compared to end-to-end training, while achieving higher performance. The Momentum Auxiliary Network thus offers a new perspective for supervised local learning. Our code is available at: https://github.com/JunhaoSuO/MAN." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 159, + 591, + 453, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 591, + 453, + 615 + ], + "spans": [ + { + "bbox": [ + 159, + 591, + 453, + 615 + ], + "type": "text", + "content": "Keywords: Local Learning " + }, + { + "bbox": [ + 159, + 591, + 453, + 615 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 591, + 453, + 615 + ], + "type": "text", + "content": " Image Classification " + }, + { + "bbox": [ + 159, + 591, + 453, + 615 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 591, + 453, + 615 + ], + "type": "text", + "content": " Momentum Auxiliary Network" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 141, + 632, + 230, + 643 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 632, + 230, + 643 + ], + "spans": [ + { + "bbox": [ + 141, + 632, + 230, + 643 + ], + "type": "text", + "content": "*Equal Contribution." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 141, + 643, + 481, + 654 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 643, + 481, + 654 + ], + "spans": [ + { + "bbox": [ + 141, + 643, + 481, + 654 + ], + "type": "text", + "content": "Corresponding Authors: Chengyang Si (chenyang.si.mail@gmail.com) and Dongzhi" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 141, + 654, + 279, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 654, + 279, + 665 + ], + "spans": [ + { + "bbox": [ + 141, + 654, + 279, + 665 + ], + "type": "text", + "content": "Guan (guandongzhi@seu.edu.cn)." + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 115, + 482, + 260 + ], + "blocks": [ + { + "bbox": [ + 133, + 115, + 482, + 260 + ], + "lines": [ + { + "bbox": [ + 133, + 115, + 482, + 260 + ], + "spans": [ + { + "bbox": [ + 133, + 115, + 482, + 260 + ], + "type": "image", + "image_path": "59658765abdb05b2d54ec752161d340c1c89205054a19306ed2110965081b8ad.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 268, + 482, + 302 + ], + "lines": [ + { + "bbox": [ + 130, + 268, + 482, + 302 + ], + "spans": [ + { + "bbox": [ + 130, + 268, + 482, + 302 + ], + "type": "text", + "content": "Fig. 1: Comparison between different methods with MAN and the original methods in terms of accuracy. Results are obtained using ResNet-110 (K=55) on the various datasets." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 326, + 230, + 339 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 326, + 230, + 339 + ], + "spans": [ + { + "bbox": [ + 133, + 326, + 230, + 339 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 354, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 354, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 482, + 668 + ], + "type": "text", + "content": "In deep learning, the prevalent use of end-to-end backpropagation [35] is essential for training complex neural networks, which requires heavy computational work for loss evaluation and successive gradient descent through network layers to refine parameters [17, 24, 29]. This method starkly contrasts with the local signal processing of biological synapses [6, 7, 40] and imposes a 'locking' constraint [19] that delays parameter updates until the entire forward and backward pass is complete. Such constraints exacerbate challenges like reduced parallelism and increased GPU memory usage [28, 38], compromising training efficiency and scalability. An emerging alternative to traditional end-to-end (E2E) training is Local Learning [2, 3, 5, 18, 19, 29, 32, 37, 41], which promises to mitigate the drawbacks of traditional methods. These methods divide the network into discrete, gradient-isolated blocks, each of which is updated by its dedicated auxiliary network according to distinct local objectives [3, 32, 37]. This training strategy enables immediate parameter updates upon local error reception, circumventing the sequential update bottleneck of E2E training, thereby significantly enhancing the efficiency of parallel training [13, 22]. Moreover, local learning reduces the demand on GPU memory by only retaining gradients for the local and auxiliary networks, thus eliminating the need to store extensive global gradient information and saving computational resources [28, 38]. However, even with these advantages, there is still a significant performance gap between local learning and traditional E2E methods, especially as the network is divided into more local blocks. Current local learning techniques mainly focus on improving the design of auxiliary network structures [3] and making local loss functions better to close this performance gap [32, 37]. Yet, these improvements do not completely address the inherent short-sightedness of local learning: the separation into blocks can make each part of the network only focus on its local objectives, possibly ig-" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 234, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 234, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 234, + 100 + ], + "type": "text", + "content": "Junhao Su et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 139 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 139 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 139 + ], + "type": "text", + "content": "noring the overall objectives of the network. This can lead to the discarding of globally beneficial information due to the lack of inter-block communication." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 391 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 391 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 391 + ], + "type": "text", + "content": "In this paper, we introduce the Momentum Auxiliary Network (MAN), a novel network architecture designed to mitigate the inherent limitations of supervised local learning by enhancing inter-block communication. Specifically, MAN not only accepts the current local block as input but also absorbs the parameters of its next block. Upon completion of a forward pass, parameter updates are performed using local gradients refined through Exponential Moving Average (EMA) techniques [14]. This novel approach enables each local block to integrate information from the following block, thereby extending the operational perspective of each block beyond its immediate objectives and further aligning with the global objective of the network. This addresses the critical limitations in local learning frameworks, ensuring a more cohesive alignment with the overarching network objectives. Furthermore, we identify that directly applying EMA parameters is constrained by the feature incongruities across gradient-isolated blocks. To mitigate this, we implement additional learnable bias terms to each auxiliary block, enhancing their ability to share information effectively. The proposed MAN requires only a minimal increase in memory use but significantly improves performance by enhancing the information sharing ability of local blocks. The efficacy of the MAN approach is validated on a suite of benchmark image classification datasets, including CIFAR-10 [21], STL-10 [9], SVHN [30], and ImageNet [12]. The results demonstrate the effectiveness of our method in surpassing the limitations of traditional supervised local learning." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 146, + 392, + 429, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 392, + 429, + 403 + ], + "spans": [ + { + "bbox": [ + 146, + 392, + 429, + 403 + ], + "type": "text", + "content": "The contributions of this paper could be summarized as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 411, + 479, + 555 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 138, + 411, + 479, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 411, + 479, + 470 + ], + "spans": [ + { + "bbox": [ + 138, + 411, + 479, + 470 + ], + "type": "text", + "content": "- We propose Momentum Auxiliary Network (MAN), designed to facilitate inter-block communication with Exponential Moving Average (EMA), mitigating the short-sightedness issue in conventional supervised local learning techniques and culminating in an elevated overall performance of the network." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 472, + 479, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 472, + 479, + 506 + ], + "spans": [ + { + "bbox": [ + 138, + 472, + 479, + 506 + ], + "type": "text", + "content": "- MAN is a versatile, plug-and-play approach that can seamlessly integrate with any supervised local learning technique, markedly amplifying their efficacy while requiring only a negligible increase in memory usage." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 508, + 479, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 508, + 479, + 555 + ], + "spans": [ + { + "bbox": [ + 138, + 508, + 479, + 555 + ], + "type": "text", + "content": "- MAN has demonstrated its effectiveness through experiments on benchmark image classification datasets, achieving state-of-the-art performance. Notably, on the ImageNet [12] dataset, it outperforms E2E training while using significantly less memory." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 583, + 237, + 596 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 583, + 237, + 596 + ], + "spans": [ + { + "bbox": [ + 132, + 583, + 237, + 596 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 609, + 236, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 609, + 236, + 621 + ], + "spans": [ + { + "bbox": [ + 132, + 609, + 236, + 621 + ], + "type": "text", + "content": "2.1 Local Learning" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 629, + 480, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 480, + 666 + ], + "type": "text", + "content": "Local learning is first proposed as an innovative deep learning algorithm with the intention of utilizing memory more efficiently and adhering more closely to principles of biological plausibility [10]. This approach emerges as a response to" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 197, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 197, + 91, + 448, + 103 + ], + "type": "text", + "content": "Momentum Auxiliary Network for Supervised Local Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 113, + 483, + 232 + ], + "blocks": [ + { + "bbox": [ + 134, + 113, + 483, + 232 + ], + "lines": [ + { + "bbox": [ + 134, + 113, + 483, + 232 + ], + "spans": [ + { + "bbox": [ + 134, + 113, + 483, + 232 + ], + "type": "image", + "image_path": "21efcd67c2d31a85b36be6d1f25ab9a282f7e5aedfdad7e1deda6dcee912e8b2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 242, + 482, + 275 + ], + "lines": [ + { + "bbox": [ + 130, + 242, + 482, + 275 + ], + "spans": [ + { + "bbox": [ + 130, + 242, + 482, + 275 + ], + "type": "text", + "content": "Fig. 2: Comparison of (a) end-to-end backpropagation, (b) other supervised local learning methods, and (c) our proposed method. Unlike E2E, supervised local learning separates the network into K gradient-isolated local blocks." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 308, + 482, + 405 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 308, + 482, + 405 + ], + "spans": [ + { + "bbox": [ + 130, + 308, + 482, + 405 + ], + "type": "text", + "content": "the limitations presented by global E2E training [16], prompting the development of alternative supervised local learning methodologies. Examples of these methodologies include a differentiable search algorithm, which decouples network blocks for block-level learning and selects a manually assisted network for each local block [33], as well as a self-supervised contrasting loss function that leverages its local learning rules [18, 41]. In supervised local learning, training primarily relies on the design of reasonable supervised local loss functions or the manual construction of auxiliary networks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 409, + 482, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 409, + 482, + 541 + ], + "spans": [ + { + "bbox": [ + 130, + 409, + 482, + 541 + ], + "type": "text", + "content": "Supervised Local Loss: InfoPro [37] presents a method that involves the reconstruction loss for each hidden layer, coupled with cross-entropy loss. This approach forces the hidden layers to preserve vital input information, thereby preventing the early layers from discarding information relevant to the task. PredSim [32] employs layer-wise loss functions for network training. It makes use of two distinct supervised loss functions to generate local error signals for hidden layers. The optimization of these losses takes place during supervised local learning, thus eliminating the necessity to propagate global errors back to the hidden layers. These methods have achieved advanced performance in the field of supervised local learning, but there is still a significant gap compared to the performance of E2E training." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "content": "Auxiliary Networks: InfoPro [37] incorporates two auxiliary networks into its approach. One network consists of a single convolutional layer followed by two fully-connected layers, which are employed for the cross-entropy loss. The other network utilizes two convolutional layers with upsampling for the reconstruction loss. Alternatively, DGL [3] designs a structure that uses three consecutive convolutional layers connected to a global pooling layer, followed by three consecutive fully connected layers. Moreover, DGL restricts the parameter of the auxiliary network to only " + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "content": " of the main network capacity, offering advantages in terms of speed and memory savings. However, these method still perform poorly when the network is divided into a large number of local blocks." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 234, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 234, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 234, + 100 + ], + "type": "text", + "content": "Junhao Su et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 131, + 116, + 383, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 116, + 383, + 129 + ], + "spans": [ + { + "bbox": [ + 131, + 116, + 383, + 129 + ], + "type": "text", + "content": "2.2 Alternative Learning Rules to E2E Training" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 136, + 482, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 136, + 482, + 304 + ], + "spans": [ + { + "bbox": [ + 130, + 136, + 482, + 304 + ], + "type": "text", + "content": "Due to certain limitations inherent in E2E training, the pursuit of alternative methods to E2E training has gradually garnered attention in recent years [27]. Several efforts have been dedicated to addressing the biologically implausible aspects of E2E training, such as the training methodology of target propagation [4, 23, 25] attempts to train a specialized backward connection by utilizing local reconstruction targets. Additionally, some recent studies have attempted to completely avoid backpropagation in neural networks through forward gradient learning [11, 34]. Meanwhile, the weight transport problem [10]. This has been tackled either by employing distinct feedback connections [1, 26] or by directly disseminating global errors to each hidden unit [8, 31]. While these methods have somewhat alleviated the biological implausibility of E2E training, they still rely on global objectives, which fundamentally differ from biological neural networks that rely on local synapses for information transmission. Furthermore, these methods currently struggle to be effective on large datasets [12]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 322, + 202, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 322, + 202, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 322, + 202, + 335 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 346, + 228, + 358 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 346, + 228, + 358 + ], + "spans": [ + { + "bbox": [ + 131, + 346, + 228, + 358 + ], + "type": "text", + "content": "3.1 Preliminaries" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 366, + 482, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 366, + 482, + 474 + ], + "spans": [ + { + "bbox": [ + 130, + 366, + 482, + 474 + ], + "type": "text", + "content": "To set the stage, we first provide a brief overview of traditional end-to-end supervised learning and backpropagation mechanisms. We denote a data sample as " + }, + { + "bbox": [ + 130, + 366, + 482, + 474 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 366, + 482, + 474 + ], + "type": "text", + "content": " and its corresponding ground-truth label as " + }, + { + "bbox": [ + 130, + 366, + 482, + 474 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 130, + 366, + 482, + 474 + ], + "type": "text", + "content": ". The entire deep network is fragmented into several local blocks. During the forward propagation process, the output from the j-th block serves as the input for the (j+1)-th block, expressed as " + }, + { + "bbox": [ + 130, + 366, + 482, + 474 + ], + "type": "inline_equation", + "content": "x_{j+1} = f_{\\theta_j}(x_j)" + }, + { + "bbox": [ + 130, + 366, + 482, + 474 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 130, + 366, + 482, + 474 + ], + "type": "inline_equation", + "content": "\\theta_j" + }, + { + "bbox": [ + 130, + 366, + 482, + 474 + ], + "type": "text", + "content": " symbolizes the parameters of the j-th local block and " + }, + { + "bbox": [ + 130, + 366, + 482, + 474 + ], + "type": "inline_equation", + "content": "f(\\cdot)" + }, + { + "bbox": [ + 130, + 366, + 482, + 474 + ], + "type": "text", + "content": " represents the forward calculation of the block. We evaluate the loss function " + }, + { + "bbox": [ + 130, + 366, + 482, + 474 + ], + "type": "inline_equation", + "content": "\\mathcal{L}(\\hat{y}, y)" + }, + { + "bbox": [ + 130, + 366, + 482, + 474 + ], + "type": "text", + "content": " between the output of the last block and the ground truth label, and propagate it back iteratively to the preceding blocks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 474, + 482, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 482, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 482, + 533 + ], + "type": "text", + "content": "Supervised local learning strategies [3, 32, 37] integrate auxiliary networks for local supervision. For each local block, an auxiliary network is affixed. The output from a local block is fed to its corresponding auxiliary network, generating the local supervisory signal as " + }, + { + "bbox": [ + 130, + 474, + 482, + 533 + ], + "type": "inline_equation", + "content": "\\hat{y}_j = g_{\\gamma_j}(x_{j + 1})" + }, + { + "bbox": [ + 130, + 474, + 482, + 533 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 130, + 474, + 482, + 533 + ], + "type": "inline_equation", + "content": "\\gamma_{j}" + }, + { + "bbox": [ + 130, + 474, + 482, + 533 + ], + "type": "text", + "content": " denotes the parameters of the j-th auxiliary network." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 534, + 482, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 534, + 482, + 559 + ], + "spans": [ + { + "bbox": [ + 130, + 534, + 482, + 559 + ], + "type": "text", + "content": "In this setup, we update the parameters of the j-th auxiliary network and local block, " + }, + { + "bbox": [ + 130, + 534, + 482, + 559 + ], + "type": "inline_equation", + "content": "\\gamma_{j},\\theta_{j}" + }, + { + "bbox": [ + 130, + 534, + 482, + 559 + ], + "type": "text", + "content": ", as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 247, + 571, + 480, + 583 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 247, + 571, + 480, + 583 + ], + "spans": [ + { + "bbox": [ + 247, + 571, + 480, + 583 + ], + "type": "interline_equation", + "content": "\\gamma_ {j} \\leftarrow \\gamma_ {j} - \\eta_ {a} \\times \\nabla_ {\\gamma_ {j}} \\mathcal {L} (\\hat {y} _ {j}, y) \\tag {1}", + "image_path": "a4e37914b034f476a6348016da4c66761631622570fa674baf0705102f624f0e.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 249, + 600, + 480, + 613 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 249, + 600, + 480, + 613 + ], + "spans": [ + { + "bbox": [ + 249, + 600, + 480, + 613 + ], + "type": "interline_equation", + "content": "\\theta_ {j} \\leftarrow \\theta_ {j} - \\eta_ {l} \\times \\nabla_ {\\theta_ {j}} \\mathcal {L} \\left(y _ {\\hat {j}}, y\\right) \\tag {2}", + "image_path": "281813987847622073a8c99a5df53b7ef187f322eaff280e99a977c54e0f5c78.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\eta_{a},\\eta_{l}" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": " are the learning rates of the auxiliary networks and local blocks, respectively. By attaching auxiliary networks, each local block becomes gradient-isolated and can be updated with local supervision rather than global backpropagation." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 197, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 197, + 91, + 448, + 103 + ], + "type": "text", + "content": "Momentum Auxiliary Network for Supervised Local Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 317, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 317, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 317, + 129 + ], + "type": "text", + "content": "3.2 Momentum Auxiliary Network" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 137, + 482, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 137, + 482, + 196 + ], + "spans": [ + { + "bbox": [ + 130, + 137, + 482, + 196 + ], + "type": "text", + "content": "Existing techniques incorporate supervision signals into individual local blocks, which allows for parallel parameter updates and reduces memory overhead. However, this approach can lead to a short-sightedness problem, where each local block neglects the information from subsequent blocks, ultimately resulting in suboptimal final accuracy." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 197, + 482, + 257 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 197, + 482, + 257 + ], + "spans": [ + { + "bbox": [ + 130, + 197, + 482, + 257 + ], + "type": "text", + "content": "To address this issue, we propose a comprehensive information exchange module—the Momentum Auxiliary Network (MAN). The MAN employs the Exponential Moving Average (EMA) mechanism [14] as a conduit for transferring information from subsequent blocks to the current block. In the context of MAN, we update the parameters of the j-th auxiliary network and local block as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 247, + 268, + 481, + 282 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 247, + 268, + 481, + 282 + ], + "spans": [ + { + "bbox": [ + 247, + 268, + 481, + 282 + ], + "type": "interline_equation", + "content": "\\gamma_ {j} \\leftarrow \\gamma_ {j} - \\eta_ {a} \\times \\nabla_ {\\gamma_ {j}} \\mathcal {L} \\left(\\hat {y} _ {j}, y\\right) \\tag {3}", + "image_path": "df27f575d0a2fe6cab6eb473f914e3a94a1386c8c72382917d402b5fae1cf829.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 260, + 299, + 481, + 312 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 260, + 299, + 481, + 312 + ], + "spans": [ + { + "bbox": [ + 260, + 299, + 481, + 312 + ], + "type": "interline_equation", + "content": "\\gamma_ {j} \\leftarrow E M A \\left(\\gamma_ {j}, \\theta_ {j + 1}\\right) \\tag {4}", + "image_path": "dbb5bd0b9d304d00b465fb2e4551bb8c1abd9d3304448c080da795007436513d.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 247, + 329, + 481, + 342 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 247, + 329, + 481, + 342 + ], + "spans": [ + { + "bbox": [ + 247, + 329, + 481, + 342 + ], + "type": "interline_equation", + "content": "\\theta_ {j} \\leftarrow \\theta_ {j} - \\eta_ {l} \\times \\nabla_ {\\theta_ {j}} \\mathcal {L} (\\hat {y} _ {j}, y) \\tag {5}", + "image_path": "64bdca8ed2056874d41b7e8a6e75051b313ade825903473c8110cbfb84d699d5.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 347, + 482, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 347, + 482, + 396 + ], + "spans": [ + { + "bbox": [ + 130, + 347, + 482, + 396 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 347, + 482, + 396 + ], + "type": "inline_equation", + "content": "\\gamma_{j}" + }, + { + "bbox": [ + 130, + 347, + 482, + 396 + ], + "type": "text", + "content": " represents the parameters of the j-th auxiliary network and " + }, + { + "bbox": [ + 130, + 347, + 482, + 396 + ], + "type": "inline_equation", + "content": "\\theta_{j}" + }, + { + "bbox": [ + 130, + 347, + 482, + 396 + ], + "type": "text", + "content": " represents those of the j-th local block. After updating with local gradients, " + }, + { + "bbox": [ + 130, + 347, + 482, + 396 + ], + "type": "inline_equation", + "content": "\\gamma_{j}" + }, + { + "bbox": [ + 130, + 347, + 482, + 396 + ], + "type": "text", + "content": " undergoes further refinement by incorporating the parameters of the subsequent local block via the EMA, which is a weighted sum operation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 396, + 298, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 396, + 298, + 586 + ], + "spans": [ + { + "bbox": [ + 130, + 396, + 298, + 586 + ], + "type": "text", + "content": "However, through experimental results, it becomes apparent that directly applying EMA parameters to update the auxiliary network has its limitations, and it provides limited improvements to each local block. Upon analyzing, we find that the features learned between each local block vary to a certain degree, which hampers the effectiveness of the EMA update method. As a consequence, a learnable bias is introduced to augment the learning capability of the hidden layers within the local block, while also compensating for the deficiencies of the EMA update method." + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 305, + 418, + 480, + 501 + ], + "blocks": [ + { + "bbox": [ + 305, + 418, + 480, + 501 + ], + "lines": [ + { + "bbox": [ + 305, + 418, + 480, + 501 + ], + "spans": [ + { + "bbox": [ + 305, + 418, + 480, + 501 + ], + "type": "image", + "image_path": "3b716f4cdbd8f877ca15468385def71a2b8632344813761894cbd88dfa245ebb.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 510, + 482, + 586 + ], + "lines": [ + { + "bbox": [ + 304, + 510, + 482, + 586 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 482, + 586 + ], + "type": "text", + "content": "Fig. 3: Details of the Momentum Auxiliary Network. Local " + }, + { + "bbox": [ + 304, + 510, + 482, + 586 + ], + "type": "inline_equation", + "content": "(\\mathrm{i} + 1)" + }, + { + "bbox": [ + 304, + 510, + 482, + 586 + ], + "type": "text", + "content": " represents the " + }, + { + "bbox": [ + 304, + 510, + 482, + 586 + ], + "type": "inline_equation", + "content": "(\\mathrm{i} + 1)" + }, + { + "bbox": [ + 304, + 510, + 482, + 586 + ], + "type": "text", + "content": "-th gradient-isolated local block, which contains layers from layer m to layer " + }, + { + "bbox": [ + 304, + 510, + 482, + 586 + ], + "type": "inline_equation", + "content": "(\\mathrm{m} + \\mathrm{n})" + }, + { + "bbox": [ + 304, + 510, + 482, + 586 + ], + "type": "text", + "content": ", totaling n+1 layers " + }, + { + "bbox": [ + 304, + 510, + 482, + 586 + ], + "type": "inline_equation", + "content": "(\\mathrm{n} \\geqslant 0)" + }, + { + "bbox": [ + 304, + 510, + 482, + 586 + ], + "type": "text", + "content": ". We only use the parameters of the first layer to ensure a balance in GPU memory usage." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 586, + 482, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 586, + 482, + 610 + ], + "spans": [ + { + "bbox": [ + 130, + 586, + 482, + 610 + ], + "type": "text", + "content": "Therefore, the parameter update method for the j-th auxiliary network can be written as:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 217, + 622, + 481, + 638 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 217, + 622, + 481, + 638 + ], + "spans": [ + { + "bbox": [ + 217, + 622, + 481, + 638 + ], + "type": "interline_equation", + "content": "\\left(\\gamma_ {j}, b _ {j}\\right) \\leftarrow \\left(\\gamma_ {j}, b _ {j}\\right) - \\eta_ {a} \\times \\nabla_ {\\left(\\gamma_ {j}, b _ {j}\\right)} \\mathcal {L} \\left(\\hat {y} _ {j}, y\\right) \\tag {6}", + "image_path": "218128c4c31f4c3dac6c5726e855a6b7ff39883587c2bac13bc9c38af6fde492.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 259, + 653, + 481, + 667 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 259, + 653, + 481, + 667 + ], + "spans": [ + { + "bbox": [ + 259, + 653, + 481, + 667 + ], + "type": "interline_equation", + "content": "\\gamma_ {j} \\leftarrow E M A \\left(\\gamma_ {j}, \\theta_ {j + 1}\\right) \\tag {7}", + "image_path": "0c588b870f3f64d66cd535562712398c4c2fa504de00fd7046e0992831e712eb.jpg" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 235, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 235, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 235, + 101 + ], + "type": "text", + "content": "Junhao Su et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "inline_equation", + "content": "(\\gamma_j, b_j)" + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": " represent the parameters of the j-th auxiliary network. During this process, we update the parameters " + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "inline_equation", + "content": "\\gamma_j" + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": " of the j-th auxiliary network jointly through the learnable bias " + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "inline_equation", + "content": "b_j" + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": " and EMA." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 479, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 479, + 223 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 479, + 223 + ], + "type": "text", + "content": "Essentially, the proposed MAN facilitates communication between subsequent blocks through the EMA mechanism [14], effectively resolving the short-sightedness problem inherent in traditional supervised local learning methods. Concurrently, it introduces an independent learnable bias to mitigate information discrepancies caused by feature variations among different local blocks, thereby aligning the output of each block closer to the global target." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 224, + 479, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 224, + 479, + 272 + ], + "spans": [ + { + "bbox": [ + 130, + 224, + 479, + 272 + ], + "type": "text", + "content": "In addition to the two innovative methods mentioned above, MAN demonstrates strong versatility. It can be seamlessly integrated into existing supervised local learning methods, and it excels across various datasets, reflecting its flexibility and generalizability." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 289, + 230, + 303 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 289, + 230, + 303 + ], + "spans": [ + { + "bbox": [ + 131, + 289, + 230, + 303 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 312, + 262, + 325 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 312, + 262, + 325 + ], + "spans": [ + { + "bbox": [ + 131, + 312, + 262, + 325 + ], + "type": "text", + "content": "4.1 Experimental Setup" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 331, + 479, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 331, + 479, + 366 + ], + "spans": [ + { + "bbox": [ + 130, + 331, + 479, + 366 + ], + "type": "text", + "content": "We conduct experiments using four widely adopted datasets: CIFAR-10 [21], SVHN [30], STL-10 [9], and ImageNet [12], with ResNets [15] of varying depths serving as the network architectures." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 367, + 480, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 367, + 480, + 463 + ], + "spans": [ + { + "bbox": [ + 130, + 367, + 480, + 463 + ], + "type": "text", + "content": "Three state-of-the-art supervised local learning methods are selected for comparison: PredSim [32], DGL [3], and InfoPro [37]. We then partition the network into K local blocks, each containing an approximately equal number of layers. Our proposed Momentum Auxiliary Network is incorporated only into the first K-1 local blocks. The K-th local block does not employ an auxiliary network and is directly connected to the output classifier. We compare these configurations against traditional E2E and original supervised local learning methods to maintain consistent training settings and eliminate confounding variables." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 479, + 254, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 479, + 254, + 491 + ], + "spans": [ + { + "bbox": [ + 131, + 479, + 254, + 491 + ], + "type": "text", + "content": "4.2 Implement Details" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 498, + 479, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 479, + 570 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 479, + 570 + ], + "type": "text", + "content": "In our experiments on CIFAR-10 [21], SVHN [30], and STL-10 [9] datasets with ResNet-32 [15] and ResNet-110 [15], we utilize the SGD optimizer with Nesterov momentum set at 0.9 and an L2 weight decay factor of 1e-4. We employ batch sizes of 1024 for CIFAR-10 and SVHN and 128 for STL-10. The training duration spans 400 epochs, starting with initial learning rates of 0.8 for CIFAR-10 / SVHN and 0.1 for STL-10, following a cosine annealing scheduler [9]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 570, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 480, + 665 + ], + "type": "text", + "content": "In our experiments with ImageNet [12], we adopt different training settings for various architectures. We train VGG13 [36] for 90 epochs with an initial learning rate of 0.025. For ResNet-101 [15], ResNet-152 [15], and ResNeXt-101,32 " + }, + { + "bbox": [ + 130, + 570, + 480, + 665 + ], + "type": "inline_equation", + "content": "\\times" + }, + { + "bbox": [ + 130, + 570, + 480, + 665 + ], + "type": "text", + "content": " 8d [39], we train them for 90 epochs as well, with initial learning rates of 0.05, 0.05, and 0.025, respectively. We set the batch sizes to 64 for VGG13, 128 for ResNet-101 and ResNet-152, and 64 for ResNeXt-101,32 " + }, + { + "bbox": [ + 130, + 570, + 480, + 665 + ], + "type": "inline_equation", + "content": "\\times" + }, + { + "bbox": [ + 130, + 570, + 480, + 665 + ], + "type": "text", + "content": " 8d. We maintain consistency with other training configurations as previously described for CIFAR-10 [21]." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 197, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 197, + 91, + 448, + 103 + ], + "type": "text", + "content": "Momentum Auxiliary Network for Supervised Local Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "content": "Based on our multiple experimental results, we select the hyperparameter momentum value as 0.995, due to its consistently stable and superior performance. We will provide the experimental results of other momentum hyperparameters in the supplementary materials." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 182, + 366, + 194 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 182, + 366, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 182, + 366, + 194 + ], + "type": "text", + "content": "4.3 Results on Image Classification Datasets" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 201, + 482, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 201, + 482, + 262 + ], + "spans": [ + { + "bbox": [ + 130, + 201, + 482, + 262 + ], + "type": "text", + "content": "Results on Image Classification Benchmarks: We start by assessing the accuracy performance of our approach using the CIFAR-10 [21], SVHN [30], and STL-10 [9] datasets. We employ ResNet-32 [15], partitioned into 8 and 16 local blocks, and ResNet-110 [15], divided into 32 and 55 local blocks. As illustrated in Table 1, our MAN significantly bolsters the accuracy of all methods." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 135, + 316, + 482, + 473 + ], + "blocks": [ + { + "bbox": [ + 132, + 283, + 481, + 308 + ], + "lines": [ + { + "bbox": [ + 132, + 283, + 481, + 308 + ], + "spans": [ + { + "bbox": [ + 132, + 283, + 481, + 308 + ], + "type": "text", + "content": "Table 1: Performance of different networks with varying numbers of local blocks. The average test error is obtained by 5 experiments. The * means addition of our MAN." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 316, + 482, + 473 + ], + "lines": [ + { + "bbox": [ + 135, + 316, + 482, + 473 + ], + "spans": [ + { + "bbox": [ + 135, + 316, + 482, + 473 + ], + "type": "table", + "html": "
DatasetMethodResNet-32ResNet-110
K = 8 (Test Error)K = 16 (Test Error)K = 32 (Test Error)K = 55 (Test Error)
CIFAR-10(E2E(ResNet-32)=6.37,E2E(ResNet-110)=5.42)PredSim [32]20.6222.7122.0824.74
PredSim*14.29(↓6.33)15.58(↓7.13)17.05(↓5.03)18.02(↓6.72)
DGL [3]11.6314.0812.5114.45
DGL*8.42(↓3.21)9.11(↓4.97)9.65(↓2.86)9.73(↓4.72)
InfoPro [37]11.5112.9312.2613.22
InfoPro*9.32(↓2.19)9.65(↓3.28)9.06(↓3.20)9.77(↓3.45)
STL-10(E2E(ResNet-32)=19.35,E2E(ResNet-110)=19.67)PredSim [32]31.9732.9032.0533.27
PredSim*29.97(↓2.00)29.99(↓2.91)30.48(↓1.57)31.55(↓1.72)
DGL [3]25.0527.1425.6728.16
DGL*20.74(↓4.31)21.37(↓5.77)22.54(↓3.13)22.69(↓5.47)
InfoPro [37]27.3229.2828.5829.20
InfoPro*23.17(↓4.15)23.54(↓5.74)24.08(↓4.50)24.74(↓4.46)
SVHN(E2E(ResNet-32)=2.99,E2E(ResNet-110)=2.92)PredSim [32]6.918.089.1210.47
PredSim*5.54(↓1.37)6.39(↓1.69)7.27(↓1.85)8.24(↓2.23)
DGL [3]4.835.055.125.36
DGL*3.80(↓1.03)4.04(↓1.01)4.08(↓1.04)4.52(↓0.84)
InfoPro [37]5.615.975.896.11
InfoPro*4.49(↓1.12)5.19(↓0.78)4.85(↓1.04)4.99(↓1.12)
", + "image_path": "4f3d82f689a3e522b8bb8528ab4cf6588240a70062adc67c2439434322184a52.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 498, + 482, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 629 + ], + "type": "text", + "content": "On the CIFAR-10 dataset [21], our method exhibits considerable improvements in diminishing test errors across various methods. In the relatively shallower network of ResNet-32 (K=16), where individual layers function as gradient-isolated local blocks, we record a reduction in test errors for PredSim [32], DGL [3], and InfoPro [37], from 22.71, 14.08, and 12.93 to 15.58, 9.11, and 9.65 respectively. This translates to a performance enhancement exceeding " + }, + { + "bbox": [ + 130, + 498, + 482, + 629 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 130, + 498, + 482, + 629 + ], + "type": "text", + "content": " for all methods. Even though the performance across all methods in the comparatively deeper network, ResNet-110 (K=55), is somewhat inferior due to the inherent need for more global information in such networks, our method still delivers exceptional performance. It achieves approximately a " + }, + { + "bbox": [ + 130, + 498, + 482, + 629 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 130, + 498, + 482, + 629 + ], + "type": "text", + "content": " improvement across all methods, underscoring the robust effectiveness of MAN in deeper networks." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "content": "When applied to other datasets, MAN can also reduce the test error of Pred-Sim [32], DGL [3], and InfoPro [37] by at least " + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "inline_equation", + "content": "12\\%" + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "inline_equation", + "content": "16\\%" + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "content": " on the STL-10 [9] dataset. On the SVHN [30] dataset, our improvements over the three" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 234, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 234, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 234, + 101 + ], + "type": "text", + "content": "Junhao Su et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 164 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 164 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 164 + ], + "type": "text", + "content": "methods also surpass " + }, + { + "bbox": [ + 130, + 116, + 479, + 164 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 130, + 116, + 479, + 164 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 116, + 479, + 164 + ], + "type": "inline_equation", + "content": "21\\%" + }, + { + "bbox": [ + 130, + 116, + 479, + 164 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 116, + 479, + 164 + ], + "type": "inline_equation", + "content": "13\\%" + }, + { + "bbox": [ + 130, + 116, + 479, + 164 + ], + "type": "text", + "content": ". As can be seen, the improvement our MAN introduces to all methods is quite remarkable—comparable even to the accuracy of E2E training—and it significantly mitigates the overwhelming performance issue that has continually plagued supervised local learning." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 166, + 480, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 166, + 480, + 262 + ], + "spans": [ + { + "bbox": [ + 130, + 166, + 480, + 262 + ], + "type": "text", + "content": "Results on ImageNet: We further validate the effectiveness of our approach on ImageNet [12] using four networks of varying depths (ResNets [15] and VGG13 [36]). As depicted in Table 2, when we employ VGG13 as the backbone and divide the network into 10 blocks, DGL [3] achieves merely a Top1-Error of 35.60 and a Top5-Error of 14.2, representing a substantial gap when compared to the E2E method. However, with the introduction of our MAN, the Top1-Error reduces by 3.61 points, and the Top5-Error decreases by 3.36 points for DGL. This significant enhancement brings the performance closer to the E2E method." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 263, + 481, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 263, + 481, + 348 + ], + "spans": [ + { + "bbox": [ + 130, + 263, + 481, + 348 + ], + "type": "text", + "content": "As illustrated in Table 2, when we use ResNet-101 [15], ResNet-152 [15], ResNeXt-101, " + }, + { + "bbox": [ + 130, + 263, + 481, + 348 + ], + "type": "inline_equation", + "content": "32 \\times 8\\mathrm{d}" + }, + { + "bbox": [ + 130, + 263, + 481, + 348 + ], + "type": "text", + "content": " [39] as backbones and divide the network into four blocks, the performance of InfoPro [37] is already below that of E2E. After incorporating our MAN, the Top-1 Error of these three backbone networks can be reduced by approximately " + }, + { + "bbox": [ + 130, + 263, + 481, + 348 + ], + "type": "inline_equation", + "content": "6\\%" + }, + { + "bbox": [ + 130, + 263, + 481, + 348 + ], + "type": "text", + "content": " compared to the original, surpassing the performance of E2E training. These results underscore the effectiveness of our MAN on the large-scale ImageNet [12] dataset, even when using deeper networks." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 143, + 393, + 470, + 552 + ], + "blocks": [ + { + "bbox": [ + 201, + 373, + 412, + 385 + ], + "lines": [ + { + "bbox": [ + 201, + 373, + 412, + 385 + ], + "spans": [ + { + "bbox": [ + 201, + 373, + 412, + 385 + ], + "type": "text", + "content": "Table 2: Results on the validation set of ImageNet" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 143, + 393, + 470, + 552 + ], + "lines": [ + { + "bbox": [ + 143, + 393, + 470, + 552 + ], + "spans": [ + { + "bbox": [ + 143, + 393, + 470, + 552 + ], + "type": "table", + "html": "
NetworkMethodTop1-ErrorTop5-Error
ResNet-101E2E22.035.93
InfoPro(K=2) [37]21.855.89
InfoPro*(K=2)21.65(↓0.20)5.49(↓0.40)
InfoPro(K=4) [37]22.816.54
InfoPro*(K=4)21.73(↓1.08)5.81(↓0.73)
ResNet-152E2E21.605.92
InfoPro(K=2) [37]21.455.84
InfoPro*(K=2)21.23(↓0.22)5.53(↓0.31)
InfoPro(K=4) [37]22.936.71
InfoPro*(K=4)21.59(↓1.34)5.89(↓0.82)
ResNeXt-101, 32 × 8dE2E20.645.40
InfoPro(K=2) [37]20.355.28
InfoPro*(K=2)20.11(↓0.24)5.18(↓0.10)
InfoPro(K=4) [37]21.696.11
InfoPro*(K=4)20.37(↓1.32)5.34(↓0.77)
VGG13E2E28.419.63
DGL [3]35.6014.20
DGL*(K=10)31.99(↓3.61)10.84(↓3.36)
", + "image_path": "580e507a3e07a6eb64b7d026a5fde080f06d3b55364e2ca96f02c81de23ec32c.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "Training-Accuracy Curve Analysis: As depicted in the accuracy-epoch curve in Fig. 4, our Momentum Auxiliary Network consistently outperforms the original method in terms of accuracy throughout the entire training process. This underscores its reliability and stability during the training process of classification tasks. Moreover, our MAN achieves a higher accuracy earlier in the later stages of training and attains stability sooner, indicating a faster convergence rate—a critical attribute in large-scale and complex tasks." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 197, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 197, + 91, + 448, + 102 + ], + "type": "text", + "content": "Momentum Auxiliary Network for Supervised Local Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 141, + 128, + 294, + 244 + ], + "blocks": [ + { + "bbox": [ + 141, + 128, + 294, + 244 + ], + "lines": [ + { + "bbox": [ + 141, + 128, + 294, + 244 + ], + "spans": [ + { + "bbox": [ + 141, + 128, + 294, + 244 + ], + "type": "image", + "image_path": "b07c01938291e0a448ebb3858e0be9c8c8bb17e68a298b6cdb7c67ccbacb0085.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 254, + 482, + 277 + ], + "lines": [ + { + "bbox": [ + 131, + 254, + 482, + 277 + ], + "spans": [ + { + "bbox": [ + 131, + 254, + 482, + 277 + ], + "type": "text", + "content": "Fig. 4: Training-Accuracy curves, the left uses ResNet-32 (K=16) as the backbone, while the right uses ResNet-110 (K=55). Both are utilizing the CIFAR-10 dataset." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 313, + 128, + 465, + 243 + ], + "blocks": [ + { + "bbox": [ + 313, + 128, + 465, + 243 + ], + "lines": [ + { + "bbox": [ + 313, + 128, + 465, + 243 + ], + "spans": [ + { + "bbox": [ + 313, + 128, + 465, + 243 + ], + "type": "image", + "image_path": "df3d2ab998231fd5d3874f9afe2e9b6e29d6c8f6cb317212026fe1b08f1dd61a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 300, + 482, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 300, + 482, + 385 + ], + "spans": [ + { + "bbox": [ + 130, + 300, + 482, + 385 + ], + "type": "text", + "content": "In summary, our Momentum Auxiliary Network significantly enhances the accuracy of traditional supervised local learning methods by promoting information exchange between local blocks. The beneficial information exchange facilitated by Momentum Auxiliary Network significantly mitigates the pervasive issue of shortsightedness within the supervised local learning domain, thereby offering advantages in terms of training stability and convergence speed. This is crucial for training models more effectively and efficiently." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 145, + 437, + 466, + 533 + ], + "blocks": [ + { + "bbox": [ + 131, + 405, + 482, + 429 + ], + "lines": [ + { + "bbox": [ + 131, + 405, + 482, + 429 + ], + "spans": [ + { + "bbox": [ + 131, + 405, + 482, + 429 + ], + "type": "text", + "content": "Table 3: Comparison of GPU memory usage using InfoPro as baseline with different backbones on the ImageNet dataset." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 145, + 437, + 466, + 533 + ], + "lines": [ + { + "bbox": [ + 145, + 437, + 466, + 533 + ], + "spans": [ + { + "bbox": [ + 145, + 437, + 466, + 533 + ], + "type": "table", + "html": "
MethodResNet-101\nGPU Memory(GB)ResNet-152\nGPU Memory(GB)ResNeXt-101, 32×8d\nGPU Memory(GB)
E2E19.7126.2919.22
InfoPro(K=2) [37]12.06(↓ 38.8%)15.53(↓ 40.9%)11.55(↓ 39.9%)
InfoPro*(K=2)12.32(↓ 37.5%)15.93(↓ 39.4%)11.73(↓ 38.9%)
InfoPro(K=4) [37]10.37(↓ 47.3%)13.48(↓ 48.7%)10.24(↓ 46.7%)
InfoPro*(K=4)10.69(↓ 45.8%)13.91(↓ 47.1%)10.49(↓ 45.5%)
", + "image_path": "ec04e57e122a5123e26bcd488997ca24aaf363e36720f62d19a5682b3eb8f0ab.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": "Results on GPU memory requirement: Supervised local learning markedly conserves GPU memory by limiting gradient propagation to within local blocks, thereby ensuring that backpropagation transpires exclusively within each respective local block and its corresponding auxiliary network. This strategy significantly diminishes the storage demands for activation parameters and gradient information, which would typically disseminate across local blocks. Consequently, this facilitates substantial savings in GPU memory utilization. Our analysis, featuring a comprehensive comparison of GPU memory usage on the ImageNet dataset, illustrates that the implementation of a Momentum Auxiliary Network" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 234, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 234, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 234, + 100 + ], + "type": "text", + "content": "Junhao Su et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 133, + 156, + 239, + 216 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 479, + 148 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 479, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 479, + 148 + ], + "type": "text", + "content": "Table 4: Abalation study of MAN. (a) Using DGL as baseline and ResNet-32 " + }, + { + "bbox": [ + 132, + 114, + 479, + 148 + ], + "type": "inline_equation", + "content": "(\\mathrm{K} = 16)" + }, + { + "bbox": [ + 132, + 114, + 479, + 148 + ], + "type": "text", + "content": " as backbone on the CIFAR-10 dataset. (b) Using InfoPro as baseline and ResNet-101 " + }, + { + "bbox": [ + 132, + 114, + 479, + 148 + ], + "type": "inline_equation", + "content": "(\\mathrm{K} = 4)" + }, + { + "bbox": [ + 132, + 114, + 479, + 148 + ], + "type": "text", + "content": " as backbone on the ImageNet dataset. LB stands for Learnable Bias." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 156, + 239, + 216 + ], + "lines": [ + { + "bbox": [ + 133, + 156, + 239, + 216 + ], + "spans": [ + { + "bbox": [ + 133, + 156, + 239, + 216 + ], + "type": "table", + "html": "
EMALBTest Error
××14.08
✓×11.07(↓3.01)
✓✓9.11(↓4.97)
(a)
", + "image_path": "e2eb95355a30cf53f58dc887b70013841e734ecb8adc3080590c32728f2664ff.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 306, + 157, + 468, + 216 + ], + "blocks": [ + { + "bbox": [ + 306, + 157, + 468, + 216 + ], + "lines": [ + { + "bbox": [ + 306, + 157, + 468, + 216 + ], + "spans": [ + { + "bbox": [ + 306, + 157, + 468, + 216 + ], + "type": "table", + "html": "
EMALBTop1-ErrorTop5-Error
××22.816.54
✓×22.09(↓ 0.72)6.07(↓ 0.47)
✓✓21.73(↓ 1.08)5.81(↓ 0.73)
(b)
", + "image_path": "cd15f42c2505970c794118f68718ff70719e972a1612ae1cbdaaf7ae424e2159.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 239, + 479, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 239, + 479, + 263 + ], + "spans": [ + { + "bbox": [ + 130, + 239, + 479, + 263 + ], + "type": "text", + "content": "can elicit notable performance enhancements while incurring only a minimal uptick in GPU memory consumption." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 264, + 480, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 264, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 130, + 264, + 480, + 407 + ], + "type": "text", + "content": "As shown in Table 3, it indicates the substantial GPU memory reduction on the ImageNet [12]. When applying our method to InfoPro [37] on ResNet-101 [15], ResNet-152 [15], and ResNeXt-101, " + }, + { + "bbox": [ + 130, + 264, + 480, + 407 + ], + "type": "inline_equation", + "content": "32 \\times 8d" + }, + { + "bbox": [ + 130, + 264, + 480, + 407 + ], + "type": "text", + "content": " [39], dividing the backbone into two local blocks results in a GPU memory saving of " + }, + { + "bbox": [ + 130, + 264, + 480, + 407 + ], + "type": "inline_equation", + "content": "37.5\\%" + }, + { + "bbox": [ + 130, + 264, + 480, + 407 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 264, + 480, + 407 + ], + "type": "inline_equation", + "content": "39.4\\%" + }, + { + "bbox": [ + 130, + 264, + 480, + 407 + ], + "type": "text", + "content": ". When the network is divided into four local blocks, the degree of GPU memory savings is further enhanced, exceeding " + }, + { + "bbox": [ + 130, + 264, + 480, + 407 + ], + "type": "inline_equation", + "content": "45\\%" + }, + { + "bbox": [ + 130, + 264, + 480, + 407 + ], + "type": "text", + "content": ". Combined with Table 2, it can be seen that we achieve better performance with almost half the GPU memory required for end-to-end training. Further analysis of the memory usage comparison with the original method shows that our approach only increases the GPU memory by about " + }, + { + "bbox": [ + 130, + 264, + 480, + 407 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 130, + 264, + 480, + 407 + ], + "type": "text", + "content": " over the original method, while achieving a performance improvement of over " + }, + { + "bbox": [ + 130, + 264, + 480, + 407 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 130, + 264, + 480, + 407 + ], + "type": "text", + "content": ". These results highlight the excellent balance that the MAN method achieves in terms of GPU memory usage and performance enhancement." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 427, + 245, + 438 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 427, + 245, + 438 + ], + "spans": [ + { + "bbox": [ + 132, + 427, + 245, + 438 + ], + "type": "text", + "content": "4.4 Ablation Studies" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 449, + 480, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 480, + 496 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 480, + 496 + ], + "type": "text", + "content": "We conduct an ablation study on the CIFAR-10 [21] dataset and ImageNet dataset to assess the impact of the EMA method [14] and learnable bias in the MAN on performance. For this analysis, we use ResNet-32 (K=16) [15] as the backbone and the original DGL [3] method as a comparison baseline." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 498, + 480, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 480, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 480, + 581 + ], + "type": "text", + "content": "As shown in Table 4(a), when we only use the EMA [14] in MAN to promote information exchange with subsequent blocks, without adding a learnable bias, the test error decreases from 14.08 to 11.07. When we further add a learnable bias, the test error decreases from 11.07 to 9.11, it is evident that both the EMA method and learnable bias contribute to performance improvement. We hypothesize that the EMA and learnable bias are somewhat complementary in terms of the features they learn. To validate this, we conduct feature visualization." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "type": "text", + "content": "As depicted in Fig. 5(a), we can observe that the original method without the addition of MAN learns very limited and chaotic features. After adding the EMA method alone, the features it learns are clearly more characteristic and defined, indicating that it indeed receives some global features from the information in the subsequent block. However, there are still many blurry features, which may be due to the EMA parameter update method [14], causing information imbalance due to feature differences between different local blocks. When we" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 197, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 197, + 91, + 448, + 102 + ], + "type": "text", + "content": "Momentum Auxiliary Network for Supervised Local Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": "add learnable bias to the original method alone, the features it learns are more concentrated and clear, indicating that the learnable bias significantly enhances the learning ability of the current local block's hidden layer, but it still lacks some details because it does not receive more global information from the subsequent block. When we add the EMA method and learnable bias simultaneously, the learned features are not only clear and specific, but also more comprehensive in detail, proving that their learning abilities are complementary, and greatly improving the performance of the original method. The feature visualization in Fig. 5 verifies our previous thoughts." + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 149, + 251, + 206, + 306 + ], + "blocks": [ + { + "bbox": [ + 149, + 251, + 206, + 306 + ], + "lines": [ + { + "bbox": [ + 149, + 251, + 206, + 306 + ], + "spans": [ + { + "bbox": [ + 149, + 251, + 206, + 306 + ], + "type": "image", + "image_path": "6a5e04a5245e930abbb02927da9156858b05b05c4cd1fdc37fa7c71ca7236623.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 171, + 312, + 184, + 323 + ], + "lines": [ + { + "bbox": [ + 171, + 312, + 184, + 323 + ], + "spans": [ + { + "bbox": [ + 171, + 312, + 184, + 323 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 235, + 251, + 291, + 306 + ], + "blocks": [ + { + "bbox": [ + 235, + 251, + 291, + 306 + ], + "lines": [ + { + "bbox": [ + 235, + 251, + 291, + 306 + ], + "spans": [ + { + "bbox": [ + 235, + 251, + 291, + 306 + ], + "type": "image", + "image_path": "e86b2ebc2f6adfda5e94bb461ad8384510581d9974c641ed47a94b62988cc62c.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 257, + 312, + 270, + 323 + ], + "lines": [ + { + "bbox": [ + 257, + 312, + 270, + 323 + ], + "spans": [ + { + "bbox": [ + 257, + 312, + 270, + 323 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 331, + 482, + 365 + ], + "lines": [ + { + "bbox": [ + 130, + 331, + 482, + 365 + ], + "spans": [ + { + "bbox": [ + 130, + 331, + 482, + 365 + ], + "type": "text", + "content": "Fig.5: Feature maps comparison of (a) Original Method without MAN, (b) MAN with only EMA, (c) MAN with only learnable bias and (d) MAN with both EMA and learnable bias. The feature map are obtained using ResNet-32 " + }, + { + "bbox": [ + 130, + 331, + 482, + 365 + ], + "type": "inline_equation", + "content": "(\\mathrm{K} = 16)" + }, + { + "bbox": [ + 130, + 331, + 482, + 365 + ], + "type": "text", + "content": " on CIFAR-10." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 321, + 251, + 377, + 306 + ], + "blocks": [ + { + "bbox": [ + 321, + 251, + 377, + 306 + ], + "lines": [ + { + "bbox": [ + 321, + 251, + 377, + 306 + ], + "spans": [ + { + "bbox": [ + 321, + 251, + 377, + 306 + ], + "type": "image", + "image_path": "cfd441b347f678abb1f0f8ef3af1fa200eff7a0e5621a1276fcbfe90fcc6b4e5.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 344, + 312, + 356, + 323 + ], + "lines": [ + { + "bbox": [ + 344, + 312, + 356, + 323 + ], + "spans": [ + { + "bbox": [ + 344, + 312, + 356, + 323 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 408, + 251, + 463, + 306 + ], + "blocks": [ + { + "bbox": [ + 408, + 251, + 463, + 306 + ], + "lines": [ + { + "bbox": [ + 408, + 251, + 463, + 306 + ], + "spans": [ + { + "bbox": [ + 408, + 251, + 463, + 306 + ], + "type": "image", + "image_path": "f99499f3083090cf91f95005d7233b93e4bcdb12d86301673c6056e6733178ae.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 429, + 311, + 443, + 323 + ], + "lines": [ + { + "bbox": [ + 429, + 311, + 443, + 323 + ], + "spans": [ + { + "bbox": [ + 429, + 311, + 443, + 323 + ], + "type": "text", + "content": "(d)" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 390, + 292, + 402 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 390, + 292, + 402 + ], + "spans": [ + { + "bbox": [ + 131, + 390, + 292, + 402 + ], + "type": "text", + "content": "4.5 The Effectiveness of EMA" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 410, + 482, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 410, + 482, + 531 + ], + "spans": [ + { + "bbox": [ + 130, + 410, + 482, + 531 + ], + "type": "text", + "content": "To verify the effectiveness of the EMA update method in MAN, we conduct further ablation experiments. As shown in Table 5, using only the parameters of the next layer in MAN results in very limited improvement. In contrast, employing the EMA method to utilize the parameters of the next layer leads to more significant enhancements. This is because the subsequent local block and the current local block learn different features and have different learning objectives, leading to minimal improvement. However, using the EMA method can gradually promote information exchange between local blocks and reduce parameter fluctuations during training, leading to smoother parameter updates. This beneficial interaction brings significant improvements to MAN." + } + ] + } + ], + "index": 13 + }, + { + "type": "table", + "bbox": [ + 133, + 604, + 251, + 652 + ], + "blocks": [ + { + "bbox": [ + 130, + 551, + 482, + 596 + ], + "lines": [ + { + "bbox": [ + 130, + 551, + 482, + 596 + ], + "spans": [ + { + "bbox": [ + 130, + 551, + 482, + 596 + ], + "type": "text", + "content": "Table 5: Abalation study of EMA. (a) Using DGL as baseline and ResNet-32 " + }, + { + "bbox": [ + 130, + 551, + 482, + 596 + ], + "type": "inline_equation", + "content": "(\\mathrm{K} = 16)" + }, + { + "bbox": [ + 130, + 551, + 482, + 596 + ], + "type": "text", + "content": " as backbone on the CIFAR-10 dataset. (b) Using InfoPro as baseline and ResNet-101 " + }, + { + "bbox": [ + 130, + 551, + 482, + 596 + ], + "type": "inline_equation", + "content": "(\\mathrm{K} = 4)" + }, + { + "bbox": [ + 130, + 551, + 482, + 596 + ], + "type": "text", + "content": " as backbone on the ImageNet dataset. Parameter signifies whether each local block utilizes the parameters of the adjacent next layer connected to it." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 604, + 251, + 652 + ], + "lines": [ + { + "bbox": [ + 133, + 604, + 251, + 652 + ], + "spans": [ + { + "bbox": [ + 133, + 604, + 251, + 652 + ], + "type": "table", + "html": "
ParameterEMATestError
××14.08
✓×12.93
✓✓11.07
", + "image_path": "46b62afa0abb81d15ce5a5cd531fdbc3b17ede73bf9922d85d3ace06fdbb81dc.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_body" + } + ], + "index": 15 + }, + { + "type": "table", + "bbox": [ + 306, + 604, + 473, + 652 + ], + "blocks": [ + { + "bbox": [ + 186, + 652, + 197, + 662 + ], + "lines": [ + { + "bbox": [ + 186, + 652, + 197, + 662 + ], + "spans": [ + { + "bbox": [ + 186, + 652, + 197, + 662 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 306, + 604, + 473, + 652 + ], + "lines": [ + { + "bbox": [ + 306, + 604, + 473, + 652 + ], + "spans": [ + { + "bbox": [ + 306, + 604, + 473, + 652 + ], + "type": "table", + "html": "
ParameterEMATop1-ErrorTop5-Error
××22.816.54
✓×22.646.41
✓✓22.096.07
", + "image_path": "da282c873bf314ae3e071496196c78c221ed3675ba80279aa6d185ca4a1fad4d.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "table_body" + } + ], + "index": 17 + }, + { + "bbox": [ + 383, + 652, + 395, + 662 + ], + "lines": [ + { + "bbox": [ + 383, + 652, + 395, + 662 + ], + "spans": [ + { + "bbox": [ + 383, + 652, + 395, + 662 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "type": "text", + "content": "Junhao Su et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 302, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 302, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 302, + 129 + ], + "type": "text", + "content": "4.6 Linear Separability Analysis" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 136, + 482, + 292 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 136, + 482, + 292 + ], + "spans": [ + { + "bbox": [ + 130, + 136, + 482, + 292 + ], + "type": "text", + "content": "To demonstrate the effectiveness of our Momentum Auxiliary Network, we freeze the parameters of the main network and train a classifier for each local block to obtain the classification accuracy of each local block. As depicted in Fig. 6, we use DGL [3] as a baseline, and after integrating our method, the accuracy of the earlier layers decreases, while the accuracy of the middle and later layers significantly improve. The decrease in accuracy of the earlier layers suggests that they have learned more generalized features. Although these features are not beneficial for optimizing local objectives, they are beneficial from a global perspective. The middle and later layers, receiving these globally beneficial features, have seen a significant increase in accuracy. These results illustrate that our proposed method has facilitated information interaction between gradient-isolated local blocks. This addresses the shortsightedness issue existing in the current supervised local learning field and significantly enhances their performance." + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 151, + 313, + 305, + 444 + ], + "blocks": [ + { + "bbox": [ + 151, + 313, + 305, + 444 + ], + "lines": [ + { + "bbox": [ + 151, + 313, + 305, + 444 + ], + "spans": [ + { + "bbox": [ + 151, + 313, + 305, + 444 + ], + "type": "image", + "image_path": "6db996f1514a558097af01f9586d67f5704f0d5d7d835416605ab9f89d57a5e7.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 453, + 482, + 475 + ], + "lines": [ + { + "bbox": [ + 130, + 453, + 482, + 475 + ], + "spans": [ + { + "bbox": [ + 130, + 453, + 482, + 475 + ], + "type": "text", + "content": "Fig. 6: Comparison of layer-wise linear separability across different learning rules on ResNet-32 and ResNet-110." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 309, + 314, + 462, + 444 + ], + "blocks": [ + { + "bbox": [ + 309, + 314, + 462, + 444 + ], + "lines": [ + { + "bbox": [ + 309, + 314, + 462, + 444 + ], + "spans": [ + { + "bbox": [ + 309, + 314, + 462, + 444 + ], + "type": "image", + "image_path": "759438b2911d841f2adcd073e32c9fb54746344980fa8fe40b9372997a11ed2c.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 502, + 336, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 502, + 336, + 514 + ], + "spans": [ + { + "bbox": [ + 132, + 502, + 336, + 514 + ], + "type": "text", + "content": "4.7 Representation Similarity Analysis" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": "To further validate the efficacy of MAN, we conduct a Centered Kernel Alignment (CKA) experiment [20]. CKA serves as a metric to assess the similarity between feature representations. If a method's CKA score is closer to 1 in relation to the E2E training method, it indicates that the method's feature learning process is more aligned with that of E2E training. As depicted in Fig. 7, we use DGL [3] as a baseline and incorporate our method. It can be observed that whether adding the EMA method or learnable bias alone, the CKA scores significantly improve compared to the original method. When both the EMA and learnable bias are used in MAN, the CKA score further improves and performs more stably. Notably, the most significant increase in the CKA score occurs in the early and late layers. In conjunction with our previous analysis of the linear separability experiment, this is because if the early layers' learning method is closer" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 197, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 197, + 91, + 447, + 102 + ], + "type": "text", + "content": "Momentum Auxiliary Network for Supervised Local Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": "to E2E training, they will focus more on learning general features to optimize the global objective, rather than focusing narrowly on local objectives. While this may result in poorer classification capabilities in the early layers, it greatly contributes to the overall performance improvement of the network. Through the analysis of images and experimental results, it can be demonstrated that MAN enables information interaction between gradient-isolated local blocks, solving the myopia problem present in current supervised local learning methods." + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 151, + 222, + 306, + 353 + ], + "blocks": [ + { + "bbox": [ + 151, + 222, + 306, + 353 + ], + "lines": [ + { + "bbox": [ + 151, + 222, + 306, + 353 + ], + "spans": [ + { + "bbox": [ + 151, + 222, + 306, + 353 + ], + "type": "image", + "image_path": "41e853fbfca688fd107444a9f7f8ce88d84afe36a94bfae8e47b9a49985d920f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 362, + 482, + 396 + ], + "lines": [ + { + "bbox": [ + 130, + 362, + 482, + 396 + ], + "spans": [ + { + "bbox": [ + 130, + 362, + 482, + 396 + ], + "type": "text", + "content": "Fig. 7: Assessment of Similarity in Layer-wise Representations. We use Centered Kernel Alignment (CKA) [20] to quantify the degree of similarity in the layer-wise representations between the E2E backpropagation (BP) and our proposed MAN." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 307, + 222, + 463, + 353 + ], + "blocks": [ + { + "bbox": [ + 307, + 222, + 463, + 353 + ], + "lines": [ + { + "bbox": [ + 307, + 222, + 463, + 353 + ], + "spans": [ + { + "bbox": [ + 307, + 222, + 463, + 353 + ], + "type": "image", + "image_path": "8773d9182a59cd09b3ff8c23529e403f422635454fb0f3e22848b615db1e5763.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 423, + 220, + 435 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 423, + 220, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 423, + 220, + 435 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 449, + 482, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 482, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 482, + 593 + ], + "type": "text", + "content": "This study addresses the performance disparities between traditional supervised local learning and end-to-end (E2E) methods in deep learning. We introduce a versatile Momentum Auxiliary Network to tackle the short-sighted problem in the early optimization work related to supervised local learning, facilitating information exchange between gradient-isolated local blocks. We integrate our Momentum Auxiliary Network into three advanced supervised local learning approaches and evaluate their performance across network architectures with varying depths on four widely adopted datasets. The results demonstrate our method's ability to significantly enhance the ultimate output performance of original supervised local learning methods. Particularly when combined with InfoPro [37], our method significantly reduces GPU memory usage while consistently maintaining performance levels closely aligned with E2E approaches." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "content": "Limitations and future works: Despite the superior performance on large-scale problems like ImageNet [12], our method still performs less accurately than E2E on some conventional image classification datasets. This may be due to the MAN using too few information interaction layers when the network is divided into a larger number of local blocks. In future work, we could explore deepening these information interaction layers to achieve better precision performance." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "type": "text", + "content": "Junhao Su et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 139, + 480, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 139, + 480, + 186 + ], + "spans": [ + { + "bbox": [ + 132, + 139, + 480, + 186 + ], + "type": "text", + "content": "This work was supported by the National Natural Science Foundation of China (52278154), the Natural Science Foundation of Jiangsu (BK20231429), the Fundamental Research Funds for the Central Universities (2242024RCB0008), and assist from the program of Zhishan Young Scholar of Southeast University." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 205, + 197, + 217 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 205, + 197, + 217 + ], + "spans": [ + { + "bbox": [ + 133, + 205, + 197, + 217 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 228, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 138, + 228, + 481, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 228, + 481, + 262 + ], + "spans": [ + { + "bbox": [ + 138, + 228, + 481, + 262 + ], + "type": "text", + "content": "1. Akrout, M., Wilson, C., Humphreys, P., Lillicrap, T., Tweed, D.B.: Deep learning without weight transport. Advances in neural information processing systems 32 (2019)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 262, + 481, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 262, + 481, + 294 + ], + "spans": [ + { + "bbox": [ + 138, + 262, + 481, + 294 + ], + "type": "text", + "content": "2. Belilovsky, E., Eickenberg, M., Oyallon, E.: Greedy layerwise learning can scale to imagenet. In: International conference on machine learning. pp. 583-593. PMLR (2019)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 295, + 481, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 295, + 481, + 316 + ], + "spans": [ + { + "bbox": [ + 138, + 295, + 481, + 316 + ], + "type": "text", + "content": "3. Belilovsky, E., Eickenberg, M., Oyallon, E.: Decoupled greedy learning of cnns. In: International Conference on Machine Learning. pp. 736-745. PMLR (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 317, + 481, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 317, + 481, + 338 + ], + "spans": [ + { + "bbox": [ + 138, + 317, + 481, + 338 + ], + "type": "text", + "content": "4. Bengio, Y.: How auto-encoders could provide credit assignment in deep networks via target propagation. arXiv preprint arXiv:1407.7906 (2014)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 339, + 481, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 339, + 481, + 360 + ], + "spans": [ + { + "bbox": [ + 138, + 339, + 481, + 360 + ], + "type": "text", + "content": "5. Bengio, Y., Lamblin, P., Popovici, D., Larochelle, H.: Greedy layer-wise training of deep networks. Advances in neural information processing systems 19 (2006)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 361, + 481, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 361, + 481, + 381 + ], + "spans": [ + { + "bbox": [ + 138, + 361, + 481, + 381 + ], + "type": "text", + "content": "6. Bengio, Y., Lee, D.H., Bornschein, J., Mesnard, T., Lin, Z.: Towards biologically plausible deep learning. arXiv preprint arXiv:1502.04156 (2015)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 383, + 481, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 383, + 481, + 403 + ], + "spans": [ + { + "bbox": [ + 138, + 383, + 481, + 403 + ], + "type": "text", + "content": "7. Caporale, N., Dan, Y.: Spike timing-dependent plasticity: a hebbian learning rule. Annu. Rev. Neurosci. 31, 25-46 (2008)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 404, + 481, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 404, + 481, + 436 + ], + "spans": [ + { + "bbox": [ + 138, + 404, + 481, + 436 + ], + "type": "text", + "content": "8. Clark, D., Abbott, L., Chung, S.: Credit assignment through broadcasting a global error vector. Advances in Neural Information Processing Systems 34, 10053-10066 (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 437, + 481, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 437, + 481, + 480 + ], + "spans": [ + { + "bbox": [ + 138, + 437, + 481, + 480 + ], + "type": "text", + "content": "9. Coates, A., Ng, A., Lee, H.: An analysis of single-layer networks in unsupervised feature learning. In: Proceedings of the fourteenth international conference on artificial intelligence and statistics. pp. 215-223. JMLR Workshop and Conference Proceedings (2011)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 481, + 481, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 481, + 481, + 502 + ], + "spans": [ + { + "bbox": [ + 138, + 481, + 481, + 502 + ], + "type": "text", + "content": "0. Crick, F.: The recent excitement about neural networks. Nature 337(6203), 129-132 (1989)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 502, + 481, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 502, + 481, + 534 + ], + "spans": [ + { + "bbox": [ + 138, + 502, + 481, + 534 + ], + "type": "text", + "content": "1. Dellafererra, G., Kreiman, G.: Error-driven input modulation: solving the credit assignment problem without a backward pass. In: International Conference on Machine Learning. pp. 4937-4955. PMLR (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 535, + 481, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 535, + 481, + 567 + ], + "spans": [ + { + "bbox": [ + 138, + 535, + 481, + 567 + ], + "type": "text", + "content": "2. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 248-255. IEEE (2009)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 567, + 481, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 567, + 481, + 600 + ], + "spans": [ + { + "bbox": [ + 138, + 567, + 481, + 600 + ], + "type": "text", + "content": "3. Gomez, A.N., Key, O., Perlin, K., Gou, S., Frosst, N., Dean, J., Gal, Y.: Interlocking backpropagation: Improving depthwise model-parallelism. The Journal of Machine Learning Research 23(1), 7714-7741 (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 601, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 601, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 601, + 481, + 632 + ], + "type": "text", + "content": "4. He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.: Momentum contrast for unsupervised visual representation learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9729-9738 (2020)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 633, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 633, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 633, + 481, + 665 + ], + "type": "text", + "content": "5. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 197, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 197, + 91, + 447, + 102 + ], + "type": "text", + "content": "Momentum Auxiliary Network for Supervised Local Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 138 + ], + "type": "text", + "content": "16. Hinton, G.E., Osindero, S., Teh, Y.W.: A fast learning algorithm for deep belief nets. Neural computation 18(7), 1527-1554 (2006)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 139, + 481, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 139, + 481, + 171 + ], + "spans": [ + { + "bbox": [ + 133, + 139, + 481, + 171 + ], + "type": "text", + "content": "17. Huang, G., Liu, Z., Van Der Maaten, L., Weinberger, K.Q.: Densely connected convolutional networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4700-4708 (2017)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 172, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 172, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 133, + 172, + 481, + 205 + ], + "type": "text", + "content": "18. Illing, B., Ventura, J., Bellec, G., Gerstner, W.: Local plasticity rules can learn deep representations using self-supervised contrastive predictions. Advances in Neural Information Processing Systems 34, 30365-30379 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 205, + 481, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 481, + 237 + ], + "type": "text", + "content": "19. Jaderberg, M., Czarnecki, W.M., Osindero, S., Vinyals, O., Graves, A., Silver, D., Kavukcuoglu, K.: Decoupled neural interfaces using synthetic gradients. In: International conference on machine learning. pp. 1627-1635. PMLR (2017)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 238, + 481, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 238, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 238, + 481, + 270 + ], + "type": "text", + "content": "20. Kornblith, S., Norouzi, M., Lee, H., Hinton, G.: Similarity of neural network representations revisited. In: International conference on machine learning. pp. 3519-3529. PMLR (2019)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 271, + 481, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 481, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 481, + 293 + ], + "type": "text", + "content": "21. Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images (2009)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 293, + 481, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 481, + 326 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 481, + 326 + ], + "type": "text", + "content": "22. Laskin, M., Metz, L., Nabarro, S., Saroufim, M., Noune, B., Luschi, C., Sohl-Dickstein, J., Abbeel, P.: Parallel training of deep networks with local updates. arXiv preprint arXiv:2012.03837 (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 326, + 481, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 481, + 347 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 481, + 347 + ], + "type": "text", + "content": "23. Le Cun, Y.: Learning process in an asymmetric threshold network. In: Disordered systems and biological organization, pp. 233-240. Springer (1986)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 348, + 481, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 348, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 348, + 481, + 369 + ], + "type": "text", + "content": "24. LeCun, Y., Bengio, Y., Hinton, G.: Deep learning. nature 521(7553), 436-444 (2015)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 369, + 481, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 481, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 481, + 413 + ], + "type": "text", + "content": "25. Lee, D.H., Zhang, S., Fischer, A., Bengio, Y.: Difference target propagation. In: Machine Learning and Knowledge Discovery in Databases: European Conference, ECML PKDD 2015, Porto, Portugal, September 7-11, 2015, Proceedings, Part I 15. pp. 498-515. Springer (2015)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 414, + 481, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 414, + 481, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 414, + 481, + 445 + ], + "type": "text", + "content": "26. Lillicrap, T.P., Cownden, D., Tweed, D.B., Akerman, C.J.: Random synaptic feedback weights support error backpropagation for deep learning. Nature communications 7(1), 13276 (2016)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 447, + 481, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 447, + 481, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 447, + 481, + 468 + ], + "type": "text", + "content": "27. Lillicrap, T.P., Santoro, A., Marris, L., Akerman, C.J., Hinton, G.: Backpropagation and the brain. Nature Reviews Neuroscience 21(6), 335-346 (2020)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 468, + 481, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 481, + 501 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 481, + 501 + ], + "type": "text", + "content": "28. Löwe, S., O'Connor, P., Veeling, B.: Putting an end to end-to-end: Gradient-isolated learning of representations. Advances in neural information processing systems 32 (2019)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 502, + 481, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 502, + 481, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 502, + 481, + 522 + ], + "type": "text", + "content": "29. Mostafa, H., Ramesh, V., Cauwenberghs, G.: Deep supervised learning using local errors. Frontiers in neuroscience 12, 608 (2018)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 523, + 481, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 523, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 523, + 481, + 544 + ], + "type": "text", + "content": "30. Netzer, Y., Wang, T., Coates, A., Bissacco, A., Wu, B., Ng, A.Y.: Reading digits in natural images with unsupervised feature learning (2011)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 545, + 481, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 545, + 481, + 567 + ], + "spans": [ + { + "bbox": [ + 132, + 545, + 481, + 567 + ], + "type": "text", + "content": "31. Nokland, A.: Direct feedback alignment provides learning in deep neural networks. Advances in neural information processing systems 29 (2016)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 567, + 481, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 481, + 589 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 481, + 589 + ], + "type": "text", + "content": "32. Nokland, A., Eidnes, L.H.: Training neural networks with local error signals. In: International conference on machine learning. pp. 4839-4850. PMLR (2019)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 590, + 481, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 590, + 481, + 621 + ], + "spans": [ + { + "bbox": [ + 132, + 590, + 481, + 621 + ], + "type": "text", + "content": "33. Pyeon, M., Moon, J., Hahn, T., Kim, G.: Sedona: Search for decoupled neural networks toward greedy block-wise learning. In: International Conference on Learning Representations (2020)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 622, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 622, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 622, + 481, + 643 + ], + "type": "text", + "content": "34. Ren, M., Kornblith, S., Liao, R., Hinton, G.: Scaling forward gradient with local losses. arXiv preprint arXiv:2210.03310 (2022)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 644, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 644, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 644, + 481, + 666 + ], + "type": "text", + "content": "35. Rumelhart, D.E., Hinton, G.E., Williams, R.J.: Learning representations by backpropagating errors. nature 323(6088), 533-536 (1986)" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "type": "text", + "content": "Junhao Su et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 293 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "type": "text", + "content": "36. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556 (2014)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 138, + 482, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 138, + 482, + 171 + ], + "spans": [ + { + "bbox": [ + 130, + 138, + 482, + 171 + ], + "type": "text", + "content": "37. Wang, Y., Ni, Z., Song, S., Yang, L., Huang, G.: Revisiting locally supervised learning: an alternative to end-to-end training. arXiv preprint arXiv:2101.10832 (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 172, + 482, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 172, + 482, + 215 + ], + "spans": [ + { + "bbox": [ + 130, + 172, + 482, + 215 + ], + "type": "text", + "content": "38. Wu, B., Nair, S., Martin-Martin, R., Fei-Fei, L., Finn, C.: Greedy hierarchical variational autoencoders for large-scale video prediction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2318-2328 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 216, + 482, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 216, + 482, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 216, + 482, + 248 + ], + "type": "text", + "content": "39. Xie, S., Girshick, R., Dollar, P., Tu, Z., He, K.: Aggregated residual transformations for deep neural networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1492-1500 (2017)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 249, + 482, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 249, + 482, + 270 + ], + "spans": [ + { + "bbox": [ + 130, + 249, + 482, + 270 + ], + "type": "text", + "content": "40. Xie, X., Seung, H.S.: Equivalence of backpropagation and contrastive hebbian learning in a layered network. Neural computation 15(2), 441-454 (2003)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 271, + 482, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 271, + 482, + 293 + ], + "spans": [ + { + "bbox": [ + 130, + 271, + 482, + 293 + ], + "type": "text", + "content": "41. Xiong, Y., Ren, M., Urtasun, R.: Loco: Local contrastive representation learning. Advances in neural information processing systems 33, 11142-11153 (2020)" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 197, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 197, + 91, + 447, + 102 + ], + "type": "text", + "content": "Momentum Auxiliary Network for Supervised Local Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Mono-ViFI_ A Unified Learning Framework for Self-supervised Single- and Multi-frame Monocular Depth Estimation/47dcf543-78f7-47b3-955c-778904eede62_content_list.json b/2024/Mono-ViFI_ A Unified Learning Framework for Self-supervised Single- and Multi-frame Monocular Depth Estimation/47dcf543-78f7-47b3-955c-778904eede62_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..7feba8a6f0f893362f8801df699d9f4e94a98477 --- /dev/null +++ b/2024/Mono-ViFI_ A Unified Learning Framework for Self-supervised Single- and Multi-frame Monocular Depth Estimation/47dcf543-78f7-47b3-955c-778904eede62_content_list.json @@ -0,0 +1,1893 @@ +[ + { + "type": "text", + "text": "Mono-ViFi: A Unified Learning Framework for Self-supervised Single- and Multi-frame Monocular Depth Estimation", + "text_level": 1, + "bbox": [ + 233, + 140, + 769, + 209 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jinfeng Liu, Lingtong Kong\\*, Bo Li, Zerong Wang, Hong Gu, and Jinwei Chen", + "bbox": [ + 215, + 234, + 785, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "vivo Mobile Communication Co., Ltd, China", + "bbox": [ + 349, + 262, + 651, + 276 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{liujinfeng, ltkong, libra, wangzerong, guhong, jinwei.chen}@vivo.com", + "bbox": [ + 254, + 277, + 750, + 291 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Self-supervised monocular depth estimation has gathered notable interest since it can liberate training from dependency on depth annotations. In monocular video training case, recent methods only conduct view synthesis between existing camera views, leading to insufficient guidance. To tackle this, we try to synthesize more virtual camera views by flow-based video frame interpolation (VFI), termed as temporal augmentation. For multi-frame inference, to sidestep the problem of dynamic objects encountered by explicit geometry-based methods like ManyDepth, we return to the feature fusion paradigm and design a VFI-assisted multi-frame fusion module to align and aggregate multi-frame features, using motion and occlusion information obtained by the flow-based VFI model. Finally, we construct a unified self-supervised learning framework, named Mono-ViFi, to bilaterally connect single- and multiframe depth. In this framework, spatial data augmentation through image affine transformation is incorporated for data diversity, along with a triplet depth consistency loss for regularization. The single- and multiframe models can share weights, making our framework compact and memory-efficient. Extensive experiments demonstrate that our method can bring significant improvements to current advanced architectures. Source code is available at https://github.com/LiuJF1226/Mono-ViFi.", + "bbox": [ + 261, + 327, + 740, + 604 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Monocular depth estimation $\\cdot$ Self-supervised learning $\\cdot$ Data augmentation $\\cdot$ Video frame interpolation", + "bbox": [ + 261, + 618, + 745, + 646 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 217, + 672, + 375, + 688 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Monocular depth estimation (MDE) is a fundamental 3D vision task, which finds applications in robotics, autonomous driving and VR/AR. To reduce the dependence on depth annotations, self-supervised MDE [1,2,15,17,46,63,66] has risen as a convenient and practical solution, which uses monocular videos, stereo pairs, or even both to generate supervision through view synthesis, where an image photometric reconstruction loss is minimized. Compared to stereo training case [15,17,59] with known camera pose, learning from monocular videos [18,58, 66] is more challenging since it requires an extra pose network to estimate camera ego-motion. Even so, it is much easier to collect monocular videos compared to", + "bbox": [ + 212, + 703, + 787, + 840 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "stereo image pairs, making monocular training more attractive. Therefore, this paper also focuses on self-supervised training with unlabeled monocular videos.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent works have made various attempts to improve accuracy, like introducing masking techniques [18, 66] and leveraging segmentation priors [28, 31]. However, they only conduct view synthesis between existing fixed camera views within video datasets, leading to insufficient guidance especially in invisible regions by occlusion or disappearance. Considering that video data possesses temporal continuity, an intuitive idea is to generate novel frames, i.e., virtual camera views in the temporal dimension. We term it as temporal augmentation, which can be achieved by video frame interpolation (VFI), as shown in Fig. 1(a).", + "bbox": [ + 212, + 181, + 785, + 303 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "From another perspective, multi-frame reasoning paradigm have recently been researched because multiple frames can be available for reasoning in many application scenarios. ManyDepth [58] proposes to use cost volume constructed from multi-frame matching features. It explicitly reasons about geometry based on static scene assumption, which can be easily affected by dynamic objects and occlusion. DynamicDepth [14] attempts to tackle dynamic scenes, but relies on depth priors and segmentation masks of moving objects when reasoning. To sidestep the issue of dynamic objects in multi-frame inference encountered by these methods, we return to the traditional paradigm [42, 49], which aggregates multi-frame features directly. However, direct feature fusion without alignment harms the learning of temporal representations, since feature embeddings at the same location in neighbor frames are usually irrelevant, making it difficult for models to leverage the correlation of multiple frames. Hence, we regard feature alignment as a reasonable and necessary step before aggregation. As revealed in several video enhancement tasks like video super-resolution [6, 7] and frame interpolation [26, 29, 34, 36], inter-frame feature alignment can be implemented by warping through optical flow. This inspires us to employ a flow-based VFI model to concurrently conduct temporal augmentation and provide explicit correspondence for feature alignment. Moreover, the VFI model can reason out occlusion mask between neighbor frames, helping to alleviate inter-frame occlusion problem in feature fusion. Based on above analysis, we design a VFI-assisted multi-frame fusion module for multi-frame depth estimation, as depicted in Fig. 2.", + "bbox": [ + 212, + 306, + 787, + 638 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Finally, we build a unified framework to bilaterally connect single- and multiframe depth, as demonstrated in Fig. 1(b). The direct way is by depth consistency, usually for geometric regularization or distillation in self-supervised MDE. ManyDepth conducts unilateral distillation from single-frame model to multiframe one. DynamicDepth extends it to a bilateral mutual reinforcement, but only in the regions of moving objects. Differently, our framework applies the bilateral consistency on the whole image. For an original target image, we use a standard view depth consistency (SVDC) loss between its two depth maps predicted from the single- and multi-frame models. We also include classic data augmentation strategies, categorized as spatial augmentation. Specifically, we combine resizing-cropping with image rotation, forming a generic affine transformation. Then, we give the formulation of rectified camera pose for affine-augmented view synthesis using the cue of relative object size [4, 55]. For the", + "bbox": [ + 212, + 643, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "J. Liu et al.", + "bbox": [ + 271, + 114, + 351, + 127 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/9a3d1e5d9591e39c21ee0f4c6d3c34d41e74bc997a12e6f691af7d204581947b.jpg", + "image_caption": [ + "(a) Temporal Augmentation with VFI" + ], + "image_footnote": [], + "bbox": [ + 217, + 141, + 521, + 294 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/ae377703be20437e014938cb17aad97ac308c67f058d7004bde80aa93374d895.jpg", + "image_caption": [ + "(b) Learning Framework $(I_{t}$ for example", + "Fig. 1: Overview of Mono-ViFI. (a) We achieve temporal augmentation by a flow-based VFI model $\\mathcal{H}$ , which also reasons out intermediate optical flow and occlusion mask for multi-frame inference. (b) For target image $I_{t}$ , we obtain its single-frame depth $D_{t}$ and multi-frame depth $D_{t}^{m}$ . Also, $I_{t}$ is augmented to $\\widetilde{I}_{t}$ through affine transformation $\\mathcal{A}$ . We then calculate the depth $\\widehat{D}_t$ of $\\widetilde{I}_t$ and inversely convert it to the original view, generating $\\widehat{D}_t$ . Finally, we enforce a triplet depth consistency loss $L_{tc}(t)$ among the three depth maps, including a standard view depth consistency loss $L_{sv}(t)$ and two scale-aware depth consistency losses, $L_{sa}(t)$ and $L_{sa}^{m}(t)$ . Note that each depth map also corresponds to a photometric loss and a smoothness loss, which are omitted here." + ], + "image_footnote": [], + "bbox": [ + 522, + 142, + 784, + 292 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "depth map from augmented target view, we enforce two scale-aware depth consistency (SADC) losses between it and above two standard view depths, respectively. SADC stems from the relative size cue, which assumes that when an image is zoomed up by a factor, the depth decreases by the same factor. It guides depth model to learn consistent geometric scale relationship from scene variations. The SVDC loss and two SADC losses are termed as a triplet depth consistency loss.", + "bbox": [ + 212, + 476, + 787, + 566 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our framework is named Mono-ViFI, since the self-supervised Monocular depth is largely boosted by Video Frame Interpolation. To our best knowledge, we are the first to leverage VFI for the improvement of self-supervised MDE. The main contributions of this paper are summarized as follows:", + "bbox": [ + 212, + 566, + 787, + 628 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose Mono-ViFi, a unified learning framework to bilaterally connect self-supervised single- and multi-frame MDE. In Mono-ViFi, the single- and multi-frame models can share weights, making the unified framework more compact and memory-efficient.", + "- We conduct temporal augmentation through flow-based VFI, and develop a novel VFI-assisted multi-frame fusion module, which utilizes motion and occlusion from flow-based VFI model to improve multi-frame depth.", + "- We incorporate image affine transformation as spatial augmentation, along with a triplet depth consistency loss for regularization and distillation.", + "- Experiments show that Mono-ViFI can achieve state-of-the-art performance. Moreover, our approach is orthogonal to architecture design, allowing current advanced depth networks to be further improved by seamlessly integrating them with Mono-ViFI." + ], + "bbox": [ + 223, + 641, + 785, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Mono-ViFi: A Unified Framework for Self-supervised Monocular Depth", + "bbox": [ + 256, + 113, + 732, + 130 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 114, + 785, + 127 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 143, + 387, + 160 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.1 Self-supervised Monocular Depth Estimation", + "text_level": 1, + "bbox": [ + 215, + 172, + 633, + 188 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Self-supervised learning paradigm for monocular depth estimation can be categorized into stereo training and monocular training according to the data source.", + "bbox": [ + 212, + 194, + 782, + 223 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Stereo training leverages the left-right consistency in stereo image pairs. Garg et al. [15] first consider the training as a view synthesis problem with a photometric consistency loss between stereo pairs. Building upon this idea, Godard et al. [17] introduce a left-right disparity consistency loss. Then, it is further extended by using depth hints [59], data augmentation [43, 55], knowledge distillation [43, 45, 55] and plane-based scene representation [3, 55].", + "bbox": [ + 212, + 224, + 784, + 314 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Monocular training leverages the temporal consistency in monocular video frames. Zhou et al. [66] propose a pioneering method which jointly learns depth and camera ego-motion from monocular videos in a self-supervised way. Then, Bian et al. [5] propose a scale-consistent depth and ego-motion estimation approach by adding a depth consistency loss, which reduces scale drift of estimated depth but decreases the accuracy. Later, Godard et al. [18] propose Monodepth2, a strong baseline for most follow-up works, which employs a minimum reprojection loss to mitigate occlusion issues and an auto-masking loss to filter out moving objects with similar velocity as the camera. Furthermore, some works introduce multi-task learning to combine depth estimation with optical flow [9, 32, 47, 60], semantic segmentation [27, 28, 40] and instance segmentation [31]. Others [21-23, 35, 37, 39, 62, 64, 65] try to improve the network architectures for depth estimation. However, these methods only utilize the geometric constrains between existing fixed camera views in video datasets. This insufficient guidance may block current depth models from unleashing their full capabilities. To address it, we conduct temporal augmentation through VFI in this paper.", + "bbox": [ + 212, + 316, + 785, + 556 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Multi-frame Monocular Depth", + "text_level": 1, + "bbox": [ + 215, + 575, + 513, + 590 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Above single-frame depth models do not leverage any temporal information for inference, which limits the performance to some degree. Regarding multi-frame depth prediction, early works adopt test-time refinement [9, 30, 38], use spatial-temporal attention [49], or combine traditional depth networks with recurrent layers [54, 61] to process sequence information. Another broad group of methods [14, 20, 56, 58] utilize cost volume from stereo matching, which explicitly encode and reason from geometric constraints by combining a multi-frame cost volume with single-frame features. On the one hand, this type of methods based on ManyDepth [58] usually meet with the problem of dynamic scenes. To sidestep this issue, we develop a VFI-assisted multi-frame fusion depth model to align and aggregate multi-frame features by optical flow and occlusion masks. On the other hand, ManyDepth uses single-frame depth prior to boost multi-frame depth estimation. However, this improvement is unidirectional. DynamicDepth [14] extends it to a mutual reinforcement, but only targets on dynamic objects. Differently, we propose a unified framework to bilaterally connect single- and multi-frame depth in a more compact and memory-efficient manner.", + "bbox": [ + 212, + 598, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 230, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "J. Liu et al.", + "bbox": [ + 271, + 114, + 351, + 127 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Problem Setup", + "text_level": 1, + "bbox": [ + 215, + 143, + 398, + 162 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Self-supervised MDE in the monocular training case leverages the geometric constraints from video sequences. Specifically, for a training triplet $\\{I_{t-2}, I_t, I_{t+2}\\}$ which contains three consecutive frames, we use $I_t$ as the target view and the other two as the source views $\\{I_s\\}_{s \\in \\{t-2, t+2\\}}$ to implement view synthesis. Here the temporal notation $t \\pm 2$ instead of $t \\pm 1$ is for the convenience of following description. Also, a pose network is utilized to estimate the camera pose from the target view to the source view, denoted as $\\pmb{T}_{t \\to s} = [\\pmb{R}_{t \\to s} | \\pmb{t}_{t \\to s}]$ , where $\\pmb{R}_{t \\to s}$ and $\\pmb{t}_{t \\to s}$ represent the rotation and translation components. Formally, the target depth $D_t$ and the camera pose $\\pmb{T}_{t \\to s}$ can be derived from:", + "bbox": [ + 212, + 176, + 787, + 313 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nD _ {t} = \\mathcal {D} \\left(\\mathcal {E} \\left(I _ {t}\\right)\\right), \\quad \\boldsymbol {T} _ {t \\rightarrow s} = \\mathcal {P} \\left(I _ {t}, I _ {s}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 377, + 321, + 784, + 335 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{E}$ and $\\mathcal{D}$ represent the encoder and decoder, respectively, of a single-frame depth network, while $\\mathcal{P}$ denotes the pose network. Then, the reconstructed view of $I_{t}$ from the source view $I_{s}$ can be obtained by:", + "bbox": [ + 212, + 342, + 784, + 387 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nI _ {s \\rightarrow t} = I _ {s} \\left\\langle \\operatorname {p r o j} \\left(D _ {t}, \\boldsymbol {T} _ {t \\rightarrow s}, \\boldsymbol {K}\\right)\\right\\rangle , \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 390, + 395, + 784, + 411 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $I_{s\\rightarrow t}$ is the reconstructed target image, $\\langle \\rangle$ means the differentiable bilinear sampling operator, $proj()$ returns the 2D coordinates of the pixels in $I_{t}$ after they are reprojected to the viewpoint of $I_{s}$ , and $K$ is the camera intrinsic matrix assumed to be known. Finally, the per-pixel photometric loss between $I_{s\\rightarrow t}$ and $I_{t}$ is used to optimize the depth and pose networks in an end-to-end manner. We follow Monodepth2 [18] to adopt the combination of L1 distance and SSIM [57] as the photometric error function and take the minimum across all the source views at each pixel to handle occlusion, which is formulated as:", + "bbox": [ + 212, + 417, + 787, + 537 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {p e} (t) = \\min _ {s} \\frac {\\alpha}{2} \\left(1 - \\operatorname {S S I M} \\left(I _ {t}, I _ {s \\rightarrow t}\\right)\\right) + (1 - \\alpha) \\| I _ {t}, I _ {s \\rightarrow t} \\| _ {1}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 295, + 544, + 784, + 568 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\alpha = 0.85$ . Besides, the edge-aware smoothness loss is employed to cope with depth discontinuities:", + "bbox": [ + 212, + 573, + 784, + 602 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {s m} (t) = \\left| \\partial_ {x} d _ {t} ^ {*} \\right| e ^ {- \\left| \\partial_ {x} I _ {t} \\right|} + \\left| \\partial_ {y} d _ {t} ^ {*} \\right| e ^ {- \\left| \\partial_ {y} I _ {t} \\right|}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 361, + 609, + 784, + 626 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $d_t^* = d_t / \\overline{d_t}$ is the mean-normalized inverse depth to discourage shrinking [53]. In summary, the standard self-supervised loss on $I_{t}$ is:", + "bbox": [ + 212, + 632, + 782, + 662 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {s s} (t) = \\mathcal {L} \\left(D _ {t}, \\boldsymbol {T} _ {t \\rightarrow s}, I _ {t}, I _ {s}\\right) = \\mu L _ {p e} (t) + \\gamma L _ {s m} (t), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 326, + 671, + 784, + 686 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mu$ is the auto-masking in [18] and $\\gamma = 0.001$ .", + "bbox": [ + 214, + 691, + 583, + 708 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4 Methodology", + "text_level": 1, + "bbox": [ + 215, + 731, + 380, + 750 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this paper, we propose novel Mono-ViFI for self-supervised monocular training, which bilaterally connects single- and multi-frame depth models into a unified learning framework. With flow-based VFI model, we introduce temporal augmentation and develop VFI-assisted multi-frame fusion depth, which are presented in Sec. 4.1 and Sec. 4.2, respectively. Besides, our framework incorporates", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Mono-ViFi: A Unified Framework for Self-supervised Monocular Depth", + "bbox": [ + 256, + 113, + 732, + 130 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "spatial data augmentation and a triplet depth consistency loss, which are separately described in Sec. 4.3 and Sec. 4.4. Finally, we give the overall training objective in Sec. 4.5. The overview of our Mono-ViFi is depicted in Fig. 1.", + "bbox": [ + 212, + 146, + 787, + 194 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1 Temporal Augmentation with VFI", + "text_level": 1, + "bbox": [ + 215, + 210, + 547, + 227 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Firstly, we generate more new frames in the temporal dimension based on existing video frames for temporal augmentation. To achieve this, we adopt a pretrained flow-based VFI model, such as IFRNet [29], which takes two temporally adjacent frames $\\{I_0,I_2\\}$ as input to synthesize an intermediate frame $I_{1}$ . The frame interpolation process can be formulated as:", + "bbox": [ + 212, + 233, + 787, + 309 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nF _ {1 \\rightarrow 0}, F _ {1 \\rightarrow 2}, M _ {1} = \\mathcal {H} \\left(I _ {0}, I _ {2}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 395, + 316, + 784, + 333 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nI _ {0 \\rightarrow 1} = \\omega \\left(I _ {0}, F _ {1 \\rightarrow 0}\\right), \\quad I _ {2 \\rightarrow 1} = \\omega \\left(I _ {2}, F _ {1 \\rightarrow 2}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 346, + 335, + 784, + 351 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nI _ {1} = M _ {1} \\odot I _ {0 \\rightarrow 1} + (1 - M _ {1}) \\odot I _ {2 \\rightarrow 1}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 372, + 354, + 784, + 369 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathcal{H}$ is the VFI model, $F_{1\\rightarrow 0}$ and $F_{1\\rightarrow 2}$ are the intermediate optical flow, $\\omega$ represents backward warping. $I_{0\\rightarrow 1}$ and $I_{2\\rightarrow 1}$ denote the warped images to the intermediate position, while $M_{1}$ ranging from 0 to 1 is the merge mask adjusting the mixing ratio according to bidirectional occlusion information. Here, $\\odot$ indicates the element-wise multiplication. Considering a triplet $\\{I_{t - 2},I_t,I_{t + 2}\\}$ with three consecutive frames as described in Sec. 3, we use the VFI model to synthesize $I_{t - 1}$ from $\\{I_{t - 2},I_t\\}$ and $I_{t + 1}$ from $\\{I_t,I_{t + 2}\\}$ , as shown in Fig. 1(a). Then, we regard $I_{t - 1}$ and $I_{t + 1}$ as another two target frames. All three target frames have common sources of $\\{I_{t - 2},I_{t + 2}\\}$ . The following operations and losses are the same for each of them, and we only take $I_{t}$ for instance.", + "bbox": [ + 212, + 376, + 787, + 527 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2 VFI-assisted Multi-frame Fusion Depth", + "text_level": 1, + "bbox": [ + 215, + 547, + 588, + 561 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Apart from temporal augmentation, our employed VFI model can also provide other inter-frame information, such as optical flow and occlusion merge masks. This enables us to design a novel VFI-assisted multi-frame fusion module for multi-frame depth prediction, as demonstrated in Fig. 2. The fusion module consists of two parts, i.e., motion-aware feature alignment (MAFA) and occlusion-alarviated feature fusion (OAFF). Taking original triplet $\\{I_{t - 2},I_t,I_{t + 2}\\}$ for example, we implement forward process of the VFI model to interpolate the middle position $t$ of $\\{I_{t - 2},I_{t + 2}\\}$ in Eq. (9), yielding intermediate flow $F_{t\\rightarrow (t - 2)},F_{t\\rightarrow (t + 2)}$ and merge mask $M_{t}$ to align and aggregate multi-frame features.", + "bbox": [ + 212, + 569, + 787, + 705 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nF _ {t \\rightarrow (t - 2)}, F _ {t \\rightarrow (t + 2)}, M _ {t} = \\mathcal {H} \\left(I _ {t - 2}, I _ {t + 2}\\right). \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 362, + 713, + 784, + 729 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Multi-frame Feature Extraction. Our fusion module accepts multi-frame features extracted by the same encoder, which is shared with the single-frame depth encoder $\\mathcal{E}$ . The extraction process can be written as:", + "bbox": [ + 212, + 734, + 787, + 780 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{\\phi_ {t - 2} ^ {k}, \\phi_ {t} ^ {k}, \\phi_ {t + 2} ^ {k} \\right\\} _ {k = 1} ^ {N} = \\left\\{\\mathcal {E} \\left(I _ {t - 2}\\right), \\mathcal {E} \\left(I _ {t}\\right), \\mathcal {E} \\left(I _ {t + 2}\\right) \\right\\}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 334, + 786, + 784, + 803 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\{\\phi_{t - 2}^k,\\phi_t^k,\\phi_{t + 2}^k\\}_{k = 1}^N$ are the multi-level features of the three frames and $N$ is the number of feature levels.", + "bbox": [ + 212, + 808, + 787, + 839 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "J. Liu et al.", + "bbox": [ + 271, + 114, + 351, + 127 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5a8c29e39bf5ffda6cab944146caef336883dae0319f1957a820946db98e53dc.jpg", + "image_caption": [ + "Fig. 2: Our VFI-assisted multi-frame fusion depth model. Motion and occlusion provided by VFI can be used to align and fuse multi-frame features in an explicit manner." + ], + "image_footnote": [], + "bbox": [ + 215, + 145, + 787, + 309 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Motion-aware Feature Alignment (MAFA). For the $k$ -th level, we use bidirectional optical flow to align the neighbor features $\\phi_{t-2}^k$ and $\\phi_{t+2}^k$ to the target position $t$ , generating $\\phi_{(t-2) \\to t}^k$ and $\\phi_{(t+2) \\to t}^k$ , respectively, formulated as:", + "bbox": [ + 214, + 385, + 782, + 434 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\phi_ {(t - 2) \\rightarrow t} ^ {k} = \\omega \\left(\\phi_ {t - 2} ^ {k}, F _ {t \\rightarrow (t - 2)} ^ {k}\\right), \\\\ \\phi_ {(t + 2) \\rightarrow t} ^ {k} = \\omega \\left(\\phi_ {t + 2} ^ {k}, F _ {t \\rightarrow (t + 2)} ^ {k}\\right), \\tag {11} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 388, + 439, + 784, + 481 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $F_{t\\rightarrow (t - 2)}^k$ and $F_{t\\rightarrow (t + 2)}^k$ denote the corresponding flows of the $k$ -th level, downsampled from $F_{t\\rightarrow (t - 2)}$ and $F_{t\\rightarrow (t + 2)}$ , respectively. Inspired by recent 2D implicit neural representation [8], we also concatenate the aligned features with encoded optical flow, which can be denoted as:", + "bbox": [ + 214, + 487, + 782, + 547 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\varphi_ {(t - 2) \\rightarrow t} ^ {k} = \\operatorname {C o n c a t} \\left(\\phi_ {(t - 2) \\rightarrow t} ^ {k}, g \\left(F _ {t \\rightarrow (t - 2)} ^ {k}\\right)\\right), \\\\ \\varphi_ {(t + 2) \\rightarrow t} ^ {k} = \\operatorname {C o n c a t} \\left(\\phi_ {(t + 2) \\rightarrow t} ^ {k}, g \\left(F _ {t \\rightarrow (t + 2)} ^ {k}\\right)\\right), \\tag {12} \\\\ \\varphi_ {t} ^ {k} = \\operatorname {C o n c a t} \\left(\\phi_ {t} ^ {k}, g \\left(F _ {0} ^ {k}\\right)\\right), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 344, + 553, + 784, + 614 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where Concat means feature concatenation, $F_0^k$ denotes zero flow and $g$ is fourier positional encoding [41,51] to transform optical flow to high-frequency signal. Actually, optical flow indicates the relative coordinate of a feature vector in the warped feature map w.r.t. the original one. When incorporating it, multi-frame model can adaptively adjust feature learning through motion-aware alignment.", + "bbox": [ + 214, + 623, + 782, + 698 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Occlusion-alleviated Feature Fusion (OAFF). Due to the ubiquitous interframe occlusion, aggregating the aligned features directly through summation or concatenation can deteriorate the performance. Therefore, we use the $k$ -th level merge mask $M_t^k$ downsampled from $M_t$ to mitigate this issue, just like the image merging process of VFI in Eq. (8):", + "bbox": [ + 214, + 702, + 782, + 777 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\chi_ {t} ^ {k} = M _ {t} ^ {k} \\odot \\varphi_ {(t - 2) \\rightarrow t} ^ {k} + (1 - M _ {t} ^ {k}) \\odot \\varphi_ {(t + 2) \\rightarrow t} ^ {k}, \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 336, + 782, + 784, + 801 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\chi_t^k$ is the occlusion-alleviated feature merged from the temporal positions $t - 2$ and $t + 2$ . It is then concatenated with the feature $\\varphi_t^k$ in the position $t$ .", + "bbox": [ + 214, + 809, + 782, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Mono-ViFi: A Unified Framework for Self-supervised Monocular Depth", + "bbox": [ + 256, + 113, + 732, + 130 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "And we can obtain the final feature after fusion by:", + "bbox": [ + 215, + 146, + 584, + 161 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\psi_ {t} ^ {k} = \\operatorname {C o n v} _ {1 \\times 1} \\left(\\operatorname {C o n c a t} \\left(\\varphi_ {t} ^ {k}, \\chi_ {t} ^ {k}\\right)\\right), \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 385, + 167, + 785, + 185 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\mathrm{Conv}_{1\\times 1}$ is a $1\\times 1$ convolutional layer to transform the feature into the same channel as the original feature $\\phi_t^k$ , enabling the multi-frame depth decoder to employ the same architecture as the single-frame one.", + "bbox": [ + 212, + 190, + 782, + 236 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Multi-frame Depth Decoder. To translate the multi-level fusion features into a depth map, a decoder is required. Concretely, we use the same architecture as the single-frame depth decoder, but the parameters are not shared. This encourages the single- and multi-frame models to decode the features in their individual ways without conflict. Finally, the fusion features $\\{\\psi_t^k\\}_{k=1}^N$ are processed by a multi-frame depth decoder $\\mathcal{D}_m$ to generate the fusion depth $D_t^m$ of the target frame $I_t$ , which can be written as:", + "bbox": [ + 212, + 238, + 784, + 342 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nD _ {t} ^ {m} = \\mathcal {D} _ {m} \\left(\\left\\{\\psi_ {t} ^ {k} \\right\\} _ {k = 1} ^ {N}\\right). \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 426, + 349, + 785, + 367 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Note that the multi-frame depth $D_{t}^{m}$ also corresponds to a self-supervised loss like Eq. (5), which is denoted as:", + "bbox": [ + 212, + 373, + 782, + 402 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nL _ {s s} ^ {m} (t) = \\mathcal {L} \\left(D _ {t} ^ {m}, \\boldsymbol {T} _ {t \\rightarrow s}, I _ {t}, I _ {s}\\right). \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 398, + 411, + 785, + 426 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3 Spatial Augmentation", + "text_level": 1, + "bbox": [ + 215, + 455, + 442, + 472 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Orthogonal to temporal augmentation, we also include classic data augmentation to spatially expand the data diversity, which is categorized as spatial augmentation in this paper. Resizing-cropping is a commonly used and effective augmentation approach. Based on this operation, we introduce image rotation, extending it to a generic affine transformation (cropping is actually image translation). Specifically, we follow [55] to consider that the augmented image and the original image are captured by the same camera system whose intrinsics are fixed. For resizing operation, it is assumed that when an image is zoomed up by a factor $f_{s} \\geq 1$ , the relative size of an object increases by the same factor $f_{s}$ and its depth decreases by $f_{s}$ simultaneously. This is inspired from a vital monocular cue [4] that the closer an object is, the larger its relative size is. Our affine transformation $\\mathcal{A}$ is applied to both the target image $I_{t}$ and the source image $I_{s}$ , generating $\\widetilde{I}_{t} = \\mathcal{A}(I_{t})$ and $\\widetilde{I}_{s} = \\mathcal{A}(I_{s})$ , respectively. For the augmented target view, we can obtain its depth through the single-frame model by $\\widetilde{D}_{t} = \\mathcal{D}(\\mathcal{E}(\\widetilde{I}_{t}))$ . To conduct view synthesis on augmented images, the original camera pose $T_{t\\rightarrow s} = [\\pmb{R}_{t\\rightarrow s}|\\pmb{t}_{t\\rightarrow s}]$ from target to source should be rectified to $\\widetilde{T}_{t\\rightarrow s} = [\\widetilde{\\pmb{R}}_{t\\rightarrow s}|\\widetilde{\\pmb{t}}_{t\\rightarrow s}]$ as done in [55]:", + "bbox": [ + 212, + 489, + 787, + 750 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\boldsymbol {R}} _ {t \\rightarrow s} = \\boldsymbol {R} _ {c} \\boldsymbol {R} _ {t \\rightarrow s} \\boldsymbol {R} _ {c} ^ {- 1}, \\quad \\tilde {\\boldsymbol {t}} _ {t \\rightarrow s} = \\boldsymbol {R} _ {c} \\boldsymbol {t} _ {t \\rightarrow s}, \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 354, + 756, + 785, + 773 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\mathbf{R}_c$ is the rectification matrix determined only by the camera intrinsics and the augmentation parameters. Different from [55] calculating $\\mathbf{R}_c$ under resizing-cropping, we give its formulation and derivation under a more general affine transformation, which can be seen in our supplementary. Therefore, we can derive", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "J. Liu et al.", + "bbox": [ + 271, + 114, + 351, + 127 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the self-supervised loss $\\widetilde{L}_{ss}(t)$ on the augmented images similarly to Eq. (5):", + "bbox": [ + 215, + 145, + 763, + 162 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\widetilde {L} _ {s s} (t) = \\mathcal {L} (\\widetilde {D} _ {t}, \\widetilde {\\boldsymbol {T}} _ {t \\rightarrow s}, \\widetilde {I} _ {t}, \\widetilde {I} _ {s}). \\tag {18}\n$$\n", + "text_format": "latex", + "bbox": [ + 401, + 167, + 784, + 186 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Note that there is a mask to exclude the possible invalid pixels in $\\widetilde{I}_t$ when calculating $\\widetilde{L}_{ss}(t)$ due to image rotation.", + "bbox": [ + 214, + 193, + 785, + 224 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.4 Triplet Depth Consistency", + "text_level": 1, + "bbox": [ + 215, + 243, + 480, + 258 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Now we have single-frame depths $D_{t}$ and $\\widetilde{D}_{t}$ of the original image $I_{t}$ and the augmented image $\\widetilde{I}_{t}$ , along with multi-frame depth $D_{t}^{m}$ of $I_{t}$ . For better knowledge distillation and regularization, our unified learning framework applies a triplet depth consistency loss among the three depth maps, including a standard view depth consistency (SVDC) loss and two scale-aware depth consistency (SADC) losses, which are shown in Fig. 1(b). The SVDC loss $L_{sv}(t)$ conducts mutual distillation between single- and multi-frame models, promoting their mutual reinforcement, which can be denoted as:", + "bbox": [ + 214, + 266, + 787, + 387 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nL _ {s v} (t) = \\operatorname {S I} \\left(D _ {t} ^ {m}, D _ {t}\\right), \\tag {19}\n$$\n", + "text_format": "latex", + "bbox": [ + 426, + 395, + 784, + 411 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\mathrm{SI}()$ is the widely used scale-invariant error function [13]. Besides, we use depth $\\widetilde{D}_t$ under spatial augmentation to enforce two SADC losses, i.e., $L_{sa}(t)$ between $\\widetilde{D}_t$ and $D_t$ , and $L_{sa}^m(t)$ between $\\widetilde{D}_t$ and $D_t^m$ . To this end, we first restore $\\widetilde{D}_t$ to the original image view, generating $\\widehat{D}_t = \\mathcal{A}^{-1}(\\widetilde{D}_t)$ in Fig. 1(b), where $\\mathcal{A}^{-1}$ is the inverse transformation of affine augmentation $\\mathcal{A}$ . To exclude the invalid pixels in $\\widehat{D}_t$ , we adopt a mask $M_{sa}$ and calculate $L_{sa}(t)$ and $L_{sa}^m(t)$ through:", + "bbox": [ + 214, + 416, + 787, + 513 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nL _ {s a} (t) = \\operatorname {S I} \\left(M _ {s a} \\odot D _ {t}, M _ {s a} \\odot f _ {s} \\widehat {D} _ {t}\\right), \\tag {20}\n$$\n", + "text_format": "latex", + "bbox": [ + 372, + 532, + 784, + 550 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nL _ {s a} ^ {m} (t) = \\operatorname {S I} \\left(M _ {s a} \\odot D _ {t} ^ {m}, M _ {s a} \\odot f _ {s} \\widehat {D} _ {t}\\right), \\tag {21}\n$$\n", + "text_format": "latex", + "bbox": [ + 369, + 555, + 784, + 571 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where the factor $f_{s}$ is due to the assumption in Sec. 4.3 that the depth of an object in $I_{t}$ is $f_{s}$ times of that in $\\widetilde{I}_{t}$ . Our SADC constrains the 3D geometric relationship between the scenes of $I_{t}$ and $\\widetilde{I}_{t}$ , providing scale-aware regularization. Combining SVDC and SADC, we have the triplet depth consistency loss as:", + "bbox": [ + 214, + 578, + 787, + 638 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nL _ {t c} (t) = L _ {s v} (t) + L _ {s a} (t) + L _ {s a} ^ {m} (t). \\tag {22}\n$$\n", + "text_format": "latex", + "bbox": [ + 382, + 646, + 784, + 662 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.5 Training Objective", + "text_level": 1, + "bbox": [ + 215, + 680, + 418, + 696 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "For target position $t$ , we have already obtained the self-supervised losses $L_{ss}(t)$ , $L_{ss}^{m}(t)$ and $\\widetilde{L}_{ss}(t)$ in Eqs. (5), (16) and (18), respectively. Besides, our triplet depth consistency loss $L_{tc}(t)$ is acquired in Eq. (22). Therefore, the total loss on position $t$ can be summed as:", + "bbox": [ + 214, + 703, + 785, + 765 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nL (t) = L _ {s s} (t) + L _ {s s} ^ {m} (t) + \\widetilde {L} _ {s s} (t) + \\lambda L _ {t c} (t), \\tag {23}\n$$\n", + "text_format": "latex", + "bbox": [ + 354, + 770, + 784, + 787 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\lambda$ is a hyper-parameter set to 0.2 in this paper. For $t - 1$ and $t + 1$ , we can derive their losses $L(t - 1)$ and $L(t + 1)$ in the same way. In summary, the overall loss function of our Mono-ViFI is $L = L(t - 1) + L(t) + L(t + 1)$ .", + "bbox": [ + 214, + 794, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Mono-ViFi: A Unified Framework for Self-supervised Monocular Depth", + "bbox": [ + 254, + 113, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Table 1: Performance comparison on KITTI [16] benchmark using the raw ground truth, with a resolution of $640 \\times 192$ . The best results in each subsection are in bold; second best are underlined. For MACs of multi-frame models, we use the formulation A/B. A denotes the MACs when predicting for a single frame; B means the average MACs per frame when processing a long video sequence. #Params and MACs of our multi-frame models include the VFI network.", + "bbox": [ + 212, + 143, + 787, + 227 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/5be4c37c825e8a0cf365029b8fd0f56453e49802107cdf6d98fe3832c9b17e38.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodTest frames#ParamsMACsAbs RelSq RelRMSERMSE logδ1δ2δ3
PackNet-SfM [21]1128M205G0.1110.7854.6010.1890.8780.9600.982
HR-Depth [39]116.9M31.6G0.1090.7924.6320.1850.8840.9620.983
R-MSFM6 [67]13.8M31.2G0.1120.8064.7040.1910.8780.9600.981
BRNet [22]119.1M33.7G0.1050.6984.4620.1790.8900.9650.984
MonoViT [64]127M-0.0990.7084.3720.1750.9000.9670.984
DIFFNet [65]112.2M23.4G0.1020.7644.4830.1800.8960.9650.983
SD-SSMDE [44]1--0.1000.6614.2640.1720.8960.9670.985
MUSTNet [48]1--0.1060.7634.5620.1820.8880.9630.983
DepthSegNet [40]130.3M-0.1040.6904.4730.1790.8860.9650.984
DualRefine [2]1--0.1030.7764.4910.1810.8940.9650.983
DaCNN [23]113M-0.0990.6614.3160.1730.8970.9670.985
Monodepth2 (ResNet18) [18]114.3M8G0.1150.9034.8630.1930.8770.9590.981
Mono-ViFI (ResNet18)114.3M8G0.1050.7084.4460.1790.8870.9650.984
Lite-Mono [62]13.1M5.1G0.1070.7654.5610.1830.8860.9630.983
Mono-ViFI (Lite-Mono)13.1M5.1G0.1030.6804.3490.1760.8910.9660.985
D-HRNet [25]110M10.8G0.1020.7604.4790.1790.8970.9650.983
RA-Depth (D-HRNet) [25]110M10.8G0.0960.6324.2160.1710.9030.9680.985
Mono-ViFI (D-HRNet)110M10.8G0.0960.6274.1790.1700.9030.9690.985
Patil et al. [42]N--/-0.1110.8214.6500.1870.8830.9610.982
ManyDepth (ResNet18) [58]2 (-1, 0)26.9M15.1G/13.7G0.0980.7704.4590.1760.9000.9650.983
DynamicDepth [14]2 (-1, 0)--/-0.0960.7204.4580.1750.8970.9640.984
DepthFormer [20]2 (-1, 0)28.7M175G/-0.0900.6614.1490.1750.9050.9670.984
MOVEDepth [56]2 (-1, 0)28.2M20.2G/19.2G0.0940.7044.3890.1750.9020.9650.983
DualRefine [2]2 (-1, 0)--/-0.0870.6984.2340.1700.9140.9670.983
TC-Depth [49]3 (-1, 0, 1)--/-0.1030.7464.4830.1850.894-0.983
Mono-ViFI (ResNet18)3 (-1, 0, 1)17.9M25.8G/16.9G0.0990.6614.3210.1740.8980.9660.985
Mono-ViFI (Lite-Mono)3 (-1, 0, 1)5.9M21.9G/13.3G0.0990.6484.2560.1720.8980.9670.985
Mono-ViFI (D-HRNet)3 (-1, 0, 1)12.9M35.7G/19.4G0.0910.5894.0880.1660.9120.9690.986
", + "bbox": [ + 220, + 250, + 787, + 537 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5 Experiments", + "text_level": 1, + "bbox": [ + 215, + 574, + 375, + 593 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5.1 Implementation Details", + "text_level": 1, + "bbox": [ + 215, + 604, + 457, + 619 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Datasets. We train and evaluate the models on two widely used outdoor urban datasets of KITTI [16] and Cityscapes [10]. The depth labels of KITTI are collected with LiDARs. For fair comparison, we use the Eigen split [12] and follow the pre-processing operation in [18], generating 39,810 monocular triplets for training and 697 frames for evaluation. To reduce the influence of noise in the sparse depth from Velodyne, we also evaluate using $93\\%$ of the Eigen split with the improved ground truth from [52], which contains 652 test frames. In regard to Cityscapes, following [58, 66], we preprocess the monocular sequences into triples, producing 69,731 training images. And we evaluate on 1,525 test images using the ground truth provided by [58]. For generalization comparison, KITTI models are evaluated on Cityscapes and Make3D [50]. Make3D is a small outdoor dataset containing 134 test images with aligned depth information.", + "bbox": [ + 212, + 628, + 787, + 809 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Flow-based VFI Model. We choose IFRNet [29] as our VFI model due to its efficiency. In practice, we use the large model of IFRNet for training and replace", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "J. Liu et al.", + "bbox": [ + 271, + 114, + 351, + 126 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/99e79450c6426335bf4c0e9386f8b6e7c1850d1ad49ca9df9568c42be55984ec.jpg", + "image_caption": [ + "Fig. 3: Qualitative results on KITTI. Error maps in columns 2 and 4 show the Abs Rel error compared to the improved ground truth [52] from good (blue) to bad (red)." + ], + "image_footnote": [], + "bbox": [ + 217, + 145, + 395, + 305 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/0da88bfe2f0d99496838e7378941a00260474aa275d552af0216217e93a6d7e8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 395, + 146, + 517, + 304 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/3c94a3e3f9f98c4879c9225deb6ca36e4b8fe91f56f484a1376167f9532d1156.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 519, + 146, + 651, + 304 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/283d98377f6cc2f608e38498ec70a8b43581e3270bbefc364a23cbbba0a2309c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 651, + 148, + 782, + 304 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "it with the small model when evaluating for multi-frame fusion depth. Please see ablation about their gap in supplementary. Both large and small IFRNets are pretrained on KITTI/Cityscapes and frozen in Mono-ViFI training. The pretraining data split is the same as self-supervised depth training. Each original triplet in depth training is directly used for VFI training. Besides, we employ another two flow-based VFI networks, RIFE [26] and AMT [34], to verify the robustness of our method to different VFI models.", + "bbox": [ + 212, + 372, + 787, + 477 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Depth and Pose Networks. We experiment with three depth architectures, including the ResNet18 [24] based network in Monodepth2 [18], medium LiteMono [62], and D-HRNet proposed in RA-Depth [25]. For the pose network, we follow Monodepth2 to use a ResNet18 based model which takes a pair of images (six channels) as input and predict a 6-DoF relative pose. Encoders of all the networks are pretrained on ImageNet [11].", + "bbox": [ + 212, + 479, + 787, + 570 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "5.2 Performance Comparison", + "text_level": 1, + "bbox": [ + 214, + 589, + 472, + 604 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "KITTI Results. Quantitative comparison results on KITTI under $640 \\times 192$ resolution are listed in Table 1 and Table 2, using raw ground truth and improved ground truth, respectively. It is obvious that our method with D-HRNet can achieve state-of-the-art performance. Besides, our Mono-ViFI brings notable improvements to the three depth architectures without increasing model complexity (see each subsection separated by dotted lines). Despite that the difference between Mono-ViFI (D-HRNet) and RA-Depth [25] under single-frame mode is slight in Table 1, it is more significant in Table 2 when using improved ground truth. For multi-frame inference, Mono-ViFI with ResNet18 has already shown advantages over ManyDepth [58] and DynamicDepth [14] on some metrics including Sq Rel, RMSE, RMSE log, $\\delta_{2}$ and $\\delta_{3}$ . When changing to D-HRNet backbone, proposed Mono-ViFI can totally outperform them. Although our models have more multiply-add computations (MACs) when predicting for a single frame, the average MACs per frame are comparable when processing a long video sequence. Fig. 3 depicts two visual comparison results on KITTI, where", + "bbox": [ + 212, + 613, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Mono-ViFi: A Unified Framework for Self-supervised Monocular Depth", + "bbox": [ + 256, + 113, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/435e95fff156baa792e511883d37b23e8e8a2f35072e483c0c45f816e4ef97a9.jpg", + "table_caption": [ + "Table 2: Performance comparison on Table 3: Performance comparison on KITTI [16] using improved ground truth Cityscapes [10] with a resolution of $416 \\times$ from [52], with a resolution of $640 \\times 192$ . 128 unless stated." + ], + "table_footnote": [], + "table_body": "
MethodAbs RelSq RelRMSERMSE logδ1δ2δ3
Single-frame Inference
PackNet-SIM [21]0.0780.4203.4850.1210.9310.9860.996
R-MSF M6 [67]0.0880.4923.8370.1350.9150.9830.995
DIFFNet [65]0.0760.4143.4050.1190.9360.9880.996
Mono-ViF [64]0.0740.3883.4140.1150.9380.9890.997
Monodepth2 (ResNet18) [18]0.0900.5453.9420.1370.9140.9830.995
Mono-ViF (ResNet18)0.0840.4313.6180.1260.9240.9860.997
Lite-Mono [62]0.0820.4553.6850.1270.9230.9850.996
Mono-ViF (Lite-Mono)0.0800.4003.4970.1210.9300.9870.997
D-HRNet [25]0.0770.4233.3960.1190.9350.9870.996
RA-Depth (D-HRNet) [25]0.0740.3633.3490.1140.9400.9900.997
Mono-ViF (D-HRNet)0.0720.3423.2540.1110.9410.9910.998
Multi-frame Inference
ManyDepth (ResNet18) [58]0.0700.3993.4550.1130.9410.9890.997
MOVEDepth [56]0.0650.3773.4490.1120.9420.9880.996
Mono-ViF (ResNet18)0.0740.3743.3810.1130.9410.9900.997
Mono-ViF (Lite-Mono)0.0740.3543.3720.1140.9390.9890.997
Mono-ViF (D-HRNet)0.0660.2943.0990.1020.9520.9920.998
", + "bbox": [ + 225, + 227, + 496, + 358 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/e9bc3f513579ccdbf6b94249042eb8622a3c7da720244114eb709a8573d7254c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodAbs RelSq RelRMSERMSE logδ1δ2δ3
Single-frame Inference
Videos in the wild [19]0.1271.3306.9600.1950.8300.9470.981
Li et al. [33]0.1191.2906.9800.1900.8460.9520.982
InstDM [31] (832 × 256)0.1111.1586.4370.1820.8680.9610.983
SD-SSMDE [44]0.1100.9885.9530.165---
Monodepth2 (ResNet18) [18]0.1291.5696.3760.1970.8490.9570.983
Mono-ViFi (ResNet18)0.1121.0035.8990.1650.8720.9660.990
Lite-Mono [62]0.1141.1946.3280.1720.8710.9670.989
Mono-ViFi (Lite-Mono)0.1070.9405.7940.1590.8760.9680.991
D-HRNet [25]0.1111.1426.2050.1690.8790.9680.989
Mono-ViFi (D-HRNet)0.1050.8805.5500.1560.8810.9690.992
Multi-frame Inference
ManyDepth (ResNet18) [58]0.1141.1936.2230.1700.8750.9670.989
DynamicDepth [14]0.1041.0115.9870.1590.8890.9720.991
TC-Depth [49]0.1100.9585.820-0.867-0.991
Mono-ViFi (ResNet18)0.1070.9345.6880.1590.8800.9690.990
Mono-ViFi (Lite-Mono)0.1050.9045.6460.1570.8800.9700.991
Mono-ViFi (D-HRNet)0.1020.8365.3950.1530.8850.9710.992
", + "bbox": [ + 506, + 227, + 779, + 358 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/20f3189acd7c8cdae6073b6e3aed93ff243e1048e4d833e70a54d3eb35192845.jpg", + "table_caption": [ + "Table 4: Evaluation results on dynamic objects and static regions of Cityscapes [10] dataset under $416\\times 128$ resolution. Note that DynamicDepth relies on depth prior and segmentation masks of dynamic objects when reasoning." + ], + "table_footnote": [], + "table_body": "
MethodTest AreaAbs RelSq RelRMSERMSE logδ1δ2δ3
Multi-frame Inference
ManyDepth (ResNet18) [58]DynamicDepth [14]Mono-ViFI (ResNet18)Dynamic Objects0.1722.2065.9240.2160.7920.9170.967
0.1441.5545.2380.1850.8310.9500.982
0.1621.4535.1140.1990.7940.9240.977
ManyDepth (ResNet18) [58]DynamicDepth [14]Mono-ViFI (ResNet18)Static Regions0.1071.0906.3140.1600.8820.9720.991
0.1000.9776.1550.1550.8920.9740.992
0.0990.9075.8350.1490.8900.9750.992
", + "bbox": [ + 261, + 440, + 743, + 523 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "we choose D-HRNet as the baseline depth model. It can be seen that depth maps look qualitatively similar among all four methods. However, the error maps reveal mistakes of D-HRNet baseline and RA-Depth in some ambiguous regions like bushes (see the red boxes). In contrast, our Mono-ViFi have dramatically lower error in these regions, especially under multi-frame reasoning.", + "bbox": [ + 212, + 566, + 784, + 642 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Cityscapes Results. Results of Cityscapes in Table 3 demonstrate the same trend as KITTI. Moreover, proposed multi-frame model with ResNet18 can beat ManyDepth on all metrics, while our Mono-ViFI is better than DynamicDepth only on partial indexes. It is because Cityscapes with more dynamic scenes is more challenging than KITTI, and DynamicDepth leverages additional information to cope with dynamic objects. For fine-grained analysis, we separately evaluate on the dynamic objects and static regions, using the segmentation masks of dynamic objects provided in [14], as listed in Table 4. On dynamic objects, Mono-ViFI with ResNet18 surpasses ManyDepth by a large margin, demonstrating that our multi-frame inference paradigm can sidestep the issue of dynamic objects to a greater extent compared to ManyDepth. DynamicDepth utilizes depth prior and segmentation masks of dynamic objects to tackle this problem for multi-frame inference. Even though, our method still excels it on Sq Rel and", + "bbox": [ + 212, + 643, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "J. Liu et al.", + "bbox": [ + 271, + 114, + 351, + 126 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/b055474670ed0ea897e8507ce5a1717e48236a09eacf6bed41360cfcb8c2708b.jpg", + "image_caption": [ + "Fig. 4: Qualitative generalization results of a YouTube video example. All depth maps are predicted under $640 \\times 360$ resolution using KITTI models trained under $640 \\times 192$ ." + ], + "image_footnote": [], + "bbox": [ + 254, + 146, + 750, + 308 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/473ae66e2f84105a645934f03f578f72612f2632967d3f61d15e1e3511764b58.jpg", + "table_caption": [ + "Table 5: Ablation results of each proposed component on Cityscapes [10]. SA is spatial augmentation. TA is temporal augmentation. SVDC denotes standard view depth consistency. SADC denotes scale-aware depth consistency. FA is feature alignment. MAFA is motion-aware feature alignment. OAFF means occlusion-alleviated feature fusion." + ], + "table_footnote": [], + "table_body": "
MethodSATASVDCSADCFAMAFAOAFFAbs RelSq RelRMSEδ1
Single-frame Inference
Baseline (Monodepth2 [18])0.1291.5696.8760.849
Baseline + TA✓0.1191.2546.5230.857
Baseline + SVDC✓0.1171.1866.4920.862
Baseline + SA✓0.1221.3866.5570.856
Baseline + SA + SADC✓✓0.1161.0796.1890.868
Mono-ViFI (Single-frame)✓✓✓✓0.1121.0035.8990.872
Multi-frame Inference
Baseline (direct feature fusion)0.1201.3606.4910.869
FA (not motion-aware)✓0.1161.2176.3870.871
MAFA✓0.1141.1506.1150.874
MAFA + OAFF✓✓0.1111.1076.0120.876
MAFA + OAFF + TA✓✓✓0.1091.0235.8920.877
Mono-ViFI (Multi-frame)✓✓✓✓✓✓0.1070.9345.6880.880
", + "bbox": [ + 263, + 431, + 743, + 594 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "RMSE. When evaluating on static regions, proposed Mono-ViFI shows absolute superiority over ManyDepth and DynamicDepth.", + "bbox": [ + 212, + 637, + 784, + 666 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Generalization Results. We evaluate the KITTI models on Cityscapes and Make3D without fine-tuning, results of which are put in the supplementary. For qualitative comparison, we give a visualization example in Fig. 4, where KITTI models are used to predict the multi-frame depth of a YouTube video frame downloaded from the Wind Walk Travel Videos channel. This indicates that our Mono-ViFI has better generalization ability than existing methods.", + "bbox": [ + 212, + 667, + 785, + 758 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5.3 Ablation Study", + "text_level": 1, + "bbox": [ + 214, + 782, + 388, + 797 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "In this part, we explore the influence of each component in Mono-ViFi. Our ablations are conducted on Cityscapes or KITTI with ResNet18 backbone.", + "bbox": [ + 212, + 809, + 784, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Mono-ViFi: A Unified Framework for Self-supervised Monocular Depth", + "bbox": [ + 256, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/93df679f27216709e4f6f97dd794cad2c50611cd94ee31e2e2b70395ae90721c.jpg", + "table_caption": [ + "Table 6: Ablation of different flow-based VFI models on KITTI [16]." + ], + "table_footnote": [], + "table_body": "
VFI ModelAbs RelSq RelRMSERMSE logδ1δ2δ3
Multi-frame Inference
IFRNet [29] (Ours)0.0990.6614.3210.1740.8980.9660.985
RIFE [26]0.1000.6724.3520.1750.8960.9650.985
AMT [34]0.0990.6534.3350.1740.8970.9660.985
", + "bbox": [ + 305, + 171, + 702, + 238 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Single-frame Ablation. We first ablate on the single-frame model. As listed in the top half part of Table 5, the standard view depth consistency and spatial augmentation can bring gains individually. Note that scale-aware depth consistency $L_{sa}$ should be based on spatial augmentation. When adding $L_{sa}$ to spatial augmentation, the accuracy can be further improved. And this leads to the most significant enhancement among all components, since $L_{sa}$ can enforce the model to learn scale relationship from scene variations. In orthogonal dimension, our proposed temporal augmentation can also reduce the prediction error by providing more guidance from newly synthesized camera views. When integrating all above elements, our single-frame model can achieve the best performance.", + "bbox": [ + 212, + 276, + 785, + 426 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Multi-frame Ablation. To probe the effect of each component in our VFI-assisted multi-frame fusion module, we conduct ablation experiments under multi-frame inference, whose results are summarized in the bottom half part of Table 5. The baseline is stacking the raw feature of three frames directly, i.e., direct feature fusion. Then, we gradually apply components of the proposed fusion module, including motion-aware feature alignment and occlusion-alleviated feature fusion. The results demonstrate that these strategies can bring cumulative improvements. Furthermore, when incorporating spatial augmentation, temporal augmentation and relevant depth consistency losses, our multi-frame model can be further promoted like the single-frame setting.", + "bbox": [ + 212, + 426, + 785, + 577 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Ablation of different VFI models. Apart from IFRNet [29], we additionally experiment with another two flow-based VFI models, RIFE [26] and AMT [34]. The results are depicted in Table 6. It can be observed that there is only a negligible performance difference among the three VFI networks, which reveals that our method is robust to different VFI models.", + "bbox": [ + 212, + 578, + 785, + 652 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 674, + 359, + 690 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In conclusion, this paper presents a novel self-supervised learning framework, named Mono-ViFI, to bilaterally connect single- and multi-frame depth estimation in a unified manner. First, a flow-based VFI model is employed for temporal augmentation. Then, we propose a VFI-assisted multi-frame fusion module for more accurate multi-frame depth. Besides, resizing-cropping is extended to affine transformation as spatial augmentation for enriching data diversity. Finally, we develop a new triplet depth consistency loss for better regularization. Experiments show that our approach can unleash the potential of existing depth models to a greater extent, but without increasing inference complexity.", + "bbox": [ + 212, + 704, + 785, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "J. Liu et al.", + "bbox": [ + 271, + 114, + 351, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Bae, J., Moon, S., Im, S.: Deep digging into the generalization of self-supervised monocular depth estimation. In: AAAI. pp. 187-196 (2023)", + "2. Bangunharcana, A., Magd, A., Kim, K.S.: Dualrefine: Self-supervised depth and pose estimation through iterative epipolar sampling and refinement toward equilibrium. In: CVPR. pp. 726-738 (2023)", + "3. Bello, J.L.G., Kim, M.: Forget about the lidar: Self-supervised depth estimators with med probability volumes. In: NeurIPS. pp. 12626-12637 (2020)", + "4. Bello, J.L.G., Moon, J., Kim, M.: Positional information is all you need: A novel pipeline for self-supervised svde from videos. arXiv:2205.08851 (2022)", + "5. Bian, J., Li, Z., Wang, N., Zhan, H., Shen, C., Cheng, M.M., Reid, I.: Unsupervised scale-consistent depth and ego-motion learning from monocular video. In: NeurIPS. pp. 35-45 (2019)", + "6. Chan, K.C., Wang, X., Yu, K., Dong, C., Loy, C.C.: Basicvsr: The search for essential components in video super-resolution and beyond. In: CVPR. pp. 4945-4954 (2021)", + "7. Chan, K.C., Zhou, S., Xu, X., Loy, C.C.: Basicvsr++: Improving video superresolution with enhanced propagation and alignment. In: CVPR. pp. 5962-5971 (2022)", + "8. Chen, Y., Liu, S., Wang, X.: Learning continuous image representation with local implicit image function. In: CVPR. pp. 8624-8634 (2021)", + "9. Chen, Y., Schmid, C., Sminchisescu, C.: Self-supervised learning with geometric constraints in monocular video: Connecting flow, depth, and camera. In: ICCV. pp. 7063-7072 (2019)", + "10. Cordts, M., Omran, M., Ramos, S., Rehfeld, T., Enzweiler, M., Benenson, R., Franke, U., Roth, S., Schiele, B.: The cityscapes dataset for semantic urban scene understanding. In: CVPR. pp. 3213-3223 (2016)", + "1. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: CVPR. pp. 248-255 (2009)", + "2. Eigen, D., Fergus, R.: Predicting depth, surface normals and semantic labels with a common multi-scale convolutional architecture. In: ICCV. pp. 2650-2658 (2015)", + "3. Eigen, D., Puhrsch, C., Fergus, R.: Depth map prediction from a single image using a multi-scale deep network. In: NeurIPS. pp. 2366-2374 (2014)", + "4. Feng, Z., Yang, L., Jing, L., Wang, H., Tian, Y., Li, B.: Disentangling object motion and occlusion for unsupervised multi-frame monocular depth. In: ECCV. pp. 228-244 (2022)", + "5. Garg, R., Kumar, B.G.V., Carneiro, G., Reid, I.: Unsupervised cnn for single view depth estimation: Geometry to the rescue. In: ECCV. pp. 740-756 (2016)", + "6. Geiger, A., Lenz, P., Stiller, C., Urtasun, R.: Vision meets robotics: The kitti dataset. *IJRR* **32**(11), 1231–1237 (2013)", + "7. Godard, C., Aodha, O.M., Brostow, G.J.: Unsupervised monocular depth estimation with left-right consistency. In: CVPR. pp. 6602-6611 (2017)", + "8. Godard, C., Aodha, O.M., Firman, M., Brostow, G.J.: Digging into self-supervised monocular depth estimation. In: CVPR. pp. 3828-3838 (2019)", + "9. Gordon, A., Li, H., Jonschkowski, R., Angelova, A.: Depth from videos in the wild: Unsupervised monocular depth learning from unknown cameras. In: ICCV. pp. 8976-8985 (2019)", + "20. Guizilini, V., Ambrus, R., Chen, D., Zakharov, S., Gaidon, A.: Multi-frame self-supervised depth with transformers. In: CVPR. pp. 160-170 (2022)" + ], + "bbox": [ + 225, + 176, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Mono-ViFi: A Unified Framework for Self-supervised Monocular Depth", + "bbox": [ + 256, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "21. Guizilini, V., Ambrus, R., Pillai, S., Gaidon, A.: 3d packing for self-supervised monocular depth estimation. In: CVPR. pp. 2482-2491 (2020)", + "22. Han, W., Yin, J., Jin, X., Dai, X., Shen, J.: Brnet: Exploring comprehensive features for monocular depth estimation. In: ECCV. pp. 586-602 (2022)", + "23. Han, W., Yin, J., Shen, J.: Self-supervised monocular depth estimation by direction-aware cumulative convolution network. In: ICCV (2023)", + "24. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR. pp. 770-778 (2016)", + "25. He, M., Hui, L., Bian, Y., Ren, J., Xie, J., Yang, J.: Ra-depth: Resolution adaptive self-supervised monocular depth estimation. In: ECCV. pp. 565-581 (2022)", + "26. Huang, Z., Zhang, T., Heng, W., Shi, B., Zhou, S.: Real-time intermediate flow estimation for video frame interpolation. In: ECCV. pp. 624-642 (2022)", + "27. Jung, H., Park, E., Yoo, S.: Fine-grained semantics-aware representation enhancement for self-supervised monocular depth estimation. In: ICCV. pp. 12622-12632 (2021)", + "28. Klingner, M., Termöhlen, J.A., Mikolajczyk, J., Fingscheidt, T.: Self-supervised monocular depth estimation: Solving the dynamic object problem by semantic guidance. In: ECCV. pp. 582-600 (2020)", + "29. Kong, L., Jiang, B., Luo, D., Chu, W., Huang, X., Tai, Y., Wang, C., Yang, J.: Ifrnet: Intermediate feature refine network for efficient frame interpolation. In: CVPR. pp. 1959-1968 (2022)", + "30. Kuznietsov, Y., Proesmans, M., Gool, L.V.: Comoda: Continuous monocular depth adaptation using past experiences. In: WACV. pp. 2906-2916 (2021)", + "31. Lee, S., Im, S., Lin, S., Kweon, I.S.: Learning monocular depth in dynamic scenes via instance-aware projection consistency. In: AAAI. pp. 1863-1872 (2021)", + "32. Lee, S., Rameau, F., Pan, F., Kweon, I.S.: Attentive and contrastive learning for joint depth and motion field estimation. In: ICCV. pp. 4862-4871 (2021)", + "33. Li, H., Gordon, A., Zhao, H., Casser, V., Angelova, A.: Unsupervised monocular depth learning in dynamic scenes. In: CoRL. pp. 1908-1917 (2020)", + "34. Li, Z., Zhu, Z.L., Han, L.H., Hou, Q., Guo, C.L., Cheng, M.M.: Amt: All-pairs multi-field transforms for efficient frame interpolation. In: CVPR. pp. 9801-9810 (2023)", + "35. Liu, J., Kong, L., Yang, J.: Designing and searching for lightweight monocular depth network. In: ICONIP. pp. 477-488 (2021)", + "36. Liu, J., Kong, L., Yang, J.: Atca: An arc trajectory based model with curvature attention for video frame interpolation. In: ICIP. pp. 1486-1490 (2022)", + "37. Liu, J., Kong, L., Yang, J., Liu, W.: Towards better data exploitation in self-supervised monocular depth estimation. IEEE Robotics and Automation Letters 9(1), 763-770 (2024)", + "38. Luo, X., Huang, J.B., Szeliski, R., Matzen, K., Kopf, J.: Consistent video depth estimation. TOG 39(71), 1-13 (2020)", + "39. Lyu, X., Liu, L., Wang, M., Kong, X., Liu, L., Liu, Y., Chen, X., Yuan, Y.: Hrdepth: High resolution self-supervised monocular depth estimation. In: AAAI. pp. 2294-2301 (2021)", + "40. Ma, J., Lei, X., Liu, N., Zhao, X., Pu, S.: Towards comprehensive representation enhancement in semantics-guided self-supervised monocular depth estimation. In: ECCV. pp. 304-321 (2022)", + "41. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: ECCV. pp. 405-421 (2020)" + ], + "bbox": [ + 215, + 146, + 784, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "J. Liu et al.", + "bbox": [ + 271, + 114, + 351, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "42. Patil, V., Gansbeke, W.V., Dai, D., Gool, L.V.: Don't forget the past: Recurrent depth estimation from monocular video. IEEE Robotics and Automation Letters 5(4), 6813-6820 (2020)", + "43. Peng, R., Wang, R., Lai, Y., Tang, L., Cai, Y.: Excavating the potential capacity of selfsupervised monocular depth estimation. In: ICCV. pp. 15560-15569 (2021)", + "44. Petrovai, A., Nedevschi, S.: Exploiting pseudo labels in a self-supervised learning framework for improved monocular depth estimation. In: CVPR. pp. 1568-1578 (2022)", + "45. Pilzer, A., Lathuilière, S., Sebe, N., Ricci, E.: Refine and distill: Exploiting cycle-inconsistency and knowledge distillation for unsupervised monocular depth estimation. In: CVPR. pp. 9760-9769 (2019)", + "46. Poggi, M., Aleotti, F., Tosi, F., Mattoccia, S.: On the uncertainty of self-supervised monocular depth estimation. In: CVPR. pp. 3224-3234 (2020)", + "47. Ranjan, A., Jampani, V., Balles, L., Kim, K., Sun, D., Wulff, J., Black, M.J.: Competitive collaboration: Joint unsupervised learning of depth, camera motion, optical flow and motion segmentation. In: CVPR. pp. 12240-12249 (2019)", + "48. Ren, W., Wang, L., Piao, Y., Zhang, M., Lu, H., Liu, T.: Adaptive co-teaching for unsupervised monocular depth estimation. In: ECCV. pp. 89-105 (2022)", + "49. Ruhkamp, P., Gao, D., Chen, H., Navab, N., Busam, B.: Attention meets geometry: Geometry guided spatial-temporal attention for consistent self-supervised monococular depth estimation. In: 3DV. pp. 837-847 (2021)", + "50. Saxena, A., Sun, M., Ng, A.Y.: Make3d: Learning 3d scene structure from a single still image. PAMI 31(5), 824-840 (2009)", + "51. Tancik, M., Srinivasan, P.P., Mildenhall, B., Fridovich-Keil, S., Raghavan, N., Singhal, U., Ramamoorthi, R., Barron, J.T., Ng, R.: Fourier features let networks learn high frequency functions in low dimensional domains. In: NeurIPS. pp. 7537-7547 (2020)", + "52. Uhrig, J., Schneider, N., Schneider, L., Franke, U., Brox, T., Geiger, A.: Sparsity invariant cnns. In: 3DV. pp. 11-20 (2017)", + "53. Wang, C., Buenaposada, J.M., Zhu, R., Lucey, S.: Learning depth from monocular videos using direct methods. In: CVPR. pp. 2022-2030 (2018)", + "54. Wang, R., Pizer, S.M., Frahm, J.M.: Recurrent neural network for (un-)supervised learning of monocular video visual odometry and depth. In: CVPR. pp. 5555-5564 (2019)", + "55. Wang, R., Yu, Z., Gao, S.: Planedepth: Self-supervised depth estimation via orthogonal planes. In: CVPR. pp. 21425-21434 (2023)", + "56. Wang, X., Zhu, Z., Huang, G., Chi, X., Ye, Y., Chen, Z., Wang, X.: Drafting monocular cues and velocity guidance for self-supervised multi-frame depth learning. In: AAAI. pp. 2689-2697 (2023)", + "57. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: From error visibility to structural similarity. IEEE TIP 13(4), 600-612 (2004)", + "58. Watson, J., Aodha, O.M., Prisacariu, V.A., Brostow, G.J., Firman, M.: The temporal opportunist: Self-supervised multi-frame monocular depth. In: CVPR. pp. 1164-1174 (2021)", + "59. Watson, J., Firman, M., Brostow, G.J., Turmukhambetov, D.: Self-supervised monocular depth hints. In: ICCV. pp. 2162-2171 (2019)", + "60. Yin, Z., Shi, J.: Geonet: Unsupervised learning of dense depth, optical flow and camera pose. In: CVPR. pp. 1983-1992 (2018)", + "61. Zhang, H., Li, Y., Cao, Y., Liu, Y., Shen, C., Yan, Y.: Exploiting temporal consistency for real-time video depth estimation. In: ICCV. pp. 1725-1734 (2019)" + ], + "bbox": [ + 215, + 146, + 785, + 840 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Mono-ViFi: A Unified Framework for Self-supervised Monocular Depth", + "bbox": [ + 256, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "62. Zhang, N., Nex, F., Vosselman, G., Kerle, N.: Lite-mono: A lightweight cnn and transformer architecture for self-supervised monocular depth estimation. In: CVPR. pp. 18537-18546 (2023)", + "63. Zhao, C., Poggi, M., Tosi, F., Zhou, L., Sun, Q., Tanga, Y., Mattoccia, S.: Gasmono: Geometry-aided self-supervised monocular depth estimation for indoor scenes. In: ICCV. pp. 16163-16174 (2023)", + "64. Zhao, C., Zhang, Y., Poggi, M., Tosi, F., Guo, X., Zhu, Z., Huang, G., Tang, Y., Mattoccia, S.: Monovit: Self-supervised monocular depth estimation with a vision transformer. In: 3DV. pp. 668-678 (2022)", + "65. Zhou, H., Greenwood, D., Taylor, S.: Self-supervised monocular depth estimation with internal feature fusion. In: BMVC (2021)", + "66. Zhou, T., Brown, M., Snavely, N., Lowe, D.G.: Unsupervised learning of depth and ego-motion from video. In: CVPR. pp. 6612-6619 (2017)", + "67. Zhou, Z., Fan, X., Shi, P., Xin, Y.: R-msfm: Recurrent multi-scale feature modulation for monocular depth estimating. In: ICCV. pp. 12757-12766 (2021)" + ], + "bbox": [ + 215, + 146, + 784, + 354 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "J. Liu et al.", + "bbox": [ + 271, + 114, + 351, + 126 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/Mono-ViFI_ A Unified Learning Framework for Self-supervised Single- and Multi-frame Monocular Depth Estimation/47dcf543-78f7-47b3-955c-778904eede62_model.json b/2024/Mono-ViFI_ A Unified Learning Framework for Self-supervised Single- and Multi-frame Monocular Depth Estimation/47dcf543-78f7-47b3-955c-778904eede62_model.json new file mode 100644 index 0000000000000000000000000000000000000000..9c18c15e1bc375f297d7d25eac811b9930ace4c6 --- /dev/null +++ b/2024/Mono-ViFI_ A Unified Learning Framework for Self-supervised Single- and Multi-frame Monocular Depth Estimation/47dcf543-78f7-47b3-955c-778904eede62_model.json @@ -0,0 +1,2645 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.235, + 0.141, + 0.771, + 0.21 + ], + "angle": 0, + "content": "Mono-ViFi: A Unified Learning Framework for Self-supervised Single- and Multi-frame Monocular Depth Estimation" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.235, + 0.786, + 0.252 + ], + "angle": 0, + "content": "Jinfeng Liu, Lingtong Kong\\*, Bo Li, Zerong Wang, Hong Gu, and Jinwei Chen" + }, + { + "type": "text", + "bbox": [ + 0.35, + 0.263, + 0.653, + 0.277 + ], + "angle": 0, + "content": "vivo Mobile Communication Co., Ltd, China" + }, + { + "type": "text", + "bbox": [ + 0.255, + 0.278, + 0.75, + 0.292 + ], + "angle": 0, + "content": "{liujinfeng, ltkong, libra, wangzerong, guhong, jinwei.chen}@vivo.com" + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.328, + 0.741, + 0.606 + ], + "angle": 0, + "content": "Abstract. Self-supervised monocular depth estimation has gathered notable interest since it can liberate training from dependency on depth annotations. In monocular video training case, recent methods only conduct view synthesis between existing camera views, leading to insufficient guidance. To tackle this, we try to synthesize more virtual camera views by flow-based video frame interpolation (VFI), termed as temporal augmentation. For multi-frame inference, to sidestep the problem of dynamic objects encountered by explicit geometry-based methods like ManyDepth, we return to the feature fusion paradigm and design a VFI-assisted multi-frame fusion module to align and aggregate multi-frame features, using motion and occlusion information obtained by the flow-based VFI model. Finally, we construct a unified self-supervised learning framework, named Mono-ViFi, to bilaterally connect single- and multiframe depth. In this framework, spatial data augmentation through image affine transformation is incorporated for data diversity, along with a triplet depth consistency loss for regularization. The single- and multiframe models can share weights, making our framework compact and memory-efficient. Extensive experiments demonstrate that our method can bring significant improvements to current advanced architectures. Source code is available at https://github.com/LiuJF1226/Mono-ViFi." + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.619, + 0.746, + 0.647 + ], + "angle": 0, + "content": "Keywords: Monocular depth estimation \\(\\cdot\\) Self-supervised learning \\(\\cdot\\) Data augmentation \\(\\cdot\\) Video frame interpolation" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.673, + 0.377, + 0.689 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Monocular depth estimation (MDE) is a fundamental 3D vision task, which finds applications in robotics, autonomous driving and VR/AR. To reduce the dependence on depth annotations, self-supervised MDE [1,2,15,17,46,63,66] has risen as a convenient and practical solution, which uses monocular videos, stereo pairs, or even both to generate supervision through view synthesis, where an image photometric reconstruction loss is minimized. Compared to stereo training case [15,17,59] with known camera pose, learning from monocular videos [18,58, 66] is more challenging since it requires an extra pose network to estimate camera ego-motion. Even so, it is much easier to collect monocular videos compared to" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.128 + ], + "angle": 0, + "content": "J. Liu et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.178 + ], + "angle": 0, + "content": "stereo image pairs, making monocular training more attractive. Therefore, this paper also focuses on self-supervised training with unlabeled monocular videos." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.182, + 0.787, + 0.304 + ], + "angle": 0, + "content": "Recent works have made various attempts to improve accuracy, like introducing masking techniques [18, 66] and leveraging segmentation priors [28, 31]. However, they only conduct view synthesis between existing fixed camera views within video datasets, leading to insufficient guidance especially in invisible regions by occlusion or disappearance. Considering that video data possesses temporal continuity, an intuitive idea is to generate novel frames, i.e., virtual camera views in the temporal dimension. We term it as temporal augmentation, which can be achieved by video frame interpolation (VFI), as shown in Fig. 1(a)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.307, + 0.788, + 0.64 + ], + "angle": 0, + "content": "From another perspective, multi-frame reasoning paradigm have recently been researched because multiple frames can be available for reasoning in many application scenarios. ManyDepth [58] proposes to use cost volume constructed from multi-frame matching features. It explicitly reasons about geometry based on static scene assumption, which can be easily affected by dynamic objects and occlusion. DynamicDepth [14] attempts to tackle dynamic scenes, but relies on depth priors and segmentation masks of moving objects when reasoning. To sidestep the issue of dynamic objects in multi-frame inference encountered by these methods, we return to the traditional paradigm [42, 49], which aggregates multi-frame features directly. However, direct feature fusion without alignment harms the learning of temporal representations, since feature embeddings at the same location in neighbor frames are usually irrelevant, making it difficult for models to leverage the correlation of multiple frames. Hence, we regard feature alignment as a reasonable and necessary step before aggregation. As revealed in several video enhancement tasks like video super-resolution [6, 7] and frame interpolation [26, 29, 34, 36], inter-frame feature alignment can be implemented by warping through optical flow. This inspires us to employ a flow-based VFI model to concurrently conduct temporal augmentation and provide explicit correspondence for feature alignment. Moreover, the VFI model can reason out occlusion mask between neighbor frames, helping to alleviate inter-frame occlusion problem in feature fusion. Based on above analysis, we design a VFI-assisted multi-frame fusion module for multi-frame depth estimation, as depicted in Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Finally, we build a unified framework to bilaterally connect single- and multiframe depth, as demonstrated in Fig. 1(b). The direct way is by depth consistency, usually for geometric regularization or distillation in self-supervised MDE. ManyDepth conducts unilateral distillation from single-frame model to multiframe one. DynamicDepth extends it to a bilateral mutual reinforcement, but only in the regions of moving objects. Differently, our framework applies the bilateral consistency on the whole image. For an original target image, we use a standard view depth consistency (SVDC) loss between its two depth maps predicted from the single- and multi-frame models. We also include classic data augmentation strategies, categorized as spatial augmentation. Specifically, we combine resizing-cropping with image rotation, forming a generic affine transformation. Then, we give the formulation of rectified camera pose for affine-augmented view synthesis using the cue of relative object size [4, 55]. For the" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.257, + 0.114, + 0.733, + 0.131 + ], + "angle": 0, + "content": "Mono-ViFi: A Unified Framework for Self-supervised Monocular Depth" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.128 + ], + "angle": 0, + "content": "3" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.142, + 0.522, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.263, + 0.298, + 0.484, + 0.311 + ], + "angle": 0, + "content": "(a) Temporal Augmentation with VFI" + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.143, + 0.785, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.529, + 0.298, + 0.766, + 0.311 + ], + "angle": 0, + "content": "(b) Learning Framework \\((I_{t}\\) for example" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.319, + 0.789, + 0.446 + ], + "angle": 0, + "content": "Fig. 1: Overview of Mono-ViFI. (a) We achieve temporal augmentation by a flow-based VFI model \\(\\mathcal{H}\\), which also reasons out intermediate optical flow and occlusion mask for multi-frame inference. (b) For target image \\(I_{t}\\), we obtain its single-frame depth \\(D_{t}\\) and multi-frame depth \\(D_{t}^{m}\\). Also, \\(I_{t}\\) is augmented to \\(\\widetilde{I}_{t}\\) through affine transformation \\(\\mathcal{A}\\). We then calculate the depth \\(\\widehat{D}_t\\) of \\(\\widetilde{I}_t\\) and inversely convert it to the original view, generating \\(\\widehat{D}_t\\). Finally, we enforce a triplet depth consistency loss \\(L_{tc}(t)\\) among the three depth maps, including a standard view depth consistency loss \\(L_{sv}(t)\\) and two scale-aware depth consistency losses, \\(L_{sa}(t)\\) and \\(L_{sa}^{m}(t)\\). Note that each depth map also corresponds to a photometric loss and a smoothness loss, which are omitted here." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.477, + 0.789, + 0.568 + ], + "angle": 0, + "content": "depth map from augmented target view, we enforce two scale-aware depth consistency (SADC) losses between it and above two standard view depths, respectively. SADC stems from the relative size cue, which assumes that when an image is zoomed up by a factor, the depth decreases by the same factor. It guides depth model to learn consistent geometric scale relationship from scene variations. The SVDC loss and two SADC losses are termed as a triplet depth consistency loss." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.568, + 0.789, + 0.63 + ], + "angle": 0, + "content": "Our framework is named Mono-ViFI, since the self-supervised Monocular depth is largely boosted by Video Frame Interpolation. To our best knowledge, we are the first to leverage VFI for the improvement of self-supervised MDE. The main contributions of this paper are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.642, + 0.787, + 0.702 + ], + "angle": 0, + "content": "- We propose Mono-ViFi, a unified learning framework to bilaterally connect self-supervised single- and multi-frame MDE. In Mono-ViFi, the single- and multi-frame models can share weights, making the unified framework more compact and memory-efficient." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.703, + 0.787, + 0.748 + ], + "angle": 0, + "content": "- We conduct temporal augmentation through flow-based VFI, and develop a novel VFI-assisted multi-frame fusion module, which utilizes motion and occlusion from flow-based VFI model to improve multi-frame depth." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.749, + 0.787, + 0.779 + ], + "angle": 0, + "content": "- We incorporate image affine transformation as spatial augmentation, along with a triplet depth consistency loss for regularization and distillation." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.78, + 0.787, + 0.84 + ], + "angle": 0, + "content": "- Experiments show that Mono-ViFI can achieve state-of-the-art performance. Moreover, our approach is orthogonal to architecture design, allowing current advanced depth networks to be further improved by seamlessly integrating them with Mono-ViFI." + }, + { + "type": "list", + "bbox": [ + 0.225, + 0.642, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.232, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.128 + ], + "angle": 0, + "content": "J. Liu et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.145, + 0.388, + 0.161 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.173, + 0.634, + 0.189 + ], + "angle": 0, + "content": "2.1 Self-supervised Monocular Depth Estimation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.195, + 0.784, + 0.224 + ], + "angle": 0, + "content": "Self-supervised learning paradigm for monocular depth estimation can be categorized into stereo training and monocular training according to the data source." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.226, + 0.785, + 0.315 + ], + "angle": 0, + "content": "Stereo training leverages the left-right consistency in stereo image pairs. Garg et al. [15] first consider the training as a view synthesis problem with a photometric consistency loss between stereo pairs. Building upon this idea, Godard et al. [17] introduce a left-right disparity consistency loss. Then, it is further extended by using depth hints [59], data augmentation [43, 55], knowledge distillation [43, 45, 55] and plane-based scene representation [3, 55]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.317, + 0.787, + 0.558 + ], + "angle": 0, + "content": "Monocular training leverages the temporal consistency in monocular video frames. Zhou et al. [66] propose a pioneering method which jointly learns depth and camera ego-motion from monocular videos in a self-supervised way. Then, Bian et al. [5] propose a scale-consistent depth and ego-motion estimation approach by adding a depth consistency loss, which reduces scale drift of estimated depth but decreases the accuracy. Later, Godard et al. [18] propose Monodepth2, a strong baseline for most follow-up works, which employs a minimum reprojection loss to mitigate occlusion issues and an auto-masking loss to filter out moving objects with similar velocity as the camera. Furthermore, some works introduce multi-task learning to combine depth estimation with optical flow [9, 32, 47, 60], semantic segmentation [27, 28, 40] and instance segmentation [31]. Others [21-23, 35, 37, 39, 62, 64, 65] try to improve the network architectures for depth estimation. However, these methods only utilize the geometric constrains between existing fixed camera views in video datasets. This insufficient guidance may block current depth models from unleashing their full capabilities. To address it, we conduct temporal augmentation through VFI in this paper." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.577, + 0.514, + 0.592 + ], + "angle": 0, + "content": "2.2 Multi-frame Monocular Depth" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Above single-frame depth models do not leverage any temporal information for inference, which limits the performance to some degree. Regarding multi-frame depth prediction, early works adopt test-time refinement [9, 30, 38], use spatial-temporal attention [49], or combine traditional depth networks with recurrent layers [54, 61] to process sequence information. Another broad group of methods [14, 20, 56, 58] utilize cost volume from stereo matching, which explicitly encode and reason from geometric constraints by combining a multi-frame cost volume with single-frame features. On the one hand, this type of methods based on ManyDepth [58] usually meet with the problem of dynamic scenes. To sidestep this issue, we develop a VFI-assisted multi-frame fusion depth model to align and aggregate multi-frame features by optical flow and occlusion masks. On the other hand, ManyDepth uses single-frame depth prior to boost multi-frame depth estimation. However, this improvement is unidirectional. DynamicDepth [14] extends it to a mutual reinforcement, but only targets on dynamic objects. Differently, we propose a unified framework to bilaterally connect single- and multi-frame depth in a more compact and memory-efficient manner." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.257, + 0.114, + 0.733, + 0.131 + ], + "angle": 0, + "content": "Mono-ViFi: A Unified Framework for Self-supervised Monocular Depth" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.145, + 0.4, + 0.163 + ], + "angle": 0, + "content": "3 Problem Setup" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.788, + 0.314 + ], + "angle": 0, + "content": "Self-supervised MDE in the monocular training case leverages the geometric constraints from video sequences. Specifically, for a training triplet \\(\\{I_{t-2}, I_t, I_{t+2}\\}\\) which contains three consecutive frames, we use \\(I_t\\) as the target view and the other two as the source views \\(\\{I_s\\}_{s \\in \\{t-2, t+2\\}}\\) to implement view synthesis. Here the temporal notation \\(t \\pm 2\\) instead of \\(t \\pm 1\\) is for the convenience of following description. Also, a pose network is utilized to estimate the camera pose from the target view to the source view, denoted as \\(\\pmb{T}_{t \\to s} = [\\pmb{R}_{t \\to s} | \\pmb{t}_{t \\to s}]\\), where \\(\\pmb{R}_{t \\to s}\\) and \\(\\pmb{t}_{t \\to s}\\) represent the rotation and translation components. Formally, the target depth \\(D_t\\) and the camera pose \\(\\pmb{T}_{t \\to s}\\) can be derived from:" + }, + { + "type": "equation", + "bbox": [ + 0.379, + 0.322, + 0.785, + 0.337 + ], + "angle": 0, + "content": "\\[\nD _ {t} = \\mathcal {D} \\left(\\mathcal {E} \\left(I _ {t}\\right)\\right), \\quad \\boldsymbol {T} _ {t \\rightarrow s} = \\mathcal {P} \\left(I _ {t}, I _ {s}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.343, + 0.785, + 0.388 + ], + "angle": 0, + "content": "where \\(\\mathcal{E}\\) and \\(\\mathcal{D}\\) represent the encoder and decoder, respectively, of a single-frame depth network, while \\(\\mathcal{P}\\) denotes the pose network. Then, the reconstructed view of \\(I_{t}\\) from the source view \\(I_{s}\\) can be obtained by:" + }, + { + "type": "equation", + "bbox": [ + 0.392, + 0.396, + 0.785, + 0.412 + ], + "angle": 0, + "content": "\\[\nI _ {s \\rightarrow t} = I _ {s} \\left\\langle \\operatorname {p r o j} \\left(D _ {t}, \\boldsymbol {T} _ {t \\rightarrow s}, \\boldsymbol {K}\\right)\\right\\rangle , \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.418, + 0.788, + 0.538 + ], + "angle": 0, + "content": "where \\( I_{s\\rightarrow t} \\) is the reconstructed target image, \\( \\langle \\rangle \\) means the differentiable bilinear sampling operator, \\( proj() \\) returns the 2D coordinates of the pixels in \\( I_{t} \\) after they are reprojected to the viewpoint of \\( I_{s} \\), and \\( K \\) is the camera intrinsic matrix assumed to be known. Finally, the per-pixel photometric loss between \\( I_{s\\rightarrow t} \\) and \\( I_{t} \\) is used to optimize the depth and pose networks in an end-to-end manner. We follow Monodepth2 [18] to adopt the combination of L1 distance and SSIM [57] as the photometric error function and take the minimum across all the source views at each pixel to handle occlusion, which is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.297, + 0.545, + 0.785, + 0.569 + ], + "angle": 0, + "content": "\\[\nL _ {p e} (t) = \\min _ {s} \\frac {\\alpha}{2} \\left(1 - \\operatorname {S S I M} \\left(I _ {t}, I _ {s \\rightarrow t}\\right)\\right) + (1 - \\alpha) \\| I _ {t}, I _ {s \\rightarrow t} \\| _ {1}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.574, + 0.785, + 0.603 + ], + "angle": 0, + "content": "where \\(\\alpha = 0.85\\). Besides, the edge-aware smoothness loss is employed to cope with depth discontinuities:" + }, + { + "type": "equation", + "bbox": [ + 0.362, + 0.61, + 0.785, + 0.627 + ], + "angle": 0, + "content": "\\[\nL _ {s m} (t) = \\left| \\partial_ {x} d _ {t} ^ {*} \\right| e ^ {- \\left| \\partial_ {x} I _ {t} \\right|} + \\left| \\partial_ {y} d _ {t} ^ {*} \\right| e ^ {- \\left| \\partial_ {y} I _ {t} \\right|}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.633, + 0.784, + 0.664 + ], + "angle": 0, + "content": "where \\( d_t^* = d_t / \\overline{d_t} \\) is the mean-normalized inverse depth to discourage shrinking [53]. In summary, the standard self-supervised loss on \\( I_{t} \\) is:" + }, + { + "type": "equation", + "bbox": [ + 0.327, + 0.672, + 0.785, + 0.687 + ], + "angle": 0, + "content": "\\[\nL _ {s s} (t) = \\mathcal {L} \\left(D _ {t}, \\boldsymbol {T} _ {t \\rightarrow s}, I _ {t}, I _ {s}\\right) = \\mu L _ {p e} (t) + \\gamma L _ {s m} (t), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.693, + 0.584, + 0.709 + ], + "angle": 0, + "content": "where \\(\\mu\\) is the auto-masking in [18] and \\(\\gamma = 0.001\\)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.732, + 0.381, + 0.751 + ], + "angle": 0, + "content": "4 Methodology" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In this paper, we propose novel Mono-ViFI for self-supervised monocular training, which bilaterally connects single- and multi-frame depth models into a unified learning framework. With flow-based VFI model, we introduce temporal augmentation and develop VFI-assisted multi-frame fusion depth, which are presented in Sec. 4.1 and Sec. 4.2, respectively. Besides, our framework incorporates" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.128 + ], + "angle": 0, + "content": "J. Liu et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.195 + ], + "angle": 0, + "content": "spatial data augmentation and a triplet depth consistency loss, which are separately described in Sec. 4.3 and Sec. 4.4. Finally, we give the overall training objective in Sec. 4.5. The overview of our Mono-ViFi is depicted in Fig. 1." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.212, + 0.548, + 0.228 + ], + "angle": 0, + "content": "4.1 Temporal Augmentation with VFI" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.234, + 0.788, + 0.31 + ], + "angle": 0, + "content": "Firstly, we generate more new frames in the temporal dimension based on existing video frames for temporal augmentation. To achieve this, we adopt a pretrained flow-based VFI model, such as IFRNet [29], which takes two temporally adjacent frames \\(\\{I_0,I_2\\}\\) as input to synthesize an intermediate frame \\(I_{1}\\). The frame interpolation process can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.397, + 0.318, + 0.785, + 0.334 + ], + "angle": 0, + "content": "\\[\nF _ {1 \\rightarrow 0}, F _ {1 \\rightarrow 2}, M _ {1} = \\mathcal {H} \\left(I _ {0}, I _ {2}\\right), \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.347, + 0.337, + 0.785, + 0.352 + ], + "angle": 0, + "content": "\\[\nI _ {0 \\rightarrow 1} = \\omega \\left(I _ {0}, F _ {1 \\rightarrow 0}\\right), \\quad I _ {2 \\rightarrow 1} = \\omega \\left(I _ {2}, F _ {1 \\rightarrow 2}\\right), \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.374, + 0.356, + 0.785, + 0.371 + ], + "angle": 0, + "content": "\\[\nI _ {1} = M _ {1} \\odot I _ {0 \\rightarrow 1} + (1 - M _ {1}) \\odot I _ {2 \\rightarrow 1}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.377, + 0.789, + 0.528 + ], + "angle": 0, + "content": "where \\(\\mathcal{H}\\) is the VFI model, \\(F_{1\\rightarrow 0}\\) and \\(F_{1\\rightarrow 2}\\) are the intermediate optical flow, \\(\\omega\\) represents backward warping. \\(I_{0\\rightarrow 1}\\) and \\(I_{2\\rightarrow 1}\\) denote the warped images to the intermediate position, while \\(M_{1}\\) ranging from 0 to 1 is the merge mask adjusting the mixing ratio according to bidirectional occlusion information. Here, \\(\\odot\\) indicates the element-wise multiplication. Considering a triplet \\(\\{I_{t - 2},I_t,I_{t + 2}\\}\\) with three consecutive frames as described in Sec. 3, we use the VFI model to synthesize \\(I_{t - 1}\\) from \\(\\{I_{t - 2},I_t\\}\\) and \\(I_{t + 1}\\) from \\(\\{I_t,I_{t + 2}\\}\\), as shown in Fig. 1(a). Then, we regard \\(I_{t - 1}\\) and \\(I_{t + 1}\\) as another two target frames. All three target frames have common sources of \\(\\{I_{t - 2},I_{t + 2}\\}\\). The following operations and losses are the same for each of them, and we only take \\(I_{t}\\) for instance." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.548, + 0.589, + 0.563 + ], + "angle": 0, + "content": "4.2 VFI-assisted Multi-frame Fusion Depth" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.57, + 0.788, + 0.707 + ], + "angle": 0, + "content": "Apart from temporal augmentation, our employed VFI model can also provide other inter-frame information, such as optical flow and occlusion merge masks. This enables us to design a novel VFI-assisted multi-frame fusion module for multi-frame depth prediction, as demonstrated in Fig. 2. The fusion module consists of two parts, i.e., motion-aware feature alignment (MAFA) and occlusion-alarviated feature fusion (OAFF). Taking original triplet \\(\\{I_{t - 2},I_t,I_{t + 2}\\}\\) for example, we implement forward process of the VFI model to interpolate the middle position \\(t\\) of \\(\\{I_{t - 2},I_{t + 2}\\}\\) in Eq. (9), yielding intermediate flow \\(F_{t\\rightarrow (t - 2)},F_{t\\rightarrow (t + 2)}\\) and merge mask \\(M_{t}\\) to align and aggregate multi-frame features." + }, + { + "type": "equation", + "bbox": [ + 0.364, + 0.714, + 0.785, + 0.73 + ], + "angle": 0, + "content": "\\[\nF _ {t \\rightarrow (t - 2)}, F _ {t \\rightarrow (t + 2)}, M _ {t} = \\mathcal {H} \\left(I _ {t - 2}, I _ {t + 2}\\right). \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.736, + 0.788, + 0.781 + ], + "angle": 0, + "content": "Multi-frame Feature Extraction. Our fusion module accepts multi-frame features extracted by the same encoder, which is shared with the single-frame depth encoder \\(\\mathcal{E}\\). The extraction process can be written as:" + }, + { + "type": "equation", + "bbox": [ + 0.336, + 0.787, + 0.785, + 0.804 + ], + "angle": 0, + "content": "\\[\n\\left\\{\\phi_ {t - 2} ^ {k}, \\phi_ {t} ^ {k}, \\phi_ {t + 2} ^ {k} \\right\\} _ {k = 1} ^ {N} = \\left\\{\\mathcal {E} \\left(I _ {t - 2}\\right), \\mathcal {E} \\left(I _ {t}\\right), \\mathcal {E} \\left(I _ {t + 2}\\right) \\right\\}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.809, + 0.788, + 0.84 + ], + "angle": 0, + "content": "where \\(\\{\\phi_{t - 2}^k,\\phi_t^k,\\phi_{t + 2}^k\\}_{k = 1}^N\\) are the multi-level features of the three frames and \\(N\\) is the number of feature levels." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.257, + 0.114, + 0.733, + 0.131 + ], + "angle": 0, + "content": "Mono-ViFi: A Unified Framework for Self-supervised Monocular Depth" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.216, + 0.146, + 0.788, + 0.31 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.32, + 0.784, + 0.348 + ], + "angle": 0, + "content": "Fig. 2: Our VFI-assisted multi-frame fusion depth model. Motion and occlusion provided by VFI can be used to align and fuse multi-frame features in an explicit manner." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.386, + 0.784, + 0.435 + ], + "angle": 0, + "content": "Motion-aware Feature Alignment (MAFA). For the \\(k\\)-th level, we use bidirectional optical flow to align the neighbor features \\(\\phi_{t-2}^k\\) and \\(\\phi_{t+2}^k\\) to the target position \\(t\\), generating \\(\\phi_{(t-2) \\to t}^k\\) and \\(\\phi_{(t+2) \\to t}^k\\), respectively, formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.39, + 0.44, + 0.785, + 0.482 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\phi_ {(t - 2) \\rightarrow t} ^ {k} = \\omega \\left(\\phi_ {t - 2} ^ {k}, F _ {t \\rightarrow (t - 2)} ^ {k}\\right), \\\\ \\phi_ {(t + 2) \\rightarrow t} ^ {k} = \\omega \\left(\\phi_ {t + 2} ^ {k}, F _ {t \\rightarrow (t + 2)} ^ {k}\\right), \\tag {11} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.488, + 0.784, + 0.548 + ], + "angle": 0, + "content": "where \\( F_{t\\rightarrow (t - 2)}^k \\) and \\( F_{t\\rightarrow (t + 2)}^k \\) denote the corresponding flows of the \\( k \\)-th level, downsampled from \\( F_{t\\rightarrow (t - 2)} \\) and \\( F_{t\\rightarrow (t + 2)} \\), respectively. Inspired by recent 2D implicit neural representation [8], we also concatenate the aligned features with encoded optical flow, which can be denoted as:" + }, + { + "type": "equation", + "bbox": [ + 0.345, + 0.554, + 0.785, + 0.616 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\varphi_ {(t - 2) \\rightarrow t} ^ {k} = \\operatorname {C o n c a t} \\left(\\phi_ {(t - 2) \\rightarrow t} ^ {k}, g \\left(F _ {t \\rightarrow (t - 2)} ^ {k}\\right)\\right), \\\\ \\varphi_ {(t + 2) \\rightarrow t} ^ {k} = \\operatorname {C o n c a t} \\left(\\phi_ {(t + 2) \\rightarrow t} ^ {k}, g \\left(F _ {t \\rightarrow (t + 2)} ^ {k}\\right)\\right), \\tag {12} \\\\ \\varphi_ {t} ^ {k} = \\operatorname {C o n c a t} \\left(\\phi_ {t} ^ {k}, g \\left(F _ {0} ^ {k}\\right)\\right), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.624, + 0.784, + 0.699 + ], + "angle": 0, + "content": "where Concat means feature concatenation, \\( F_0^k \\) denotes zero flow and \\( g \\) is fourier positional encoding [41,51] to transform optical flow to high-frequency signal. Actually, optical flow indicates the relative coordinate of a feature vector in the warped feature map w.r.t. the original one. When incorporating it, multi-frame model can adaptively adjust feature learning through motion-aware alignment." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.703, + 0.784, + 0.779 + ], + "angle": 0, + "content": "Occlusion-alleviated Feature Fusion (OAFF). Due to the ubiquitous interframe occlusion, aggregating the aligned features directly through summation or concatenation can deteriorate the performance. Therefore, we use the \\(k\\)-th level merge mask \\(M_t^k\\) downsampled from \\(M_t\\) to mitigate this issue, just like the image merging process of VFI in Eq. (8):" + }, + { + "type": "equation", + "bbox": [ + 0.338, + 0.783, + 0.785, + 0.803 + ], + "angle": 0, + "content": "\\[\n\\chi_ {t} ^ {k} = M _ {t} ^ {k} \\odot \\varphi_ {(t - 2) \\rightarrow t} ^ {k} + (1 - M _ {t} ^ {k}) \\odot \\varphi_ {(t + 2) \\rightarrow t} ^ {k}, \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.784, + 0.84 + ], + "angle": 0, + "content": "where \\(\\chi_t^k\\) is the occlusion-alleviated feature merged from the temporal positions \\(t - 2\\) and \\(t + 2\\). It is then concatenated with the feature \\(\\varphi_t^k\\) in the position \\(t\\)." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.128 + ], + "angle": 0, + "content": "J. Liu et al." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.147, + 0.586, + 0.162 + ], + "angle": 0, + "content": "And we can obtain the final feature after fusion by:" + }, + { + "type": "equation", + "bbox": [ + 0.386, + 0.168, + 0.786, + 0.186 + ], + "angle": 0, + "content": "\\[\n\\psi_ {t} ^ {k} = \\operatorname {C o n v} _ {1 \\times 1} \\left(\\operatorname {C o n c a t} \\left(\\varphi_ {t} ^ {k}, \\chi_ {t} ^ {k}\\right)\\right), \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.191, + 0.784, + 0.237 + ], + "angle": 0, + "content": "where \\(\\mathrm{Conv}_{1\\times 1}\\) is a \\(1\\times 1\\) convolutional layer to transform the feature into the same channel as the original feature \\(\\phi_t^k\\), enabling the multi-frame depth decoder to employ the same architecture as the single-frame one." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.239, + 0.785, + 0.343 + ], + "angle": 0, + "content": "Multi-frame Depth Decoder. To translate the multi-level fusion features into a depth map, a decoder is required. Concretely, we use the same architecture as the single-frame depth decoder, but the parameters are not shared. This encourages the single- and multi-frame models to decode the features in their individual ways without conflict. Finally, the fusion features \\(\\{\\psi_t^k\\}_{k=1}^N\\) are processed by a multi-frame depth decoder \\(\\mathcal{D}_m\\) to generate the fusion depth \\(D_t^m\\) of the target frame \\(I_t\\), which can be written as:" + }, + { + "type": "equation", + "bbox": [ + 0.428, + 0.35, + 0.786, + 0.368 + ], + "angle": 0, + "content": "\\[\nD _ {t} ^ {m} = \\mathcal {D} _ {m} \\left(\\left\\{\\psi_ {t} ^ {k} \\right\\} _ {k = 1} ^ {N}\\right). \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.374, + 0.784, + 0.403 + ], + "angle": 0, + "content": "Note that the multi-frame depth \\( D_{t}^{m} \\) also corresponds to a self-supervised loss like Eq. (5), which is denoted as:" + }, + { + "type": "equation", + "bbox": [ + 0.399, + 0.412, + 0.786, + 0.428 + ], + "angle": 0, + "content": "\\[\nL _ {s s} ^ {m} (t) = \\mathcal {L} \\left(D _ {t} ^ {m}, \\boldsymbol {T} _ {t \\rightarrow s}, I _ {t}, I _ {s}\\right). \\tag {16}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.457, + 0.443, + 0.473 + ], + "angle": 0, + "content": "4.3 Spatial Augmentation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.49, + 0.788, + 0.75 + ], + "angle": 0, + "content": "Orthogonal to temporal augmentation, we also include classic data augmentation to spatially expand the data diversity, which is categorized as spatial augmentation in this paper. Resizing-cropping is a commonly used and effective augmentation approach. Based on this operation, we introduce image rotation, extending it to a generic affine transformation (cropping is actually image translation). Specifically, we follow [55] to consider that the augmented image and the original image are captured by the same camera system whose intrinsics are fixed. For resizing operation, it is assumed that when an image is zoomed up by a factor \\( f_{s} \\geq 1 \\), the relative size of an object increases by the same factor \\( f_{s} \\) and its depth decreases by \\( f_{s} \\) simultaneously. This is inspired from a vital monocular cue [4] that the closer an object is, the larger its relative size is. Our affine transformation \\( \\mathcal{A} \\) is applied to both the target image \\( I_{t} \\) and the source image \\( I_{s} \\), generating \\( \\widetilde{I}_{t} = \\mathcal{A}(I_{t}) \\) and \\( \\widetilde{I}_{s} = \\mathcal{A}(I_{s}) \\), respectively. For the augmented target view, we can obtain its depth through the single-frame model by \\( \\widetilde{D}_{t} = \\mathcal{D}(\\mathcal{E}(\\widetilde{I}_{t})) \\). To conduct view synthesis on augmented images, the original camera pose \\( T_{t\\rightarrow s} = [\\pmb{R}_{t\\rightarrow s}|\\pmb{t}_{t\\rightarrow s}] \\) from target to source should be rectified to \\( \\widetilde{T}_{t\\rightarrow s} = [\\widetilde{\\pmb{R}}_{t\\rightarrow s}|\\widetilde{\\pmb{t}}_{t\\rightarrow s}] \\) as done in [55]:" + }, + { + "type": "equation", + "bbox": [ + 0.356, + 0.757, + 0.786, + 0.774 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\boldsymbol {R}} _ {t \\rightarrow s} = \\boldsymbol {R} _ {c} \\boldsymbol {R} _ {t \\rightarrow s} \\boldsymbol {R} _ {c} ^ {- 1}, \\quad \\tilde {\\boldsymbol {t}} _ {t \\rightarrow s} = \\boldsymbol {R} _ {c} \\boldsymbol {t} _ {t \\rightarrow s}, \\tag {17}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "where \\(\\mathbf{R}_c\\) is the rectification matrix determined only by the camera intrinsics and the augmentation parameters. Different from [55] calculating \\(\\mathbf{R}_c\\) under resizing-cropping, we give its formulation and derivation under a more general affine transformation, which can be seen in our supplementary. Therefore, we can derive" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.256, + 0.114, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Mono-ViFi: A Unified Framework for Self-supervised Monocular Depth" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.146, + 0.764, + 0.163 + ], + "angle": 0, + "content": "the self-supervised loss \\(\\widetilde{L}_{ss}(t)\\) on the augmented images similarly to Eq. (5):" + }, + { + "type": "equation", + "bbox": [ + 0.403, + 0.169, + 0.785, + 0.187 + ], + "angle": 0, + "content": "\\[\n\\widetilde {L} _ {s s} (t) = \\mathcal {L} (\\widetilde {D} _ {t}, \\widetilde {\\boldsymbol {T}} _ {t \\rightarrow s}, \\widetilde {I} _ {t}, \\widetilde {I} _ {s}). \\tag {18}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.194, + 0.786, + 0.226 + ], + "angle": 0, + "content": "Note that there is a mask to exclude the possible invalid pixels in \\(\\widetilde{I}_t\\) when calculating \\(\\widetilde{L}_{ss}(t)\\) due to image rotation." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.244, + 0.482, + 0.26 + ], + "angle": 0, + "content": "4.4 Triplet Depth Consistency" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.267, + 0.788, + 0.388 + ], + "angle": 0, + "content": "Now we have single-frame depths \\( D_{t} \\) and \\( \\widetilde{D}_{t} \\) of the original image \\( I_{t} \\) and the augmented image \\( \\widetilde{I}_{t} \\), along with multi-frame depth \\( D_{t}^{m} \\) of \\( I_{t} \\). For better knowledge distillation and regularization, our unified learning framework applies a triplet depth consistency loss among the three depth maps, including a standard view depth consistency (SVDC) loss and two scale-aware depth consistency (SADC) losses, which are shown in Fig. 1(b). The SVDC loss \\( L_{sv}(t) \\) conducts mutual distillation between single- and multi-frame models, promoting their mutual reinforcement, which can be denoted as:" + }, + { + "type": "equation", + "bbox": [ + 0.428, + 0.396, + 0.785, + 0.412 + ], + "angle": 0, + "content": "\\[\nL _ {s v} (t) = \\operatorname {S I} \\left(D _ {t} ^ {m}, D _ {t}\\right), \\tag {19}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.417, + 0.788, + 0.514 + ], + "angle": 0, + "content": "where \\(\\mathrm{SI}()\\) is the widely used scale-invariant error function [13]. Besides, we use depth \\(\\widetilde{D}_t\\) under spatial augmentation to enforce two SADC losses, i.e., \\(L_{sa}(t)\\) between \\(\\widetilde{D}_t\\) and \\(D_t\\), and \\(L_{sa}^m(t)\\) between \\(\\widetilde{D}_t\\) and \\(D_t^m\\). To this end, we first restore \\(\\widetilde{D}_t\\) to the original image view, generating \\(\\widehat{D}_t = \\mathcal{A}^{-1}(\\widetilde{D}_t)\\) in Fig. 1(b), where \\(\\mathcal{A}^{-1}\\) is the inverse transformation of affine augmentation \\(\\mathcal{A}\\). To exclude the invalid pixels in \\(\\widehat{D}_t\\), we adopt a mask \\(M_{sa}\\) and calculate \\(L_{sa}(t)\\) and \\(L_{sa}^m(t)\\) through:" + }, + { + "type": "equation", + "bbox": [ + 0.374, + 0.534, + 0.785, + 0.551 + ], + "angle": 0, + "content": "\\[\nL _ {s a} (t) = \\operatorname {S I} \\left(M _ {s a} \\odot D _ {t}, M _ {s a} \\odot f _ {s} \\widehat {D} _ {t}\\right), \\tag {20}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.37, + 0.556, + 0.785, + 0.573 + ], + "angle": 0, + "content": "\\[\nL _ {s a} ^ {m} (t) = \\operatorname {S I} \\left(M _ {s a} \\odot D _ {t} ^ {m}, M _ {s a} \\odot f _ {s} \\widehat {D} _ {t}\\right), \\tag {21}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.579, + 0.788, + 0.64 + ], + "angle": 0, + "content": "where the factor \\( f_{s} \\) is due to the assumption in Sec. 4.3 that the depth of an object in \\( I_{t} \\) is \\( f_{s} \\) times of that in \\( \\widetilde{I}_{t} \\). Our SADC constrains the 3D geometric relationship between the scenes of \\( I_{t} \\) and \\( \\widetilde{I}_{t} \\), providing scale-aware regularization. Combining SVDC and SADC, we have the triplet depth consistency loss as:" + }, + { + "type": "equation", + "bbox": [ + 0.383, + 0.647, + 0.785, + 0.663 + ], + "angle": 0, + "content": "\\[\nL _ {t c} (t) = L _ {s v} (t) + L _ {s a} (t) + L _ {s a} ^ {m} (t). \\tag {22}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.681, + 0.419, + 0.697 + ], + "angle": 0, + "content": "4.5 Training Objective" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.704, + 0.787, + 0.766 + ], + "angle": 0, + "content": "For target position \\( t \\), we have already obtained the self-supervised losses \\( L_{ss}(t) \\), \\( L_{ss}^{m}(t) \\) and \\( \\widetilde{L}_{ss}(t) \\) in Eqs. (5), (16) and (18), respectively. Besides, our triplet depth consistency loss \\( L_{tc}(t) \\) is acquired in Eq. (22). Therefore, the total loss on position \\( t \\) can be summed as:" + }, + { + "type": "equation", + "bbox": [ + 0.355, + 0.771, + 0.785, + 0.789 + ], + "angle": 0, + "content": "\\[\nL (t) = L _ {s s} (t) + L _ {s s} ^ {m} (t) + \\widetilde {L} _ {s s} (t) + \\lambda L _ {t c} (t), \\tag {23}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.795, + 0.788, + 0.841 + ], + "angle": 0, + "content": "where \\(\\lambda\\) is a hyper-parameter set to 0.2 in this paper. For \\(t - 1\\) and \\(t + 1\\), we can derive their losses \\(L(t - 1)\\) and \\(L(t + 1)\\) in the same way. In summary, the overall loss function of our Mono-ViFI is \\(L = L(t - 1) + L(t) + L(t + 1)\\)." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.127 + ], + "angle": 0, + "content": "J. Liu et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.228 + ], + "angle": 0, + "content": "Table 1: Performance comparison on KITTI [16] benchmark using the raw ground truth, with a resolution of \\(640 \\times 192\\). The best results in each subsection are in bold; second best are underlined. For MACs of multi-frame models, we use the formulation A/B. A denotes the MACs when predicting for a single frame; B means the average MACs per frame when processing a long video sequence. #Params and MACs of our multi-frame models include the VFI network." + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.25, + 0.788, + 0.538 + ], + "angle": 0, + "content": "
MethodTest frames#ParamsMACsAbs RelSq RelRMSERMSE logδ1δ2δ3
PackNet-SfM [21]1128M205G0.1110.7854.6010.1890.8780.9600.982
HR-Depth [39]116.9M31.6G0.1090.7924.6320.1850.8840.9620.983
R-MSFM6 [67]13.8M31.2G0.1120.8064.7040.1910.8780.9600.981
BRNet [22]119.1M33.7G0.1050.6984.4620.1790.8900.9650.984
MonoViT [64]127M-0.0990.7084.3720.1750.9000.9670.984
DIFFNet [65]112.2M23.4G0.1020.7644.4830.1800.8960.9650.983
SD-SSMDE [44]1--0.1000.6614.2640.1720.8960.9670.985
MUSTNet [48]1--0.1060.7634.5620.1820.8880.9630.983
DepthSegNet [40]130.3M-0.1040.6904.4730.1790.8860.9650.984
DualRefine [2]1--0.1030.7764.4910.1810.8940.9650.983
DaCNN [23]113M-0.0990.6614.3160.1730.8970.9670.985
Monodepth2 (ResNet18) [18]114.3M8G0.1150.9034.8630.1930.8770.9590.981
Mono-ViFI (ResNet18)114.3M8G0.1050.7084.4460.1790.8870.9650.984
Lite-Mono [62]13.1M5.1G0.1070.7654.5610.1830.8860.9630.983
Mono-ViFI (Lite-Mono)13.1M5.1G0.1030.6804.3490.1760.8910.9660.985
D-HRNet [25]110M10.8G0.1020.7604.4790.1790.8970.9650.983
RA-Depth (D-HRNet) [25]110M10.8G0.0960.6324.2160.1710.9030.9680.985
Mono-ViFI (D-HRNet)110M10.8G0.0960.6274.1790.1700.9030.9690.985
Patil et al. [42]N--/-0.1110.8214.6500.1870.8830.9610.982
ManyDepth (ResNet18) [58]2 (-1, 0)26.9M15.1G/13.7G0.0980.7704.4590.1760.9000.9650.983
DynamicDepth [14]2 (-1, 0)--/-0.0960.7204.4580.1750.8970.9640.984
DepthFormer [20]2 (-1, 0)28.7M175G/-0.0900.6614.1490.1750.9050.9670.984
MOVEDepth [56]2 (-1, 0)28.2M20.2G/19.2G0.0940.7044.3890.1750.9020.9650.983
DualRefine [2]2 (-1, 0)--/-0.0870.6984.2340.1700.9140.9670.983
TC-Depth [49]3 (-1, 0, 1)--/-0.1030.7464.4830.1850.894-0.983
Mono-ViFI (ResNet18)3 (-1, 0, 1)17.9M25.8G/16.9G0.0990.6614.3210.1740.8980.9660.985
Mono-ViFI (Lite-Mono)3 (-1, 0, 1)5.9M21.9G/13.3G0.0990.6484.2560.1720.8980.9670.985
Mono-ViFI (D-HRNet)3 (-1, 0, 1)12.9M35.7G/19.4G0.0910.5894.0880.1660.9120.9690.986
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.575, + 0.377, + 0.594 + ], + "angle": 0, + "content": "5 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.606, + 0.458, + 0.621 + ], + "angle": 0, + "content": "5.1 Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.788, + 0.81 + ], + "angle": 0, + "content": "Datasets. We train and evaluate the models on two widely used outdoor urban datasets of KITTI [16] and Cityscapes [10]. The depth labels of KITTI are collected with LiDARs. For fair comparison, we use the Eigen split [12] and follow the pre-processing operation in [18], generating 39,810 monocular triplets for training and 697 frames for evaluation. To reduce the influence of noise in the sparse depth from Velodyne, we also evaluate using \\(93\\%\\) of the Eigen split with the improved ground truth from [52], which contains 652 test frames. In regard to Cityscapes, following [58, 66], we preprocess the monocular sequences into triples, producing 69,731 training images. And we evaluate on 1,525 test images using the ground truth provided by [58]. For generalization comparison, KITTI models are evaluated on Cityscapes and Make3D [50]. Make3D is a small outdoor dataset containing 134 test images with aligned depth information." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Flow-based VFI Model. We choose IFRNet [29] as our VFI model due to its efficiency. In practice, we use the large model of IFRNet for training and replace" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.257, + 0.114, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Mono-ViFi: A Unified Framework for Self-supervised Monocular Depth" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.146, + 0.396, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.396, + 0.147, + 0.519, + 0.305 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.147, + 0.653, + 0.305 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.653, + 0.15, + 0.783, + 0.305 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.317, + 0.785, + 0.346 + ], + "angle": 0, + "content": "Fig. 3: Qualitative results on KITTI. Error maps in columns 2 and 4 show the Abs Rel error compared to the improved ground truth [52] from good (blue) to bad (red)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.373, + 0.788, + 0.478 + ], + "angle": 0, + "content": "it with the small model when evaluating for multi-frame fusion depth. Please see ablation about their gap in supplementary. Both large and small IFRNets are pretrained on KITTI/Cityscapes and frozen in Mono-ViFI training. The pretraining data split is the same as self-supervised depth training. Each original triplet in depth training is directly used for VFI training. Besides, we employ another two flow-based VFI networks, RIFE [26] and AMT [34], to verify the robustness of our method to different VFI models." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.48, + 0.788, + 0.571 + ], + "angle": 0, + "content": "Depth and Pose Networks. We experiment with three depth architectures, including the ResNet18 [24] based network in Monodepth2 [18], medium LiteMono [62], and D-HRNet proposed in RA-Depth [25]. For the pose network, we follow Monodepth2 to use a ResNet18 based model which takes a pair of images (six channels) as input and predict a 6-DoF relative pose. Encoders of all the networks are pretrained on ImageNet [11]." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.59, + 0.473, + 0.606 + ], + "angle": 0, + "content": "5.2 Performance Comparison" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.789, + 0.842 + ], + "angle": 0, + "content": "KITTI Results. Quantitative comparison results on KITTI under \\(640 \\times 192\\) resolution are listed in Table 1 and Table 2, using raw ground truth and improved ground truth, respectively. It is obvious that our method with D-HRNet can achieve state-of-the-art performance. Besides, our Mono-ViFI brings notable improvements to the three depth architectures without increasing model complexity (see each subsection separated by dotted lines). Despite that the difference between Mono-ViFI (D-HRNet) and RA-Depth [25] under single-frame mode is slight in Table 1, it is more significant in Table 2 when using improved ground truth. For multi-frame inference, Mono-ViFI with ResNet18 has already shown advantages over ManyDepth [58] and DynamicDepth [14] on some metrics including Sq Rel, RMSE, RMSE log, \\(\\delta_{2}\\) and \\(\\delta_{3}\\). When changing to D-HRNet backbone, proposed Mono-ViFI can totally outperform them. Although our models have more multiply-add computations (MACs) when predicting for a single frame, the average MACs per frame are comparable when processing a long video sequence. Fig. 3 depicts two visual comparison results on KITTI, where" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.127 + ], + "angle": 0, + "content": "J. Liu et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.218, + 0.174, + 0.784, + 0.216 + ], + "angle": 0, + "content": "Table 2: Performance comparison on Table 3: Performance comparison on KITTI [16] using improved ground truth Cityscapes [10] with a resolution of \\(416 \\times\\) from [52], with a resolution of \\(640 \\times 192\\). 128 unless stated." + }, + { + "type": "table", + "bbox": [ + 0.226, + 0.228, + 0.498, + 0.359 + ], + "angle": 0, + "content": "
MethodAbs RelSq RelRMSERMSE logδ1δ2δ3
Single-frame Inference
PackNet-SIM [21]0.0780.4203.4850.1210.9310.9860.996
R-MSF M6 [67]0.0880.4923.8370.1350.9150.9830.995
DIFFNet [65]0.0760.4143.4050.1190.9360.9880.996
Mono-ViF [64]0.0740.3883.4140.1150.9380.9890.997
Monodepth2 (ResNet18) [18]0.0900.5453.9420.1370.9140.9830.995
Mono-ViF (ResNet18)0.0840.4313.6180.1260.9240.9860.997
Lite-Mono [62]0.0820.4553.6850.1270.9230.9850.996
Mono-ViF (Lite-Mono)0.0800.4003.4970.1210.9300.9870.997
D-HRNet [25]0.0770.4233.3960.1190.9350.9870.996
RA-Depth (D-HRNet) [25]0.0740.3633.3490.1140.9400.9900.997
Mono-ViF (D-HRNet)0.0720.3423.2540.1110.9410.9910.998
Multi-frame Inference
ManyDepth (ResNet18) [58]0.0700.3993.4550.1130.9410.9890.997
MOVEDepth [56]0.0650.3773.4490.1120.9420.9880.996
Mono-ViF (ResNet18)0.0740.3743.3810.1130.9410.9900.997
Mono-ViF (Lite-Mono)0.0740.3543.3720.1140.9390.9890.997
Mono-ViF (D-HRNet)0.0660.2943.0990.1020.9520.9920.998
" + }, + { + "type": "table", + "bbox": [ + 0.507, + 0.228, + 0.781, + 0.359 + ], + "angle": 0, + "content": "
MethodAbs RelSq RelRMSERMSE logδ1δ2δ3
Single-frame Inference
Videos in the wild [19]0.1271.3306.9600.1950.8300.9470.981
Li et al. [33]0.1191.2906.9800.1900.8460.9520.982
InstDM [31] (832 × 256)0.1111.1586.4370.1820.8680.9610.983
SD-SSMDE [44]0.1100.9885.9530.165---
Monodepth2 (ResNet18) [18]0.1291.5696.3760.1970.8490.9570.983
Mono-ViFi (ResNet18)0.1121.0035.8990.1650.8720.9660.990
Lite-Mono [62]0.1141.1946.3280.1720.8710.9670.989
Mono-ViFi (Lite-Mono)0.1070.9405.7940.1590.8760.9680.991
D-HRNet [25]0.1111.1426.2050.1690.8790.9680.989
Mono-ViFi (D-HRNet)0.1050.8805.5500.1560.8810.9690.992
Multi-frame Inference
ManyDepth (ResNet18) [58]0.1141.1936.2230.1700.8750.9670.989
DynamicDepth [14]0.1041.0115.9870.1590.8890.9720.991
TC-Depth [49]0.1100.9585.820-0.867-0.991
Mono-ViFi (ResNet18)0.1070.9345.6880.1590.8800.9690.990
Mono-ViFi (Lite-Mono)0.1050.9045.6460.1570.8800.9700.991
Mono-ViFi (D-HRNet)0.1020.8365.3950.1530.8850.9710.992
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.387, + 0.787, + 0.43 + ], + "angle": 0, + "content": "Table 4: Evaluation results on dynamic objects and static regions of Cityscapes [10] dataset under \\(416\\times 128\\) resolution. Note that DynamicDepth relies on depth prior and segmentation masks of dynamic objects when reasoning." + }, + { + "type": "table", + "bbox": [ + 0.263, + 0.441, + 0.744, + 0.524 + ], + "angle": 0, + "content": "
MethodTest AreaAbs RelSq RelRMSERMSE logδ1δ2δ3
Multi-frame Inference
ManyDepth (ResNet18) [58]DynamicDepth [14]Mono-ViFI (ResNet18)Dynamic Objects0.1722.2065.9240.2160.7920.9170.967
0.1441.5545.2380.1850.8310.9500.982
0.1621.4535.1140.1990.7940.9240.977
ManyDepth (ResNet18) [58]DynamicDepth [14]Mono-ViFI (ResNet18)Static Regions0.1071.0906.3140.1600.8820.9720.991
0.1000.9776.1550.1550.8920.9740.992
0.0990.9075.8350.1490.8900.9750.992
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.567, + 0.785, + 0.643 + ], + "angle": 0, + "content": "we choose D-HRNet as the baseline depth model. It can be seen that depth maps look qualitatively similar among all four methods. However, the error maps reveal mistakes of D-HRNet baseline and RA-Depth in some ambiguous regions like bushes (see the red boxes). In contrast, our Mono-ViFi have dramatically lower error in these regions, especially under multi-frame reasoning." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Cityscapes Results. Results of Cityscapes in Table 3 demonstrate the same trend as KITTI. Moreover, proposed multi-frame model with ResNet18 can beat ManyDepth on all metrics, while our Mono-ViFI is better than DynamicDepth only on partial indexes. It is because Cityscapes with more dynamic scenes is more challenging than KITTI, and DynamicDepth leverages additional information to cope with dynamic objects. For fine-grained analysis, we separately evaluate on the dynamic objects and static regions, using the segmentation masks of dynamic objects provided in [14], as listed in Table 4. On dynamic objects, Mono-ViFI with ResNet18 surpasses ManyDepth by a large margin, demonstrating that our multi-frame inference paradigm can sidestep the issue of dynamic objects to a greater extent compared to ManyDepth. DynamicDepth utilizes depth prior and segmentation masks of dynamic objects to tackle this problem for multi-frame inference. Even though, our method still excels it on Sq Rel and" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.257, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "Mono-ViFi: A Unified Framework for Self-supervised Monocular Depth" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.256, + 0.147, + 0.75, + 0.309 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.32, + 0.785, + 0.348 + ], + "angle": 0, + "content": "Fig. 4: Qualitative generalization results of a YouTube video example. All depth maps are predicted under \\(640 \\times 360\\) resolution using KITTI models trained under \\(640 \\times 192\\)." + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.363, + 0.785, + 0.418 + ], + "angle": 0, + "content": "Table 5: Ablation results of each proposed component on Cityscapes [10]. SA is spatial augmentation. TA is temporal augmentation. SVDC denotes standard view depth consistency. SADC denotes scale-aware depth consistency. FA is feature alignment. MAFA is motion-aware feature alignment. OAFF means occlusion-alleviated feature fusion." + }, + { + "type": "table", + "bbox": [ + 0.264, + 0.432, + 0.744, + 0.595 + ], + "angle": 0, + "content": "
MethodSATASVDCSADCFAMAFAOAFFAbs RelSq RelRMSEδ1
Single-frame Inference
Baseline (Monodepth2 [18])0.1291.5696.8760.849
Baseline + TA✓0.1191.2546.5230.857
Baseline + SVDC✓0.1171.1866.4920.862
Baseline + SA✓0.1221.3866.5570.856
Baseline + SA + SADC✓✓0.1161.0796.1890.868
Mono-ViFI (Single-frame)✓✓✓✓0.1121.0035.8990.872
Multi-frame Inference
Baseline (direct feature fusion)0.1201.3606.4910.869
FA (not motion-aware)✓0.1161.2176.3870.871
MAFA✓0.1141.1506.1150.874
MAFA + OAFF✓✓0.1111.1076.0120.876
MAFA + OAFF + TA✓✓✓0.1091.0235.8920.877
Mono-ViFI (Multi-frame)✓✓✓✓✓✓0.1070.9345.6880.880
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.638, + 0.785, + 0.667 + ], + "angle": 0, + "content": "RMSE. When evaluating on static regions, proposed Mono-ViFI shows absolute superiority over ManyDepth and DynamicDepth." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.669, + 0.787, + 0.759 + ], + "angle": 0, + "content": "Generalization Results. We evaluate the KITTI models on Cityscapes and Make3D without fine-tuning, results of which are put in the supplementary. For qualitative comparison, we give a visualization example in Fig. 4, where KITTI models are used to predict the multi-frame depth of a YouTube video frame downloaded from the Wind Walk Travel Videos channel. This indicates that our Mono-ViFI has better generalization ability than existing methods." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.783, + 0.389, + 0.799 + ], + "angle": 0, + "content": "5.3 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.785, + 0.84 + ], + "angle": 0, + "content": "In this part, we explore the influence of each component in Mono-ViFi. Our ablations are conducted on Cityscapes or KITTI with ResNet18 backbone." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.127 + ], + "angle": 0, + "content": "J. Liu et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.27, + 0.145, + 0.732, + 0.16 + ], + "angle": 0, + "content": "Table 6: Ablation of different flow-based VFI models on KITTI [16]." + }, + { + "type": "table", + "bbox": [ + 0.306, + 0.172, + 0.703, + 0.239 + ], + "angle": 0, + "content": "
VFI ModelAbs RelSq RelRMSERMSE logδ1δ2δ3
Multi-frame Inference
IFRNet [29] (Ours)0.0990.6614.3210.1740.8980.9660.985
RIFE [26]0.1000.6724.3520.1750.8960.9650.985
AMT [34]0.0990.6534.3350.1740.8970.9660.985
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.277, + 0.787, + 0.428 + ], + "angle": 0, + "content": "Single-frame Ablation. We first ablate on the single-frame model. As listed in the top half part of Table 5, the standard view depth consistency and spatial augmentation can bring gains individually. Note that scale-aware depth consistency \\( L_{sa} \\) should be based on spatial augmentation. When adding \\( L_{sa} \\) to spatial augmentation, the accuracy can be further improved. And this leads to the most significant enhancement among all components, since \\( L_{sa} \\) can enforce the model to learn scale relationship from scene variations. In orthogonal dimension, our proposed temporal augmentation can also reduce the prediction error by providing more guidance from newly synthesized camera views. When integrating all above elements, our single-frame model can achieve the best performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.428, + 0.787, + 0.578 + ], + "angle": 0, + "content": "Multi-frame Ablation. To probe the effect of each component in our VFI-assisted multi-frame fusion module, we conduct ablation experiments under multi-frame inference, whose results are summarized in the bottom half part of Table 5. The baseline is stacking the raw feature of three frames directly, i.e., direct feature fusion. Then, we gradually apply components of the proposed fusion module, including motion-aware feature alignment and occlusion-alleviated feature fusion. The results demonstrate that these strategies can bring cumulative improvements. Furthermore, when incorporating spatial augmentation, temporal augmentation and relevant depth consistency losses, our multi-frame model can be further promoted like the single-frame setting." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.579, + 0.787, + 0.654 + ], + "angle": 0, + "content": "Ablation of different VFI models. Apart from IFRNet [29], we additionally experiment with another two flow-based VFI models, RIFE [26] and AMT [34]. The results are depicted in Table 6. It can be observed that there is only a negligible performance difference among the three VFI networks, which reveals that our method is robust to different VFI models." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.675, + 0.36, + 0.691 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.787, + 0.84 + ], + "angle": 0, + "content": "In conclusion, this paper presents a novel self-supervised learning framework, named Mono-ViFI, to bilaterally connect single- and multi-frame depth estimation in a unified manner. First, a flow-based VFI model is employed for temporal augmentation. Then, we propose a VFI-assisted multi-frame fusion module for more accurate multi-frame depth. Besides, resizing-cropping is extended to affine transformation as spatial augmentation for enriching data diversity. Finally, we develop a new triplet depth consistency loss for better regularization. Experiments show that our approach can unleash the potential of existing depth models to a greater extent, but without increasing inference complexity." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.257, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Mono-ViFi: A Unified Framework for Self-supervised Monocular Depth" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.177, + 0.785, + 0.204 + ], + "angle": 0, + "content": "1. Bae, J., Moon, S., Im, S.: Deep digging into the generalization of self-supervised monocular depth estimation. In: AAAI. pp. 187-196 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.205, + 0.785, + 0.245 + ], + "angle": 0, + "content": "2. Bangunharcana, A., Magd, A., Kim, K.S.: Dualrefine: Self-supervised depth and pose estimation through iterative epipolar sampling and refinement toward equilibrium. In: CVPR. pp. 726-738 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.247, + 0.785, + 0.273 + ], + "angle": 0, + "content": "3. Bello, J.L.G., Kim, M.: Forget about the lidar: Self-supervised depth estimators with med probability volumes. In: NeurIPS. pp. 12626-12637 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.274, + 0.785, + 0.3 + ], + "angle": 0, + "content": "4. Bello, J.L.G., Moon, J., Kim, M.: Positional information is all you need: A novel pipeline for self-supervised svde from videos. arXiv:2205.08851 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.302, + 0.785, + 0.342 + ], + "angle": 0, + "content": "5. Bian, J., Li, Z., Wang, N., Zhan, H., Shen, C., Cheng, M.M., Reid, I.: Unsupervised scale-consistent depth and ego-motion learning from monocular video. In: NeurIPS. pp. 35-45 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.343, + 0.785, + 0.383 + ], + "angle": 0, + "content": "6. Chan, K.C., Wang, X., Yu, K., Dong, C., Loy, C.C.: Basicvsr: The search for essential components in video super-resolution and beyond. In: CVPR. pp. 4945-4954 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.385, + 0.785, + 0.425 + ], + "angle": 0, + "content": "7. Chan, K.C., Zhou, S., Xu, X., Loy, C.C.: Basicvsr++: Improving video superresolution with enhanced propagation and alignment. In: CVPR. pp. 5962-5971 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.426, + 0.785, + 0.453 + ], + "angle": 0, + "content": "8. Chen, Y., Liu, S., Wang, X.: Learning continuous image representation with local implicit image function. In: CVPR. pp. 8624-8634 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.455, + 0.785, + 0.494 + ], + "angle": 0, + "content": "9. Chen, Y., Schmid, C., Sminchisescu, C.: Self-supervised learning with geometric constraints in monocular video: Connecting flow, depth, and camera. In: ICCV. pp. 7063-7072 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.496, + 0.785, + 0.536 + ], + "angle": 0, + "content": "10. Cordts, M., Omran, M., Ramos, S., Rehfeld, T., Enzweiler, M., Benenson, R., Franke, U., Roth, S., Schiele, B.: The cityscapes dataset for semantic urban scene understanding. In: CVPR. pp. 3213-3223 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.537, + 0.785, + 0.564 + ], + "angle": 0, + "content": "1. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: CVPR. pp. 248-255 (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.565, + 0.785, + 0.592 + ], + "angle": 0, + "content": "2. Eigen, D., Fergus, R.: Predicting depth, surface normals and semantic labels with a common multi-scale convolutional architecture. In: ICCV. pp. 2650-2658 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.593, + 0.785, + 0.619 + ], + "angle": 0, + "content": "3. Eigen, D., Puhrsch, C., Fergus, R.: Depth map prediction from a single image using a multi-scale deep network. In: NeurIPS. pp. 2366-2374 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.62, + 0.785, + 0.66 + ], + "angle": 0, + "content": "4. Feng, Z., Yang, L., Jing, L., Wang, H., Tian, Y., Li, B.: Disentangling object motion and occlusion for unsupervised multi-frame monocular depth. In: ECCV. pp. 228-244 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.661, + 0.785, + 0.688 + ], + "angle": 0, + "content": "5. Garg, R., Kumar, B.G.V., Carneiro, G., Reid, I.: Unsupervised cnn for single view depth estimation: Geometry to the rescue. In: ECCV. pp. 740-756 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.689, + 0.785, + 0.716 + ], + "angle": 0, + "content": "6. Geiger, A., Lenz, P., Stiller, C., Urtasun, R.: Vision meets robotics: The kitti dataset. *IJRR* **32**(11), 1231–1237 (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.717, + 0.785, + 0.744 + ], + "angle": 0, + "content": "7. Godard, C., Aodha, O.M., Brostow, G.J.: Unsupervised monocular depth estimation with left-right consistency. In: CVPR. pp. 6602-6611 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.745, + 0.785, + 0.771 + ], + "angle": 0, + "content": "8. Godard, C., Aodha, O.M., Firman, M., Brostow, G.J.: Digging into self-supervised monocular depth estimation. In: CVPR. pp. 3828-3838 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.772, + 0.785, + 0.812 + ], + "angle": 0, + "content": "9. Gordon, A., Li, H., Jonschkowski, R., Angelova, A.: Depth from videos in the wild: Unsupervised monocular depth learning from unknown cameras. In: ICCV. pp. 8976-8985 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "20. Guizilini, V., Ambrus, R., Chen, D., Zakharov, S., Gaidon, A.: Multi-frame self-supervised depth with transformers. In: CVPR. pp. 160-170 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.177, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.127 + ], + "angle": 0, + "content": "J. Liu et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "21. Guizilini, V., Ambrus, R., Pillai, S., Gaidon, A.: 3d packing for self-supervised monocular depth estimation. In: CVPR. pp. 2482-2491 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.177, + 0.785, + 0.203 + ], + "angle": 0, + "content": "22. Han, W., Yin, J., Jin, X., Dai, X., Shen, J.: Brnet: Exploring comprehensive features for monocular depth estimation. In: ECCV. pp. 586-602 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.204, + 0.785, + 0.231 + ], + "angle": 0, + "content": "23. Han, W., Yin, J., Shen, J.: Self-supervised monocular depth estimation by direction-aware cumulative convolution network. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.232, + 0.785, + 0.259 + ], + "angle": 0, + "content": "24. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR. pp. 770-778 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.26, + 0.785, + 0.286 + ], + "angle": 0, + "content": "25. He, M., Hui, L., Bian, Y., Ren, J., Xie, J., Yang, J.: Ra-depth: Resolution adaptive self-supervised monocular depth estimation. In: ECCV. pp. 565-581 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.785, + 0.314 + ], + "angle": 0, + "content": "26. Huang, Z., Zhang, T., Heng, W., Shi, B., Zhou, S.: Real-time intermediate flow estimation for video frame interpolation. In: ECCV. pp. 624-642 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.785, + 0.354 + ], + "angle": 0, + "content": "27. Jung, H., Park, E., Yoo, S.: Fine-grained semantics-aware representation enhancement for self-supervised monocular depth estimation. In: ICCV. pp. 12622-12632 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.355, + 0.785, + 0.397 + ], + "angle": 0, + "content": "28. Klingner, M., Termöhlen, J.A., Mikolajczyk, J., Fingscheidt, T.: Self-supervised monocular depth estimation: Solving the dynamic object problem by semantic guidance. In: ECCV. pp. 582-600 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.785, + 0.438 + ], + "angle": 0, + "content": "29. Kong, L., Jiang, B., Luo, D., Chu, W., Huang, X., Tai, Y., Wang, C., Yang, J.: Ifrnet: Intermediate feature refine network for efficient frame interpolation. In: CVPR. pp. 1959-1968 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.785, + 0.466 + ], + "angle": 0, + "content": "30. Kuznietsov, Y., Proesmans, M., Gool, L.V.: Comoda: Continuous monocular depth adaptation using past experiences. In: WACV. pp. 2906-2916 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.467, + 0.785, + 0.494 + ], + "angle": 0, + "content": "31. Lee, S., Im, S., Lin, S., Kweon, I.S.: Learning monocular depth in dynamic scenes via instance-aware projection consistency. In: AAAI. pp. 1863-1872 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.495, + 0.785, + 0.522 + ], + "angle": 0, + "content": "32. Lee, S., Rameau, F., Pan, F., Kweon, I.S.: Attentive and contrastive learning for joint depth and motion field estimation. In: ICCV. pp. 4862-4871 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.523, + 0.785, + 0.549 + ], + "angle": 0, + "content": "33. Li, H., Gordon, A., Zhao, H., Casser, V., Angelova, A.: Unsupervised monocular depth learning in dynamic scenes. In: CoRL. pp. 1908-1917 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.55, + 0.785, + 0.591 + ], + "angle": 0, + "content": "34. Li, Z., Zhu, Z.L., Han, L.H., Hou, Q., Guo, C.L., Cheng, M.M.: Amt: All-pairs multi-field transforms for efficient frame interpolation. In: CVPR. pp. 9801-9810 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.592, + 0.785, + 0.619 + ], + "angle": 0, + "content": "35. Liu, J., Kong, L., Yang, J.: Designing and searching for lightweight monocular depth network. In: ICONIP. pp. 477-488 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.62, + 0.785, + 0.646 + ], + "angle": 0, + "content": "36. Liu, J., Kong, L., Yang, J.: Atca: An arc trajectory based model with curvature attention for video frame interpolation. In: ICIP. pp. 1486-1490 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.647, + 0.785, + 0.688 + ], + "angle": 0, + "content": "37. Liu, J., Kong, L., Yang, J., Liu, W.: Towards better data exploitation in self-supervised monocular depth estimation. IEEE Robotics and Automation Letters 9(1), 763-770 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.689, + 0.785, + 0.716 + ], + "angle": 0, + "content": "38. Luo, X., Huang, J.B., Szeliski, R., Matzen, K., Kopf, J.: Consistent video depth estimation. TOG 39(71), 1-13 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.717, + 0.785, + 0.757 + ], + "angle": 0, + "content": "39. Lyu, X., Liu, L., Wang, M., Kong, X., Liu, L., Liu, Y., Chen, X., Yuan, Y.: Hrdepth: High resolution self-supervised monocular depth estimation. In: AAAI. pp. 2294-2301 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.785, + 0.799 + ], + "angle": 0, + "content": "40. Ma, J., Lei, X., Liu, N., Zhao, X., Pu, S.: Towards comprehensive representation enhancement in semantics-guided self-supervised monocular depth estimation. In: ECCV. pp. 304-321 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.8, + 0.785, + 0.841 + ], + "angle": 0, + "content": "41. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: ECCV. pp. 405-421 (2020)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.257, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Mono-ViFi: A Unified Framework for Self-supervised Monocular Depth" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.786, + 0.189 + ], + "angle": 0, + "content": "42. Patil, V., Gansbeke, W.V., Dai, D., Gool, L.V.: Don't forget the past: Recurrent depth estimation from monocular video. IEEE Robotics and Automation Letters 5(4), 6813-6820 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.787, + 0.218 + ], + "angle": 0, + "content": "43. Peng, R., Wang, R., Lai, Y., Tang, L., Cai, Y.: Excavating the potential capacity of selfsupervised monocular depth estimation. In: ICCV. pp. 15560-15569 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.218, + 0.787, + 0.259 + ], + "angle": 0, + "content": "44. Petrovai, A., Nedevschi, S.: Exploiting pseudo labels in a self-supervised learning framework for improved monocular depth estimation. In: CVPR. pp. 1568-1578 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.259, + 0.787, + 0.3 + ], + "angle": 0, + "content": "45. Pilzer, A., Lathuilière, S., Sebe, N., Ricci, E.: Refine and distill: Exploiting cycle-inconsistency and knowledge distillation for unsupervised monocular depth estimation. In: CVPR. pp. 9760-9769 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.3, + 0.787, + 0.328 + ], + "angle": 0, + "content": "46. Poggi, M., Aleotti, F., Tosi, F., Mattoccia, S.: On the uncertainty of self-supervised monocular depth estimation. In: CVPR. pp. 3224-3234 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.787, + 0.37 + ], + "angle": 0, + "content": "47. Ranjan, A., Jampani, V., Balles, L., Kim, K., Sun, D., Wulff, J., Black, M.J.: Competitive collaboration: Joint unsupervised learning of depth, camera motion, optical flow and motion segmentation. In: CVPR. pp. 12240-12249 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.371, + 0.787, + 0.398 + ], + "angle": 0, + "content": "48. Ren, W., Wang, L., Piao, Y., Zhang, M., Lu, H., Liu, T.: Adaptive co-teaching for unsupervised monocular depth estimation. In: ECCV. pp. 89-105 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.787, + 0.439 + ], + "angle": 0, + "content": "49. Ruhkamp, P., Gao, D., Chen, H., Navab, N., Busam, B.: Attention meets geometry: Geometry guided spatial-temporal attention for consistent self-supervised monococular depth estimation. In: 3DV. pp. 837-847 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.787, + 0.467 + ], + "angle": 0, + "content": "50. Saxena, A., Sun, M., Ng, A.Y.: Make3d: Learning 3d scene structure from a single still image. PAMI 31(5), 824-840 (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.467, + 0.787, + 0.522 + ], + "angle": 0, + "content": "51. Tancik, M., Srinivasan, P.P., Mildenhall, B., Fridovich-Keil, S., Raghavan, N., Singhal, U., Ramamoorthi, R., Barron, J.T., Ng, R.: Fourier features let networks learn high frequency functions in low dimensional domains. In: NeurIPS. pp. 7537-7547 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.522, + 0.787, + 0.55 + ], + "angle": 0, + "content": "52. Uhrig, J., Schneider, N., Schneider, L., Franke, U., Brox, T., Geiger, A.: Sparsity invariant cnns. In: 3DV. pp. 11-20 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.55, + 0.787, + 0.577 + ], + "angle": 0, + "content": "53. Wang, C., Buenaposada, J.M., Zhu, R., Lucey, S.: Learning depth from monocular videos using direct methods. In: CVPR. pp. 2022-2030 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.577, + 0.787, + 0.618 + ], + "angle": 0, + "content": "54. Wang, R., Pizer, S.M., Frahm, J.M.: Recurrent neural network for (un-)supervised learning of monocular video visual odometry and depth. In: CVPR. pp. 5555-5564 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.619, + 0.787, + 0.647 + ], + "angle": 0, + "content": "55. Wang, R., Yu, Z., Gao, S.: Planedepth: Self-supervised depth estimation via orthogonal planes. In: CVPR. pp. 21425-21434 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.647, + 0.787, + 0.688 + ], + "angle": 0, + "content": "56. Wang, X., Zhu, Z., Huang, G., Chi, X., Ye, Y., Chen, Z., Wang, X.: Drafting monocular cues and velocity guidance for self-supervised multi-frame depth learning. In: AAAI. pp. 2689-2697 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.688, + 0.787, + 0.716 + ], + "angle": 0, + "content": "57. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: From error visibility to structural similarity. IEEE TIP 13(4), 600-612 (2004)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.716, + 0.787, + 0.758 + ], + "angle": 0, + "content": "58. Watson, J., Aodha, O.M., Prisacariu, V.A., Brostow, G.J., Firman, M.: The temporal opportunist: Self-supervised multi-frame monocular depth. In: CVPR. pp. 1164-1174 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.787, + 0.786 + ], + "angle": 0, + "content": "59. Watson, J., Firman, M., Brostow, G.J., Turmukhambetov, D.: Self-supervised monocular depth hints. In: ICCV. pp. 2162-2171 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.786, + 0.787, + 0.813 + ], + "angle": 0, + "content": "60. Yin, Z., Shi, J.: Geonet: Unsupervised learning of dense depth, optical flow and camera pose. In: CVPR. pp. 1983-1992 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.787, + 0.841 + ], + "angle": 0, + "content": "61. Zhang, H., Li, Y., Cao, Y., Liu, Y., Shen, C., Yan, Y.: Exploiting temporal consistency for real-time video depth estimation. In: ICCV. pp. 1725-1734 (2019)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.127 + ], + "angle": 0, + "content": "J. Liu et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "62. Zhang, N., Nex, F., Vosselman, G., Kerle, N.: Lite-mono: A lightweight cnn and transformer architecture for self-supervised monocular depth estimation. In: CVPR. pp. 18537-18546 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.785, + 0.232 + ], + "angle": 0, + "content": "63. Zhao, C., Poggi, M., Tosi, F., Zhou, L., Sun, Q., Tanga, Y., Mattoccia, S.: Gasmono: Geometry-aided self-supervised monocular depth estimation for indoor scenes. In: ICCV. pp. 16163-16174 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.785, + 0.273 + ], + "angle": 0, + "content": "64. Zhao, C., Zhang, Y., Poggi, M., Tosi, F., Guo, X., Zhu, Z., Huang, G., Tang, Y., Mattoccia, S.: Monovit: Self-supervised monocular depth estimation with a vision transformer. In: 3DV. pp. 668-678 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.785, + 0.3 + ], + "angle": 0, + "content": "65. Zhou, H., Greenwood, D., Taylor, S.: Self-supervised monocular depth estimation with internal feature fusion. In: BMVC (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.3, + 0.785, + 0.328 + ], + "angle": 0, + "content": "66. Zhou, T., Brown, M., Snavely, N., Lowe, D.G.: Unsupervised learning of depth and ego-motion from video. In: CVPR. pp. 6612-6619 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.785, + 0.356 + ], + "angle": 0, + "content": "67. Zhou, Z., Fan, X., Shi, P., Xin, Y.: R-msfm: Recurrent multi-scale feature modulation for monocular depth estimating. In: ICCV. pp. 12757-12766 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.356 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Mono-ViFI_ A Unified Learning Framework for Self-supervised Single- and Multi-frame Monocular Depth Estimation/47dcf543-78f7-47b3-955c-778904eede62_origin.pdf b/2024/Mono-ViFI_ A Unified Learning Framework for Self-supervised Single- and Multi-frame Monocular Depth Estimation/47dcf543-78f7-47b3-955c-778904eede62_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..d4d3f7ee3d0712a62c000199054c527a78ecb12c --- /dev/null +++ b/2024/Mono-ViFI_ A Unified Learning Framework for Self-supervised Single- and Multi-frame Monocular Depth Estimation/47dcf543-78f7-47b3-955c-778904eede62_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9bf5b6ab3ed2b8499ae8b7d4bda67e4fe164537d53d2966e2fb4d597ce9d7107 +size 1590008 diff --git a/2024/Mono-ViFI_ A Unified Learning Framework for Self-supervised Single- and Multi-frame Monocular Depth Estimation/full.md b/2024/Mono-ViFI_ A Unified Learning Framework for Self-supervised Single- and Multi-frame Monocular Depth Estimation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..983d3f13ad6601e4a36afb6e03586f597b49219b --- /dev/null +++ b/2024/Mono-ViFI_ A Unified Learning Framework for Self-supervised Single- and Multi-frame Monocular Depth Estimation/full.md @@ -0,0 +1,369 @@ +# Mono-ViFi: A Unified Learning Framework for Self-supervised Single- and Multi-frame Monocular Depth Estimation + +Jinfeng Liu, Lingtong Kong\*, Bo Li, Zerong Wang, Hong Gu, and Jinwei Chen + +vivo Mobile Communication Co., Ltd, China + +{liujinfeng, ltkong, libra, wangzerong, guhong, jinwei.chen}@vivo.com + +Abstract. Self-supervised monocular depth estimation has gathered notable interest since it can liberate training from dependency on depth annotations. In monocular video training case, recent methods only conduct view synthesis between existing camera views, leading to insufficient guidance. To tackle this, we try to synthesize more virtual camera views by flow-based video frame interpolation (VFI), termed as temporal augmentation. For multi-frame inference, to sidestep the problem of dynamic objects encountered by explicit geometry-based methods like ManyDepth, we return to the feature fusion paradigm and design a VFI-assisted multi-frame fusion module to align and aggregate multi-frame features, using motion and occlusion information obtained by the flow-based VFI model. Finally, we construct a unified self-supervised learning framework, named Mono-ViFi, to bilaterally connect single- and multiframe depth. In this framework, spatial data augmentation through image affine transformation is incorporated for data diversity, along with a triplet depth consistency loss for regularization. The single- and multiframe models can share weights, making our framework compact and memory-efficient. Extensive experiments demonstrate that our method can bring significant improvements to current advanced architectures. Source code is available at https://github.com/LiuJF1226/Mono-ViFi. + +Keywords: Monocular depth estimation $\cdot$ Self-supervised learning $\cdot$ Data augmentation $\cdot$ Video frame interpolation + +# 1 Introduction + +Monocular depth estimation (MDE) is a fundamental 3D vision task, which finds applications in robotics, autonomous driving and VR/AR. To reduce the dependence on depth annotations, self-supervised MDE [1,2,15,17,46,63,66] has risen as a convenient and practical solution, which uses monocular videos, stereo pairs, or even both to generate supervision through view synthesis, where an image photometric reconstruction loss is minimized. Compared to stereo training case [15,17,59] with known camera pose, learning from monocular videos [18,58, 66] is more challenging since it requires an extra pose network to estimate camera ego-motion. Even so, it is much easier to collect monocular videos compared to + +stereo image pairs, making monocular training more attractive. Therefore, this paper also focuses on self-supervised training with unlabeled monocular videos. + +Recent works have made various attempts to improve accuracy, like introducing masking techniques [18, 66] and leveraging segmentation priors [28, 31]. However, they only conduct view synthesis between existing fixed camera views within video datasets, leading to insufficient guidance especially in invisible regions by occlusion or disappearance. Considering that video data possesses temporal continuity, an intuitive idea is to generate novel frames, i.e., virtual camera views in the temporal dimension. We term it as temporal augmentation, which can be achieved by video frame interpolation (VFI), as shown in Fig. 1(a). + +From another perspective, multi-frame reasoning paradigm have recently been researched because multiple frames can be available for reasoning in many application scenarios. ManyDepth [58] proposes to use cost volume constructed from multi-frame matching features. It explicitly reasons about geometry based on static scene assumption, which can be easily affected by dynamic objects and occlusion. DynamicDepth [14] attempts to tackle dynamic scenes, but relies on depth priors and segmentation masks of moving objects when reasoning. To sidestep the issue of dynamic objects in multi-frame inference encountered by these methods, we return to the traditional paradigm [42, 49], which aggregates multi-frame features directly. However, direct feature fusion without alignment harms the learning of temporal representations, since feature embeddings at the same location in neighbor frames are usually irrelevant, making it difficult for models to leverage the correlation of multiple frames. Hence, we regard feature alignment as a reasonable and necessary step before aggregation. As revealed in several video enhancement tasks like video super-resolution [6, 7] and frame interpolation [26, 29, 34, 36], inter-frame feature alignment can be implemented by warping through optical flow. This inspires us to employ a flow-based VFI model to concurrently conduct temporal augmentation and provide explicit correspondence for feature alignment. Moreover, the VFI model can reason out occlusion mask between neighbor frames, helping to alleviate inter-frame occlusion problem in feature fusion. Based on above analysis, we design a VFI-assisted multi-frame fusion module for multi-frame depth estimation, as depicted in Fig. 2. + +Finally, we build a unified framework to bilaterally connect single- and multiframe depth, as demonstrated in Fig. 1(b). The direct way is by depth consistency, usually for geometric regularization or distillation in self-supervised MDE. ManyDepth conducts unilateral distillation from single-frame model to multiframe one. DynamicDepth extends it to a bilateral mutual reinforcement, but only in the regions of moving objects. Differently, our framework applies the bilateral consistency on the whole image. For an original target image, we use a standard view depth consistency (SVDC) loss between its two depth maps predicted from the single- and multi-frame models. We also include classic data augmentation strategies, categorized as spatial augmentation. Specifically, we combine resizing-cropping with image rotation, forming a generic affine transformation. Then, we give the formulation of rectified camera pose for affine-augmented view synthesis using the cue of relative object size [4, 55]. For the + +![](images/9a3d1e5d9591e39c21ee0f4c6d3c34d41e74bc997a12e6f691af7d204581947b.jpg) +(a) Temporal Augmentation with VFI + +![](images/ae377703be20437e014938cb17aad97ac308c67f058d7004bde80aa93374d895.jpg) +(b) Learning Framework $(I_{t}$ for example +Fig. 1: Overview of Mono-ViFI. (a) We achieve temporal augmentation by a flow-based VFI model $\mathcal{H}$ , which also reasons out intermediate optical flow and occlusion mask for multi-frame inference. (b) For target image $I_{t}$ , we obtain its single-frame depth $D_{t}$ and multi-frame depth $D_{t}^{m}$ . Also, $I_{t}$ is augmented to $\widetilde{I}_{t}$ through affine transformation $\mathcal{A}$ . We then calculate the depth $\widehat{D}_t$ of $\widetilde{I}_t$ and inversely convert it to the original view, generating $\widehat{D}_t$ . Finally, we enforce a triplet depth consistency loss $L_{tc}(t)$ among the three depth maps, including a standard view depth consistency loss $L_{sv}(t)$ and two scale-aware depth consistency losses, $L_{sa}(t)$ and $L_{sa}^{m}(t)$ . Note that each depth map also corresponds to a photometric loss and a smoothness loss, which are omitted here. + +depth map from augmented target view, we enforce two scale-aware depth consistency (SADC) losses between it and above two standard view depths, respectively. SADC stems from the relative size cue, which assumes that when an image is zoomed up by a factor, the depth decreases by the same factor. It guides depth model to learn consistent geometric scale relationship from scene variations. The SVDC loss and two SADC losses are termed as a triplet depth consistency loss. + +Our framework is named Mono-ViFI, since the self-supervised Monocular depth is largely boosted by Video Frame Interpolation. To our best knowledge, we are the first to leverage VFI for the improvement of self-supervised MDE. The main contributions of this paper are summarized as follows: + +- We propose Mono-ViFi, a unified learning framework to bilaterally connect self-supervised single- and multi-frame MDE. In Mono-ViFi, the single- and multi-frame models can share weights, making the unified framework more compact and memory-efficient. +- We conduct temporal augmentation through flow-based VFI, and develop a novel VFI-assisted multi-frame fusion module, which utilizes motion and occlusion from flow-based VFI model to improve multi-frame depth. +- We incorporate image affine transformation as spatial augmentation, along with a triplet depth consistency loss for regularization and distillation. +- Experiments show that Mono-ViFI can achieve state-of-the-art performance. Moreover, our approach is orthogonal to architecture design, allowing current advanced depth networks to be further improved by seamlessly integrating them with Mono-ViFI. + +# 2 Related Work + +# 2.1 Self-supervised Monocular Depth Estimation + +Self-supervised learning paradigm for monocular depth estimation can be categorized into stereo training and monocular training according to the data source. + +Stereo training leverages the left-right consistency in stereo image pairs. Garg et al. [15] first consider the training as a view synthesis problem with a photometric consistency loss between stereo pairs. Building upon this idea, Godard et al. [17] introduce a left-right disparity consistency loss. Then, it is further extended by using depth hints [59], data augmentation [43, 55], knowledge distillation [43, 45, 55] and plane-based scene representation [3, 55]. + +Monocular training leverages the temporal consistency in monocular video frames. Zhou et al. [66] propose a pioneering method which jointly learns depth and camera ego-motion from monocular videos in a self-supervised way. Then, Bian et al. [5] propose a scale-consistent depth and ego-motion estimation approach by adding a depth consistency loss, which reduces scale drift of estimated depth but decreases the accuracy. Later, Godard et al. [18] propose Monodepth2, a strong baseline for most follow-up works, which employs a minimum reprojection loss to mitigate occlusion issues and an auto-masking loss to filter out moving objects with similar velocity as the camera. Furthermore, some works introduce multi-task learning to combine depth estimation with optical flow [9, 32, 47, 60], semantic segmentation [27, 28, 40] and instance segmentation [31]. Others [21-23, 35, 37, 39, 62, 64, 65] try to improve the network architectures for depth estimation. However, these methods only utilize the geometric constrains between existing fixed camera views in video datasets. This insufficient guidance may block current depth models from unleashing their full capabilities. To address it, we conduct temporal augmentation through VFI in this paper. + +# 2.2 Multi-frame Monocular Depth + +Above single-frame depth models do not leverage any temporal information for inference, which limits the performance to some degree. Regarding multi-frame depth prediction, early works adopt test-time refinement [9, 30, 38], use spatial-temporal attention [49], or combine traditional depth networks with recurrent layers [54, 61] to process sequence information. Another broad group of methods [14, 20, 56, 58] utilize cost volume from stereo matching, which explicitly encode and reason from geometric constraints by combining a multi-frame cost volume with single-frame features. On the one hand, this type of methods based on ManyDepth [58] usually meet with the problem of dynamic scenes. To sidestep this issue, we develop a VFI-assisted multi-frame fusion depth model to align and aggregate multi-frame features by optical flow and occlusion masks. On the other hand, ManyDepth uses single-frame depth prior to boost multi-frame depth estimation. However, this improvement is unidirectional. DynamicDepth [14] extends it to a mutual reinforcement, but only targets on dynamic objects. Differently, we propose a unified framework to bilaterally connect single- and multi-frame depth in a more compact and memory-efficient manner. + +# 3 Problem Setup + +Self-supervised MDE in the monocular training case leverages the geometric constraints from video sequences. Specifically, for a training triplet $\{I_{t-2}, I_t, I_{t+2}\}$ which contains three consecutive frames, we use $I_t$ as the target view and the other two as the source views $\{I_s\}_{s \in \{t-2, t+2\}}$ to implement view synthesis. Here the temporal notation $t \pm 2$ instead of $t \pm 1$ is for the convenience of following description. Also, a pose network is utilized to estimate the camera pose from the target view to the source view, denoted as $\pmb{T}_{t \to s} = [\pmb{R}_{t \to s} | \pmb{t}_{t \to s}]$ , where $\pmb{R}_{t \to s}$ and $\pmb{t}_{t \to s}$ represent the rotation and translation components. Formally, the target depth $D_t$ and the camera pose $\pmb{T}_{t \to s}$ can be derived from: + +$$ +D _ {t} = \mathcal {D} \left(\mathcal {E} \left(I _ {t}\right)\right), \quad \boldsymbol {T} _ {t \rightarrow s} = \mathcal {P} \left(I _ {t}, I _ {s}\right), \tag {1} +$$ + +where $\mathcal{E}$ and $\mathcal{D}$ represent the encoder and decoder, respectively, of a single-frame depth network, while $\mathcal{P}$ denotes the pose network. Then, the reconstructed view of $I_{t}$ from the source view $I_{s}$ can be obtained by: + +$$ +I _ {s \rightarrow t} = I _ {s} \left\langle \operatorname {p r o j} \left(D _ {t}, \boldsymbol {T} _ {t \rightarrow s}, \boldsymbol {K}\right)\right\rangle , \tag {2} +$$ + +where $I_{s\rightarrow t}$ is the reconstructed target image, $\langle \rangle$ means the differentiable bilinear sampling operator, $proj()$ returns the 2D coordinates of the pixels in $I_{t}$ after they are reprojected to the viewpoint of $I_{s}$ , and $K$ is the camera intrinsic matrix assumed to be known. Finally, the per-pixel photometric loss between $I_{s\rightarrow t}$ and $I_{t}$ is used to optimize the depth and pose networks in an end-to-end manner. We follow Monodepth2 [18] to adopt the combination of L1 distance and SSIM [57] as the photometric error function and take the minimum across all the source views at each pixel to handle occlusion, which is formulated as: + +$$ +L _ {p e} (t) = \min _ {s} \frac {\alpha}{2} \left(1 - \operatorname {S S I M} \left(I _ {t}, I _ {s \rightarrow t}\right)\right) + (1 - \alpha) \| I _ {t}, I _ {s \rightarrow t} \| _ {1}, \tag {3} +$$ + +where $\alpha = 0.85$ . Besides, the edge-aware smoothness loss is employed to cope with depth discontinuities: + +$$ +L _ {s m} (t) = \left| \partial_ {x} d _ {t} ^ {*} \right| e ^ {- \left| \partial_ {x} I _ {t} \right|} + \left| \partial_ {y} d _ {t} ^ {*} \right| e ^ {- \left| \partial_ {y} I _ {t} \right|}, \tag {4} +$$ + +where $d_t^* = d_t / \overline{d_t}$ is the mean-normalized inverse depth to discourage shrinking [53]. In summary, the standard self-supervised loss on $I_{t}$ is: + +$$ +L _ {s s} (t) = \mathcal {L} \left(D _ {t}, \boldsymbol {T} _ {t \rightarrow s}, I _ {t}, I _ {s}\right) = \mu L _ {p e} (t) + \gamma L _ {s m} (t), \tag {5} +$$ + +where $\mu$ is the auto-masking in [18] and $\gamma = 0.001$ . + +# 4 Methodology + +In this paper, we propose novel Mono-ViFI for self-supervised monocular training, which bilaterally connects single- and multi-frame depth models into a unified learning framework. With flow-based VFI model, we introduce temporal augmentation and develop VFI-assisted multi-frame fusion depth, which are presented in Sec. 4.1 and Sec. 4.2, respectively. Besides, our framework incorporates + +spatial data augmentation and a triplet depth consistency loss, which are separately described in Sec. 4.3 and Sec. 4.4. Finally, we give the overall training objective in Sec. 4.5. The overview of our Mono-ViFi is depicted in Fig. 1. + +# 4.1 Temporal Augmentation with VFI + +Firstly, we generate more new frames in the temporal dimension based on existing video frames for temporal augmentation. To achieve this, we adopt a pretrained flow-based VFI model, such as IFRNet [29], which takes two temporally adjacent frames $\{I_0,I_2\}$ as input to synthesize an intermediate frame $I_{1}$ . The frame interpolation process can be formulated as: + +$$ +F _ {1 \rightarrow 0}, F _ {1 \rightarrow 2}, M _ {1} = \mathcal {H} \left(I _ {0}, I _ {2}\right), \tag {6} +$$ + +$$ +I _ {0 \rightarrow 1} = \omega \left(I _ {0}, F _ {1 \rightarrow 0}\right), \quad I _ {2 \rightarrow 1} = \omega \left(I _ {2}, F _ {1 \rightarrow 2}\right), \tag {7} +$$ + +$$ +I _ {1} = M _ {1} \odot I _ {0 \rightarrow 1} + (1 - M _ {1}) \odot I _ {2 \rightarrow 1}, \tag {8} +$$ + +where $\mathcal{H}$ is the VFI model, $F_{1\rightarrow 0}$ and $F_{1\rightarrow 2}$ are the intermediate optical flow, $\omega$ represents backward warping. $I_{0\rightarrow 1}$ and $I_{2\rightarrow 1}$ denote the warped images to the intermediate position, while $M_{1}$ ranging from 0 to 1 is the merge mask adjusting the mixing ratio according to bidirectional occlusion information. Here, $\odot$ indicates the element-wise multiplication. Considering a triplet $\{I_{t - 2},I_t,I_{t + 2}\}$ with three consecutive frames as described in Sec. 3, we use the VFI model to synthesize $I_{t - 1}$ from $\{I_{t - 2},I_t\}$ and $I_{t + 1}$ from $\{I_t,I_{t + 2}\}$ , as shown in Fig. 1(a). Then, we regard $I_{t - 1}$ and $I_{t + 1}$ as another two target frames. All three target frames have common sources of $\{I_{t - 2},I_{t + 2}\}$ . The following operations and losses are the same for each of them, and we only take $I_{t}$ for instance. + +# 4.2 VFI-assisted Multi-frame Fusion Depth + +Apart from temporal augmentation, our employed VFI model can also provide other inter-frame information, such as optical flow and occlusion merge masks. This enables us to design a novel VFI-assisted multi-frame fusion module for multi-frame depth prediction, as demonstrated in Fig. 2. The fusion module consists of two parts, i.e., motion-aware feature alignment (MAFA) and occlusion-alarviated feature fusion (OAFF). Taking original triplet $\{I_{t - 2},I_t,I_{t + 2}\}$ for example, we implement forward process of the VFI model to interpolate the middle position $t$ of $\{I_{t - 2},I_{t + 2}\}$ in Eq. (9), yielding intermediate flow $F_{t\rightarrow (t - 2)},F_{t\rightarrow (t + 2)}$ and merge mask $M_{t}$ to align and aggregate multi-frame features. + +$$ +F _ {t \rightarrow (t - 2)}, F _ {t \rightarrow (t + 2)}, M _ {t} = \mathcal {H} \left(I _ {t - 2}, I _ {t + 2}\right). \tag {9} +$$ + +Multi-frame Feature Extraction. Our fusion module accepts multi-frame features extracted by the same encoder, which is shared with the single-frame depth encoder $\mathcal{E}$ . The extraction process can be written as: + +$$ +\left\{\phi_ {t - 2} ^ {k}, \phi_ {t} ^ {k}, \phi_ {t + 2} ^ {k} \right\} _ {k = 1} ^ {N} = \left\{\mathcal {E} \left(I _ {t - 2}\right), \mathcal {E} \left(I _ {t}\right), \mathcal {E} \left(I _ {t + 2}\right) \right\}, \tag {10} +$$ + +where $\{\phi_{t - 2}^k,\phi_t^k,\phi_{t + 2}^k\}_{k = 1}^N$ are the multi-level features of the three frames and $N$ is the number of feature levels. + +![](images/5a8c29e39bf5ffda6cab944146caef336883dae0319f1957a820946db98e53dc.jpg) +Fig. 2: Our VFI-assisted multi-frame fusion depth model. Motion and occlusion provided by VFI can be used to align and fuse multi-frame features in an explicit manner. + +Motion-aware Feature Alignment (MAFA). For the $k$ -th level, we use bidirectional optical flow to align the neighbor features $\phi_{t-2}^k$ and $\phi_{t+2}^k$ to the target position $t$ , generating $\phi_{(t-2) \to t}^k$ and $\phi_{(t+2) \to t}^k$ , respectively, formulated as: + +$$ +\begin{array}{l} \phi_ {(t - 2) \rightarrow t} ^ {k} = \omega \left(\phi_ {t - 2} ^ {k}, F _ {t \rightarrow (t - 2)} ^ {k}\right), \\ \phi_ {(t + 2) \rightarrow t} ^ {k} = \omega \left(\phi_ {t + 2} ^ {k}, F _ {t \rightarrow (t + 2)} ^ {k}\right), \tag {11} \\ \end{array} +$$ + +where $F_{t\rightarrow (t - 2)}^k$ and $F_{t\rightarrow (t + 2)}^k$ denote the corresponding flows of the $k$ -th level, downsampled from $F_{t\rightarrow (t - 2)}$ and $F_{t\rightarrow (t + 2)}$ , respectively. Inspired by recent 2D implicit neural representation [8], we also concatenate the aligned features with encoded optical flow, which can be denoted as: + +$$ +\begin{array}{l} \varphi_ {(t - 2) \rightarrow t} ^ {k} = \operatorname {C o n c a t} \left(\phi_ {(t - 2) \rightarrow t} ^ {k}, g \left(F _ {t \rightarrow (t - 2)} ^ {k}\right)\right), \\ \varphi_ {(t + 2) \rightarrow t} ^ {k} = \operatorname {C o n c a t} \left(\phi_ {(t + 2) \rightarrow t} ^ {k}, g \left(F _ {t \rightarrow (t + 2)} ^ {k}\right)\right), \tag {12} \\ \varphi_ {t} ^ {k} = \operatorname {C o n c a t} \left(\phi_ {t} ^ {k}, g \left(F _ {0} ^ {k}\right)\right), \\ \end{array} +$$ + +where Concat means feature concatenation, $F_0^k$ denotes zero flow and $g$ is fourier positional encoding [41,51] to transform optical flow to high-frequency signal. Actually, optical flow indicates the relative coordinate of a feature vector in the warped feature map w.r.t. the original one. When incorporating it, multi-frame model can adaptively adjust feature learning through motion-aware alignment. + +Occlusion-alleviated Feature Fusion (OAFF). Due to the ubiquitous interframe occlusion, aggregating the aligned features directly through summation or concatenation can deteriorate the performance. Therefore, we use the $k$ -th level merge mask $M_t^k$ downsampled from $M_t$ to mitigate this issue, just like the image merging process of VFI in Eq. (8): + +$$ +\chi_ {t} ^ {k} = M _ {t} ^ {k} \odot \varphi_ {(t - 2) \rightarrow t} ^ {k} + (1 - M _ {t} ^ {k}) \odot \varphi_ {(t + 2) \rightarrow t} ^ {k}, \tag {13} +$$ + +where $\chi_t^k$ is the occlusion-alleviated feature merged from the temporal positions $t - 2$ and $t + 2$ . It is then concatenated with the feature $\varphi_t^k$ in the position $t$ . + +And we can obtain the final feature after fusion by: + +$$ +\psi_ {t} ^ {k} = \operatorname {C o n v} _ {1 \times 1} \left(\operatorname {C o n c a t} \left(\varphi_ {t} ^ {k}, \chi_ {t} ^ {k}\right)\right), \tag {14} +$$ + +where $\mathrm{Conv}_{1\times 1}$ is a $1\times 1$ convolutional layer to transform the feature into the same channel as the original feature $\phi_t^k$ , enabling the multi-frame depth decoder to employ the same architecture as the single-frame one. + +Multi-frame Depth Decoder. To translate the multi-level fusion features into a depth map, a decoder is required. Concretely, we use the same architecture as the single-frame depth decoder, but the parameters are not shared. This encourages the single- and multi-frame models to decode the features in their individual ways without conflict. Finally, the fusion features $\{\psi_t^k\}_{k=1}^N$ are processed by a multi-frame depth decoder $\mathcal{D}_m$ to generate the fusion depth $D_t^m$ of the target frame $I_t$ , which can be written as: + +$$ +D _ {t} ^ {m} = \mathcal {D} _ {m} \left(\left\{\psi_ {t} ^ {k} \right\} _ {k = 1} ^ {N}\right). \tag {15} +$$ + +Note that the multi-frame depth $D_{t}^{m}$ also corresponds to a self-supervised loss like Eq. (5), which is denoted as: + +$$ +L _ {s s} ^ {m} (t) = \mathcal {L} \left(D _ {t} ^ {m}, \boldsymbol {T} _ {t \rightarrow s}, I _ {t}, I _ {s}\right). \tag {16} +$$ + +# 4.3 Spatial Augmentation + +Orthogonal to temporal augmentation, we also include classic data augmentation to spatially expand the data diversity, which is categorized as spatial augmentation in this paper. Resizing-cropping is a commonly used and effective augmentation approach. Based on this operation, we introduce image rotation, extending it to a generic affine transformation (cropping is actually image translation). Specifically, we follow [55] to consider that the augmented image and the original image are captured by the same camera system whose intrinsics are fixed. For resizing operation, it is assumed that when an image is zoomed up by a factor $f_{s} \geq 1$ , the relative size of an object increases by the same factor $f_{s}$ and its depth decreases by $f_{s}$ simultaneously. This is inspired from a vital monocular cue [4] that the closer an object is, the larger its relative size is. Our affine transformation $\mathcal{A}$ is applied to both the target image $I_{t}$ and the source image $I_{s}$ , generating $\widetilde{I}_{t} = \mathcal{A}(I_{t})$ and $\widetilde{I}_{s} = \mathcal{A}(I_{s})$ , respectively. For the augmented target view, we can obtain its depth through the single-frame model by $\widetilde{D}_{t} = \mathcal{D}(\mathcal{E}(\widetilde{I}_{t}))$ . To conduct view synthesis on augmented images, the original camera pose $T_{t\rightarrow s} = [\pmb{R}_{t\rightarrow s}|\pmb{t}_{t\rightarrow s}]$ from target to source should be rectified to $\widetilde{T}_{t\rightarrow s} = [\widetilde{\pmb{R}}_{t\rightarrow s}|\widetilde{\pmb{t}}_{t\rightarrow s}]$ as done in [55]: + +$$ +\tilde {\boldsymbol {R}} _ {t \rightarrow s} = \boldsymbol {R} _ {c} \boldsymbol {R} _ {t \rightarrow s} \boldsymbol {R} _ {c} ^ {- 1}, \quad \tilde {\boldsymbol {t}} _ {t \rightarrow s} = \boldsymbol {R} _ {c} \boldsymbol {t} _ {t \rightarrow s}, \tag {17} +$$ + +where $\mathbf{R}_c$ is the rectification matrix determined only by the camera intrinsics and the augmentation parameters. Different from [55] calculating $\mathbf{R}_c$ under resizing-cropping, we give its formulation and derivation under a more general affine transformation, which can be seen in our supplementary. Therefore, we can derive + +the self-supervised loss $\widetilde{L}_{ss}(t)$ on the augmented images similarly to Eq. (5): + +$$ +\widetilde {L} _ {s s} (t) = \mathcal {L} (\widetilde {D} _ {t}, \widetilde {\boldsymbol {T}} _ {t \rightarrow s}, \widetilde {I} _ {t}, \widetilde {I} _ {s}). \tag {18} +$$ + +Note that there is a mask to exclude the possible invalid pixels in $\widetilde{I}_t$ when calculating $\widetilde{L}_{ss}(t)$ due to image rotation. + +# 4.4 Triplet Depth Consistency + +Now we have single-frame depths $D_{t}$ and $\widetilde{D}_{t}$ of the original image $I_{t}$ and the augmented image $\widetilde{I}_{t}$ , along with multi-frame depth $D_{t}^{m}$ of $I_{t}$ . For better knowledge distillation and regularization, our unified learning framework applies a triplet depth consistency loss among the three depth maps, including a standard view depth consistency (SVDC) loss and two scale-aware depth consistency (SADC) losses, which are shown in Fig. 1(b). The SVDC loss $L_{sv}(t)$ conducts mutual distillation between single- and multi-frame models, promoting their mutual reinforcement, which can be denoted as: + +$$ +L _ {s v} (t) = \operatorname {S I} \left(D _ {t} ^ {m}, D _ {t}\right), \tag {19} +$$ + +where $\mathrm{SI}()$ is the widely used scale-invariant error function [13]. Besides, we use depth $\widetilde{D}_t$ under spatial augmentation to enforce two SADC losses, i.e., $L_{sa}(t)$ between $\widetilde{D}_t$ and $D_t$ , and $L_{sa}^m(t)$ between $\widetilde{D}_t$ and $D_t^m$ . To this end, we first restore $\widetilde{D}_t$ to the original image view, generating $\widehat{D}_t = \mathcal{A}^{-1}(\widetilde{D}_t)$ in Fig. 1(b), where $\mathcal{A}^{-1}$ is the inverse transformation of affine augmentation $\mathcal{A}$ . To exclude the invalid pixels in $\widehat{D}_t$ , we adopt a mask $M_{sa}$ and calculate $L_{sa}(t)$ and $L_{sa}^m(t)$ through: + +$$ +L _ {s a} (t) = \operatorname {S I} \left(M _ {s a} \odot D _ {t}, M _ {s a} \odot f _ {s} \widehat {D} _ {t}\right), \tag {20} +$$ + +$$ +L _ {s a} ^ {m} (t) = \operatorname {S I} \left(M _ {s a} \odot D _ {t} ^ {m}, M _ {s a} \odot f _ {s} \widehat {D} _ {t}\right), \tag {21} +$$ + +where the factor $f_{s}$ is due to the assumption in Sec. 4.3 that the depth of an object in $I_{t}$ is $f_{s}$ times of that in $\widetilde{I}_{t}$ . Our SADC constrains the 3D geometric relationship between the scenes of $I_{t}$ and $\widetilde{I}_{t}$ , providing scale-aware regularization. Combining SVDC and SADC, we have the triplet depth consistency loss as: + +$$ +L _ {t c} (t) = L _ {s v} (t) + L _ {s a} (t) + L _ {s a} ^ {m} (t). \tag {22} +$$ + +# 4.5 Training Objective + +For target position $t$ , we have already obtained the self-supervised losses $L_{ss}(t)$ , $L_{ss}^{m}(t)$ and $\widetilde{L}_{ss}(t)$ in Eqs. (5), (16) and (18), respectively. Besides, our triplet depth consistency loss $L_{tc}(t)$ is acquired in Eq. (22). Therefore, the total loss on position $t$ can be summed as: + +$$ +L (t) = L _ {s s} (t) + L _ {s s} ^ {m} (t) + \widetilde {L} _ {s s} (t) + \lambda L _ {t c} (t), \tag {23} +$$ + +where $\lambda$ is a hyper-parameter set to 0.2 in this paper. For $t - 1$ and $t + 1$ , we can derive their losses $L(t - 1)$ and $L(t + 1)$ in the same way. In summary, the overall loss function of our Mono-ViFI is $L = L(t - 1) + L(t) + L(t + 1)$ . + +Table 1: Performance comparison on KITTI [16] benchmark using the raw ground truth, with a resolution of $640 \times 192$ . The best results in each subsection are in bold; second best are underlined. For MACs of multi-frame models, we use the formulation A/B. A denotes the MACs when predicting for a single frame; B means the average MACs per frame when processing a long video sequence. #Params and MACs of our multi-frame models include the VFI network. + +
MethodTest frames#ParamsMACsAbs RelSq RelRMSERMSE logδ1δ2δ3
PackNet-SfM [21]1128M205G0.1110.7854.6010.1890.8780.9600.982
HR-Depth [39]116.9M31.6G0.1090.7924.6320.1850.8840.9620.983
R-MSFM6 [67]13.8M31.2G0.1120.8064.7040.1910.8780.9600.981
BRNet [22]119.1M33.7G0.1050.6984.4620.1790.8900.9650.984
MonoViT [64]127M-0.0990.7084.3720.1750.9000.9670.984
DIFFNet [65]112.2M23.4G0.1020.7644.4830.1800.8960.9650.983
SD-SSMDE [44]1--0.1000.6614.2640.1720.8960.9670.985
MUSTNet [48]1--0.1060.7634.5620.1820.8880.9630.983
DepthSegNet [40]130.3M-0.1040.6904.4730.1790.8860.9650.984
DualRefine [2]1--0.1030.7764.4910.1810.8940.9650.983
DaCNN [23]113M-0.0990.6614.3160.1730.8970.9670.985
Monodepth2 (ResNet18) [18]114.3M8G0.1150.9034.8630.1930.8770.9590.981
Mono-ViFI (ResNet18)114.3M8G0.1050.7084.4460.1790.8870.9650.984
Lite-Mono [62]13.1M5.1G0.1070.7654.5610.1830.8860.9630.983
Mono-ViFI (Lite-Mono)13.1M5.1G0.1030.6804.3490.1760.8910.9660.985
D-HRNet [25]110M10.8G0.1020.7604.4790.1790.8970.9650.983
RA-Depth (D-HRNet) [25]110M10.8G0.0960.6324.2160.1710.9030.9680.985
Mono-ViFI (D-HRNet)110M10.8G0.0960.6274.1790.1700.9030.9690.985
Patil et al. [42]N--/-0.1110.8214.6500.1870.8830.9610.982
ManyDepth (ResNet18) [58]2 (-1, 0)26.9M15.1G/13.7G0.0980.7704.4590.1760.9000.9650.983
DynamicDepth [14]2 (-1, 0)--/-0.0960.7204.4580.1750.8970.9640.984
DepthFormer [20]2 (-1, 0)28.7M175G/-0.0900.6614.1490.1750.9050.9670.984
MOVEDepth [56]2 (-1, 0)28.2M20.2G/19.2G0.0940.7044.3890.1750.9020.9650.983
DualRefine [2]2 (-1, 0)--/-0.0870.6984.2340.1700.9140.9670.983
TC-Depth [49]3 (-1, 0, 1)--/-0.1030.7464.4830.1850.894-0.983
Mono-ViFI (ResNet18)3 (-1, 0, 1)17.9M25.8G/16.9G0.0990.6614.3210.1740.8980.9660.985
Mono-ViFI (Lite-Mono)3 (-1, 0, 1)5.9M21.9G/13.3G0.0990.6484.2560.1720.8980.9670.985
Mono-ViFI (D-HRNet)3 (-1, 0, 1)12.9M35.7G/19.4G0.0910.5894.0880.1660.9120.9690.986
+ +# 5 Experiments + +# 5.1 Implementation Details + +Datasets. We train and evaluate the models on two widely used outdoor urban datasets of KITTI [16] and Cityscapes [10]. The depth labels of KITTI are collected with LiDARs. For fair comparison, we use the Eigen split [12] and follow the pre-processing operation in [18], generating 39,810 monocular triplets for training and 697 frames for evaluation. To reduce the influence of noise in the sparse depth from Velodyne, we also evaluate using $93\%$ of the Eigen split with the improved ground truth from [52], which contains 652 test frames. In regard to Cityscapes, following [58, 66], we preprocess the monocular sequences into triples, producing 69,731 training images. And we evaluate on 1,525 test images using the ground truth provided by [58]. For generalization comparison, KITTI models are evaluated on Cityscapes and Make3D [50]. Make3D is a small outdoor dataset containing 134 test images with aligned depth information. + +Flow-based VFI Model. We choose IFRNet [29] as our VFI model due to its efficiency. In practice, we use the large model of IFRNet for training and replace + +![](images/99e79450c6426335bf4c0e9386f8b6e7c1850d1ad49ca9df9568c42be55984ec.jpg) +Fig. 3: Qualitative results on KITTI. Error maps in columns 2 and 4 show the Abs Rel error compared to the improved ground truth [52] from good (blue) to bad (red). + +![](images/0da88bfe2f0d99496838e7378941a00260474aa275d552af0216217e93a6d7e8.jpg) + +![](images/3c94a3e3f9f98c4879c9225deb6ca36e4b8fe91f56f484a1376167f9532d1156.jpg) + +![](images/283d98377f6cc2f608e38498ec70a8b43581e3270bbefc364a23cbbba0a2309c.jpg) + +it with the small model when evaluating for multi-frame fusion depth. Please see ablation about their gap in supplementary. Both large and small IFRNets are pretrained on KITTI/Cityscapes and frozen in Mono-ViFI training. The pretraining data split is the same as self-supervised depth training. Each original triplet in depth training is directly used for VFI training. Besides, we employ another two flow-based VFI networks, RIFE [26] and AMT [34], to verify the robustness of our method to different VFI models. + +Depth and Pose Networks. We experiment with three depth architectures, including the ResNet18 [24] based network in Monodepth2 [18], medium LiteMono [62], and D-HRNet proposed in RA-Depth [25]. For the pose network, we follow Monodepth2 to use a ResNet18 based model which takes a pair of images (six channels) as input and predict a 6-DoF relative pose. Encoders of all the networks are pretrained on ImageNet [11]. + +# 5.2 Performance Comparison + +KITTI Results. Quantitative comparison results on KITTI under $640 \times 192$ resolution are listed in Table 1 and Table 2, using raw ground truth and improved ground truth, respectively. It is obvious that our method with D-HRNet can achieve state-of-the-art performance. Besides, our Mono-ViFI brings notable improvements to the three depth architectures without increasing model complexity (see each subsection separated by dotted lines). Despite that the difference between Mono-ViFI (D-HRNet) and RA-Depth [25] under single-frame mode is slight in Table 1, it is more significant in Table 2 when using improved ground truth. For multi-frame inference, Mono-ViFI with ResNet18 has already shown advantages over ManyDepth [58] and DynamicDepth [14] on some metrics including Sq Rel, RMSE, RMSE log, $\delta_{2}$ and $\delta_{3}$ . When changing to D-HRNet backbone, proposed Mono-ViFI can totally outperform them. Although our models have more multiply-add computations (MACs) when predicting for a single frame, the average MACs per frame are comparable when processing a long video sequence. Fig. 3 depicts two visual comparison results on KITTI, where + +Table 2: Performance comparison on Table 3: Performance comparison on KITTI [16] using improved ground truth Cityscapes [10] with a resolution of $416 \times$ from [52], with a resolution of $640 \times 192$ . 128 unless stated. + +
MethodAbs RelSq RelRMSERMSE logδ1δ2δ3
Single-frame Inference
PackNet-SIM [21]0.0780.4203.4850.1210.9310.9860.996
R-MSF M6 [67]0.0880.4923.8370.1350.9150.9830.995
DIFFNet [65]0.0760.4143.4050.1190.9360.9880.996
Mono-ViF [64]0.0740.3883.4140.1150.9380.9890.997
Monodepth2 (ResNet18) [18]0.0900.5453.9420.1370.9140.9830.995
Mono-ViF (ResNet18)0.0840.4313.6180.1260.9240.9860.997
Lite-Mono [62]0.0820.4553.6850.1270.9230.9850.996
Mono-ViF (Lite-Mono)0.0800.4003.4970.1210.9300.9870.997
D-HRNet [25]0.0770.4233.3960.1190.9350.9870.996
RA-Depth (D-HRNet) [25]0.0740.3633.3490.1140.9400.9900.997
Mono-ViF (D-HRNet)0.0720.3423.2540.1110.9410.9910.998
Multi-frame Inference
ManyDepth (ResNet18) [58]0.0700.3993.4550.1130.9410.9890.997
MOVEDepth [56]0.0650.3773.4490.1120.9420.9880.996
Mono-ViF (ResNet18)0.0740.3743.3810.1130.9410.9900.997
Mono-ViF (Lite-Mono)0.0740.3543.3720.1140.9390.9890.997
Mono-ViF (D-HRNet)0.0660.2943.0990.1020.9520.9920.998
+ +
MethodAbs RelSq RelRMSERMSE logδ1δ2δ3
Single-frame Inference
Videos in the wild [19]0.1271.3306.9600.1950.8300.9470.981
Li et al. [33]0.1191.2906.9800.1900.8460.9520.982
InstDM [31] (832 × 256)0.1111.1586.4370.1820.8680.9610.983
SD-SSMDE [44]0.1100.9885.9530.165---
Monodepth2 (ResNet18) [18]0.1291.5696.3760.1970.8490.9570.983
Mono-ViFi (ResNet18)0.1121.0035.8990.1650.8720.9660.990
Lite-Mono [62]0.1141.1946.3280.1720.8710.9670.989
Mono-ViFi (Lite-Mono)0.1070.9405.7940.1590.8760.9680.991
D-HRNet [25]0.1111.1426.2050.1690.8790.9680.989
Mono-ViFi (D-HRNet)0.1050.8805.5500.1560.8810.9690.992
Multi-frame Inference
ManyDepth (ResNet18) [58]0.1141.1936.2230.1700.8750.9670.989
DynamicDepth [14]0.1041.0115.9870.1590.8890.9720.991
TC-Depth [49]0.1100.9585.820-0.867-0.991
Mono-ViFi (ResNet18)0.1070.9345.6880.1590.8800.9690.990
Mono-ViFi (Lite-Mono)0.1050.9045.6460.1570.8800.9700.991
Mono-ViFi (D-HRNet)0.1020.8365.3950.1530.8850.9710.992
+ +Table 4: Evaluation results on dynamic objects and static regions of Cityscapes [10] dataset under $416\times 128$ resolution. Note that DynamicDepth relies on depth prior and segmentation masks of dynamic objects when reasoning. + +
MethodTest AreaAbs RelSq RelRMSERMSE logδ1δ2δ3
Multi-frame Inference
ManyDepth (ResNet18) [58]DynamicDepth [14]Mono-ViFI (ResNet18)Dynamic Objects0.1722.2065.9240.2160.7920.9170.967
0.1441.5545.2380.1850.8310.9500.982
0.1621.4535.1140.1990.7940.9240.977
ManyDepth (ResNet18) [58]DynamicDepth [14]Mono-ViFI (ResNet18)Static Regions0.1071.0906.3140.1600.8820.9720.991
0.1000.9776.1550.1550.8920.9740.992
0.0990.9075.8350.1490.8900.9750.992
+ +we choose D-HRNet as the baseline depth model. It can be seen that depth maps look qualitatively similar among all four methods. However, the error maps reveal mistakes of D-HRNet baseline and RA-Depth in some ambiguous regions like bushes (see the red boxes). In contrast, our Mono-ViFi have dramatically lower error in these regions, especially under multi-frame reasoning. + +Cityscapes Results. Results of Cityscapes in Table 3 demonstrate the same trend as KITTI. Moreover, proposed multi-frame model with ResNet18 can beat ManyDepth on all metrics, while our Mono-ViFI is better than DynamicDepth only on partial indexes. It is because Cityscapes with more dynamic scenes is more challenging than KITTI, and DynamicDepth leverages additional information to cope with dynamic objects. For fine-grained analysis, we separately evaluate on the dynamic objects and static regions, using the segmentation masks of dynamic objects provided in [14], as listed in Table 4. On dynamic objects, Mono-ViFI with ResNet18 surpasses ManyDepth by a large margin, demonstrating that our multi-frame inference paradigm can sidestep the issue of dynamic objects to a greater extent compared to ManyDepth. DynamicDepth utilizes depth prior and segmentation masks of dynamic objects to tackle this problem for multi-frame inference. Even though, our method still excels it on Sq Rel and + +![](images/b055474670ed0ea897e8507ce5a1717e48236a09eacf6bed41360cfcb8c2708b.jpg) +Fig. 4: Qualitative generalization results of a YouTube video example. All depth maps are predicted under $640 \times 360$ resolution using KITTI models trained under $640 \times 192$ . + +Table 5: Ablation results of each proposed component on Cityscapes [10]. SA is spatial augmentation. TA is temporal augmentation. SVDC denotes standard view depth consistency. SADC denotes scale-aware depth consistency. FA is feature alignment. MAFA is motion-aware feature alignment. OAFF means occlusion-alleviated feature fusion. + +
MethodSATASVDCSADCFAMAFAOAFFAbs RelSq RelRMSEδ1
Single-frame Inference
Baseline (Monodepth2 [18])0.1291.5696.8760.849
Baseline + TA✓0.1191.2546.5230.857
Baseline + SVDC✓0.1171.1866.4920.862
Baseline + SA✓0.1221.3866.5570.856
Baseline + SA + SADC✓✓0.1161.0796.1890.868
Mono-ViFI (Single-frame)✓✓✓✓0.1121.0035.8990.872
Multi-frame Inference
Baseline (direct feature fusion)0.1201.3606.4910.869
FA (not motion-aware)✓0.1161.2176.3870.871
MAFA✓0.1141.1506.1150.874
MAFA + OAFF✓✓0.1111.1076.0120.876
MAFA + OAFF + TA✓✓✓0.1091.0235.8920.877
Mono-ViFI (Multi-frame)✓✓✓✓✓✓0.1070.9345.6880.880
+ +RMSE. When evaluating on static regions, proposed Mono-ViFI shows absolute superiority over ManyDepth and DynamicDepth. + +Generalization Results. We evaluate the KITTI models on Cityscapes and Make3D without fine-tuning, results of which are put in the supplementary. For qualitative comparison, we give a visualization example in Fig. 4, where KITTI models are used to predict the multi-frame depth of a YouTube video frame downloaded from the Wind Walk Travel Videos channel. This indicates that our Mono-ViFI has better generalization ability than existing methods. + +# 5.3 Ablation Study + +In this part, we explore the influence of each component in Mono-ViFi. Our ablations are conducted on Cityscapes or KITTI with ResNet18 backbone. + +Table 6: Ablation of different flow-based VFI models on KITTI [16]. + +
VFI ModelAbs RelSq RelRMSERMSE logδ1δ2δ3
Multi-frame Inference
IFRNet [29] (Ours)0.0990.6614.3210.1740.8980.9660.985
RIFE [26]0.1000.6724.3520.1750.8960.9650.985
AMT [34]0.0990.6534.3350.1740.8970.9660.985
+ +Single-frame Ablation. We first ablate on the single-frame model. As listed in the top half part of Table 5, the standard view depth consistency and spatial augmentation can bring gains individually. Note that scale-aware depth consistency $L_{sa}$ should be based on spatial augmentation. When adding $L_{sa}$ to spatial augmentation, the accuracy can be further improved. And this leads to the most significant enhancement among all components, since $L_{sa}$ can enforce the model to learn scale relationship from scene variations. In orthogonal dimension, our proposed temporal augmentation can also reduce the prediction error by providing more guidance from newly synthesized camera views. When integrating all above elements, our single-frame model can achieve the best performance. + +Multi-frame Ablation. To probe the effect of each component in our VFI-assisted multi-frame fusion module, we conduct ablation experiments under multi-frame inference, whose results are summarized in the bottom half part of Table 5. The baseline is stacking the raw feature of three frames directly, i.e., direct feature fusion. Then, we gradually apply components of the proposed fusion module, including motion-aware feature alignment and occlusion-alleviated feature fusion. The results demonstrate that these strategies can bring cumulative improvements. Furthermore, when incorporating spatial augmentation, temporal augmentation and relevant depth consistency losses, our multi-frame model can be further promoted like the single-frame setting. + +Ablation of different VFI models. Apart from IFRNet [29], we additionally experiment with another two flow-based VFI models, RIFE [26] and AMT [34]. The results are depicted in Table 6. It can be observed that there is only a negligible performance difference among the three VFI networks, which reveals that our method is robust to different VFI models. + +# 6 Conclusion + +In conclusion, this paper presents a novel self-supervised learning framework, named Mono-ViFI, to bilaterally connect single- and multi-frame depth estimation in a unified manner. First, a flow-based VFI model is employed for temporal augmentation. Then, we propose a VFI-assisted multi-frame fusion module for more accurate multi-frame depth. Besides, resizing-cropping is extended to affine transformation as spatial augmentation for enriching data diversity. Finally, we develop a new triplet depth consistency loss for better regularization. Experiments show that our approach can unleash the potential of existing depth models to a greater extent, but without increasing inference complexity. + +# References + +1. Bae, J., Moon, S., Im, S.: Deep digging into the generalization of self-supervised monocular depth estimation. In: AAAI. pp. 187-196 (2023) +2. Bangunharcana, A., Magd, A., Kim, K.S.: Dualrefine: Self-supervised depth and pose estimation through iterative epipolar sampling and refinement toward equilibrium. In: CVPR. pp. 726-738 (2023) +3. Bello, J.L.G., Kim, M.: Forget about the lidar: Self-supervised depth estimators with med probability volumes. In: NeurIPS. pp. 12626-12637 (2020) +4. Bello, J.L.G., Moon, J., Kim, M.: Positional information is all you need: A novel pipeline for self-supervised svde from videos. arXiv:2205.08851 (2022) +5. Bian, J., Li, Z., Wang, N., Zhan, H., Shen, C., Cheng, M.M., Reid, I.: Unsupervised scale-consistent depth and ego-motion learning from monocular video. In: NeurIPS. pp. 35-45 (2019) +6. Chan, K.C., Wang, X., Yu, K., Dong, C., Loy, C.C.: Basicvsr: The search for essential components in video super-resolution and beyond. In: CVPR. pp. 4945-4954 (2021) +7. Chan, K.C., Zhou, S., Xu, X., Loy, C.C.: Basicvsr++: Improving video superresolution with enhanced propagation and alignment. In: CVPR. pp. 5962-5971 (2022) +8. Chen, Y., Liu, S., Wang, X.: Learning continuous image representation with local implicit image function. In: CVPR. pp. 8624-8634 (2021) +9. Chen, Y., Schmid, C., Sminchisescu, C.: Self-supervised learning with geometric constraints in monocular video: Connecting flow, depth, and camera. In: ICCV. pp. 7063-7072 (2019) +10. Cordts, M., Omran, M., Ramos, S., Rehfeld, T., Enzweiler, M., Benenson, R., Franke, U., Roth, S., Schiele, B.: The cityscapes dataset for semantic urban scene understanding. In: CVPR. pp. 3213-3223 (2016) +1. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: CVPR. pp. 248-255 (2009) +2. Eigen, D., Fergus, R.: Predicting depth, surface normals and semantic labels with a common multi-scale convolutional architecture. In: ICCV. pp. 2650-2658 (2015) +3. Eigen, D., Puhrsch, C., Fergus, R.: Depth map prediction from a single image using a multi-scale deep network. In: NeurIPS. pp. 2366-2374 (2014) +4. Feng, Z., Yang, L., Jing, L., Wang, H., Tian, Y., Li, B.: Disentangling object motion and occlusion for unsupervised multi-frame monocular depth. In: ECCV. pp. 228-244 (2022) +5. Garg, R., Kumar, B.G.V., Carneiro, G., Reid, I.: Unsupervised cnn for single view depth estimation: Geometry to the rescue. In: ECCV. pp. 740-756 (2016) +6. Geiger, A., Lenz, P., Stiller, C., Urtasun, R.: Vision meets robotics: The kitti dataset. *IJRR* **32**(11), 1231–1237 (2013) +7. Godard, C., Aodha, O.M., Brostow, G.J.: Unsupervised monocular depth estimation with left-right consistency. In: CVPR. pp. 6602-6611 (2017) +8. Godard, C., Aodha, O.M., Firman, M., Brostow, G.J.: Digging into self-supervised monocular depth estimation. In: CVPR. pp. 3828-3838 (2019) +9. Gordon, A., Li, H., Jonschkowski, R., Angelova, A.: Depth from videos in the wild: Unsupervised monocular depth learning from unknown cameras. In: ICCV. pp. 8976-8985 (2019) +20. Guizilini, V., Ambrus, R., Chen, D., Zakharov, S., Gaidon, A.: Multi-frame self-supervised depth with transformers. In: CVPR. pp. 160-170 (2022) + +21. Guizilini, V., Ambrus, R., Pillai, S., Gaidon, A.: 3d packing for self-supervised monocular depth estimation. In: CVPR. pp. 2482-2491 (2020) +22. Han, W., Yin, J., Jin, X., Dai, X., Shen, J.: Brnet: Exploring comprehensive features for monocular depth estimation. In: ECCV. pp. 586-602 (2022) +23. Han, W., Yin, J., Shen, J.: Self-supervised monocular depth estimation by direction-aware cumulative convolution network. In: ICCV (2023) +24. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR. pp. 770-778 (2016) +25. He, M., Hui, L., Bian, Y., Ren, J., Xie, J., Yang, J.: Ra-depth: Resolution adaptive self-supervised monocular depth estimation. In: ECCV. pp. 565-581 (2022) +26. Huang, Z., Zhang, T., Heng, W., Shi, B., Zhou, S.: Real-time intermediate flow estimation for video frame interpolation. In: ECCV. pp. 624-642 (2022) +27. Jung, H., Park, E., Yoo, S.: Fine-grained semantics-aware representation enhancement for self-supervised monocular depth estimation. In: ICCV. pp. 12622-12632 (2021) +28. Klingner, M., Termöhlen, J.A., Mikolajczyk, J., Fingscheidt, T.: Self-supervised monocular depth estimation: Solving the dynamic object problem by semantic guidance. In: ECCV. pp. 582-600 (2020) +29. Kong, L., Jiang, B., Luo, D., Chu, W., Huang, X., Tai, Y., Wang, C., Yang, J.: Ifrnet: Intermediate feature refine network for efficient frame interpolation. In: CVPR. pp. 1959-1968 (2022) +30. Kuznietsov, Y., Proesmans, M., Gool, L.V.: Comoda: Continuous monocular depth adaptation using past experiences. In: WACV. pp. 2906-2916 (2021) +31. Lee, S., Im, S., Lin, S., Kweon, I.S.: Learning monocular depth in dynamic scenes via instance-aware projection consistency. In: AAAI. pp. 1863-1872 (2021) +32. Lee, S., Rameau, F., Pan, F., Kweon, I.S.: Attentive and contrastive learning for joint depth and motion field estimation. In: ICCV. pp. 4862-4871 (2021) +33. Li, H., Gordon, A., Zhao, H., Casser, V., Angelova, A.: Unsupervised monocular depth learning in dynamic scenes. In: CoRL. pp. 1908-1917 (2020) +34. Li, Z., Zhu, Z.L., Han, L.H., Hou, Q., Guo, C.L., Cheng, M.M.: Amt: All-pairs multi-field transforms for efficient frame interpolation. In: CVPR. pp. 9801-9810 (2023) +35. Liu, J., Kong, L., Yang, J.: Designing and searching for lightweight monocular depth network. In: ICONIP. pp. 477-488 (2021) +36. Liu, J., Kong, L., Yang, J.: Atca: An arc trajectory based model with curvature attention for video frame interpolation. In: ICIP. pp. 1486-1490 (2022) +37. Liu, J., Kong, L., Yang, J., Liu, W.: Towards better data exploitation in self-supervised monocular depth estimation. IEEE Robotics and Automation Letters 9(1), 763-770 (2024) +38. Luo, X., Huang, J.B., Szeliski, R., Matzen, K., Kopf, J.: Consistent video depth estimation. TOG 39(71), 1-13 (2020) +39. Lyu, X., Liu, L., Wang, M., Kong, X., Liu, L., Liu, Y., Chen, X., Yuan, Y.: Hrdepth: High resolution self-supervised monocular depth estimation. In: AAAI. pp. 2294-2301 (2021) +40. Ma, J., Lei, X., Liu, N., Zhao, X., Pu, S.: Towards comprehensive representation enhancement in semantics-guided self-supervised monocular depth estimation. In: ECCV. pp. 304-321 (2022) +41. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: ECCV. pp. 405-421 (2020) + +42. Patil, V., Gansbeke, W.V., Dai, D., Gool, L.V.: Don't forget the past: Recurrent depth estimation from monocular video. IEEE Robotics and Automation Letters 5(4), 6813-6820 (2020) +43. Peng, R., Wang, R., Lai, Y., Tang, L., Cai, Y.: Excavating the potential capacity of selfsupervised monocular depth estimation. In: ICCV. pp. 15560-15569 (2021) +44. Petrovai, A., Nedevschi, S.: Exploiting pseudo labels in a self-supervised learning framework for improved monocular depth estimation. In: CVPR. pp. 1568-1578 (2022) +45. Pilzer, A., Lathuilière, S., Sebe, N., Ricci, E.: Refine and distill: Exploiting cycle-inconsistency and knowledge distillation for unsupervised monocular depth estimation. In: CVPR. pp. 9760-9769 (2019) +46. Poggi, M., Aleotti, F., Tosi, F., Mattoccia, S.: On the uncertainty of self-supervised monocular depth estimation. In: CVPR. pp. 3224-3234 (2020) +47. Ranjan, A., Jampani, V., Balles, L., Kim, K., Sun, D., Wulff, J., Black, M.J.: Competitive collaboration: Joint unsupervised learning of depth, camera motion, optical flow and motion segmentation. In: CVPR. pp. 12240-12249 (2019) +48. Ren, W., Wang, L., Piao, Y., Zhang, M., Lu, H., Liu, T.: Adaptive co-teaching for unsupervised monocular depth estimation. In: ECCV. pp. 89-105 (2022) +49. Ruhkamp, P., Gao, D., Chen, H., Navab, N., Busam, B.: Attention meets geometry: Geometry guided spatial-temporal attention for consistent self-supervised monococular depth estimation. In: 3DV. pp. 837-847 (2021) +50. Saxena, A., Sun, M., Ng, A.Y.: Make3d: Learning 3d scene structure from a single still image. PAMI 31(5), 824-840 (2009) +51. Tancik, M., Srinivasan, P.P., Mildenhall, B., Fridovich-Keil, S., Raghavan, N., Singhal, U., Ramamoorthi, R., Barron, J.T., Ng, R.: Fourier features let networks learn high frequency functions in low dimensional domains. In: NeurIPS. pp. 7537-7547 (2020) +52. Uhrig, J., Schneider, N., Schneider, L., Franke, U., Brox, T., Geiger, A.: Sparsity invariant cnns. In: 3DV. pp. 11-20 (2017) +53. Wang, C., Buenaposada, J.M., Zhu, R., Lucey, S.: Learning depth from monocular videos using direct methods. In: CVPR. pp. 2022-2030 (2018) +54. Wang, R., Pizer, S.M., Frahm, J.M.: Recurrent neural network for (un-)supervised learning of monocular video visual odometry and depth. In: CVPR. pp. 5555-5564 (2019) +55. Wang, R., Yu, Z., Gao, S.: Planedepth: Self-supervised depth estimation via orthogonal planes. In: CVPR. pp. 21425-21434 (2023) +56. Wang, X., Zhu, Z., Huang, G., Chi, X., Ye, Y., Chen, Z., Wang, X.: Drafting monocular cues and velocity guidance for self-supervised multi-frame depth learning. In: AAAI. pp. 2689-2697 (2023) +57. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: From error visibility to structural similarity. IEEE TIP 13(4), 600-612 (2004) +58. Watson, J., Aodha, O.M., Prisacariu, V.A., Brostow, G.J., Firman, M.: The temporal opportunist: Self-supervised multi-frame monocular depth. In: CVPR. pp. 1164-1174 (2021) +59. Watson, J., Firman, M., Brostow, G.J., Turmukhambetov, D.: Self-supervised monocular depth hints. In: ICCV. pp. 2162-2171 (2019) +60. Yin, Z., Shi, J.: Geonet: Unsupervised learning of dense depth, optical flow and camera pose. In: CVPR. pp. 1983-1992 (2018) +61. Zhang, H., Li, Y., Cao, Y., Liu, Y., Shen, C., Yan, Y.: Exploiting temporal consistency for real-time video depth estimation. In: ICCV. pp. 1725-1734 (2019) + +62. Zhang, N., Nex, F., Vosselman, G., Kerle, N.: Lite-mono: A lightweight cnn and transformer architecture for self-supervised monocular depth estimation. In: CVPR. pp. 18537-18546 (2023) +63. Zhao, C., Poggi, M., Tosi, F., Zhou, L., Sun, Q., Tanga, Y., Mattoccia, S.: Gasmono: Geometry-aided self-supervised monocular depth estimation for indoor scenes. In: ICCV. pp. 16163-16174 (2023) +64. Zhao, C., Zhang, Y., Poggi, M., Tosi, F., Guo, X., Zhu, Z., Huang, G., Tang, Y., Mattoccia, S.: Monovit: Self-supervised monocular depth estimation with a vision transformer. In: 3DV. pp. 668-678 (2022) +65. Zhou, H., Greenwood, D., Taylor, S.: Self-supervised monocular depth estimation with internal feature fusion. In: BMVC (2021) +66. Zhou, T., Brown, M., Snavely, N., Lowe, D.G.: Unsupervised learning of depth and ego-motion from video. In: CVPR. pp. 6612-6619 (2017) +67. Zhou, Z., Fan, X., Shi, P., Xin, Y.: R-msfm: Recurrent multi-scale feature modulation for monocular depth estimating. In: ICCV. pp. 12757-12766 (2021) \ No newline at end of file diff --git a/2024/Mono-ViFI_ A Unified Learning Framework for Self-supervised Single- and Multi-frame Monocular Depth Estimation/images.zip b/2024/Mono-ViFI_ A Unified Learning Framework for Self-supervised Single- and Multi-frame Monocular Depth Estimation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..404018b60cc8a726a86a56b3608f7197fe1c31c0 --- /dev/null +++ b/2024/Mono-ViFI_ A Unified Learning Framework for Self-supervised Single- and Multi-frame Monocular Depth Estimation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a5bcf89f9c911ff4e9152b1aeb8b6f359e6b9122a2091d4bcbc9b9716565442c +size 730527 diff --git a/2024/Mono-ViFI_ A Unified Learning Framework for Self-supervised Single- and Multi-frame Monocular Depth Estimation/layout.json b/2024/Mono-ViFI_ A Unified Learning Framework for Self-supervised Single- and Multi-frame Monocular Depth Estimation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..f9ecd3eab5ba3a9d545ff647442364dc9618c344 --- /dev/null +++ b/2024/Mono-ViFI_ A Unified Learning Framework for Self-supervised Single- and Multi-frame Monocular Depth Estimation/layout.json @@ -0,0 +1,11566 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 143, + 111, + 471, + 166 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 111, + 471, + 166 + ], + "spans": [ + { + "bbox": [ + 143, + 111, + 471, + 166 + ], + "type": "text", + "content": "Mono-ViFi: A Unified Learning Framework for Self-supervised Single- and Multi-frame Monocular Depth Estimation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 132, + 186, + 481, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 186, + 481, + 199 + ], + "spans": [ + { + "bbox": [ + 132, + 186, + 481, + 199 + ], + "type": "text", + "content": "Jinfeng Liu, Lingtong Kong\\*, Bo Li, Zerong Wang, Hong Gu, and Jinwei Chen" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 214, + 208, + 399, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 214, + 208, + 399, + 219 + ], + "spans": [ + { + "bbox": [ + 214, + 208, + 399, + 219 + ], + "type": "text", + "content": "vivo Mobile Communication Co., Ltd, China" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 156, + 220, + 459, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 156, + 220, + 459, + 231 + ], + "spans": [ + { + "bbox": [ + 156, + 220, + 459, + 231 + ], + "type": "text", + "content": "{liujinfeng, ltkong, libra, wangzerong, guhong, jinwei.chen}@vivo.com" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 259, + 453, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 259, + 453, + 479 + ], + "spans": [ + { + "bbox": [ + 160, + 259, + 453, + 479 + ], + "type": "text", + "content": "Abstract. Self-supervised monocular depth estimation has gathered notable interest since it can liberate training from dependency on depth annotations. In monocular video training case, recent methods only conduct view synthesis between existing camera views, leading to insufficient guidance. To tackle this, we try to synthesize more virtual camera views by flow-based video frame interpolation (VFI), termed as temporal augmentation. For multi-frame inference, to sidestep the problem of dynamic objects encountered by explicit geometry-based methods like ManyDepth, we return to the feature fusion paradigm and design a VFI-assisted multi-frame fusion module to align and aggregate multi-frame features, using motion and occlusion information obtained by the flow-based VFI model. Finally, we construct a unified self-supervised learning framework, named Mono-ViFi, to bilaterally connect single- and multiframe depth. In this framework, spatial data augmentation through image affine transformation is incorporated for data diversity, along with a triplet depth consistency loss for regularization. The single- and multiframe models can share weights, making our framework compact and memory-efficient. Extensive experiments demonstrate that our method can bring significant improvements to current advanced architectures. Source code is available at https://github.com/LiuJF1226/Mono-ViFi." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 160, + 490, + 456, + 512 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 490, + 456, + 512 + ], + "spans": [ + { + "bbox": [ + 160, + 490, + 456, + 512 + ], + "type": "text", + "content": "Keywords: Monocular depth estimation " + }, + { + "bbox": [ + 160, + 490, + 456, + 512 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 490, + 456, + 512 + ], + "type": "text", + "content": " Self-supervised learning " + }, + { + "bbox": [ + 160, + 490, + 456, + 512 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 490, + 456, + 512 + ], + "type": "text", + "content": " Data augmentation " + }, + { + "bbox": [ + 160, + 490, + 456, + 512 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 490, + 456, + 512 + ], + "type": "text", + "content": " Video frame interpolation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 533, + 230, + 545 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 533, + 230, + 545 + ], + "spans": [ + { + "bbox": [ + 133, + 533, + 230, + 545 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": "Monocular depth estimation (MDE) is a fundamental 3D vision task, which finds applications in robotics, autonomous driving and VR/AR. To reduce the dependence on depth annotations, self-supervised MDE [1,2,15,17,46,63,66] has risen as a convenient and practical solution, which uses monocular videos, stereo pairs, or even both to generate supervision through view synthesis, where an image photometric reconstruction loss is minimized. Compared to stereo training case [15,17,59] with known camera pose, learning from monocular videos [18,58, 66] is more challenging since it requires an extra pose network to estimate camera ego-motion. Even so, it is much easier to collect monocular videos compared to" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "stereo image pairs, making monocular training more attractive. Therefore, this paper also focuses on self-supervised training with unlabeled monocular videos." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 144, + 481, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 144, + 481, + 240 + ], + "spans": [ + { + "bbox": [ + 130, + 144, + 481, + 240 + ], + "type": "text", + "content": "Recent works have made various attempts to improve accuracy, like introducing masking techniques [18, 66] and leveraging segmentation priors [28, 31]. However, they only conduct view synthesis between existing fixed camera views within video datasets, leading to insufficient guidance especially in invisible regions by occlusion or disappearance. Considering that video data possesses temporal continuity, an intuitive idea is to generate novel frames, i.e., virtual camera views in the temporal dimension. We term it as temporal augmentation, which can be achieved by video frame interpolation (VFI), as shown in Fig. 1(a)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 243, + 482, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 243, + 482, + 506 + ], + "spans": [ + { + "bbox": [ + 130, + 243, + 482, + 506 + ], + "type": "text", + "content": "From another perspective, multi-frame reasoning paradigm have recently been researched because multiple frames can be available for reasoning in many application scenarios. ManyDepth [58] proposes to use cost volume constructed from multi-frame matching features. It explicitly reasons about geometry based on static scene assumption, which can be easily affected by dynamic objects and occlusion. DynamicDepth [14] attempts to tackle dynamic scenes, but relies on depth priors and segmentation masks of moving objects when reasoning. To sidestep the issue of dynamic objects in multi-frame inference encountered by these methods, we return to the traditional paradigm [42, 49], which aggregates multi-frame features directly. However, direct feature fusion without alignment harms the learning of temporal representations, since feature embeddings at the same location in neighbor frames are usually irrelevant, making it difficult for models to leverage the correlation of multiple frames. Hence, we regard feature alignment as a reasonable and necessary step before aggregation. As revealed in several video enhancement tasks like video super-resolution [6, 7] and frame interpolation [26, 29, 34, 36], inter-frame feature alignment can be implemented by warping through optical flow. This inspires us to employ a flow-based VFI model to concurrently conduct temporal augmentation and provide explicit correspondence for feature alignment. Moreover, the VFI model can reason out occlusion mask between neighbor frames, helping to alleviate inter-frame occlusion problem in feature fusion. Based on above analysis, we design a VFI-assisted multi-frame fusion module for multi-frame depth estimation, as depicted in Fig. 2." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": "Finally, we build a unified framework to bilaterally connect single- and multiframe depth, as demonstrated in Fig. 1(b). The direct way is by depth consistency, usually for geometric regularization or distillation in self-supervised MDE. ManyDepth conducts unilateral distillation from single-frame model to multiframe one. DynamicDepth extends it to a bilateral mutual reinforcement, but only in the regions of moving objects. Differently, our framework applies the bilateral consistency on the whole image. For an original target image, we use a standard view depth consistency (SVDC) loss between its two depth maps predicted from the single- and multi-frame models. We also include classic data augmentation strategies, categorized as spatial augmentation. Specifically, we combine resizing-cropping with image rotation, forming a generic affine transformation. Then, we give the formulation of rectified camera pose for affine-augmented view synthesis using the cue of relative object size [4, 55]. For the" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "text", + "content": "J. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 112, + 319, + 233 + ], + "blocks": [ + { + "bbox": [ + 133, + 112, + 319, + 233 + ], + "lines": [ + { + "bbox": [ + 133, + 112, + 319, + 233 + ], + "spans": [ + { + "bbox": [ + 133, + 112, + 319, + 233 + ], + "type": "image", + "image_path": "9a3d1e5d9591e39c21ee0f4c6d3c34d41e74bc997a12e6f691af7d204581947b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 160, + 236, + 296, + 246 + ], + "lines": [ + { + "bbox": [ + 160, + 236, + 296, + 246 + ], + "spans": [ + { + "bbox": [ + 160, + 236, + 296, + 246 + ], + "type": "text", + "content": "(a) Temporal Augmentation with VFI" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 320, + 113, + 480, + 232 + ], + "blocks": [ + { + "bbox": [ + 320, + 113, + 480, + 232 + ], + "lines": [ + { + "bbox": [ + 320, + 113, + 480, + 232 + ], + "spans": [ + { + "bbox": [ + 320, + 113, + 480, + 232 + ], + "type": "image", + "image_path": "ae377703be20437e014938cb17aad97ac308c67f058d7004bde80aa93374d895.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 323, + 236, + 468, + 246 + ], + "lines": [ + { + "bbox": [ + 323, + 236, + 468, + 246 + ], + "spans": [ + { + "bbox": [ + 323, + 236, + 468, + 246 + ], + "type": "text", + "content": "(b) Learning Framework " + }, + { + "bbox": [ + 323, + 236, + 468, + 246 + ], + "type": "inline_equation", + "content": "(I_{t}" + }, + { + "bbox": [ + 323, + 236, + 468, + 246 + ], + "type": "text", + "content": " for example" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 252, + 482, + 353 + ], + "lines": [ + { + "bbox": [ + 130, + 252, + 482, + 353 + ], + "spans": [ + { + "bbox": [ + 130, + 252, + 482, + 353 + ], + "type": "text", + "content": "Fig. 1: Overview of Mono-ViFI. (a) We achieve temporal augmentation by a flow-based VFI model " + }, + { + "bbox": [ + 130, + 252, + 482, + 353 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 130, + 252, + 482, + 353 + ], + "type": "text", + "content": ", which also reasons out intermediate optical flow and occlusion mask for multi-frame inference. (b) For target image " + }, + { + "bbox": [ + 130, + 252, + 482, + 353 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 130, + 252, + 482, + 353 + ], + "type": "text", + "content": ", we obtain its single-frame depth " + }, + { + "bbox": [ + 130, + 252, + 482, + 353 + ], + "type": "inline_equation", + "content": "D_{t}" + }, + { + "bbox": [ + 130, + 252, + 482, + 353 + ], + "type": "text", + "content": " and multi-frame depth " + }, + { + "bbox": [ + 130, + 252, + 482, + 353 + ], + "type": "inline_equation", + "content": "D_{t}^{m}" + }, + { + "bbox": [ + 130, + 252, + 482, + 353 + ], + "type": "text", + "content": ". Also, " + }, + { + "bbox": [ + 130, + 252, + 482, + 353 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 130, + 252, + 482, + 353 + ], + "type": "text", + "content": " is augmented to " + }, + { + "bbox": [ + 130, + 252, + 482, + 353 + ], + "type": "inline_equation", + "content": "\\widetilde{I}_{t}" + }, + { + "bbox": [ + 130, + 252, + 482, + 353 + ], + "type": "text", + "content": " through affine transformation " + }, + { + "bbox": [ + 130, + 252, + 482, + 353 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 130, + 252, + 482, + 353 + ], + "type": "text", + "content": ". We then calculate the depth " + }, + { + "bbox": [ + 130, + 252, + 482, + 353 + ], + "type": "inline_equation", + "content": "\\widehat{D}_t" + }, + { + "bbox": [ + 130, + 252, + 482, + 353 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 130, + 252, + 482, + 353 + ], + "type": "inline_equation", + "content": "\\widetilde{I}_t" + }, + { + "bbox": [ + 130, + 252, + 482, + 353 + ], + "type": "text", + "content": " and inversely convert it to the original view, generating " + }, + { + "bbox": [ + 130, + 252, + 482, + 353 + ], + "type": "inline_equation", + "content": "\\widehat{D}_t" + }, + { + "bbox": [ + 130, + 252, + 482, + 353 + ], + "type": "text", + "content": ". Finally, we enforce a triplet depth consistency loss " + }, + { + "bbox": [ + 130, + 252, + 482, + 353 + ], + "type": "inline_equation", + "content": "L_{tc}(t)" + }, + { + "bbox": [ + 130, + 252, + 482, + 353 + ], + "type": "text", + "content": " among the three depth maps, including a standard view depth consistency loss " + }, + { + "bbox": [ + 130, + 252, + 482, + 353 + ], + "type": "inline_equation", + "content": "L_{sv}(t)" + }, + { + "bbox": [ + 130, + 252, + 482, + 353 + ], + "type": "text", + "content": " and two scale-aware depth consistency losses, " + }, + { + "bbox": [ + 130, + 252, + 482, + 353 + ], + "type": "inline_equation", + "content": "L_{sa}(t)" + }, + { + "bbox": [ + 130, + 252, + 482, + 353 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 252, + 482, + 353 + ], + "type": "inline_equation", + "content": "L_{sa}^{m}(t)" + }, + { + "bbox": [ + 130, + 252, + 482, + 353 + ], + "type": "text", + "content": ". Note that each depth map also corresponds to a photometric loss and a smoothness loss, which are omitted here." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 377, + 482, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 377, + 482, + 449 + ], + "spans": [ + { + "bbox": [ + 130, + 377, + 482, + 449 + ], + "type": "text", + "content": "depth map from augmented target view, we enforce two scale-aware depth consistency (SADC) losses between it and above two standard view depths, respectively. SADC stems from the relative size cue, which assumes that when an image is zoomed up by a factor, the depth decreases by the same factor. It guides depth model to learn consistent geometric scale relationship from scene variations. The SVDC loss and two SADC losses are termed as a triplet depth consistency loss." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 449, + 482, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 482, + 498 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 482, + 498 + ], + "type": "text", + "content": "Our framework is named Mono-ViFI, since the self-supervised Monocular depth is largely boosted by Video Frame Interpolation. To our best knowledge, we are the first to leverage VFI for the improvement of self-supervised MDE. The main contributions of this paper are summarized as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 137, + 508, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 138, + 508, + 481, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 508, + 481, + 555 + ], + "spans": [ + { + "bbox": [ + 138, + 508, + 481, + 555 + ], + "type": "text", + "content": "- We propose Mono-ViFi, a unified learning framework to bilaterally connect self-supervised single- and multi-frame MDE. In Mono-ViFi, the single- and multi-frame models can share weights, making the unified framework more compact and memory-efficient." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 556, + 481, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 556, + 481, + 592 + ], + "spans": [ + { + "bbox": [ + 138, + 556, + 481, + 592 + ], + "type": "text", + "content": "- We conduct temporal augmentation through flow-based VFI, and develop a novel VFI-assisted multi-frame fusion module, which utilizes motion and occlusion from flow-based VFI model to improve multi-frame depth." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 137, + 593, + 481, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 593, + 481, + 616 + ], + "spans": [ + { + "bbox": [ + 137, + 593, + 481, + 616 + ], + "type": "text", + "content": "- We incorporate image affine transformation as spatial augmentation, along with a triplet depth consistency loss for regularization and distillation." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 137, + 617, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 617, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 137, + 617, + 481, + 665 + ], + "type": "text", + "content": "- Experiments show that Mono-ViFI can achieve state-of-the-art performance. Moreover, our approach is orthogonal to architecture design, allowing current advanced depth networks to be further improved by seamlessly integrating them with Mono-ViFI." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 157, + 90, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 90, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 157, + 90, + 448, + 103 + ], + "type": "text", + "content": "Mono-ViFi: A Unified Framework for Self-supervised Monocular Depth" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 101 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 101 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 101 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 237, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 237, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 237, + 127 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 137, + 388, + 149 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 137, + 388, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 137, + 388, + 149 + ], + "type": "text", + "content": "2.1 Self-supervised Monocular Depth Estimation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 154, + 479, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 154, + 479, + 177 + ], + "spans": [ + { + "bbox": [ + 130, + 154, + 479, + 177 + ], + "type": "text", + "content": "Self-supervised learning paradigm for monocular depth estimation can be categorized into stereo training and monocular training according to the data source." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 178, + 480, + 249 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 178, + 480, + 249 + ], + "spans": [ + { + "bbox": [ + 130, + 178, + 480, + 249 + ], + "type": "text", + "content": "Stereo training leverages the left-right consistency in stereo image pairs. Garg et al. [15] first consider the training as a view synthesis problem with a photometric consistency loss between stereo pairs. Building upon this idea, Godard et al. [17] introduce a left-right disparity consistency loss. Then, it is further extended by using depth hints [59], data augmentation [43, 55], knowledge distillation [43, 45, 55] and plane-based scene representation [3, 55]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 251, + 481, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 251, + 481, + 441 + ], + "spans": [ + { + "bbox": [ + 130, + 251, + 481, + 441 + ], + "type": "text", + "content": "Monocular training leverages the temporal consistency in monocular video frames. Zhou et al. [66] propose a pioneering method which jointly learns depth and camera ego-motion from monocular videos in a self-supervised way. Then, Bian et al. [5] propose a scale-consistent depth and ego-motion estimation approach by adding a depth consistency loss, which reduces scale drift of estimated depth but decreases the accuracy. Later, Godard et al. [18] propose Monodepth2, a strong baseline for most follow-up works, which employs a minimum reprojection loss to mitigate occlusion issues and an auto-masking loss to filter out moving objects with similar velocity as the camera. Furthermore, some works introduce multi-task learning to combine depth estimation with optical flow [9, 32, 47, 60], semantic segmentation [27, 28, 40] and instance segmentation [31]. Others [21-23, 35, 37, 39, 62, 64, 65] try to improve the network architectures for depth estimation. However, these methods only utilize the geometric constrains between existing fixed camera views in video datasets. This insufficient guidance may block current depth models from unleashing their full capabilities. To address it, we conduct temporal augmentation through VFI in this paper." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 456, + 314, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 456, + 314, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 456, + 314, + 468 + ], + "type": "text", + "content": "2.2 Multi-frame Monocular Depth" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 474, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 482, + 666 + ], + "type": "text", + "content": "Above single-frame depth models do not leverage any temporal information for inference, which limits the performance to some degree. Regarding multi-frame depth prediction, early works adopt test-time refinement [9, 30, 38], use spatial-temporal attention [49], or combine traditional depth networks with recurrent layers [54, 61] to process sequence information. Another broad group of methods [14, 20, 56, 58] utilize cost volume from stereo matching, which explicitly encode and reason from geometric constraints by combining a multi-frame cost volume with single-frame features. On the one hand, this type of methods based on ManyDepth [58] usually meet with the problem of dynamic scenes. To sidestep this issue, we develop a VFI-assisted multi-frame fusion depth model to align and aggregate multi-frame features by optical flow and occlusion masks. On the other hand, ManyDepth uses single-frame depth prior to boost multi-frame depth estimation. However, this improvement is unidirectional. DynamicDepth [14] extends it to a mutual reinforcement, but only targets on dynamic objects. Differently, we propose a unified framework to bilaterally connect single- and multi-frame depth in a more compact and memory-efficient manner." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "text", + "content": "J. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 244, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 244, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 244, + 129 + ], + "type": "text", + "content": "3 Problem Setup" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 248 + ], + "type": "text", + "content": "Self-supervised MDE in the monocular training case leverages the geometric constraints from video sequences. Specifically, for a training triplet " + }, + { + "bbox": [ + 130, + 140, + 482, + 248 + ], + "type": "inline_equation", + "content": "\\{I_{t-2}, I_t, I_{t+2}\\}" + }, + { + "bbox": [ + 130, + 140, + 482, + 248 + ], + "type": "text", + "content": " which contains three consecutive frames, we use " + }, + { + "bbox": [ + 130, + 140, + 482, + 248 + ], + "type": "inline_equation", + "content": "I_t" + }, + { + "bbox": [ + 130, + 140, + 482, + 248 + ], + "type": "text", + "content": " as the target view and the other two as the source views " + }, + { + "bbox": [ + 130, + 140, + 482, + 248 + ], + "type": "inline_equation", + "content": "\\{I_s\\}_{s \\in \\{t-2, t+2\\}}" + }, + { + "bbox": [ + 130, + 140, + 482, + 248 + ], + "type": "text", + "content": " to implement view synthesis. Here the temporal notation " + }, + { + "bbox": [ + 130, + 140, + 482, + 248 + ], + "type": "inline_equation", + "content": "t \\pm 2" + }, + { + "bbox": [ + 130, + 140, + 482, + 248 + ], + "type": "text", + "content": " instead of " + }, + { + "bbox": [ + 130, + 140, + 482, + 248 + ], + "type": "inline_equation", + "content": "t \\pm 1" + }, + { + "bbox": [ + 130, + 140, + 482, + 248 + ], + "type": "text", + "content": " is for the convenience of following description. Also, a pose network is utilized to estimate the camera pose from the target view to the source view, denoted as " + }, + { + "bbox": [ + 130, + 140, + 482, + 248 + ], + "type": "inline_equation", + "content": "\\pmb{T}_{t \\to s} = [\\pmb{R}_{t \\to s} | \\pmb{t}_{t \\to s}]" + }, + { + "bbox": [ + 130, + 140, + 482, + 248 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 140, + 482, + 248 + ], + "type": "inline_equation", + "content": "\\pmb{R}_{t \\to s}" + }, + { + "bbox": [ + 130, + 140, + 482, + 248 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 140, + 482, + 248 + ], + "type": "inline_equation", + "content": "\\pmb{t}_{t \\to s}" + }, + { + "bbox": [ + 130, + 140, + 482, + 248 + ], + "type": "text", + "content": " represent the rotation and translation components. Formally, the target depth " + }, + { + "bbox": [ + 130, + 140, + 482, + 248 + ], + "type": "inline_equation", + "content": "D_t" + }, + { + "bbox": [ + 130, + 140, + 482, + 248 + ], + "type": "text", + "content": " and the camera pose " + }, + { + "bbox": [ + 130, + 140, + 482, + 248 + ], + "type": "inline_equation", + "content": "\\pmb{T}_{t \\to s}" + }, + { + "bbox": [ + 130, + 140, + 482, + 248 + ], + "type": "text", + "content": " can be derived from:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 231, + 255, + 480, + 266 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 255, + 480, + 266 + ], + "spans": [ + { + "bbox": [ + 231, + 255, + 480, + 266 + ], + "type": "interline_equation", + "content": "D _ {t} = \\mathcal {D} \\left(\\mathcal {E} \\left(I _ {t}\\right)\\right), \\quad \\boldsymbol {T} _ {t \\rightarrow s} = \\mathcal {P} \\left(I _ {t}, I _ {s}\\right), \\tag {1}", + "image_path": "431c705a59632a3cc581be3d1bf470c335ba421947b4aa7587f2113cbbf701b4.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 271, + 480, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 271, + 480, + 307 + ], + "spans": [ + { + "bbox": [ + 130, + 271, + 480, + 307 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 271, + 480, + 307 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 130, + 271, + 480, + 307 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 271, + 480, + 307 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 130, + 271, + 480, + 307 + ], + "type": "text", + "content": " represent the encoder and decoder, respectively, of a single-frame depth network, while " + }, + { + "bbox": [ + 130, + 271, + 480, + 307 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 130, + 271, + 480, + 307 + ], + "type": "text", + "content": " denotes the pose network. Then, the reconstructed view of " + }, + { + "bbox": [ + 130, + 271, + 480, + 307 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 130, + 271, + 480, + 307 + ], + "type": "text", + "content": " from the source view " + }, + { + "bbox": [ + 130, + 271, + 480, + 307 + ], + "type": "inline_equation", + "content": "I_{s}" + }, + { + "bbox": [ + 130, + 271, + 480, + 307 + ], + "type": "text", + "content": " can be obtained by:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 239, + 313, + 480, + 326 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 239, + 313, + 480, + 326 + ], + "spans": [ + { + "bbox": [ + 239, + 313, + 480, + 326 + ], + "type": "interline_equation", + "content": "I _ {s \\rightarrow t} = I _ {s} \\left\\langle \\operatorname {p r o j} \\left(D _ {t}, \\boldsymbol {T} _ {t \\rightarrow s}, \\boldsymbol {K}\\right)\\right\\rangle , \\tag {2}", + "image_path": "b775b342116ebde2272f73af5cd6b6c60271956e5a466122d4e39361b63c91c2.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 331, + 482, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 331, + 482, + 426 + ], + "spans": [ + { + "bbox": [ + 130, + 331, + 482, + 426 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 331, + 482, + 426 + ], + "type": "inline_equation", + "content": "I_{s\\rightarrow t}" + }, + { + "bbox": [ + 130, + 331, + 482, + 426 + ], + "type": "text", + "content": " is the reconstructed target image, " + }, + { + "bbox": [ + 130, + 331, + 482, + 426 + ], + "type": "inline_equation", + "content": "\\langle \\rangle" + }, + { + "bbox": [ + 130, + 331, + 482, + 426 + ], + "type": "text", + "content": " means the differentiable bilinear sampling operator, " + }, + { + "bbox": [ + 130, + 331, + 482, + 426 + ], + "type": "inline_equation", + "content": "proj()" + }, + { + "bbox": [ + 130, + 331, + 482, + 426 + ], + "type": "text", + "content": " returns the 2D coordinates of the pixels in " + }, + { + "bbox": [ + 130, + 331, + 482, + 426 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 130, + 331, + 482, + 426 + ], + "type": "text", + "content": " after they are reprojected to the viewpoint of " + }, + { + "bbox": [ + 130, + 331, + 482, + 426 + ], + "type": "inline_equation", + "content": "I_{s}" + }, + { + "bbox": [ + 130, + 331, + 482, + 426 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 331, + 482, + 426 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 331, + 482, + 426 + ], + "type": "text", + "content": " is the camera intrinsic matrix assumed to be known. Finally, the per-pixel photometric loss between " + }, + { + "bbox": [ + 130, + 331, + 482, + 426 + ], + "type": "inline_equation", + "content": "I_{s\\rightarrow t}" + }, + { + "bbox": [ + 130, + 331, + 482, + 426 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 331, + 482, + 426 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 130, + 331, + 482, + 426 + ], + "type": "text", + "content": " is used to optimize the depth and pose networks in an end-to-end manner. We follow Monodepth2 [18] to adopt the combination of L1 distance and SSIM [57] as the photometric error function and take the minimum across all the source views at each pixel to handle occlusion, which is formulated as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 181, + 431, + 480, + 450 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 181, + 431, + 480, + 450 + ], + "spans": [ + { + "bbox": [ + 181, + 431, + 480, + 450 + ], + "type": "interline_equation", + "content": "L _ {p e} (t) = \\min _ {s} \\frac {\\alpha}{2} \\left(1 - \\operatorname {S S I M} \\left(I _ {t}, I _ {s \\rightarrow t}\\right)\\right) + (1 - \\alpha) \\| I _ {t}, I _ {s \\rightarrow t} \\| _ {1}, \\tag {3}", + "image_path": "3480ac3c6338d666a8ba536a9b5e6581e0c571ffc4c74716362517abba588578.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 454, + 480, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 454, + 480, + 477 + ], + "spans": [ + { + "bbox": [ + 130, + 454, + 480, + 477 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 454, + 480, + 477 + ], + "type": "inline_equation", + "content": "\\alpha = 0.85" + }, + { + "bbox": [ + 130, + 454, + 480, + 477 + ], + "type": "text", + "content": ". Besides, the edge-aware smoothness loss is employed to cope with depth discontinuities:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 221, + 483, + 480, + 496 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 221, + 483, + 480, + 496 + ], + "spans": [ + { + "bbox": [ + 221, + 483, + 480, + 496 + ], + "type": "interline_equation", + "content": "L _ {s m} (t) = \\left| \\partial_ {x} d _ {t} ^ {*} \\right| e ^ {- \\left| \\partial_ {x} I _ {t} \\right|} + \\left| \\partial_ {y} d _ {t} ^ {*} \\right| e ^ {- \\left| \\partial_ {y} I _ {t} \\right|}, \\tag {4}", + "image_path": "2592c3310c7c639c0319d137e3826a6cc3d799f3c1974b465ecc6baad2212a36.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 501, + 479, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 501, + 479, + 525 + ], + "spans": [ + { + "bbox": [ + 130, + 501, + 479, + 525 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 501, + 479, + 525 + ], + "type": "inline_equation", + "content": "d_t^* = d_t / \\overline{d_t}" + }, + { + "bbox": [ + 130, + 501, + 479, + 525 + ], + "type": "text", + "content": " is the mean-normalized inverse depth to discourage shrinking [53]. In summary, the standard self-supervised loss on " + }, + { + "bbox": [ + 130, + 501, + 479, + 525 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 130, + 501, + 479, + 525 + ], + "type": "text", + "content": " is:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 200, + 532, + 480, + 544 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 532, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 200, + 532, + 480, + 544 + ], + "type": "interline_equation", + "content": "L _ {s s} (t) = \\mathcal {L} \\left(D _ {t}, \\boldsymbol {T} _ {t \\rightarrow s}, I _ {t}, I _ {s}\\right) = \\mu L _ {p e} (t) + \\gamma L _ {s m} (t), \\tag {5}", + "image_path": "e21dc33a81f145d28e4cc761af48cc60199714a655e2d7dcf4eacce50c6832a7.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 548, + 357, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 548, + 357, + 561 + ], + "spans": [ + { + "bbox": [ + 131, + 548, + 357, + 561 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 548, + 357, + 561 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 131, + 548, + 357, + 561 + ], + "type": "text", + "content": " is the auto-masking in [18] and " + }, + { + "bbox": [ + 131, + 548, + 357, + 561 + ], + "type": "inline_equation", + "content": "\\gamma = 0.001" + }, + { + "bbox": [ + 131, + 548, + 357, + 561 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 579, + 233, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 579, + 233, + 594 + ], + "spans": [ + { + "bbox": [ + 132, + 579, + 233, + 594 + ], + "type": "text", + "content": "4 Methodology" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "In this paper, we propose novel Mono-ViFI for self-supervised monocular training, which bilaterally connects single- and multi-frame depth models into a unified learning framework. With flow-based VFI model, we introduce temporal augmentation and develop VFI-assisted multi-frame fusion depth, which are presented in Sec. 4.1 and Sec. 4.2, respectively. Besides, our framework incorporates" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 157, + 90, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 90, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 157, + 90, + 448, + 103 + ], + "type": "text", + "content": "Mono-ViFi: A Unified Framework for Self-supervised Monocular Depth" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 154 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 154 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 154 + ], + "type": "text", + "content": "spatial data augmentation and a triplet depth consistency loss, which are separately described in Sec. 4.3 and Sec. 4.4. Finally, we give the overall training objective in Sec. 4.5. The overview of our Mono-ViFi is depicted in Fig. 1." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 167, + 335, + 180 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 167, + 335, + 180 + ], + "spans": [ + { + "bbox": [ + 132, + 167, + 335, + 180 + ], + "type": "text", + "content": "4.1 Temporal Augmentation with VFI" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 185, + 482, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 185, + 482, + 245 + ], + "spans": [ + { + "bbox": [ + 130, + 185, + 482, + 245 + ], + "type": "text", + "content": "Firstly, we generate more new frames in the temporal dimension based on existing video frames for temporal augmentation. To achieve this, we adopt a pretrained flow-based VFI model, such as IFRNet [29], which takes two temporally adjacent frames " + }, + { + "bbox": [ + 130, + 185, + 482, + 245 + ], + "type": "inline_equation", + "content": "\\{I_0,I_2\\}" + }, + { + "bbox": [ + 130, + 185, + 482, + 245 + ], + "type": "text", + "content": " as input to synthesize an intermediate frame " + }, + { + "bbox": [ + 130, + 185, + 482, + 245 + ], + "type": "inline_equation", + "content": "I_{1}" + }, + { + "bbox": [ + 130, + 185, + 482, + 245 + ], + "type": "text", + "content": ". The frame interpolation process can be formulated as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 242, + 251, + 480, + 264 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 242, + 251, + 480, + 264 + ], + "spans": [ + { + "bbox": [ + 242, + 251, + 480, + 264 + ], + "type": "interline_equation", + "content": "F _ {1 \\rightarrow 0}, F _ {1 \\rightarrow 2}, M _ {1} = \\mathcal {H} \\left(I _ {0}, I _ {2}\\right), \\tag {6}", + "image_path": "bbe099d3dc4becefce4105f4838ed6919dc8dbe3a2fd824e06c37f2c991d0c20.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 212, + 266, + 480, + 278 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 212, + 266, + 480, + 278 + ], + "spans": [ + { + "bbox": [ + 212, + 266, + 480, + 278 + ], + "type": "interline_equation", + "content": "I _ {0 \\rightarrow 1} = \\omega \\left(I _ {0}, F _ {1 \\rightarrow 0}\\right), \\quad I _ {2 \\rightarrow 1} = \\omega \\left(I _ {2}, F _ {1 \\rightarrow 2}\\right), \\tag {7}", + "image_path": "64063d77d26e6b603df2281be229df5f678020b6c5d2d31c7413bcecdf941037.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 228, + 281, + 480, + 293 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 281, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 228, + 281, + 480, + 293 + ], + "type": "interline_equation", + "content": "I _ {1} = M _ {1} \\odot I _ {0 \\rightarrow 1} + (1 - M _ {1}) \\odot I _ {2 \\rightarrow 1}, \\tag {8}", + "image_path": "f6d08853939d06c1867119555346d6793907eef36fbd4a0de8b36dc04776f12d.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 298, + 482, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 298, + 482, + 418 + ], + "spans": [ + { + "bbox": [ + 130, + 298, + 482, + 418 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 298, + 482, + 418 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 130, + 298, + 482, + 418 + ], + "type": "text", + "content": " is the VFI model, " + }, + { + "bbox": [ + 130, + 298, + 482, + 418 + ], + "type": "inline_equation", + "content": "F_{1\\rightarrow 0}" + }, + { + "bbox": [ + 130, + 298, + 482, + 418 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 298, + 482, + 418 + ], + "type": "inline_equation", + "content": "F_{1\\rightarrow 2}" + }, + { + "bbox": [ + 130, + 298, + 482, + 418 + ], + "type": "text", + "content": " are the intermediate optical flow, " + }, + { + "bbox": [ + 130, + 298, + 482, + 418 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 130, + 298, + 482, + 418 + ], + "type": "text", + "content": " represents backward warping. " + }, + { + "bbox": [ + 130, + 298, + 482, + 418 + ], + "type": "inline_equation", + "content": "I_{0\\rightarrow 1}" + }, + { + "bbox": [ + 130, + 298, + 482, + 418 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 298, + 482, + 418 + ], + "type": "inline_equation", + "content": "I_{2\\rightarrow 1}" + }, + { + "bbox": [ + 130, + 298, + 482, + 418 + ], + "type": "text", + "content": " denote the warped images to the intermediate position, while " + }, + { + "bbox": [ + 130, + 298, + 482, + 418 + ], + "type": "inline_equation", + "content": "M_{1}" + }, + { + "bbox": [ + 130, + 298, + 482, + 418 + ], + "type": "text", + "content": " ranging from 0 to 1 is the merge mask adjusting the mixing ratio according to bidirectional occlusion information. Here, " + }, + { + "bbox": [ + 130, + 298, + 482, + 418 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 130, + 298, + 482, + 418 + ], + "type": "text", + "content": " indicates the element-wise multiplication. Considering a triplet " + }, + { + "bbox": [ + 130, + 298, + 482, + 418 + ], + "type": "inline_equation", + "content": "\\{I_{t - 2},I_t,I_{t + 2}\\}" + }, + { + "bbox": [ + 130, + 298, + 482, + 418 + ], + "type": "text", + "content": " with three consecutive frames as described in Sec. 3, we use the VFI model to synthesize " + }, + { + "bbox": [ + 130, + 298, + 482, + 418 + ], + "type": "inline_equation", + "content": "I_{t - 1}" + }, + { + "bbox": [ + 130, + 298, + 482, + 418 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 130, + 298, + 482, + 418 + ], + "type": "inline_equation", + "content": "\\{I_{t - 2},I_t\\}" + }, + { + "bbox": [ + 130, + 298, + 482, + 418 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 298, + 482, + 418 + ], + "type": "inline_equation", + "content": "I_{t + 1}" + }, + { + "bbox": [ + 130, + 298, + 482, + 418 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 130, + 298, + 482, + 418 + ], + "type": "inline_equation", + "content": "\\{I_t,I_{t + 2}\\}" + }, + { + "bbox": [ + 130, + 298, + 482, + 418 + ], + "type": "text", + "content": ", as shown in Fig. 1(a). Then, we regard " + }, + { + "bbox": [ + 130, + 298, + 482, + 418 + ], + "type": "inline_equation", + "content": "I_{t - 1}" + }, + { + "bbox": [ + 130, + 298, + 482, + 418 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 298, + 482, + 418 + ], + "type": "inline_equation", + "content": "I_{t + 1}" + }, + { + "bbox": [ + 130, + 298, + 482, + 418 + ], + "type": "text", + "content": " as another two target frames. All three target frames have common sources of " + }, + { + "bbox": [ + 130, + 298, + 482, + 418 + ], + "type": "inline_equation", + "content": "\\{I_{t - 2},I_{t + 2}\\}" + }, + { + "bbox": [ + 130, + 298, + 482, + 418 + ], + "type": "text", + "content": ". The following operations and losses are the same for each of them, and we only take " + }, + { + "bbox": [ + 130, + 298, + 482, + 418 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 130, + 298, + 482, + 418 + ], + "type": "text", + "content": " for instance." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 434, + 360, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 434, + 360, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 434, + 360, + 445 + ], + "type": "text", + "content": "4.2 VFI-assisted Multi-frame Fusion Depth" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 451, + 482, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 451, + 482, + 559 + ], + "spans": [ + { + "bbox": [ + 130, + 451, + 482, + 559 + ], + "type": "text", + "content": "Apart from temporal augmentation, our employed VFI model can also provide other inter-frame information, such as optical flow and occlusion merge masks. This enables us to design a novel VFI-assisted multi-frame fusion module for multi-frame depth prediction, as demonstrated in Fig. 2. The fusion module consists of two parts, i.e., motion-aware feature alignment (MAFA) and occlusion-alarviated feature fusion (OAFF). Taking original triplet " + }, + { + "bbox": [ + 130, + 451, + 482, + 559 + ], + "type": "inline_equation", + "content": "\\{I_{t - 2},I_t,I_{t + 2}\\}" + }, + { + "bbox": [ + 130, + 451, + 482, + 559 + ], + "type": "text", + "content": " for example, we implement forward process of the VFI model to interpolate the middle position " + }, + { + "bbox": [ + 130, + 451, + 482, + 559 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 451, + 482, + 559 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 130, + 451, + 482, + 559 + ], + "type": "inline_equation", + "content": "\\{I_{t - 2},I_{t + 2}\\}" + }, + { + "bbox": [ + 130, + 451, + 482, + 559 + ], + "type": "text", + "content": " in Eq. (9), yielding intermediate flow " + }, + { + "bbox": [ + 130, + 451, + 482, + 559 + ], + "type": "inline_equation", + "content": "F_{t\\rightarrow (t - 2)},F_{t\\rightarrow (t + 2)}" + }, + { + "bbox": [ + 130, + 451, + 482, + 559 + ], + "type": "text", + "content": " and merge mask " + }, + { + "bbox": [ + 130, + 451, + 482, + 559 + ], + "type": "inline_equation", + "content": "M_{t}" + }, + { + "bbox": [ + 130, + 451, + 482, + 559 + ], + "type": "text", + "content": " to align and aggregate multi-frame features." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 222, + 565, + 480, + 578 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 222, + 565, + 480, + 578 + ], + "spans": [ + { + "bbox": [ + 222, + 565, + 480, + 578 + ], + "type": "interline_equation", + "content": "F _ {t \\rightarrow (t - 2)}, F _ {t \\rightarrow (t + 2)}, M _ {t} = \\mathcal {H} \\left(I _ {t - 2}, I _ {t + 2}\\right). \\tag {9}", + "image_path": "1d524654800247a8bb7bc5f2a38bbc01c50d58e76c72d5f0f02cdcf171ebf2f8.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 582, + 482, + 618 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 618 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 618 + ], + "type": "text", + "content": "Multi-frame Feature Extraction. Our fusion module accepts multi-frame features extracted by the same encoder, which is shared with the single-frame depth encoder " + }, + { + "bbox": [ + 130, + 582, + 482, + 618 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 130, + 582, + 482, + 618 + ], + "type": "text", + "content": ". The extraction process can be written as:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 205, + 623, + 480, + 636 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 623, + 480, + 636 + ], + "spans": [ + { + "bbox": [ + 205, + 623, + 480, + 636 + ], + "type": "interline_equation", + "content": "\\left\\{\\phi_ {t - 2} ^ {k}, \\phi_ {t} ^ {k}, \\phi_ {t + 2} ^ {k} \\right\\} _ {k = 1} ^ {N} = \\left\\{\\mathcal {E} \\left(I _ {t - 2}\\right), \\mathcal {E} \\left(I _ {t}\\right), \\mathcal {E} \\left(I _ {t + 2}\\right) \\right\\}, \\tag {10}", + "image_path": "b57bcd652a87bc3a2e0b9566f2c25a9adc913584faa97a042fe51c7dfd256690.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 640, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 640, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 640, + 482, + 665 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 640, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\{\\phi_{t - 2}^k,\\phi_t^k,\\phi_{t + 2}^k\\}_{k = 1}^N" + }, + { + "bbox": [ + 130, + 640, + 482, + 665 + ], + "type": "text", + "content": " are the multi-level features of the three frames and " + }, + { + "bbox": [ + 130, + 640, + 482, + 665 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 640, + 482, + 665 + ], + "type": "text", + "content": " is the number of feature levels." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "text", + "content": "J. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 115, + 482, + 245 + ], + "blocks": [ + { + "bbox": [ + 132, + 115, + 482, + 245 + ], + "lines": [ + { + "bbox": [ + 132, + 115, + 482, + 245 + ], + "spans": [ + { + "bbox": [ + 132, + 115, + 482, + 245 + ], + "type": "image", + "image_path": "5a8c29e39bf5ffda6cab944146caef336883dae0319f1957a820946db98e53dc.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 253, + 479, + 275 + ], + "lines": [ + { + "bbox": [ + 131, + 253, + 479, + 275 + ], + "spans": [ + { + "bbox": [ + 131, + 253, + 479, + 275 + ], + "type": "text", + "content": "Fig. 2: Our VFI-assisted multi-frame fusion depth model. Motion and occlusion provided by VFI can be used to align and fuse multi-frame features in an explicit manner." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 305, + 479, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 305, + 479, + 344 + ], + "spans": [ + { + "bbox": [ + 131, + 305, + 479, + 344 + ], + "type": "text", + "content": "Motion-aware Feature Alignment (MAFA). For the " + }, + { + "bbox": [ + 131, + 305, + 479, + 344 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 131, + 305, + 479, + 344 + ], + "type": "text", + "content": "-th level, we use bidirectional optical flow to align the neighbor features " + }, + { + "bbox": [ + 131, + 305, + 479, + 344 + ], + "type": "inline_equation", + "content": "\\phi_{t-2}^k" + }, + { + "bbox": [ + 131, + 305, + 479, + 344 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 305, + 479, + 344 + ], + "type": "inline_equation", + "content": "\\phi_{t+2}^k" + }, + { + "bbox": [ + 131, + 305, + 479, + 344 + ], + "type": "text", + "content": " to the target position " + }, + { + "bbox": [ + 131, + 305, + 479, + 344 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 131, + 305, + 479, + 344 + ], + "type": "text", + "content": ", generating " + }, + { + "bbox": [ + 131, + 305, + 479, + 344 + ], + "type": "inline_equation", + "content": "\\phi_{(t-2) \\to t}^k" + }, + { + "bbox": [ + 131, + 305, + 479, + 344 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 305, + 479, + 344 + ], + "type": "inline_equation", + "content": "\\phi_{(t+2) \\to t}^k" + }, + { + "bbox": [ + 131, + 305, + 479, + 344 + ], + "type": "text", + "content": ", respectively, formulated as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 238, + 348, + 480, + 381 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 238, + 348, + 480, + 381 + ], + "spans": [ + { + "bbox": [ + 238, + 348, + 480, + 381 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\phi_ {(t - 2) \\rightarrow t} ^ {k} = \\omega \\left(\\phi_ {t - 2} ^ {k}, F _ {t \\rightarrow (t - 2)} ^ {k}\\right), \\\\ \\phi_ {(t + 2) \\rightarrow t} ^ {k} = \\omega \\left(\\phi_ {t + 2} ^ {k}, F _ {t \\rightarrow (t + 2)} ^ {k}\\right), \\tag {11} \\\\ \\end{array}", + "image_path": "5190a24c8be96e02a2149337331d04a43721426f6bd9fe9efe3481c441a5263d.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 386, + 479, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 386, + 479, + 434 + ], + "spans": [ + { + "bbox": [ + 131, + 386, + 479, + 434 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 386, + 479, + 434 + ], + "type": "inline_equation", + "content": "F_{t\\rightarrow (t - 2)}^k" + }, + { + "bbox": [ + 131, + 386, + 479, + 434 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 386, + 479, + 434 + ], + "type": "inline_equation", + "content": "F_{t\\rightarrow (t + 2)}^k" + }, + { + "bbox": [ + 131, + 386, + 479, + 434 + ], + "type": "text", + "content": " denote the corresponding flows of the " + }, + { + "bbox": [ + 131, + 386, + 479, + 434 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 131, + 386, + 479, + 434 + ], + "type": "text", + "content": "-th level, downsampled from " + }, + { + "bbox": [ + 131, + 386, + 479, + 434 + ], + "type": "inline_equation", + "content": "F_{t\\rightarrow (t - 2)}" + }, + { + "bbox": [ + 131, + 386, + 479, + 434 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 386, + 479, + 434 + ], + "type": "inline_equation", + "content": "F_{t\\rightarrow (t + 2)}" + }, + { + "bbox": [ + 131, + 386, + 479, + 434 + ], + "type": "text", + "content": ", respectively. Inspired by recent 2D implicit neural representation [8], we also concatenate the aligned features with encoded optical flow, which can be denoted as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 211, + 438, + 480, + 487 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 211, + 438, + 480, + 487 + ], + "spans": [ + { + "bbox": [ + 211, + 438, + 480, + 487 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\varphi_ {(t - 2) \\rightarrow t} ^ {k} = \\operatorname {C o n c a t} \\left(\\phi_ {(t - 2) \\rightarrow t} ^ {k}, g \\left(F _ {t \\rightarrow (t - 2)} ^ {k}\\right)\\right), \\\\ \\varphi_ {(t + 2) \\rightarrow t} ^ {k} = \\operatorname {C o n c a t} \\left(\\phi_ {(t + 2) \\rightarrow t} ^ {k}, g \\left(F _ {t \\rightarrow (t + 2)} ^ {k}\\right)\\right), \\tag {12} \\\\ \\varphi_ {t} ^ {k} = \\operatorname {C o n c a t} \\left(\\phi_ {t} ^ {k}, g \\left(F _ {0} ^ {k}\\right)\\right), \\\\ \\end{array}", + "image_path": "e58e454e2c4dfc819c8379ebf2aab715ea084b549d79dce7c4376a05b78da1b4.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 494, + 479, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 494, + 479, + 553 + ], + "spans": [ + { + "bbox": [ + 131, + 494, + 479, + 553 + ], + "type": "text", + "content": "where Concat means feature concatenation, " + }, + { + "bbox": [ + 131, + 494, + 479, + 553 + ], + "type": "inline_equation", + "content": "F_0^k" + }, + { + "bbox": [ + 131, + 494, + 479, + 553 + ], + "type": "text", + "content": " denotes zero flow and " + }, + { + "bbox": [ + 131, + 494, + 479, + 553 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 131, + 494, + 479, + 553 + ], + "type": "text", + "content": " is fourier positional encoding [41,51] to transform optical flow to high-frequency signal. Actually, optical flow indicates the relative coordinate of a feature vector in the warped feature map w.r.t. the original one. When incorporating it, multi-frame model can adaptively adjust feature learning through motion-aware alignment." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 556, + 479, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 556, + 479, + 616 + ], + "spans": [ + { + "bbox": [ + 131, + 556, + 479, + 616 + ], + "type": "text", + "content": "Occlusion-alleviated Feature Fusion (OAFF). Due to the ubiquitous interframe occlusion, aggregating the aligned features directly through summation or concatenation can deteriorate the performance. Therefore, we use the " + }, + { + "bbox": [ + 131, + 556, + 479, + 616 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 131, + 556, + 479, + 616 + ], + "type": "text", + "content": "-th level merge mask " + }, + { + "bbox": [ + 131, + 556, + 479, + 616 + ], + "type": "inline_equation", + "content": "M_t^k" + }, + { + "bbox": [ + 131, + 556, + 479, + 616 + ], + "type": "text", + "content": " downsampled from " + }, + { + "bbox": [ + 131, + 556, + 479, + 616 + ], + "type": "inline_equation", + "content": "M_t" + }, + { + "bbox": [ + 131, + 556, + 479, + 616 + ], + "type": "text", + "content": " to mitigate this issue, just like the image merging process of VFI in Eq. (8):" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 206, + 620, + 480, + 635 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 206, + 620, + 480, + 635 + ], + "spans": [ + { + "bbox": [ + 206, + 620, + 480, + 635 + ], + "type": "interline_equation", + "content": "\\chi_ {t} ^ {k} = M _ {t} ^ {k} \\odot \\varphi_ {(t - 2) \\rightarrow t} ^ {k} + (1 - M _ {t} ^ {k}) \\odot \\varphi_ {(t + 2) \\rightarrow t} ^ {k}, \\tag {13}", + "image_path": "b7898f45771c28163c24dd751ba083f81e055c12ea926fd428af000bbae8bd88.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 641, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 479, + 665 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 641, + 479, + 665 + ], + "type": "inline_equation", + "content": "\\chi_t^k" + }, + { + "bbox": [ + 131, + 641, + 479, + 665 + ], + "type": "text", + "content": " is the occlusion-alleviated feature merged from the temporal positions " + }, + { + "bbox": [ + 131, + 641, + 479, + 665 + ], + "type": "inline_equation", + "content": "t - 2" + }, + { + "bbox": [ + 131, + 641, + 479, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 641, + 479, + 665 + ], + "type": "inline_equation", + "content": "t + 2" + }, + { + "bbox": [ + 131, + 641, + 479, + 665 + ], + "type": "text", + "content": ". It is then concatenated with the feature " + }, + { + "bbox": [ + 131, + 641, + 479, + 665 + ], + "type": "inline_equation", + "content": "\\varphi_t^k" + }, + { + "bbox": [ + 131, + 641, + 479, + 665 + ], + "type": "text", + "content": " in the position " + }, + { + "bbox": [ + 131, + 641, + 479, + 665 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 131, + 641, + 479, + 665 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 157, + 90, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 90, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 157, + 90, + 448, + 103 + ], + "type": "text", + "content": "Mono-ViFi: A Unified Framework for Self-supervised Monocular Depth" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 358, + 128 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 358, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 358, + 128 + ], + "type": "text", + "content": "And we can obtain the final feature after fusion by:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 236, + 133, + 481, + 147 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 236, + 133, + 481, + 147 + ], + "spans": [ + { + "bbox": [ + 236, + 133, + 481, + 147 + ], + "type": "interline_equation", + "content": "\\psi_ {t} ^ {k} = \\operatorname {C o n v} _ {1 \\times 1} \\left(\\operatorname {C o n c a t} \\left(\\varphi_ {t} ^ {k}, \\chi_ {t} ^ {k}\\right)\\right), \\tag {14}", + "image_path": "bf743360c85cfc00936b10f2401e70dfca4394b278b7507358a36a923a032c16.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 151, + 479, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 151, + 479, + 187 + ], + "spans": [ + { + "bbox": [ + 130, + 151, + 479, + 187 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 151, + 479, + 187 + ], + "type": "inline_equation", + "content": "\\mathrm{Conv}_{1\\times 1}" + }, + { + "bbox": [ + 130, + 151, + 479, + 187 + ], + "type": "text", + "content": " is a " + }, + { + "bbox": [ + 130, + 151, + 479, + 187 + ], + "type": "inline_equation", + "content": "1\\times 1" + }, + { + "bbox": [ + 130, + 151, + 479, + 187 + ], + "type": "text", + "content": " convolutional layer to transform the feature into the same channel as the original feature " + }, + { + "bbox": [ + 130, + 151, + 479, + 187 + ], + "type": "inline_equation", + "content": "\\phi_t^k" + }, + { + "bbox": [ + 130, + 151, + 479, + 187 + ], + "type": "text", + "content": ", enabling the multi-frame depth decoder to employ the same architecture as the single-frame one." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 189, + 480, + 271 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 189, + 480, + 271 + ], + "spans": [ + { + "bbox": [ + 130, + 189, + 480, + 271 + ], + "type": "text", + "content": "Multi-frame Depth Decoder. To translate the multi-level fusion features into a depth map, a decoder is required. Concretely, we use the same architecture as the single-frame depth decoder, but the parameters are not shared. This encourages the single- and multi-frame models to decode the features in their individual ways without conflict. Finally, the fusion features " + }, + { + "bbox": [ + 130, + 189, + 480, + 271 + ], + "type": "inline_equation", + "content": "\\{\\psi_t^k\\}_{k=1}^N" + }, + { + "bbox": [ + 130, + 189, + 480, + 271 + ], + "type": "text", + "content": " are processed by a multi-frame depth decoder " + }, + { + "bbox": [ + 130, + 189, + 480, + 271 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_m" + }, + { + "bbox": [ + 130, + 189, + 480, + 271 + ], + "type": "text", + "content": " to generate the fusion depth " + }, + { + "bbox": [ + 130, + 189, + 480, + 271 + ], + "type": "inline_equation", + "content": "D_t^m" + }, + { + "bbox": [ + 130, + 189, + 480, + 271 + ], + "type": "text", + "content": " of the target frame " + }, + { + "bbox": [ + 130, + 189, + 480, + 271 + ], + "type": "inline_equation", + "content": "I_t" + }, + { + "bbox": [ + 130, + 189, + 480, + 271 + ], + "type": "text", + "content": ", which can be written as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 261, + 277, + 481, + 291 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 261, + 277, + 481, + 291 + ], + "spans": [ + { + "bbox": [ + 261, + 277, + 481, + 291 + ], + "type": "interline_equation", + "content": "D _ {t} ^ {m} = \\mathcal {D} _ {m} \\left(\\left\\{\\psi_ {t} ^ {k} \\right\\} _ {k = 1} ^ {N}\\right). \\tag {15}", + "image_path": "e2a2e4d5125669594b35c648f9287a905e7ec4bc0a885c0b6db99bc36bf419ff.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 296, + 479, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 296, + 479, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 296, + 479, + 319 + ], + "type": "text", + "content": "Note that the multi-frame depth " + }, + { + "bbox": [ + 130, + 296, + 479, + 319 + ], + "type": "inline_equation", + "content": "D_{t}^{m}" + }, + { + "bbox": [ + 130, + 296, + 479, + 319 + ], + "type": "text", + "content": " also corresponds to a self-supervised loss like Eq. (5), which is denoted as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 244, + 326, + 481, + 338 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 244, + 326, + 481, + 338 + ], + "spans": [ + { + "bbox": [ + 244, + 326, + 481, + 338 + ], + "type": "interline_equation", + "content": "L _ {s s} ^ {m} (t) = \\mathcal {L} \\left(D _ {t} ^ {m}, \\boldsymbol {T} _ {t \\rightarrow s}, I _ {t}, I _ {s}\\right). \\tag {16}", + "image_path": "0f44a39ade6b092d7b356dbffe230d061b66afc1ebf7729a7101f7892f251b8a.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 361, + 271, + 374 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 361, + 271, + 374 + ], + "spans": [ + { + "bbox": [ + 132, + 361, + 271, + 374 + ], + "type": "text", + "content": "4.3 Spatial Augmentation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 388, + 482, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 388, + 482, + 594 + ], + "spans": [ + { + "bbox": [ + 130, + 388, + 482, + 594 + ], + "type": "text", + "content": "Orthogonal to temporal augmentation, we also include classic data augmentation to spatially expand the data diversity, which is categorized as spatial augmentation in this paper. Resizing-cropping is a commonly used and effective augmentation approach. Based on this operation, we introduce image rotation, extending it to a generic affine transformation (cropping is actually image translation). Specifically, we follow [55] to consider that the augmented image and the original image are captured by the same camera system whose intrinsics are fixed. For resizing operation, it is assumed that when an image is zoomed up by a factor " + }, + { + "bbox": [ + 130, + 388, + 482, + 594 + ], + "type": "inline_equation", + "content": "f_{s} \\geq 1" + }, + { + "bbox": [ + 130, + 388, + 482, + 594 + ], + "type": "text", + "content": ", the relative size of an object increases by the same factor " + }, + { + "bbox": [ + 130, + 388, + 482, + 594 + ], + "type": "inline_equation", + "content": "f_{s}" + }, + { + "bbox": [ + 130, + 388, + 482, + 594 + ], + "type": "text", + "content": " and its depth decreases by " + }, + { + "bbox": [ + 130, + 388, + 482, + 594 + ], + "type": "inline_equation", + "content": "f_{s}" + }, + { + "bbox": [ + 130, + 388, + 482, + 594 + ], + "type": "text", + "content": " simultaneously. This is inspired from a vital monocular cue [4] that the closer an object is, the larger its relative size is. Our affine transformation " + }, + { + "bbox": [ + 130, + 388, + 482, + 594 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 130, + 388, + 482, + 594 + ], + "type": "text", + "content": " is applied to both the target image " + }, + { + "bbox": [ + 130, + 388, + 482, + 594 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 130, + 388, + 482, + 594 + ], + "type": "text", + "content": " and the source image " + }, + { + "bbox": [ + 130, + 388, + 482, + 594 + ], + "type": "inline_equation", + "content": "I_{s}" + }, + { + "bbox": [ + 130, + 388, + 482, + 594 + ], + "type": "text", + "content": ", generating " + }, + { + "bbox": [ + 130, + 388, + 482, + 594 + ], + "type": "inline_equation", + "content": "\\widetilde{I}_{t} = \\mathcal{A}(I_{t})" + }, + { + "bbox": [ + 130, + 388, + 482, + 594 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 388, + 482, + 594 + ], + "type": "inline_equation", + "content": "\\widetilde{I}_{s} = \\mathcal{A}(I_{s})" + }, + { + "bbox": [ + 130, + 388, + 482, + 594 + ], + "type": "text", + "content": ", respectively. For the augmented target view, we can obtain its depth through the single-frame model by " + }, + { + "bbox": [ + 130, + 388, + 482, + 594 + ], + "type": "inline_equation", + "content": "\\widetilde{D}_{t} = \\mathcal{D}(\\mathcal{E}(\\widetilde{I}_{t}))" + }, + { + "bbox": [ + 130, + 388, + 482, + 594 + ], + "type": "text", + "content": ". To conduct view synthesis on augmented images, the original camera pose " + }, + { + "bbox": [ + 130, + 388, + 482, + 594 + ], + "type": "inline_equation", + "content": "T_{t\\rightarrow s} = [\\pmb{R}_{t\\rightarrow s}|\\pmb{t}_{t\\rightarrow s}]" + }, + { + "bbox": [ + 130, + 388, + 482, + 594 + ], + "type": "text", + "content": " from target to source should be rectified to " + }, + { + "bbox": [ + 130, + 388, + 482, + 594 + ], + "type": "inline_equation", + "content": "\\widetilde{T}_{t\\rightarrow s} = [\\widetilde{\\pmb{R}}_{t\\rightarrow s}|\\widetilde{\\pmb{t}}_{t\\rightarrow s}]" + }, + { + "bbox": [ + 130, + 388, + 482, + 594 + ], + "type": "text", + "content": " as done in [55]:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 217, + 599, + 481, + 613 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 217, + 599, + 481, + 613 + ], + "spans": [ + { + "bbox": [ + 217, + 599, + 481, + 613 + ], + "type": "interline_equation", + "content": "\\tilde {\\boldsymbol {R}} _ {t \\rightarrow s} = \\boldsymbol {R} _ {c} \\boldsymbol {R} _ {t \\rightarrow s} \\boldsymbol {R} _ {c} ^ {- 1}, \\quad \\tilde {\\boldsymbol {t}} _ {t \\rightarrow s} = \\boldsymbol {R} _ {c} \\boldsymbol {t} _ {t \\rightarrow s}, \\tag {17}", + "image_path": "1118e7402ab2b37f2e0f334f7bf08c4c20c54c04589c0c1a5082937894c97a85.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_c" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": " is the rectification matrix determined only by the camera intrinsics and the augmentation parameters. Different from [55] calculating " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_c" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": " under resizing-cropping, we give its formulation and derivation under a more general affine transformation, which can be seen in our supplementary. Therefore, we can derive" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "text", + "content": "J. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 115, + 467, + 129 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 115, + 467, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 115, + 467, + 129 + ], + "type": "text", + "content": "the self-supervised loss " + }, + { + "bbox": [ + 132, + 115, + 467, + 129 + ], + "type": "inline_equation", + "content": "\\widetilde{L}_{ss}(t)" + }, + { + "bbox": [ + 132, + 115, + 467, + 129 + ], + "type": "text", + "content": " on the augmented images similarly to Eq. (5):" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 246, + 133, + 480, + 148 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 246, + 133, + 480, + 148 + ], + "spans": [ + { + "bbox": [ + 246, + 133, + 480, + 148 + ], + "type": "interline_equation", + "content": "\\widetilde {L} _ {s s} (t) = \\mathcal {L} (\\widetilde {D} _ {t}, \\widetilde {\\boldsymbol {T}} _ {t \\rightarrow s}, \\widetilde {I} _ {t}, \\widetilde {I} _ {s}). \\tag {18}", + "image_path": "23888e97bf1d253f7cd68c19e88eb30da7d3b80300e877a7618e67ef105b8483.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 153, + 481, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 153, + 481, + 178 + ], + "spans": [ + { + "bbox": [ + 131, + 153, + 481, + 178 + ], + "type": "text", + "content": "Note that there is a mask to exclude the possible invalid pixels in " + }, + { + "bbox": [ + 131, + 153, + 481, + 178 + ], + "type": "inline_equation", + "content": "\\widetilde{I}_t" + }, + { + "bbox": [ + 131, + 153, + 481, + 178 + ], + "type": "text", + "content": " when calculating " + }, + { + "bbox": [ + 131, + 153, + 481, + 178 + ], + "type": "inline_equation", + "content": "\\widetilde{L}_{ss}(t)" + }, + { + "bbox": [ + 131, + 153, + 481, + 178 + ], + "type": "text", + "content": " due to image rotation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 193, + 294, + 205 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 193, + 294, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 193, + 294, + 205 + ], + "type": "text", + "content": "4.4 Triplet Depth Consistency" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 211, + 482, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 211, + 482, + 307 + ], + "spans": [ + { + "bbox": [ + 131, + 211, + 482, + 307 + ], + "type": "text", + "content": "Now we have single-frame depths " + }, + { + "bbox": [ + 131, + 211, + 482, + 307 + ], + "type": "inline_equation", + "content": "D_{t}" + }, + { + "bbox": [ + 131, + 211, + 482, + 307 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 211, + 482, + 307 + ], + "type": "inline_equation", + "content": "\\widetilde{D}_{t}" + }, + { + "bbox": [ + 131, + 211, + 482, + 307 + ], + "type": "text", + "content": " of the original image " + }, + { + "bbox": [ + 131, + 211, + 482, + 307 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 131, + 211, + 482, + 307 + ], + "type": "text", + "content": " and the augmented image " + }, + { + "bbox": [ + 131, + 211, + 482, + 307 + ], + "type": "inline_equation", + "content": "\\widetilde{I}_{t}" + }, + { + "bbox": [ + 131, + 211, + 482, + 307 + ], + "type": "text", + "content": ", along with multi-frame depth " + }, + { + "bbox": [ + 131, + 211, + 482, + 307 + ], + "type": "inline_equation", + "content": "D_{t}^{m}" + }, + { + "bbox": [ + 131, + 211, + 482, + 307 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 131, + 211, + 482, + 307 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 131, + 211, + 482, + 307 + ], + "type": "text", + "content": ". For better knowledge distillation and regularization, our unified learning framework applies a triplet depth consistency loss among the three depth maps, including a standard view depth consistency (SVDC) loss and two scale-aware depth consistency (SADC) losses, which are shown in Fig. 1(b). The SVDC loss " + }, + { + "bbox": [ + 131, + 211, + 482, + 307 + ], + "type": "inline_equation", + "content": "L_{sv}(t)" + }, + { + "bbox": [ + 131, + 211, + 482, + 307 + ], + "type": "text", + "content": " conducts mutual distillation between single- and multi-frame models, promoting their mutual reinforcement, which can be denoted as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 261, + 313, + 480, + 326 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 261, + 313, + 480, + 326 + ], + "spans": [ + { + "bbox": [ + 261, + 313, + 480, + 326 + ], + "type": "interline_equation", + "content": "L _ {s v} (t) = \\operatorname {S I} \\left(D _ {t} ^ {m}, D _ {t}\\right), \\tag {19}", + "image_path": "bcda4b18e4a22292d538d455d1a2db4e514766a7a0dd5bc5c4ae6be33bef8854.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 330, + 482, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 330, + 482, + 407 + ], + "spans": [ + { + "bbox": [ + 131, + 330, + 482, + 407 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 330, + 482, + 407 + ], + "type": "inline_equation", + "content": "\\mathrm{SI}()" + }, + { + "bbox": [ + 131, + 330, + 482, + 407 + ], + "type": "text", + "content": " is the widely used scale-invariant error function [13]. Besides, we use depth " + }, + { + "bbox": [ + 131, + 330, + 482, + 407 + ], + "type": "inline_equation", + "content": "\\widetilde{D}_t" + }, + { + "bbox": [ + 131, + 330, + 482, + 407 + ], + "type": "text", + "content": " under spatial augmentation to enforce two SADC losses, i.e., " + }, + { + "bbox": [ + 131, + 330, + 482, + 407 + ], + "type": "inline_equation", + "content": "L_{sa}(t)" + }, + { + "bbox": [ + 131, + 330, + 482, + 407 + ], + "type": "text", + "content": " between " + }, + { + "bbox": [ + 131, + 330, + 482, + 407 + ], + "type": "inline_equation", + "content": "\\widetilde{D}_t" + }, + { + "bbox": [ + 131, + 330, + 482, + 407 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 330, + 482, + 407 + ], + "type": "inline_equation", + "content": "D_t" + }, + { + "bbox": [ + 131, + 330, + 482, + 407 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 131, + 330, + 482, + 407 + ], + "type": "inline_equation", + "content": "L_{sa}^m(t)" + }, + { + "bbox": [ + 131, + 330, + 482, + 407 + ], + "type": "text", + "content": " between " + }, + { + "bbox": [ + 131, + 330, + 482, + 407 + ], + "type": "inline_equation", + "content": "\\widetilde{D}_t" + }, + { + "bbox": [ + 131, + 330, + 482, + 407 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 330, + 482, + 407 + ], + "type": "inline_equation", + "content": "D_t^m" + }, + { + "bbox": [ + 131, + 330, + 482, + 407 + ], + "type": "text", + "content": ". To this end, we first restore " + }, + { + "bbox": [ + 131, + 330, + 482, + 407 + ], + "type": "inline_equation", + "content": "\\widetilde{D}_t" + }, + { + "bbox": [ + 131, + 330, + 482, + 407 + ], + "type": "text", + "content": " to the original image view, generating " + }, + { + "bbox": [ + 131, + 330, + 482, + 407 + ], + "type": "inline_equation", + "content": "\\widehat{D}_t = \\mathcal{A}^{-1}(\\widetilde{D}_t)" + }, + { + "bbox": [ + 131, + 330, + 482, + 407 + ], + "type": "text", + "content": " in Fig. 1(b), where " + }, + { + "bbox": [ + 131, + 330, + 482, + 407 + ], + "type": "inline_equation", + "content": "\\mathcal{A}^{-1}" + }, + { + "bbox": [ + 131, + 330, + 482, + 407 + ], + "type": "text", + "content": " is the inverse transformation of affine augmentation " + }, + { + "bbox": [ + 131, + 330, + 482, + 407 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 131, + 330, + 482, + 407 + ], + "type": "text", + "content": ". To exclude the invalid pixels in " + }, + { + "bbox": [ + 131, + 330, + 482, + 407 + ], + "type": "inline_equation", + "content": "\\widehat{D}_t" + }, + { + "bbox": [ + 131, + 330, + 482, + 407 + ], + "type": "text", + "content": ", we adopt a mask " + }, + { + "bbox": [ + 131, + 330, + 482, + 407 + ], + "type": "inline_equation", + "content": "M_{sa}" + }, + { + "bbox": [ + 131, + 330, + 482, + 407 + ], + "type": "text", + "content": " and calculate " + }, + { + "bbox": [ + 131, + 330, + 482, + 407 + ], + "type": "inline_equation", + "content": "L_{sa}(t)" + }, + { + "bbox": [ + 131, + 330, + 482, + 407 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 330, + 482, + 407 + ], + "type": "inline_equation", + "content": "L_{sa}^m(t)" + }, + { + "bbox": [ + 131, + 330, + 482, + 407 + ], + "type": "text", + "content": " through:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 228, + 422, + 480, + 436 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 422, + 480, + 436 + ], + "spans": [ + { + "bbox": [ + 228, + 422, + 480, + 436 + ], + "type": "interline_equation", + "content": "L _ {s a} (t) = \\operatorname {S I} \\left(M _ {s a} \\odot D _ {t}, M _ {s a} \\odot f _ {s} \\widehat {D} _ {t}\\right), \\tag {20}", + "image_path": "e17d5d2d89520b539c8048d75439850d1aec1f531a2e0f8a6d99ae6726f0ff57.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 226, + 440, + 480, + 453 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 226, + 440, + 480, + 453 + ], + "spans": [ + { + "bbox": [ + 226, + 440, + 480, + 453 + ], + "type": "interline_equation", + "content": "L _ {s a} ^ {m} (t) = \\operatorname {S I} \\left(M _ {s a} \\odot D _ {t} ^ {m}, M _ {s a} \\odot f _ {s} \\widehat {D} _ {t}\\right), \\tag {21}", + "image_path": "b1dc20292b066c91b45cf8e448afac2eb97960bb27a193727bc93ae69b161a35.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 458, + 482, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 458, + 482, + 506 + ], + "spans": [ + { + "bbox": [ + 131, + 458, + 482, + 506 + ], + "type": "text", + "content": "where the factor " + }, + { + "bbox": [ + 131, + 458, + 482, + 506 + ], + "type": "inline_equation", + "content": "f_{s}" + }, + { + "bbox": [ + 131, + 458, + 482, + 506 + ], + "type": "text", + "content": " is due to the assumption in Sec. 4.3 that the depth of an object in " + }, + { + "bbox": [ + 131, + 458, + 482, + 506 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 131, + 458, + 482, + 506 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 131, + 458, + 482, + 506 + ], + "type": "inline_equation", + "content": "f_{s}" + }, + { + "bbox": [ + 131, + 458, + 482, + 506 + ], + "type": "text", + "content": " times of that in " + }, + { + "bbox": [ + 131, + 458, + 482, + 506 + ], + "type": "inline_equation", + "content": "\\widetilde{I}_{t}" + }, + { + "bbox": [ + 131, + 458, + 482, + 506 + ], + "type": "text", + "content": ". Our SADC constrains the 3D geometric relationship between the scenes of " + }, + { + "bbox": [ + 131, + 458, + 482, + 506 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 131, + 458, + 482, + 506 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 458, + 482, + 506 + ], + "type": "inline_equation", + "content": "\\widetilde{I}_{t}" + }, + { + "bbox": [ + 131, + 458, + 482, + 506 + ], + "type": "text", + "content": ", providing scale-aware regularization. Combining SVDC and SADC, we have the triplet depth consistency loss as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 234, + 512, + 480, + 525 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 234, + 512, + 480, + 525 + ], + "spans": [ + { + "bbox": [ + 234, + 512, + 480, + 525 + ], + "type": "interline_equation", + "content": "L _ {t c} (t) = L _ {s v} (t) + L _ {s a} (t) + L _ {s a} ^ {m} (t). \\tag {22}", + "image_path": "df2eee3762957a1b3a7073c8dd9eba98109290005a942a1e41d4e78a688fae02.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 539, + 256, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 539, + 256, + 552 + ], + "spans": [ + { + "bbox": [ + 132, + 539, + 256, + 552 + ], + "type": "text", + "content": "4.5 Training Objective" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 131, + 557, + 481, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 557, + 481, + 606 + ], + "spans": [ + { + "bbox": [ + 131, + 557, + 481, + 606 + ], + "type": "text", + "content": "For target position " + }, + { + "bbox": [ + 131, + 557, + 481, + 606 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 131, + 557, + 481, + 606 + ], + "type": "text", + "content": ", we have already obtained the self-supervised losses " + }, + { + "bbox": [ + 131, + 557, + 481, + 606 + ], + "type": "inline_equation", + "content": "L_{ss}(t)" + }, + { + "bbox": [ + 131, + 557, + 481, + 606 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 557, + 481, + 606 + ], + "type": "inline_equation", + "content": "L_{ss}^{m}(t)" + }, + { + "bbox": [ + 131, + 557, + 481, + 606 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 557, + 481, + 606 + ], + "type": "inline_equation", + "content": "\\widetilde{L}_{ss}(t)" + }, + { + "bbox": [ + 131, + 557, + 481, + 606 + ], + "type": "text", + "content": " in Eqs. (5), (16) and (18), respectively. Besides, our triplet depth consistency loss " + }, + { + "bbox": [ + 131, + 557, + 481, + 606 + ], + "type": "inline_equation", + "content": "L_{tc}(t)" + }, + { + "bbox": [ + 131, + 557, + 481, + 606 + ], + "type": "text", + "content": " is acquired in Eq. (22). Therefore, the total loss on position " + }, + { + "bbox": [ + 131, + 557, + 481, + 606 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 131, + 557, + 481, + 606 + ], + "type": "text", + "content": " can be summed as:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 217, + 610, + 480, + 624 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 217, + 610, + 480, + 624 + ], + "spans": [ + { + "bbox": [ + 217, + 610, + 480, + 624 + ], + "type": "interline_equation", + "content": "L (t) = L _ {s s} (t) + L _ {s s} ^ {m} (t) + \\widetilde {L} _ {s s} (t) + \\lambda L _ {t c} (t), \\tag {23}", + "image_path": "ed03ccfe1fb8d3a613895035d89477e3f3939011c0d8a9f38c10ec31f7f1e4ed.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 131, + 629, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 629, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 629, + 482, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 629, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 131, + 629, + 482, + 666 + ], + "type": "text", + "content": " is a hyper-parameter set to 0.2 in this paper. For " + }, + { + "bbox": [ + 131, + 629, + 482, + 666 + ], + "type": "inline_equation", + "content": "t - 1" + }, + { + "bbox": [ + 131, + 629, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 629, + 482, + 666 + ], + "type": "inline_equation", + "content": "t + 1" + }, + { + "bbox": [ + 131, + 629, + 482, + 666 + ], + "type": "text", + "content": ", we can derive their losses " + }, + { + "bbox": [ + 131, + 629, + 482, + 666 + ], + "type": "inline_equation", + "content": "L(t - 1)" + }, + { + "bbox": [ + 131, + 629, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 629, + 482, + 666 + ], + "type": "inline_equation", + "content": "L(t + 1)" + }, + { + "bbox": [ + 131, + 629, + 482, + 666 + ], + "type": "text", + "content": " in the same way. In summary, the overall loss function of our Mono-ViFI is " + }, + { + "bbox": [ + 131, + 629, + 482, + 666 + ], + "type": "inline_equation", + "content": "L = L(t - 1) + L(t) + L(t + 1)" + }, + { + "bbox": [ + 131, + 629, + 482, + 666 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 156, + 90, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 156, + 90, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 156, + 90, + 447, + 102 + ], + "type": "text", + "content": "Mono-ViFi: A Unified Framework for Self-supervised Monocular Depth" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 180 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 180 + ], + "type": "text", + "content": "Table 1: Performance comparison on KITTI [16] benchmark using the raw ground truth, with a resolution of " + }, + { + "bbox": [ + 130, + 114, + 482, + 180 + ], + "type": "inline_equation", + "content": "640 \\times 192" + }, + { + "bbox": [ + 130, + 114, + 482, + 180 + ], + "type": "text", + "content": ". The best results in each subsection are in bold; second best are underlined. For MACs of multi-frame models, we use the formulation A/B. A denotes the MACs when predicting for a single frame; B means the average MACs per frame when processing a long video sequence. #Params and MACs of our multi-frame models include the VFI network." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 135, + 198, + 482, + 426 + ], + "blocks": [ + { + "bbox": [ + 135, + 198, + 482, + 426 + ], + "lines": [ + { + "bbox": [ + 135, + 198, + 482, + 426 + ], + "spans": [ + { + "bbox": [ + 135, + 198, + 482, + 426 + ], + "type": "table", + "html": "
MethodTest frames#ParamsMACsAbs RelSq RelRMSERMSE logδ1δ2δ3
PackNet-SfM [21]1128M205G0.1110.7854.6010.1890.8780.9600.982
HR-Depth [39]116.9M31.6G0.1090.7924.6320.1850.8840.9620.983
R-MSFM6 [67]13.8M31.2G0.1120.8064.7040.1910.8780.9600.981
BRNet [22]119.1M33.7G0.1050.6984.4620.1790.8900.9650.984
MonoViT [64]127M-0.0990.7084.3720.1750.9000.9670.984
DIFFNet [65]112.2M23.4G0.1020.7644.4830.1800.8960.9650.983
SD-SSMDE [44]1--0.1000.6614.2640.1720.8960.9670.985
MUSTNet [48]1--0.1060.7634.5620.1820.8880.9630.983
DepthSegNet [40]130.3M-0.1040.6904.4730.1790.8860.9650.984
DualRefine [2]1--0.1030.7764.4910.1810.8940.9650.983
DaCNN [23]113M-0.0990.6614.3160.1730.8970.9670.985
Monodepth2 (ResNet18) [18]114.3M8G0.1150.9034.8630.1930.8770.9590.981
Mono-ViFI (ResNet18)114.3M8G0.1050.7084.4460.1790.8870.9650.984
Lite-Mono [62]13.1M5.1G0.1070.7654.5610.1830.8860.9630.983
Mono-ViFI (Lite-Mono)13.1M5.1G0.1030.6804.3490.1760.8910.9660.985
D-HRNet [25]110M10.8G0.1020.7604.4790.1790.8970.9650.983
RA-Depth (D-HRNet) [25]110M10.8G0.0960.6324.2160.1710.9030.9680.985
Mono-ViFI (D-HRNet)110M10.8G0.0960.6274.1790.1700.9030.9690.985
Patil et al. [42]N--/-0.1110.8214.6500.1870.8830.9610.982
ManyDepth (ResNet18) [58]2 (-1, 0)26.9M15.1G/13.7G0.0980.7704.4590.1760.9000.9650.983
DynamicDepth [14]2 (-1, 0)--/-0.0960.7204.4580.1750.8970.9640.984
DepthFormer [20]2 (-1, 0)28.7M175G/-0.0900.6614.1490.1750.9050.9670.984
MOVEDepth [56]2 (-1, 0)28.2M20.2G/19.2G0.0940.7044.3890.1750.9020.9650.983
DualRefine [2]2 (-1, 0)--/-0.0870.6984.2340.1700.9140.9670.983
TC-Depth [49]3 (-1, 0, 1)--/-0.1030.7464.4830.1850.894-0.983
Mono-ViFI (ResNet18)3 (-1, 0, 1)17.9M25.8G/16.9G0.0990.6614.3210.1740.8980.9660.985
Mono-ViFI (Lite-Mono)3 (-1, 0, 1)5.9M21.9G/13.3G0.0990.6484.2560.1720.8980.9670.985
Mono-ViFI (D-HRNet)3 (-1, 0, 1)12.9M35.7G/19.4G0.0910.5894.0880.1660.9120.9690.986
", + "image_path": "5be4c37c825e8a0cf365029b8fd0f56453e49802107cdf6d98fe3832c9b17e38.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 455, + 230, + 470 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 455, + 230, + 470 + ], + "spans": [ + { + "bbox": [ + 132, + 455, + 230, + 470 + ], + "type": "text", + "content": "5 Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 479, + 280, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 280, + 491 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 280, + 491 + ], + "type": "text", + "content": "5.1 Implementation Details" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 498, + 482, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 641 + ], + "type": "text", + "content": "Datasets. We train and evaluate the models on two widely used outdoor urban datasets of KITTI [16] and Cityscapes [10]. The depth labels of KITTI are collected with LiDARs. For fair comparison, we use the Eigen split [12] and follow the pre-processing operation in [18], generating 39,810 monocular triplets for training and 697 frames for evaluation. To reduce the influence of noise in the sparse depth from Velodyne, we also evaluate using " + }, + { + "bbox": [ + 130, + 498, + 482, + 641 + ], + "type": "inline_equation", + "content": "93\\%" + }, + { + "bbox": [ + 130, + 498, + 482, + 641 + ], + "type": "text", + "content": " of the Eigen split with the improved ground truth from [52], which contains 652 test frames. In regard to Cityscapes, following [58, 66], we preprocess the monocular sequences into triples, producing 69,731 training images. And we evaluate on 1,525 test images using the ground truth provided by [58]. For generalization comparison, KITTI models are evaluated on Cityscapes and Make3D [50]. Make3D is a small outdoor dataset containing 134 test images with aligned depth information." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "Flow-based VFI Model. We choose IFRNet [29] as our VFI model due to its efficiency. In practice, we use the large model of IFRNet for training and replace" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "type": "text", + "content": "J. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 115, + 242, + 242 + ], + "blocks": [ + { + "bbox": [ + 133, + 115, + 242, + 242 + ], + "lines": [ + { + "bbox": [ + 133, + 115, + 242, + 242 + ], + "spans": [ + { + "bbox": [ + 133, + 115, + 242, + 242 + ], + "type": "image", + "image_path": "99e79450c6426335bf4c0e9386f8b6e7c1850d1ad49ca9df9568c42be55984ec.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 251, + 480, + 274 + ], + "lines": [ + { + "bbox": [ + 131, + 251, + 480, + 274 + ], + "spans": [ + { + "bbox": [ + 131, + 251, + 480, + 274 + ], + "type": "text", + "content": "Fig. 3: Qualitative results on KITTI. Error maps in columns 2 and 4 show the Abs Rel error compared to the improved ground truth [52] from good (blue) to bad (red)." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 242, + 116, + 317, + 241 + ], + "blocks": [ + { + "bbox": [ + 242, + 116, + 317, + 241 + ], + "lines": [ + { + "bbox": [ + 242, + 116, + 317, + 241 + ], + "spans": [ + { + "bbox": [ + 242, + 116, + 317, + 241 + ], + "type": "image", + "image_path": "0da88bfe2f0d99496838e7378941a00260474aa275d552af0216217e93a6d7e8.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 318, + 116, + 399, + 241 + ], + "blocks": [ + { + "bbox": [ + 318, + 116, + 399, + 241 + ], + "lines": [ + { + "bbox": [ + 318, + 116, + 399, + 241 + ], + "spans": [ + { + "bbox": [ + 318, + 116, + 399, + 241 + ], + "type": "image", + "image_path": "3c94a3e3f9f98c4879c9225deb6ca36e4b8fe91f56f484a1376167f9532d1156.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 399, + 118, + 479, + 241 + ], + "blocks": [ + { + "bbox": [ + 399, + 118, + 479, + 241 + ], + "lines": [ + { + "bbox": [ + 399, + 118, + 479, + 241 + ], + "spans": [ + { + "bbox": [ + 399, + 118, + 479, + 241 + ], + "type": "image", + "image_path": "283d98377f6cc2f608e38498ec70a8b43581e3270bbefc364a23cbbba0a2309c.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 295, + 482, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 295, + 482, + 378 + ], + "spans": [ + { + "bbox": [ + 130, + 295, + 482, + 378 + ], + "type": "text", + "content": "it with the small model when evaluating for multi-frame fusion depth. Please see ablation about their gap in supplementary. Both large and small IFRNets are pretrained on KITTI/Cityscapes and frozen in Mono-ViFI training. The pretraining data split is the same as self-supervised depth training. Each original triplet in depth training is directly used for VFI training. Besides, we employ another two flow-based VFI networks, RIFE [26] and AMT [34], to verify the robustness of our method to different VFI models." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 380, + 482, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 380, + 482, + 452 + ], + "spans": [ + { + "bbox": [ + 130, + 380, + 482, + 452 + ], + "type": "text", + "content": "Depth and Pose Networks. We experiment with three depth architectures, including the ResNet18 [24] based network in Monodepth2 [18], medium LiteMono [62], and D-HRNet proposed in RA-Depth [25]. For the pose network, we follow Monodepth2 to use a ResNet18 based model which takes a pair of images (six channels) as input and predict a 6-DoF relative pose. Encoders of all the networks are pretrained on ImageNet [11]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 467, + 289, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 467, + 289, + 479 + ], + "spans": [ + { + "bbox": [ + 131, + 467, + 289, + 479 + ], + "type": "text", + "content": "5.2 Performance Comparison" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": "KITTI Results. Quantitative comparison results on KITTI under " + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "inline_equation", + "content": "640 \\times 192" + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": " resolution are listed in Table 1 and Table 2, using raw ground truth and improved ground truth, respectively. It is obvious that our method with D-HRNet can achieve state-of-the-art performance. Besides, our Mono-ViFI brings notable improvements to the three depth architectures without increasing model complexity (see each subsection separated by dotted lines). Despite that the difference between Mono-ViFI (D-HRNet) and RA-Depth [25] under single-frame mode is slight in Table 1, it is more significant in Table 2 when using improved ground truth. For multi-frame inference, Mono-ViFI with ResNet18 has already shown advantages over ManyDepth [58] and DynamicDepth [14] on some metrics including Sq Rel, RMSE, RMSE log, " + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\delta_{2}" + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\delta_{3}" + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": ". When changing to D-HRNet backbone, proposed Mono-ViFI can totally outperform them. Although our models have more multiply-add computations (MACs) when predicting for a single frame, the average MACs per frame are comparable when processing a long video sequence. Fig. 3 depicts two visual comparison results on KITTI, where" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 157, + 90, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 90, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 157, + 90, + 447, + 102 + ], + "type": "text", + "content": "Mono-ViFi: A Unified Framework for Self-supervised Monocular Depth" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 138, + 180, + 304, + 284 + ], + "blocks": [ + { + "bbox": [ + 133, + 137, + 479, + 171 + ], + "lines": [ + { + "bbox": [ + 133, + 137, + 479, + 171 + ], + "spans": [ + { + "bbox": [ + 133, + 137, + 479, + 171 + ], + "type": "text", + "content": "Table 2: Performance comparison on Table 3: Performance comparison on KITTI [16] using improved ground truth Cityscapes [10] with a resolution of " + }, + { + "bbox": [ + 133, + 137, + 479, + 171 + ], + "type": "inline_equation", + "content": "416 \\times" + }, + { + "bbox": [ + 133, + 137, + 479, + 171 + ], + "type": "text", + "content": " from [52], with a resolution of " + }, + { + "bbox": [ + 133, + 137, + 479, + 171 + ], + "type": "inline_equation", + "content": "640 \\times 192" + }, + { + "bbox": [ + 133, + 137, + 479, + 171 + ], + "type": "text", + "content": ". 128 unless stated." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 138, + 180, + 304, + 284 + ], + "lines": [ + { + "bbox": [ + 138, + 180, + 304, + 284 + ], + "spans": [ + { + "bbox": [ + 138, + 180, + 304, + 284 + ], + "type": "table", + "html": "
MethodAbs RelSq RelRMSERMSE logδ1δ2δ3
Single-frame Inference
PackNet-SIM [21]0.0780.4203.4850.1210.9310.9860.996
R-MSF M6 [67]0.0880.4923.8370.1350.9150.9830.995
DIFFNet [65]0.0760.4143.4050.1190.9360.9880.996
Mono-ViF [64]0.0740.3883.4140.1150.9380.9890.997
Monodepth2 (ResNet18) [18]0.0900.5453.9420.1370.9140.9830.995
Mono-ViF (ResNet18)0.0840.4313.6180.1260.9240.9860.997
Lite-Mono [62]0.0820.4553.6850.1270.9230.9850.996
Mono-ViF (Lite-Mono)0.0800.4003.4970.1210.9300.9870.997
D-HRNet [25]0.0770.4233.3960.1190.9350.9870.996
RA-Depth (D-HRNet) [25]0.0740.3633.3490.1140.9400.9900.997
Mono-ViF (D-HRNet)0.0720.3423.2540.1110.9410.9910.998
Multi-frame Inference
ManyDepth (ResNet18) [58]0.0700.3993.4550.1130.9410.9890.997
MOVEDepth [56]0.0650.3773.4490.1120.9420.9880.996
Mono-ViF (ResNet18)0.0740.3743.3810.1130.9410.9900.997
Mono-ViF (Lite-Mono)0.0740.3543.3720.1140.9390.9890.997
Mono-ViF (D-HRNet)0.0660.2943.0990.1020.9520.9920.998
", + "image_path": "435e95fff156baa792e511883d37b23e8e8a2f35072e483c0c45f816e4ef97a9.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 310, + 180, + 477, + 284 + ], + "blocks": [ + { + "bbox": [ + 310, + 180, + 477, + 284 + ], + "lines": [ + { + "bbox": [ + 310, + 180, + 477, + 284 + ], + "spans": [ + { + "bbox": [ + 310, + 180, + 477, + 284 + ], + "type": "table", + "html": "
MethodAbs RelSq RelRMSERMSE logδ1δ2δ3
Single-frame Inference
Videos in the wild [19]0.1271.3306.9600.1950.8300.9470.981
Li et al. [33]0.1191.2906.9800.1900.8460.9520.982
InstDM [31] (832 × 256)0.1111.1586.4370.1820.8680.9610.983
SD-SSMDE [44]0.1100.9885.9530.165---
Monodepth2 (ResNet18) [18]0.1291.5696.3760.1970.8490.9570.983
Mono-ViFi (ResNet18)0.1121.0035.8990.1650.8720.9660.990
Lite-Mono [62]0.1141.1946.3280.1720.8710.9670.989
Mono-ViFi (Lite-Mono)0.1070.9405.7940.1590.8760.9680.991
D-HRNet [25]0.1111.1426.2050.1690.8790.9680.989
Mono-ViFi (D-HRNet)0.1050.8805.5500.1560.8810.9690.992
Multi-frame Inference
ManyDepth (ResNet18) [58]0.1141.1936.2230.1700.8750.9670.989
DynamicDepth [14]0.1041.0115.9870.1590.8890.9720.991
TC-Depth [49]0.1100.9585.820-0.867-0.991
Mono-ViFi (ResNet18)0.1070.9345.6880.1590.8800.9690.990
Mono-ViFi (Lite-Mono)0.1050.9045.6460.1570.8800.9700.991
Mono-ViFi (D-HRNet)0.1020.8365.3950.1530.8850.9710.992
", + "image_path": "e9bc3f513579ccdbf6b94249042eb8622a3c7da720244114eb709a8573d7254c.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 160, + 349, + 455, + 415 + ], + "blocks": [ + { + "bbox": [ + 130, + 306, + 481, + 340 + ], + "lines": [ + { + "bbox": [ + 130, + 306, + 481, + 340 + ], + "spans": [ + { + "bbox": [ + 130, + 306, + 481, + 340 + ], + "type": "text", + "content": "Table 4: Evaluation results on dynamic objects and static regions of Cityscapes [10] dataset under " + }, + { + "bbox": [ + 130, + 306, + 481, + 340 + ], + "type": "inline_equation", + "content": "416\\times 128" + }, + { + "bbox": [ + 130, + 306, + 481, + 340 + ], + "type": "text", + "content": " resolution. Note that DynamicDepth relies on depth prior and segmentation masks of dynamic objects when reasoning." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 160, + 349, + 455, + 415 + ], + "lines": [ + { + "bbox": [ + 160, + 349, + 455, + 415 + ], + "spans": [ + { + "bbox": [ + 160, + 349, + 455, + 415 + ], + "type": "table", + "html": "
MethodTest AreaAbs RelSq RelRMSERMSE logδ1δ2δ3
Multi-frame Inference
ManyDepth (ResNet18) [58]DynamicDepth [14]Mono-ViFI (ResNet18)Dynamic Objects0.1722.2065.9240.2160.7920.9170.967
0.1441.5545.2380.1850.8310.9500.982
0.1621.4535.1140.1990.7940.9240.977
ManyDepth (ResNet18) [58]DynamicDepth [14]Mono-ViFI (ResNet18)Static Regions0.1071.0906.3140.1600.8820.9720.991
0.1000.9776.1550.1550.8920.9740.992
0.0990.9075.8350.1490.8900.9750.992
", + "image_path": "20f3189acd7c8cdae6073b6e3aed93ff243e1048e4d833e70a54d3eb35192845.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 449, + 480, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 480, + 509 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 480, + 509 + ], + "type": "text", + "content": "we choose D-HRNet as the baseline depth model. It can be seen that depth maps look qualitatively similar among all four methods. However, the error maps reveal mistakes of D-HRNet baseline and RA-Depth in some ambiguous regions like bushes (see the red boxes). In contrast, our Mono-ViFi have dramatically lower error in these regions, especially under multi-frame reasoning." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": "Cityscapes Results. Results of Cityscapes in Table 3 demonstrate the same trend as KITTI. Moreover, proposed multi-frame model with ResNet18 can beat ManyDepth on all metrics, while our Mono-ViFI is better than DynamicDepth only on partial indexes. It is because Cityscapes with more dynamic scenes is more challenging than KITTI, and DynamicDepth leverages additional information to cope with dynamic objects. For fine-grained analysis, we separately evaluate on the dynamic objects and static regions, using the segmentation masks of dynamic objects provided in [14], as listed in Table 4. On dynamic objects, Mono-ViFI with ResNet18 surpasses ManyDepth by a large margin, demonstrating that our multi-frame inference paradigm can sidestep the issue of dynamic objects to a greater extent compared to ManyDepth. DynamicDepth utilizes depth prior and segmentation masks of dynamic objects to tackle this problem for multi-frame inference. Even though, our method still excels it on Sq Rel and" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "type": "text", + "content": "J. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 156, + 116, + 459, + 244 + ], + "blocks": [ + { + "bbox": [ + 156, + 116, + 459, + 244 + ], + "lines": [ + { + "bbox": [ + 156, + 116, + 459, + 244 + ], + "spans": [ + { + "bbox": [ + 156, + 116, + 459, + 244 + ], + "type": "image", + "image_path": "b055474670ed0ea897e8507ce5a1717e48236a09eacf6bed41360cfcb8c2708b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 253, + 480, + 275 + ], + "lines": [ + { + "bbox": [ + 131, + 253, + 480, + 275 + ], + "spans": [ + { + "bbox": [ + 131, + 253, + 480, + 275 + ], + "type": "text", + "content": "Fig. 4: Qualitative generalization results of a YouTube video example. All depth maps are predicted under " + }, + { + "bbox": [ + 131, + 253, + 480, + 275 + ], + "type": "inline_equation", + "content": "640 \\times 360" + }, + { + "bbox": [ + 131, + 253, + 480, + 275 + ], + "type": "text", + "content": " resolution using KITTI models trained under " + }, + { + "bbox": [ + 131, + 253, + 480, + 275 + ], + "type": "inline_equation", + "content": "640 \\times 192" + }, + { + "bbox": [ + 131, + 253, + 480, + 275 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 161, + 342, + 455, + 471 + ], + "blocks": [ + { + "bbox": [ + 131, + 287, + 480, + 331 + ], + "lines": [ + { + "bbox": [ + 131, + 287, + 480, + 331 + ], + "spans": [ + { + "bbox": [ + 131, + 287, + 480, + 331 + ], + "type": "text", + "content": "Table 5: Ablation results of each proposed component on Cityscapes [10]. SA is spatial augmentation. TA is temporal augmentation. SVDC denotes standard view depth consistency. SADC denotes scale-aware depth consistency. FA is feature alignment. MAFA is motion-aware feature alignment. OAFF means occlusion-alleviated feature fusion." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 161, + 342, + 455, + 471 + ], + "lines": [ + { + "bbox": [ + 161, + 342, + 455, + 471 + ], + "spans": [ + { + "bbox": [ + 161, + 342, + 455, + 471 + ], + "type": "table", + "html": "
MethodSATASVDCSADCFAMAFAOAFFAbs RelSq RelRMSEδ1
Single-frame Inference
Baseline (Monodepth2 [18])0.1291.5696.8760.849
Baseline + TA✓0.1191.2546.5230.857
Baseline + SVDC✓0.1171.1866.4920.862
Baseline + SA✓0.1221.3866.5570.856
Baseline + SA + SADC✓✓0.1161.0796.1890.868
Mono-ViFI (Single-frame)✓✓✓✓0.1121.0035.8990.872
Multi-frame Inference
Baseline (direct feature fusion)0.1201.3606.4910.869
FA (not motion-aware)✓0.1161.2176.3870.871
MAFA✓0.1141.1506.1150.874
MAFA + OAFF✓✓0.1111.1076.0120.876
MAFA + OAFF + TA✓✓✓0.1091.0235.8920.877
Mono-ViFI (Multi-frame)✓✓✓✓✓✓0.1070.9345.6880.880
", + "image_path": "473ae66e2f84105a645934f03f578f72612f2632967d3f61d15e1e3511764b58.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 505, + 480, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 505, + 480, + 528 + ], + "spans": [ + { + "bbox": [ + 130, + 505, + 480, + 528 + ], + "type": "text", + "content": "RMSE. When evaluating on static regions, proposed Mono-ViFI shows absolute superiority over ManyDepth and DynamicDepth." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 529, + 481, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 529, + 481, + 601 + ], + "spans": [ + { + "bbox": [ + 130, + 529, + 481, + 601 + ], + "type": "text", + "content": "Generalization Results. We evaluate the KITTI models on Cityscapes and Make3D without fine-tuning, results of which are put in the supplementary. For qualitative comparison, we give a visualization example in Fig. 4, where KITTI models are used to predict the multi-frame depth of a YouTube video frame downloaded from the Wind Walk Travel Videos channel. This indicates that our Mono-ViFI has better generalization ability than existing methods." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 620, + 238, + 632 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 620, + 238, + 632 + ], + "spans": [ + { + "bbox": [ + 131, + 620, + 238, + 632 + ], + "type": "text", + "content": "5.3 Ablation Study" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "type": "text", + "content": "In this part, we explore the influence of each component in Mono-ViFi. Our ablations are conducted on Cityscapes or KITTI with ResNet18 backbone." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 157, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 157, + 91, + 447, + 102 + ], + "type": "text", + "content": "Mono-ViFi: A Unified Framework for Self-supervised Monocular Depth" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 187, + 136, + 430, + 189 + ], + "blocks": [ + { + "bbox": [ + 165, + 114, + 447, + 126 + ], + "lines": [ + { + "bbox": [ + 165, + 114, + 447, + 126 + ], + "spans": [ + { + "bbox": [ + 165, + 114, + 447, + 126 + ], + "type": "text", + "content": "Table 6: Ablation of different flow-based VFI models on KITTI [16]." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 187, + 136, + 430, + 189 + ], + "lines": [ + { + "bbox": [ + 187, + 136, + 430, + 189 + ], + "spans": [ + { + "bbox": [ + 187, + 136, + 430, + 189 + ], + "type": "table", + "html": "
VFI ModelAbs RelSq RelRMSERMSE logδ1δ2δ3
Multi-frame Inference
IFRNet [29] (Ours)0.0990.6614.3210.1740.8980.9660.985
RIFE [26]0.1000.6724.3520.1750.8960.9650.985
AMT [34]0.0990.6534.3350.1740.8970.9660.985
", + "image_path": "93df679f27216709e4f6f97dd794cad2c50611cd94ee31e2e2b70395ae90721c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 219, + 481, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 219, + 481, + 338 + ], + "spans": [ + { + "bbox": [ + 130, + 219, + 481, + 338 + ], + "type": "text", + "content": "Single-frame Ablation. We first ablate on the single-frame model. As listed in the top half part of Table 5, the standard view depth consistency and spatial augmentation can bring gains individually. Note that scale-aware depth consistency " + }, + { + "bbox": [ + 130, + 219, + 481, + 338 + ], + "type": "inline_equation", + "content": "L_{sa}" + }, + { + "bbox": [ + 130, + 219, + 481, + 338 + ], + "type": "text", + "content": " should be based on spatial augmentation. When adding " + }, + { + "bbox": [ + 130, + 219, + 481, + 338 + ], + "type": "inline_equation", + "content": "L_{sa}" + }, + { + "bbox": [ + 130, + 219, + 481, + 338 + ], + "type": "text", + "content": " to spatial augmentation, the accuracy can be further improved. And this leads to the most significant enhancement among all components, since " + }, + { + "bbox": [ + 130, + 219, + 481, + 338 + ], + "type": "inline_equation", + "content": "L_{sa}" + }, + { + "bbox": [ + 130, + 219, + 481, + 338 + ], + "type": "text", + "content": " can enforce the model to learn scale relationship from scene variations. In orthogonal dimension, our proposed temporal augmentation can also reduce the prediction error by providing more guidance from newly synthesized camera views. When integrating all above elements, our single-frame model can achieve the best performance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 338, + 481, + 457 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 338, + 481, + 457 + ], + "spans": [ + { + "bbox": [ + 130, + 338, + 481, + 457 + ], + "type": "text", + "content": "Multi-frame Ablation. To probe the effect of each component in our VFI-assisted multi-frame fusion module, we conduct ablation experiments under multi-frame inference, whose results are summarized in the bottom half part of Table 5. The baseline is stacking the raw feature of three frames directly, i.e., direct feature fusion. Then, we gradually apply components of the proposed fusion module, including motion-aware feature alignment and occlusion-alleviated feature fusion. The results demonstrate that these strategies can bring cumulative improvements. Furthermore, when incorporating spatial augmentation, temporal augmentation and relevant depth consistency losses, our multi-frame model can be further promoted like the single-frame setting." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 458, + 481, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 458, + 481, + 517 + ], + "spans": [ + { + "bbox": [ + 130, + 458, + 481, + 517 + ], + "type": "text", + "content": "Ablation of different VFI models. Apart from IFRNet [29], we additionally experiment with another two flow-based VFI models, RIFE [26] and AMT [34]. The results are depicted in Table 6. It can be observed that there is only a negligible performance difference among the three VFI networks, which reveals that our method is robust to different VFI models." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 534, + 220, + 547 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 220, + 547 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 220, + 547 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 558, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 481, + 665 + ], + "type": "text", + "content": "In conclusion, this paper presents a novel self-supervised learning framework, named Mono-ViFI, to bilaterally connect single- and multi-frame depth estimation in a unified manner. First, a flow-based VFI model is employed for temporal augmentation. Then, we propose a VFI-assisted multi-frame fusion module for more accurate multi-frame depth. Besides, resizing-cropping is extended to affine transformation as spatial augmentation for enriching data diversity. Finally, we develop a new triplet depth consistency loss for better regularization. Experiments show that our approach can unleash the potential of existing depth models to a greater extent, but without increasing inference complexity." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "type": "text", + "content": "J. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 140, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 138, + 140, + 480, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 140, + 480, + 161 + ], + "spans": [ + { + "bbox": [ + 138, + 140, + 480, + 161 + ], + "type": "text", + "content": "1. Bae, J., Moon, S., Im, S.: Deep digging into the generalization of self-supervised monocular depth estimation. In: AAAI. pp. 187-196 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 162, + 480, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 162, + 480, + 194 + ], + "spans": [ + { + "bbox": [ + 138, + 162, + 480, + 194 + ], + "type": "text", + "content": "2. Bangunharcana, A., Magd, A., Kim, K.S.: Dualrefine: Self-supervised depth and pose estimation through iterative epipolar sampling and refinement toward equilibrium. In: CVPR. pp. 726-738 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 195, + 480, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 195, + 480, + 216 + ], + "spans": [ + { + "bbox": [ + 138, + 195, + 480, + 216 + ], + "type": "text", + "content": "3. Bello, J.L.G., Kim, M.: Forget about the lidar: Self-supervised depth estimators with med probability volumes. In: NeurIPS. pp. 12626-12637 (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 217, + 480, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 217, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 138, + 217, + 480, + 237 + ], + "type": "text", + "content": "4. Bello, J.L.G., Moon, J., Kim, M.: Positional information is all you need: A novel pipeline for self-supervised svde from videos. arXiv:2205.08851 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 239, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 239, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 138, + 239, + 480, + 270 + ], + "type": "text", + "content": "5. Bian, J., Li, Z., Wang, N., Zhan, H., Shen, C., Cheng, M.M., Reid, I.: Unsupervised scale-consistent depth and ego-motion learning from monocular video. In: NeurIPS. pp. 35-45 (2019)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 271, + 480, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 271, + 480, + 303 + ], + "spans": [ + { + "bbox": [ + 138, + 271, + 480, + 303 + ], + "type": "text", + "content": "6. Chan, K.C., Wang, X., Yu, K., Dong, C., Loy, C.C.: Basicvsr: The search for essential components in video super-resolution and beyond. In: CVPR. pp. 4945-4954 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 304, + 480, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 304, + 480, + 336 + ], + "spans": [ + { + "bbox": [ + 138, + 304, + 480, + 336 + ], + "type": "text", + "content": "7. Chan, K.C., Zhou, S., Xu, X., Loy, C.C.: Basicvsr++: Improving video superresolution with enhanced propagation and alignment. In: CVPR. pp. 5962-5971 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 337, + 480, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 337, + 480, + 358 + ], + "spans": [ + { + "bbox": [ + 138, + 337, + 480, + 358 + ], + "type": "text", + "content": "8. Chen, Y., Liu, S., Wang, X.: Learning continuous image representation with local implicit image function. In: CVPR. pp. 8624-8634 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 360, + 480, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 360, + 480, + 391 + ], + "spans": [ + { + "bbox": [ + 138, + 360, + 480, + 391 + ], + "type": "text", + "content": "9. Chen, Y., Schmid, C., Sminchisescu, C.: Self-supervised learning with geometric constraints in monocular video: Connecting flow, depth, and camera. In: ICCV. pp. 7063-7072 (2019)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 392, + 480, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 392, + 480, + 424 + ], + "spans": [ + { + "bbox": [ + 138, + 392, + 480, + 424 + ], + "type": "text", + "content": "10. Cordts, M., Omran, M., Ramos, S., Rehfeld, T., Enzweiler, M., Benenson, R., Franke, U., Roth, S., Schiele, B.: The cityscapes dataset for semantic urban scene understanding. In: CVPR. pp. 3213-3223 (2016)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 425, + 480, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 425, + 480, + 446 + ], + "spans": [ + { + "bbox": [ + 138, + 425, + 480, + 446 + ], + "type": "text", + "content": "1. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: CVPR. pp. 248-255 (2009)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 447, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 447, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 138, + 447, + 480, + 468 + ], + "type": "text", + "content": "2. Eigen, D., Fergus, R.: Predicting depth, surface normals and semantic labels with a common multi-scale convolutional architecture. In: ICCV. pp. 2650-2658 (2015)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 469, + 480, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 469, + 480, + 490 + ], + "spans": [ + { + "bbox": [ + 138, + 469, + 480, + 490 + ], + "type": "text", + "content": "3. Eigen, D., Puhrsch, C., Fergus, R.: Depth map prediction from a single image using a multi-scale deep network. In: NeurIPS. pp. 2366-2374 (2014)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 491, + 480, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 491, + 480, + 522 + ], + "spans": [ + { + "bbox": [ + 138, + 491, + 480, + 522 + ], + "type": "text", + "content": "4. Feng, Z., Yang, L., Jing, L., Wang, H., Tian, Y., Li, B.: Disentangling object motion and occlusion for unsupervised multi-frame monocular depth. In: ECCV. pp. 228-244 (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 523, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 523, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 138, + 523, + 480, + 544 + ], + "type": "text", + "content": "5. Garg, R., Kumar, B.G.V., Carneiro, G., Reid, I.: Unsupervised cnn for single view depth estimation: Geometry to the rescue. In: ECCV. pp. 740-756 (2016)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 545, + 480, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 545, + 480, + 567 + ], + "spans": [ + { + "bbox": [ + 138, + 545, + 480, + 567 + ], + "type": "text", + "content": "6. Geiger, A., Lenz, P., Stiller, C., Urtasun, R.: Vision meets robotics: The kitti dataset. *IJRR* **32**(11), 1231–1237 (2013)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 567, + 480, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 567, + 480, + 589 + ], + "spans": [ + { + "bbox": [ + 138, + 567, + 480, + 589 + ], + "type": "text", + "content": "7. Godard, C., Aodha, O.M., Brostow, G.J.: Unsupervised monocular depth estimation with left-right consistency. In: CVPR. pp. 6602-6611 (2017)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 590, + 480, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 590, + 480, + 610 + ], + "spans": [ + { + "bbox": [ + 138, + 590, + 480, + 610 + ], + "type": "text", + "content": "8. Godard, C., Aodha, O.M., Firman, M., Brostow, G.J.: Digging into self-supervised monocular depth estimation. In: CVPR. pp. 3828-3838 (2019)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 138, + 611, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 611, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 138, + 611, + 480, + 643 + ], + "type": "text", + "content": "9. Gordon, A., Li, H., Jonschkowski, R., Angelova, A.: Depth from videos in the wild: Unsupervised monocular depth learning from unknown cameras. In: ICCV. pp. 8976-8985 (2019)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "text", + "content": "20. Guizilini, V., Ambrus, R., Chen, D., Zakharov, S., Gaidon, A.: Multi-frame self-supervised depth with transformers. In: CVPR. pp. 160-170 (2022)" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 157, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 157, + 91, + 447, + 102 + ], + "type": "text", + "content": "Mono-ViFi: A Unified Framework for Self-supervised Monocular Depth" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 666 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "text", + "content": "21. Guizilini, V., Ambrus, R., Pillai, S., Gaidon, A.: 3d packing for self-supervised monocular depth estimation. In: CVPR. pp. 2482-2491 (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 480, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 480, + 160 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 480, + 160 + ], + "type": "text", + "content": "22. Han, W., Yin, J., Jin, X., Dai, X., Shen, J.: Brnet: Exploring comprehensive features for monocular depth estimation. In: ECCV. pp. 586-602 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 161, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 161, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 133, + 161, + 480, + 182 + ], + "type": "text", + "content": "23. Han, W., Yin, J., Shen, J.: Self-supervised monocular depth estimation by direction-aware cumulative convolution network. In: ICCV (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 183, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 183, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 133, + 183, + 480, + 205 + ], + "type": "text", + "content": "24. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR. pp. 770-778 (2016)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 205, + 480, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 480, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 480, + 226 + ], + "type": "text", + "content": "25. He, M., Hui, L., Bian, Y., Ren, J., Xie, J., Yang, J.: Ra-depth: Resolution adaptive self-supervised monocular depth estimation. In: ECCV. pp. 565-581 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 227, + 480, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 480, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 480, + 248 + ], + "type": "text", + "content": "26. Huang, Z., Zhang, T., Heng, W., Shi, B., Zhou, S.: Real-time intermediate flow estimation for video frame interpolation. In: ECCV. pp. 624-642 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 249, + 480, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 480, + 280 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 480, + 280 + ], + "type": "text", + "content": "27. Jung, H., Park, E., Yoo, S.: Fine-grained semantics-aware representation enhancement for self-supervised monocular depth estimation. In: ICCV. pp. 12622-12632 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 281, + 480, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 281, + 480, + 314 + ], + "spans": [ + { + "bbox": [ + 132, + 281, + 480, + 314 + ], + "type": "text", + "content": "28. Klingner, M., Termöhlen, J.A., Mikolajczyk, J., Fingscheidt, T.: Self-supervised monocular depth estimation: Solving the dynamic object problem by semantic guidance. In: ECCV. pp. 582-600 (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 315, + 480, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 480, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 480, + 346 + ], + "type": "text", + "content": "29. Kong, L., Jiang, B., Luo, D., Chu, W., Huang, X., Tai, Y., Wang, C., Yang, J.: Ifrnet: Intermediate feature refine network for efficient frame interpolation. In: CVPR. pp. 1959-1968 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 347, + 480, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 480, + 369 + ], + "type": "text", + "content": "30. Kuznietsov, Y., Proesmans, M., Gool, L.V.: Comoda: Continuous monocular depth adaptation using past experiences. In: WACV. pp. 2906-2916 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 369, + 480, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 480, + 391 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 480, + 391 + ], + "type": "text", + "content": "31. Lee, S., Im, S., Lin, S., Kweon, I.S.: Learning monocular depth in dynamic scenes via instance-aware projection consistency. In: AAAI. pp. 1863-1872 (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 392, + 480, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 480, + 413 + ], + "type": "text", + "content": "32. Lee, S., Rameau, F., Pan, F., Kweon, I.S.: Attentive and contrastive learning for joint depth and motion field estimation. In: ICCV. pp. 4862-4871 (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 414, + 480, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 414, + 480, + 434 + ], + "spans": [ + { + "bbox": [ + 132, + 414, + 480, + 434 + ], + "type": "text", + "content": "33. Li, H., Gordon, A., Zhao, H., Casser, V., Angelova, A.: Unsupervised monocular depth learning in dynamic scenes. In: CoRL. pp. 1908-1917 (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 435, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 480, + 468 + ], + "type": "text", + "content": "34. Li, Z., Zhu, Z.L., Han, L.H., Hou, Q., Guo, C.L., Cheng, M.M.: Amt: All-pairs multi-field transforms for efficient frame interpolation. In: CVPR. pp. 9801-9810 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 468, + 480, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 480, + 490 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 480, + 490 + ], + "type": "text", + "content": "35. Liu, J., Kong, L., Yang, J.: Designing and searching for lightweight monocular depth network. In: ICONIP. pp. 477-488 (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 491, + 480, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 491, + 480, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 491, + 480, + 511 + ], + "type": "text", + "content": "36. Liu, J., Kong, L., Yang, J.: Atca: An arc trajectory based model with curvature attention for video frame interpolation. In: ICIP. pp. 1486-1490 (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 512, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 480, + 544 + ], + "type": "text", + "content": "37. Liu, J., Kong, L., Yang, J., Liu, W.: Towards better data exploitation in self-supervised monocular depth estimation. IEEE Robotics and Automation Letters 9(1), 763-770 (2024)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 545, + 480, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 545, + 480, + 567 + ], + "spans": [ + { + "bbox": [ + 132, + 545, + 480, + 567 + ], + "type": "text", + "content": "38. Luo, X., Huang, J.B., Szeliski, R., Matzen, K., Kopf, J.: Consistent video depth estimation. TOG 39(71), 1-13 (2020)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 567, + 480, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 480, + 599 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 480, + 599 + ], + "type": "text", + "content": "39. Lyu, X., Liu, L., Wang, M., Kong, X., Liu, L., Liu, Y., Chen, X., Yuan, Y.: Hrdepth: High resolution self-supervised monocular depth estimation. In: AAAI. pp. 2294-2301 (2021)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "type": "text", + "content": "40. Ma, J., Lei, X., Liu, N., Zhao, X., Pu, S.: Towards comprehensive representation enhancement in semantics-guided self-supervised monocular depth estimation. In: ECCV. pp. 304-321 (2022)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 633, + 480, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 633, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 633, + 480, + 666 + ], + "type": "text", + "content": "41. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: ECCV. pp. 405-421 (2020)" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "type": "text", + "content": "J. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 481, + 149 + ], + "type": "text", + "content": "42. Patil, V., Gansbeke, W.V., Dai, D., Gool, L.V.: Don't forget the past: Recurrent depth estimation from monocular video. IEEE Robotics and Automation Letters 5(4), 6813-6820 (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 481, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 481, + 172 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 481, + 172 + ], + "type": "text", + "content": "43. Peng, R., Wang, R., Lai, Y., Tang, L., Cai, Y.: Excavating the potential capacity of selfsupervised monocular depth estimation. In: ICCV. pp. 15560-15569 (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 481, + 205 + ], + "type": "text", + "content": "44. Petrovai, A., Nedevschi, S.: Exploiting pseudo labels in a self-supervised learning framework for improved monocular depth estimation. In: CVPR. pp. 1568-1578 (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 205, + 481, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 481, + 237 + ], + "type": "text", + "content": "45. Pilzer, A., Lathuilière, S., Sebe, N., Ricci, E.: Refine and distill: Exploiting cycle-inconsistency and knowledge distillation for unsupervised monocular depth estimation. In: CVPR. pp. 9760-9769 (2019)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 237, + 481, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 237, + 481, + 259 + ], + "spans": [ + { + "bbox": [ + 132, + 237, + 481, + 259 + ], + "type": "text", + "content": "46. Poggi, M., Aleotti, F., Tosi, F., Mattoccia, S.: On the uncertainty of self-supervised monocular depth estimation. In: CVPR. pp. 3224-3234 (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 259, + 481, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 481, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 481, + 293 + ], + "type": "text", + "content": "47. Ranjan, A., Jampani, V., Balles, L., Kim, K., Sun, D., Wulff, J., Black, M.J.: Competitive collaboration: Joint unsupervised learning of depth, camera motion, optical flow and motion segmentation. In: CVPR. pp. 12240-12249 (2019)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 293, + 481, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 481, + 315 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 481, + 315 + ], + "type": "text", + "content": "48. Ren, W., Wang, L., Piao, Y., Zhang, M., Lu, H., Liu, T.: Adaptive co-teaching for unsupervised monocular depth estimation. In: ECCV. pp. 89-105 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 315, + 481, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 481, + 347 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 481, + 347 + ], + "type": "text", + "content": "49. Ruhkamp, P., Gao, D., Chen, H., Navab, N., Busam, B.: Attention meets geometry: Geometry guided spatial-temporal attention for consistent self-supervised monococular depth estimation. In: 3DV. pp. 837-847 (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 347, + 481, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 481, + 369 + ], + "type": "text", + "content": "50. Saxena, A., Sun, M., Ng, A.Y.: Make3d: Learning 3d scene structure from a single still image. PAMI 31(5), 824-840 (2009)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 369, + 481, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 481, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 481, + 413 + ], + "type": "text", + "content": "51. Tancik, M., Srinivasan, P.P., Mildenhall, B., Fridovich-Keil, S., Raghavan, N., Singhal, U., Ramamoorthi, R., Barron, J.T., Ng, R.: Fourier features let networks learn high frequency functions in low dimensional domains. In: NeurIPS. pp. 7537-7547 (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 413, + 481, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 413, + 481, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 413, + 481, + 435 + ], + "type": "text", + "content": "52. Uhrig, J., Schneider, N., Schneider, L., Franke, U., Brox, T., Geiger, A.: Sparsity invariant cnns. In: 3DV. pp. 11-20 (2017)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 435, + 481, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 481, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 481, + 456 + ], + "type": "text", + "content": "53. Wang, C., Buenaposada, J.M., Zhu, R., Lucey, S.: Learning depth from monocular videos using direct methods. In: CVPR. pp. 2022-2030 (2018)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 456, + 481, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 456, + 481, + 489 + ], + "spans": [ + { + "bbox": [ + 132, + 456, + 481, + 489 + ], + "type": "text", + "content": "54. Wang, R., Pizer, S.M., Frahm, J.M.: Recurrent neural network for (un-)supervised learning of monocular video visual odometry and depth. In: CVPR. pp. 5555-5564 (2019)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 490, + 481, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 490, + 481, + 512 + ], + "spans": [ + { + "bbox": [ + 132, + 490, + 481, + 512 + ], + "type": "text", + "content": "55. Wang, R., Yu, Z., Gao, S.: Planedepth: Self-supervised depth estimation via orthogonal planes. In: CVPR. pp. 21425-21434 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 512, + 481, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 481, + 544 + ], + "type": "text", + "content": "56. Wang, X., Zhu, Z., Huang, G., Chi, X., Ye, Y., Chen, Z., Wang, X.: Drafting monocular cues and velocity guidance for self-supervised multi-frame depth learning. In: AAAI. pp. 2689-2697 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 544, + 481, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 544, + 481, + 567 + ], + "spans": [ + { + "bbox": [ + 132, + 544, + 481, + 567 + ], + "type": "text", + "content": "57. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: From error visibility to structural similarity. IEEE TIP 13(4), 600-612 (2004)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 567, + 481, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 481, + 600 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 481, + 600 + ], + "type": "text", + "content": "58. Watson, J., Aodha, O.M., Prisacariu, V.A., Brostow, G.J., Firman, M.: The temporal opportunist: Self-supervised multi-frame monocular depth. In: CVPR. pp. 1164-1174 (2021)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 600, + 481, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 481, + 622 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 481, + 622 + ], + "type": "text", + "content": "59. Watson, J., Firman, M., Brostow, G.J., Turmukhambetov, D.: Self-supervised monocular depth hints. In: ICCV. pp. 2162-2171 (2019)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 622, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 622, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 622, + 481, + 643 + ], + "type": "text", + "content": "60. Yin, Z., Shi, J.: Geonet: Unsupervised learning of dense depth, optical flow and camera pose. In: CVPR. pp. 1983-1992 (2018)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 643, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 481, + 666 + ], + "type": "text", + "content": "61. Zhang, H., Li, Y., Cao, Y., Liu, Y., Shen, C., Yan, Y.: Exploiting temporal consistency for real-time video depth estimation. In: ICCV. pp. 1725-1734 (2019)" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 157, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 157, + 91, + 447, + 102 + ], + "type": "text", + "content": "Mono-ViFi: A Unified Framework for Self-supervised Monocular Depth" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 281 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "text", + "content": "62. Zhang, N., Nex, F., Vosselman, G., Kerle, N.: Lite-mono: A lightweight cnn and transformer architecture for self-supervised monocular depth estimation. In: CVPR. pp. 18537-18546 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 480, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 480, + 183 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 480, + 183 + ], + "type": "text", + "content": "63. Zhao, C., Poggi, M., Tosi, F., Zhou, L., Sun, Q., Tanga, Y., Mattoccia, S.: Gasmono: Geometry-aided self-supervised monocular depth estimation for indoor scenes. In: ICCV. pp. 16163-16174 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 480, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 480, + 216 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 480, + 216 + ], + "type": "text", + "content": "64. Zhao, C., Zhang, Y., Poggi, M., Tosi, F., Guo, X., Zhu, Z., Huang, G., Tang, Y., Mattoccia, S.: Monovit: Self-supervised monocular depth estimation with a vision transformer. In: 3DV. pp. 668-678 (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 216, + 480, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 480, + 237 + ], + "type": "text", + "content": "65. Zhou, H., Greenwood, D., Taylor, S.: Self-supervised monocular depth estimation with internal feature fusion. In: BMVC (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 237, + 480, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 237, + 480, + 259 + ], + "spans": [ + { + "bbox": [ + 132, + 237, + 480, + 259 + ], + "type": "text", + "content": "66. Zhou, T., Brown, M., Snavely, N., Lowe, D.G.: Unsupervised learning of depth and ego-motion from video. In: CVPR. pp. 6612-6619 (2017)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 259, + 480, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 480, + 281 + ], + "type": "text", + "content": "67. Zhou, Z., Fan, X., Shi, P., Xin, Y.: R-msfm: Recurrent multi-scale feature modulation for monocular depth estimating. In: ICCV. pp. 12757-12766 (2021)" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "type": "text", + "content": "J. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/MonoTTA_ Fully Test-Time Adaptation for Monocular 3D Object Detection/cfa6dde5-bde3-4186-b95b-8bbf6c77c339_content_list.json b/2024/MonoTTA_ Fully Test-Time Adaptation for Monocular 3D Object Detection/cfa6dde5-bde3-4186-b95b-8bbf6c77c339_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..f664e48732d57a9eeb5d1e5db4a1c2e5ac777c3f --- /dev/null +++ b/2024/MonoTTA_ Fully Test-Time Adaptation for Monocular 3D Object Detection/cfa6dde5-bde3-4186-b95b-8bbf6c77c339_content_list.json @@ -0,0 +1,1590 @@ +[ + { + "type": "text", + "text": "MonoTTA: Fully Test-Time Adaptation for Monocular 3D Object Detection", + "text_level": 1, + "bbox": [ + 256, + 140, + 750, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hongbin Lin $^{1,2*}$ , Yifan Zhang $^{3,4*}$ , Shuaicheng Niu $^{5*}$ , Shuguang Cui $^{2,1}$ , Zhen Li $^{2,1\\dagger}$", + "bbox": [ + 292, + 210, + 709, + 243 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ FNii-Shenzhen, $^{2}$ SSE, CUHK-Shenzhen, $^{3}$ NUS, $^{4}$ Skywork AI, $^{5}$ NTU {hongbinlin@link.,shuguangcui@,lizhen@}cuhk.edu.cn, yifan.zhang@u.nus.edu, shuaicheng.niu@ntu.edu.sg", + "bbox": [ + 259, + 253, + 743, + 297 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Monocular 3D object detection (Mono 3Det) aims to identify 3D objects from a single RGB image. However, existing methods often assume training and test data follow the same distribution, which may not hold in real-world test scenarios. To address the out-of-distribution (OOD) problems, we explore a new adaptation paradigm for Mono 3Det, termed Fully Test-time Adaptation which aims to adapt a well-trained model to unlabeled test data by handling potential data distribution shifts at test time. However, applying this paradigm in Mono 3Det poses significant challenges due to OOD test data causing a remarkable decline in object detection scores. This decline conflicts with the pre-defined score thresholds of existing detection methods, leading to severe object omissions (i.e., rare positive detections and many false negatives). Consequently, the limited positive detection and plenty of noisy predictions cause test-time adaptation to fail in Mono 3Det. To handle this problem, we propose a novel Monocular Test-Time Adaptation (MonoTTA) method, based on two new strategies. 1) Reliability-driven adaptation: we empirically find that high-score objects are still reliable and the optimization of high-score objects can enhance confidence across all detections. Thus, we devise a self-adaptive strategy to identify reliable objects for model adaptation, which discovers potential objects and alleviates omissions. 2) Noise-guard adaptation: since high-score objects may be scarce, we develop a negative regularization term to exploit the numerous low-score objects via negative learning, preventing overfitting to noise and trivial solutions. Experimental results show that MonoTTA brings significant performance gains for Mono 3Det models in OOD test scenarios, approximately $190\\%$ gains by average on KITTI and $198\\%$ gains on nuScenes. The source code is now available at Hongbin98/MonoTTA.", + "bbox": [ + 261, + 323, + 743, + 698 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Test-time Adaptation $\\cdot$ Monocular 3D Object Detection", + "bbox": [ + 261, + 710, + 723, + 724 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 746, + 375, + 762 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Three-dimensional (3D) Object Detection is a significant computer vision task, with the objective of identifying objects and determining their spatial and di", + "bbox": [ + 212, + 773, + 785, + 805 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Authors contributed equally.", + "bbox": [ + 217, + 810, + 426, + 825 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding authors.", + "bbox": [ + 222, + 825, + 392, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8eb188ab9bc574c6820cd170f345d2d622008005abc8b11024dcb6d4b96ad52a.jpg", + "image_caption": [ + "Fig. 1: An illustration of the generalizability issue of Mono 3Det models. Compared with in-distribution (In-dis) scenarios (e.g., sunny), the detection scores within out-of-distribution (OOD) test data suffer severe degradation when the well-trained model (MonoFlex [48]) is directly applied to test scenarios affected by common natural disruptions, like weather changes (e.g., snow and fog). Since existing Mono 3Det methods mainly adopt a pre-defined score threshold (e.g., 0.2) for object detection, it leads to severe omissions and unreliable detections, thereby suffering serious performance degradation. Note that test images are the same but under different weather conditions." + ], + "image_footnote": [], + "bbox": [ + 222, + 143, + 782, + 296 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "mensional attributes through diverse sensor inputs [3,5,16,37,38]. To reduce the cost of sensors, there is an increasing trend towards implementing autonomous driving systems via Monocular 3D Object Detection (Mono 3Det) [2,42], where only one single RGB image and the camera calibration information are given. Even if this practical task is challenging, Mono 3Det methods have achieved promising results across various tasks and datasets [4,22,29,40,48]. Behind the success, a common presupposition is assuming that test images have the same distribution as the training images. However, this assumption could be possibly invalidated in many real-world scenarios due to prevalent natural corruptions such as weather changes, diminished sharpness, and other factors that introduce noise and contribute to uncalibrated cameras. In such circumstances, the well-trained model often suffers substantial performance degradation as a consequence of the data distributional shifts between the training images and the unlabeled test images. As shown in Fig. 1, the model performance degrades from $46.2\\mathrm{mAP}$ in in-distribution data to $0.3\\mathrm{mAP}$ in Snow and $7.2\\mathrm{mAP}$ in Fog. Considering the widespread application of Mono 3Det in autonomous driving, the severe performance degradation due to out-of-distribution (OOD) test data may lead to unexpected traffic accidents and pose serious safety risks. Therefore, it is crucial to deal with the OOD generalization problem for Mono 3Det.", + "bbox": [ + 212, + 443, + 787, + 731 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In addressing the OOD challenges specifically in test scenarios, one paradigm that has emerged as highly promising and gaining traction is Test-Time Adaptation (TTA), which seeks to tackle data distribution shifts by adapting a well-trained model to unlabeled test images in real time. Test-time training (TTT) [34] represents an initial approach of TTA in classification tasks, by adjusting the well-trained model to predict rotations through additional model training, while its computation demands at the adaptation stage are prohibitive", + "bbox": [ + 212, + 734, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Hongbin Lin, Yifan Zhang and et al.", + "bbox": [ + 271, + 114, + 514, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "in Mono 3Det applications, particularly in autonomous driving. To enhance efficiency, Tent [36] and EATA [26] have been developed for Fully Test-Time Adaptation (Fully TTA) where only unlabeled test images and a well-trained model are provided. Besides, Ev-TTA [12] and SOD [35] devise TTA methods to handle the event-based object recognition and weakly supervised salient object detection, respectively. Considering the constraints on time of Mono 3Det, we explore the fully TTA paradigm which seeks to deal with OOD test data in real time.", + "bbox": [ + 212, + 146, + 787, + 253 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To investigate this paradigm for Mono 3Det, we dig into the detection outcomes for objects within test scenarios with variations or corruptions that are commonly caused by weather or cameras. Specifically, we directly apply the well-trained model to the validation set of KITTI which has been artificially injected with four distinct types of weather-related corruptions, namely Brightness, Frost, Fog, and Snow. Subsequently, we plot their distributions of detection scores (c.f. Fig. 1). It is observed that the detection scores of test objects tend to markedly decline as well as the high-score objects are scarce in the extreme scenario (Snow) when the well-trained model is directly applied to the scenarios with corruptions. This phenomenon indicates that: 1) The pre-trained Mono 3Det model struggles to discriminate between objects and the background within OOD test data, presenting as quantities of omissions and unconfident detections. 2) Directly applying existing fully TTA methods to Mono 3Det could only get suboptimal performance since they struggle to optimize the model without enough high-score (positive) detections, especially in certain extreme scenarios.", + "bbox": [ + 212, + 258, + 789, + 484 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To handle it in Mono 3Det, we propose a Monocular Test-Time Adaptation (MonoTTA) method, consisting of the reliability-driven adaptation and noise-guard adaptation strategies: 1) Reliability-driven adaptation. Specifically, data distribution shifts lead to omissions and noisy detections while our empirical analysis suggests that high detection score objects are still reliable (c.f. Fig. 3 (a)). Moreover, even if we only optimize the model via high-score objects ( $e.g., \\geq 0.5$ ), both the numbers of low-score and high-score objects increase (c.f. Fig. 3 (b)). These investigations motivate us that exploiting high-score objects rather than all objects for model adaptation would be a more reliable way to alleviate data distribution shifts and discover potential objects. Hence, we develop a self-adaptive strategy for the identification of reliable high-score objects in test images and devise the adaptive optimization loss $\\mathcal{L}_{AO}$ to exploit the reliable subset for model adaptation, alleviating the detection score decline issue of OOD test data and digging out more potential objects. 2) Noise-guard adaptation. In addition, data distribution shifts may also result in a scarcity of high-score objects, i.e., the majority of objects presenting low scores as the 'Snow' scenario in Fig. 3 (a). To this end, we develop a negative regularization term to make rational use of the numerous low-scoring objects in the Negative Learning manner [13]. On the one hand, the negative regularization term $\\mathcal{L}_{Nreg}$ allows the model to conduct adaptation via numerous noisy low-scoring objects. Thus, the model can achieve more high-score objects after alleviating distribution shifts. On the other hand, this term also prevents the model from overfitting to noise and trivial solutions, i.e., assigning all classes of one object with high scores.", + "bbox": [ + 212, + 492, + 789, + 843 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "MonoTTA", + "bbox": [ + 658, + 114, + 730, + 127 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We summarize the main contributions as follows:", + "bbox": [ + 240, + 146, + 593, + 161 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- To the best of our knowledge, we are the first to explore Fully Test-Time Adaptation to address OOD generalization problems for Mono 3Det. We show that the explored novel paradigm can bring significant improvements to Mono 3Det models in OOD test scenarios, e.g., $137\\%$ and $244\\%$ average performance gains across 13 types of OOD shifts on KITTI.", + "- Our empirical investigation reveals an important insight that high-score objects maintain their reliability amidst various corruptions, while optimizing these high-score objects significantly boosts model confidence across all detections. This motivates the first effective test-time adaptation method (i.e., our MonoTTA) in Mono 3Det.", + "- Extensive experiments on 13 types of corruptions of KITTI and 2 real scenarios (daytime $\\leftrightarrow$ night) of nuScenes demonstrate the effectiveness of our MonoTTA in boosting existing Mono 3Det methods [29,48] to handle test-time OOD problems. Even for instance-level methods [40], MonoTTA also maintains sufficient improvement, which further confirms its applicability." + ], + "bbox": [ + 225, + 169, + 782, + 393 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 416, + 387, + 431 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We first review the literature on Monocular 3D Object Detection, and then discuss Source-free Domain Adaptation and Test-Time Adaptation methods. More discussions on Unsupervised Domain Adaptation [18,41,47] are in Appendix A. Monocular 3D Object Detection aims to perceive 3D objects from a single 2D image. Existing Mono 3Det methods could be divided into two groups according to the use of extra information. On the one hand, some existing methods leverage extra pre-trained depth estimation modules [6,39,49] to solve one of the most difficult problems in Mono 3Det, i.e., depth estimation from a single image. Other methods utilize LiDAR information, e.g., generating pseudo-LiDAR [23,31,37]. It is worth noting that Monoground [29] proposes to introduce the ground plane as prior information, and MonoNeRD [40] proposes to utilize scene geometric clues to enhance the detector's performance in the implicit reconstruction manner. On the other hand, some Mono 3Det methods try to detect 3D objects without extra data. For example, SMOKE [20] proposes to detect 3D objects as the key points estimation task. Then, Monoflex [48] improves this idea by providing a flexible definition of object centers, which unifies the centers of regular and truncated objects. GrooMeD-NMS [15] proposes a grouped mathematically differentiable Non-Maximal Suppression for Mono 3Det.", + "bbox": [ + 212, + 446, + 784, + 715 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Source-free Domain Adaptation (SFDA) aims to adapt the pre-trained source model to an unlabeled target domain without using the source data due to privacy issues [17, 30]. SF-UDA $^3$ D [32] first explores the SFDA framework to adapt the PointRCNN 3D detector to target domains, which consists of pseudolabeling, reversible scale-transformations and motion coherency. Recently, the authors [10] seek to exploit the source model more reliably and propose an uncertainty-aware teacher-student framework to filter incorrect pseudo labels during model adaptation, alleviating the negative impact of label noise.", + "bbox": [ + 212, + 719, + 782, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Hongbin Lin, Yifan Zhang and et al.", + "bbox": [ + 271, + 114, + 514, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Nonetheless, SFDA assumes all target data to be known in advance and makes predictions after multiple epochs of optimization, which may not be viable for real-time applications due to computational or time constraints.", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Test-Time Adaptation (TTA) seeks to improve model performance on test data via model adaptation through test samples even if data shifts exist. Early TTA methods [19, 34] endeavor to conduct additional model optimization on training data by self-supervised objectives, and then adapt the well-trained model to the test data via self-supervised objectives. However, in Mono 3Det applications like autonomous driving, the computation demands of such methods are prohibitive. To solve this, Fully Test-Time Adaptation methods are developed to adapt the well-trained model, where only unlabeled test images are available. Specifically, certain methods [25, 27, 33] tackle data distribution shifts by adapting the batch normalization layer statistics, while others alleviate this issue either by the entropy minimization of test data [8, 36] or maximizing the prediction consistency of different augmentations [43-46]. As for object detection tasks, Ev-TTA [12] and SOD [35] try to handle the event-based object recognition and weakly supervised salient object detection offline, respectively.", + "bbox": [ + 212, + 191, + 784, + 402 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "However, existing fully TTA methods struggle to optimize the model and solve distribution shifts in Mono 3Det due to numerous false negative detections. To the best of our knowledge, our MonoTTA stands as the first fully TTA method that handles distribution shifts for Mono 3Det models in real time.", + "bbox": [ + 212, + 404, + 784, + 464 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Monocular Test-Time Adaptation", + "text_level": 1, + "bbox": [ + 214, + 489, + 578, + 508 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Problem Statement", + "text_level": 1, + "bbox": [ + 214, + 523, + 423, + 539 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Without loss of generality, we denote the pre-trained (or well-trained) model as $f_{\\Theta_0}(\\mathbf{x})$ , which is achieved via training on labeled training images $\\{(\\mathbf{x}_i^s,\\mathbf{y}_i^s)\\}_{i = 1}^N$ . The training images follow the training distribution $P(\\mathbf{x})$ (i.e., $\\mathbf{x}^s\\sim P(\\mathbf{x})$ ). Here, $\\Theta_0$ represents the parameters of the pre-trained model and $N$ is the number of training data. During the training stage, the model is optimized to fit (or overfit) the training data. Then, at the test stage, the model will be able to perform well if the unlabeled test images $\\mathcal{D}_t = \\{\\mathbf{x}_i\\}_{i = 1}^{N_t}$ follows the identical data distribution, i.e., $\\mathbf{x}\\sim P(\\mathbf{x})$ where $N_{t}$ is the total number of test images. However, in real applications, it is possible for the pre-trained model to encounter Out-Of-Distribution (OOD) test samples due to prevalent natural corruptions, namely distribution shifts, i.e., $\\mathbf{x}\\sim Q(\\mathbf{x})$ and $P(\\mathbf{x})\\neq Q(\\mathbf{x})$ .", + "bbox": [ + 212, + 551, + 784, + 718 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To address this issue, fully test-time adaptation [36] seeks to tackle distribution shifts by adapting the pre-trained model $f_{\\Theta_0}(\\mathbf{x})$ to unlabeled test images $\\{\\mathbf{x}_i\\}_{i=1}^{N_t}$ in real time. To achieve this goal, existing methods typically endeavor to update the model through the minimization of unsupervised objectives defined on test samples by $\\min_{\\hat{\\Theta}} \\mathcal{L}(\\mathbf{x}; \\Theta)$ , where $\\mathbf{x} \\sim Q(\\mathbf{x})$ and $\\hat{\\Theta} \\subseteq \\Theta$ . Here, $\\hat{\\Theta}$ denotes the subset of model parameters that should be updated (i.e., batch normalization layers following existing methods [26,36]). Most existing fully TTA methods focus on classification tasks, heavily relying on sufficient positive predictions for", + "bbox": [ + 212, + 719, + 785, + 839 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "MonoTTA", + "bbox": [ + 658, + 114, + 730, + 127 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/7ba916103de92e85b6fa56879c52dc92c2b51841c794c47ecc76865a941152c8.jpg", + "image_caption": [ + "Fig. 2: An illustration of our MonoTTA. During the test phase, only the pre-trained model $f_{\\Theta_0}(\\mathbf{x})$ and unlabeled test images $\\{\\mathbf{x}_i\\}_{i=1}^{N_t}$ are given. To conduct model adaptation, we initialize the model $f_{\\Theta}(\\mathbf{x})$ by $\\Theta_0$ and only update the parameters of batch normalization layers. When a batch of test images arrives, we first compute test object scores and refine the adaptive threshold $\\alpha$ to select the reliable high-score objects, thereby optimizing $\\Theta$ via the adaptive optimization loss $\\mathcal{L}_{AO}$ . Meanwhile, we devise a negative regularization term $\\mathcal{L}_{Nreg}$ to facilitate the model to avoid overfitting to noise and trivial solutions. Here, Ped. and Cyc. represent Pedestrian and Cyclist in KITTI." + ], + "image_footnote": [], + "bbox": [ + 236, + 143, + 774, + 373 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "model adaptation. Nevertheless, there is a significant difference between conventional classification tasks and Mono 3Det. As previously indicated, the detection scores of test images $\\mathbf{x}$ derived from $f_{\\Theta_0}(\\mathbf{x})$ are prone to markedly decrease in the presence of corruptions as shown in Fig. 1, leading to severe omissions (numerous false negatives) in Mono 3Det. In such circumstances, the scarcity of positive detections presents a significant challenge for model adaptation to test distributions while adapting the model with unreliable low-score detections may significantly introduce the noise. Therefore, existing fully TTA methods tend to fail in the OOD generalization problems of Mono 3Det.", + "bbox": [ + 212, + 517, + 787, + 654 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Overall Scheme", + "text_level": 1, + "bbox": [ + 215, + 676, + 390, + 691 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "After thoroughly examining the characteristics and challenges of Mono 3Det, we introduce a Monocular Test-Time Adaptation (MonoTTA) method to address the OOD problems for Mono 3Det models, which seeks to solve the object score declining issue within unlabeled OOD test data. As shown in Fig. 2, MonoTTA consists of two strategies: 1) Reliability-driven adaptation and 2) Noise-guard adaptation. We first briefly introduce the two strategies below.", + "bbox": [ + 212, + 703, + 787, + 794 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "First, we develop a reliability-driven adaptation strategy (c.f. Section 3.3) to conduct reliable model adaptation for OOD test data based on dependable test objects. Our empirical investigations inspire us to exploit those relatively reliable", + "bbox": [ + 212, + 795, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Hongbin Lin, Yifan Zhang and et al.", + "bbox": [ + 271, + 114, + 514, + 128 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3acbb989c5da1a458cc9645f9fc38d02e6b56bbe28819edb4829bbe02c59e89f.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 256, + 143, + 478, + 277 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b71877ed35d254d833d907c9acdf1fabec7900d73d66895bba8a8d1b23eeaa23.jpg", + "image_caption": [ + "(b)", + "Fig. 3: Based on MonoGround [29], we conduct two empirical studies (Car, KITTI), with the 3D IoU threshold of 0.5. (a) We visualize the accuracy of the objects across varied scoring ranges, which shows that the accuracy of objects with high scores remains relatively stable even in the presence of diverse corruptions (Ideal means in-distribution scenarios). (b) We visualize the number of low & high-score objects before and after optimization. Although only high-score objects are optimized, the model treats low-score objects with more confidence." + ], + "image_footnote": [], + "bbox": [ + 486, + 148, + 738, + 277 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "test objects for alleviating distribution shifts, thereby discovering more potential objects. To this end, MonoTTA excludes unreliable test objects involving an adaptive threshold $\\alpha$ for any unlabeled test data. Subsequently, the model is optimized by the adaptive optimization loss $\\mathcal{L}_{AO}$ via the selected reliable objects.", + "bbox": [ + 212, + 411, + 784, + 470 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Second, we tend to utilize plenty of low-score objects to adapt the model in an indirect manner instead of directly optimizing the model since low-score objects are noisy. Hence, we devise a noise-guard adaptation strategy (c.f. Section 3.4) to prevent the model from overfitting to noisy predictions and falling into trivial solutions. Specifically, we randomly choose one of the negative classes of low-score objects and minimize the scores (e.g., score 0.03 of [0.10, 0.12, 0.03] c.f. Fig. 2) after simply filtering out extremely low-score detections. Even though the positive class is noisy (i.e., score 0.12), this term is capable of optimizing the model indirectly, i.e., learn to deny the negative category of the object.", + "bbox": [ + 212, + 472, + 787, + 607 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Overall, the training scheme of MonoTTA is as follows:", + "bbox": [ + 238, + 608, + 640, + 623 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\hat {\\Theta}} \\mathcal {L} _ {A O} (\\hat {\\Theta}) + \\lambda \\mathcal {L} _ {N r e g} (\\hat {\\Theta}), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 401, + 633, + 785, + 659 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\lambda$ is the balance hyper-parameter. The pseudo-code of MonoTTA is summarized in Algorithm 1.", + "bbox": [ + 212, + 670, + 787, + 702 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Reliability-Driven Adaptation", + "text_level": 1, + "bbox": [ + 214, + 723, + 509, + 739 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To identify dependable test objects and conduct test-time model adaptation, we propose a reliability-driven adaptation strategy that consists of two components: 1) Reliable object identification and 2) Adaptive model optimization.", + "bbox": [ + 212, + 750, + 784, + 794 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Reliable object identification. When unlabeled test images arrive, it is difficult for the pre-trained model to get accurate detections on OOD test data due to the decline in detection scores. To resolve this, we dig into the detection", + "bbox": [ + 212, + 794, + 785, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "MonoTTA", + "bbox": [ + 658, + 114, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 125 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "accuracy of the pre-trained model in test scenarios with various corruptions. Specifically, based on MonoGround [29], we visualize the detection precision for the Car category within the KITTI dataset across varied scoring ranges, with the 3D Intersection over Union (IoU) threshold of 0.5. As shown in Fig. 3 (a), we find that high-score objects are more reliable and relatively stable even in the presence of diverse corruptions. Following this, we propose to select reliable high-score objects to conduct model adaptation via an adaptive threshold $\\alpha$ . Specifically, we exploit the model $f_{\\Theta}(\\cdot)$ (initialized by $\\Theta_0$ ) to infer a batch of test images $\\{\\mathbf{x}_b\\}_{b=1}^B$ and obtain the heatmap $\\mathbf{h}_i$ of each image $\\mathbf{x}_i$ by $\\mathbf{h}_i = f_{\\Theta}(\\mathbf{x}_i)$ , where $i$ ranges from 0 to $B-1$ and $B$ denotes the batch size. Then, we could achieve object score maps $\\mathbf{s}_i \\in \\mathbb{R}^{B \\times N_m}$ like the peaks in $\\mathbf{h}_i$ [7] after normalization, where $K$ and $N_m$ denotes the number of classes and the maximum of detected objects, respectively. With the score map $\\mathbf{s}_i$ , we update the adaptive threshold $\\alpha_t$ at the iteration $t$ in the exponential moving average manner by:", + "bbox": [ + 212, + 146, + 787, + 359 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\alpha_ {t} = \\left\\{ \\begin{array}{l l} \\gamma , & \\text {i f} t = 1 \\\\ \\beta \\bar {m} _ {t} + (1 - \\beta) \\alpha_ {t - 1}, & \\text {i f} t > 1 \\end{array} \\right., \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 367, + 376, + 785, + 411 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {m} _ {t} = \\frac {1}{B} \\sum_ {i = 1} ^ {B} \\frac {\\sum_ {j = 1} ^ {N _ {m}} s _ {i j} \\cdot \\mathbb {I} \\left(s _ {i j} \\geq \\gamma\\right)}{\\sum_ {j = 1} ^ {N _ {m}} \\mathbb {I} \\left(s _ {i j} \\geq \\gamma\\right)}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 377, + 435, + 785, + 479 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Eqn. (2), $\\bar{m}_t$ denotes the average score of all detected objects in a single batch of $B$ test samples at iteration $t$ , while $\\beta \\in [0,1]$ is a decay coefficient. As for Eqn. (3), $s_{ij} \\in (0,1)$ denotes the score of the $j$ -th object in the $i$ -th test image while $\\mathbb{I}(\\cdot)$ is the indicator function. Note that $\\gamma \\in \\mathbb{R}^1$ is a pre-defined object detection threshold adopted from existing methods at their original inference stage as well as $N_m$ .", + "bbox": [ + 212, + 491, + 787, + 582 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Adaptive model optimization. As shown in Fig. 3 (b), the optimization of high-score objects can also enhance the confidence of the model for relatively low-score objects. It motivates us that exploiting high-score objects rather than all objects for model adaptation would be a more reliable way to learn from OOD test data. Therefore, with the adaptive threshold $\\alpha_{t}$ , we select the reliable subset of high scores from $\\mathbf{s}_i$ and calculate the adaptive optimization loss $\\mathcal{L}_{AO}$ to adapt the model by:", + "bbox": [ + 212, + 584, + 787, + 690 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {A O} = - \\frac {1}{B} \\sum_ {i = 1} ^ {B} \\sum_ {j = 1} ^ {N _ {m}} \\log \\left(s _ {i j} \\cdot \\mathbb {I} \\left(s _ {i j} \\geq \\alpha_ {t}\\right)\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 356, + 705, + 785, + 750 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The adaptive optimization loss $\\mathcal{L}_{AO}$ alleviates the potential data distribution shifts in OOD test data by allowing the model to confidently identify high-score test objects. As we mentioned, it solves the score decline issue of OOD test data in a more reliable way rather than directly optimizing all test objects, thereby avoiding overfitting to noise and discovering more potential test objects.", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Hongbin Lin, Yifan Zhang and et al.", + "bbox": [ + 271, + 114, + 514, + 128 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Algorithm 1 The pipeline of the proposed MonoTTA", + "text_level": 1, + "bbox": [ + 217, + 146, + 609, + 161 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Require: Unlabeled test data $\\mathcal{D}_t = \\{\\mathbf{x}_i\\}_{i=1}^{N_t}$ ; Pre-trained model $f_{\\Theta_0}(\\mathbf{x})$ ; Batch size $B$ ; Parameters $\\lambda, \\beta, \\eta$ .", + "bbox": [ + 215, + 166, + 785, + 195 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1: for a batch images $\\{\\mathbf{x}_b\\}_{b = 1}^B$ in $\\mathcal{D}_t$ do", + "2: Update the adaptive threshold $\\alpha$ based on Eqn. (2);", + "3: Calculate adaptive optimization loss $\\mathcal{L}_{AO}$ based on Eqn. (4);", + "4: Calculate negative regularization term $\\mathcal{L}_{Nreg}$ based on Eqn. (6);", + "5: Update $\\hat{\\Theta}$ by optimizing Eqn. (1)", + "6: end for", + "7: return Detection Results for all $\\mathbf{x} \\in \\mathcal{D}_t$ ." + ], + "bbox": [ + 222, + 195, + 686, + 292 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "3.4 Noise-Guard Adaptation", + "text_level": 1, + "bbox": [ + 215, + 325, + 465, + 339 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Through the optimization of adaptive optimization loss $\\mathcal{L}_{AO}$ , the model is refined to yield more confident detection outcomes. Nevertheless, high-scoring objects may be scarce due to the distribution shifts, making the adaptation procedure difficult. Meanwhile, exclusive reliance on $\\mathcal{L}_{AO}$ for adaptation may result in trivial solutions, whereby the model indiscriminately assigns high scores to all categories. Previous studies [13, 14] indicate that deep neural networks could learn from noisy pseudo labels in classification tasks through negative learning. Thus, MonoTTA proposes to learn from noisy low-score objects in a negative learning manner for Mono 3Det. Specifically, we denote $\\hat{\\mathbf{s}}_i\\in \\mathbb{R}^{B\\times N_m\\times K}$ as the multi-class score map for the test image $\\mathbf{x}_i$ , i.e. the multi-class score map $\\hat{\\mathbf{s}}_i$ contains not only the highest score of objects but also the scores for other classes. Here $\\mathbf{s}_i = \\arg \\max_k\\hat{\\mathbf{s}}_{ik}$ where $k$ is the class index. Employing a simple constant threshold $\\eta$ , we filter out extremely low-score objects and then randomly select a negative class $\\bar{k}$ for each object. Subsequently, we compute the regularization loss $e_k$ for each class $k$ with relatively low object scores $s_{ij}\\in [\\eta ,\\alpha_t)$ by:", + "bbox": [ + 212, + 349, + 787, + 578 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\ne _ {k} = - \\sum_ {i = 1} ^ {B} \\sum_ {j = 1} ^ {N _ {m}} \\bar {y} _ {i j} \\log \\left(1 - s _ {i j \\bar {k}} \\cdot \\mathbb {I} (\\bar {k} = k)\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 349, + 587, + 785, + 631 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\bar{y}_{ij} = 1 - s_{ij\\bar{k}}$ is a constant weight. We further define $n_k$ as the frequency of negative scores corresponding to class $k$ in the test batch and balance $e_k$ by:", + "bbox": [ + 212, + 641, + 785, + 672 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {N r e g} = \\sum_ {k = 1} ^ {K} \\frac {e _ {k}}{n _ {k}}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 439, + 681, + 785, + 724 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "As we mentioned before, detections with low and intermediate scores tend to be more noisy (c.f. Fig. 3 (a)), which is unreliable for direct model adaptation. To this end, we introduce the regularization term $\\mathcal{L}_{Nreg}$ to leverage noisy low-score detections and improve the model for assimilating potentially accurate information (c.f. Term 2 in Fig. 2). Moreover, this term also prevents the model from trivial solutions, i.e., indiscriminately assigning high scores to all classes of a single object (e.g., TENT [36]). In addition, high-score objects may be absent", + "bbox": [ + 212, + 734, + 787, + 842 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "MonoTTA", + "bbox": [ + 658, + 114, + 730, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/63d80a18755b299d0077d6e63c29c38fd292bb69547e45ed5e25d1416a9e0bdb.jpg", + "table_caption": [ + "Table 1: Comparison with baselines on the KITTI-C validation set, severity level 1 regarding Mean $AP_{3D|R_{40}}$ . The bold number indicates the best result." + ], + "table_footnote": [], + "table_body": "
Car, IoU @ 0.7, 0.5, 0.5
MethodNoiseBlurWeatherDigitalAvg.
Gauss.ShotImpul.Defoc.GlassMotionSnowFrostFogBrit.Contr.PixelSat.
Monoflex [48]3.847.485.312.593.7311.050.237.777.5724.876.9228.1631.4610.84
• BN adaptation [33]13.5821.9318.7815.878.5924.325.4221.4524.6331.8030.5841.0430.7122.21
• TENT [36]17.8027.0923.1821.6611.9028.756.8426.5830.7835.6534.7241.7135.9126.35
• EATA [26]16.6726.4225.0722.5413.2327.737.8726.5831.1035.3935.2841.4036.7226.62
• ActMAD [24]12.2621.4920.4711.107.5323.795.5719.1126.5426.2423.9238.5133.7520.79
• MonoTTA (Ours)21.1528.6526.6425.9119.2631.4812.4330.2433.7536.8436.8341.9738.1329.48
MonoGround [29]2.404.103.313.712.678.130.225.544.5925.374.0033.5728.089.67
• BN adaptation [33]13.4923.5219.6916.337.6123.997.9820.7124.0031.3429.0343.0632.9922.60
• TENT [36]17.9429.6019.9023.4513.9029.3910.3226.6533.3535.9636.3943.3537.7927.54
• EATA [26]16.0326.0818.0820.2812.4027.379.2223.7929.4933.6532.5843.6136.0025.28
• ActMAD [24]13.6522.3920.5614.998.1721.099.1318.7719.5228.3425.3242.6331.5821.24
• MonoTTA (Ours)26.1333.1128.6030.3825.4832.4418.7232.6037.7537.8739.5743.6737.9832.64
", + "bbox": [ + 241, + 179, + 761, + 287 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/4e086364b9fcd8756e6d9347724bdf5e142b88d84507bf9576eb6564ab9951ad.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Pedestrian, IoU @ 0.5, 0.25
MethodNoiseBlurWeatherDigitalAvg.
Gauss.ShotImpul.Defoc.GlassMotionSnowFrostFogBrit.Contr.PixelSat.
Monoflex0.191.620.323.728.476.220.004.272.259.192.081.839.113.79
• BN adaptation [33]6.218.209.207.835.357.522.896.479.249.129.9312.739.768.03
• TENT [36]6.027.969.577.756.068.632.636.719.9110.2610.5512.3310.278.36
• EATA [26]6.057.969.747.936.068.012.486.249.949.7010.0212.4110.128.21
• ActMAD [24]2.414.274.681.461.626.481.174.254.846.426.198.297.074.55
• MonotTA (Ours)6.548.419.397.637.128.993.257.6410.2610.5510.0613.2810.668.75
MonoGround0.610.930.738.417.578.030.003.191.3913.821.833.707.264.42
• BN adaptation [33]5.097.087.777.636.639.452.246.728.409.7211.0616.7012.438.53
• TENT [36]7.2710.1010.028.538.3011.033.548.739.1111.7412.1217.7015.0310.25
• EATA [26]5.928.058.588.127.5910.953.317.9110.1011.0210.9517.3014.349.55
• ActMAD [24]3.965.756.485.615.737.141.585.375.747.278.1013.129.276.55
• MonotTA (Ours)8.5811.1811.799.229.4013.204.839.9514.4612.8513.2517.1314.8511.59
", + "bbox": [ + 241, + 287, + 761, + 393 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/44412cd6db4e954e0783727b5a869eb8bf6953821f7a71c2f6828db9bd1d6114.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Cyclist, IoU @ 0.5, 0.25, 0.25
MethodNoiseBlurWeatherDigitalAvg.
Gauss.ShotImpul.Defoc.GlassMotionSnowFrostFogBrit.Contr.PixelSat.
Monoflex0.281.640.470.594.973.600.007.423.8113.073.793.808.393.99
• BN adaptation [33]2.396.264.365.786.769.091.708.539.1612.9111.2610.5511.027.67
• TENT [36]2.727.945.636.277.209.491.078.9410.9612.7512.729.6411.288.20
• EATA [26]2.337.465.467.197.237.511.248.6010.1812.8611.0210.8910.307.87
• ActMAD [24]2.225.324.362.291.575.640.485.686.4610.569.319.977.805.51
• MonotTA (Ours)3.017.245.987.006.099.511.4510.6311.2212.8511.8511.5910.818.40
MonoGround0.120.460.480.330.722.030.000.560.355.550.522.085.241.42
• BN adaptation [33]1.763.582.083.613.055.410.573.824.476.305.6011.028.874.63
• TENT [36]1.794.853.003.363.496.050.494.436.207.196.5010.439.235.15
• EATA [26]1.894.202.414.133.175.730.384.066.276.416.1410.937.724.88
• ActMAD [24]1.142.841.533.33.083.150.322.743.476.365.759.696.953.87
• MonotTA (Ours)3.935.784.555.434.706.090.694.667.537.697.7411.719.436.15
", + "bbox": [ + 241, + 393, + 761, + 498 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "in certain extreme scenarios. For instance, the Snow scenario (c.f. Fig. 3 (a)) lacks objects with scores exceeding 0.4 when the pre-trained model is directly applied. Under such a circumstance, $\\mathcal{L}_{Nreg}$ plays a more important role in model adaptation since it can alleviate distribution shifts even only low-score objects, i.e., deny the negative category. In other words, $\\mathcal{L}_{Nreg}$ enables the model to alleviate distribution shifts and achieve more relatively high-score objects, thereby laying a crucial foundation for $\\mathcal{L}_{AO}$ in extremely challenging scenarios.", + "bbox": [ + 212, + 518, + 787, + 625 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 652, + 375, + 670 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We conduct experiments based on KITTI [9] and nuScenes [1]. The results presented in this manuscript represent the average value across three difficulty levels, i.e., Easy, Moderate, Hard. Note that we provide more results of higher severity levels of KITTI and more detailed results in Appendix $D$ .", + "bbox": [ + 212, + 686, + 785, + 748 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Datasets. For KITTI, we adopt the protocol established by Monoflex [48] to split the training images into the training set (3712) and validation set (3769) to perform model training and adaptation, respectively. Following the official KITTI evaluation criteria, we evaluate detection results on three levels of difficulty. Then, we construct the KITTI-C dataset through the incorporation of 13 distinct types of data corruptions [11] to the validation set and each corruption", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Hongbin Lin, Yifan Zhang and et al.", + "bbox": [ + 271, + 114, + 514, + 128 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/1016ad6fb21f14f8f303f665c45098a31a4d72442d89e7b833a4fc5e03d381d3.jpg", + "image_caption": [ + "Average performance of 13 corruptions on KITTI-C Validation set, severity level 2", + "Fig. 4: We visualize the comparison with baselines on the KITTI-C validation set, severity level 2 regarding Mean $AP_{3D|R_{40}}$ . The bold number indicates the best result." + ], + "image_footnote": [], + "bbox": [ + 272, + 156, + 725, + 282 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "has 5 severity levels. The model is trained on the original training set and then tested on the KITTI-C validation set within one of the corruptions.", + "bbox": [ + 214, + 347, + 782, + 378 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "As for nuScenes, we adopt the front-view images and construct the Daytime and Night scenarios via their scene descriptions following [21]. Specifically, there exist $24.7\\mathrm{k} / 5.4\\mathrm{k}$ train/test images in Daytime while $3.3\\mathrm{k} / 0.6\\mathrm{k}$ train/test images in Night. Based on these splits, we construct two real-world adaptation tasks, i.e., Daytime $\\rightarrow$ Night and Night $\\rightarrow$ Daytime. For simplification, we transfer the nuScenes dataset into the KITTI format and only consider the Car category. More details of data construction are provided in Appendix $\\pmb{B}$ .", + "bbox": [ + 212, + 387, + 784, + 492 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Implementation details. We implement our method and other baselines in PyTorch [28]. In MonoTTA, we conduct model adaptation based on the public pre-trained weights and the parameter settings provided by their authors [29, 40, 48]. Besides, we employ the Stochastic Gradient Descent (SGD) optimizer with a half learning rate of the initial rate used in base training over different methods, a momentum of 0.9 and a batch size of 16 for KITTI, 4 for nuScenes. Parameters $\\lambda$ , $\\beta$ , $\\eta$ are assigned default values of 1, 0.1, and 0.05, respectively. More training details of MonoTTA are provided in Appendix $C$ .", + "bbox": [ + 212, + 502, + 784, + 623 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Compared methods. Based on three typical or state-of-the-art (SOTA) Mono 3Det methods [29, 40, 48], we fully compare MonoTTA with following methods: 1) source-only, i.e., directly apply the pre-trained model to the test data within corruptions; 2) BN adaptation [33] updates batch normalization statistics via target data; 3) TENT [36] minimizes the entropy loss of test data; 4) EATA [26] identifies reliable samples to update the model by entropy loss minimization. Here we compare MonoTTA with its variant Efficient Test-time Adaptation. 5) ActMAD [24] matches activation statistics for the distribution alignment.", + "bbox": [ + 214, + 633, + 784, + 753 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Evaluation protocols. In order to fully evaluate the proposed method, we report our experimental results in the Average Precision (AP) for 3D bounding boxes, denoted as $AP_{3D|R_{40}}$ . The results present the mean values across three levels of difficulty and the Intersection over Union (IoU) thresholds are set to 0.7, 0.5, 0.5 for Cars and 0.5, 0.25, 0.25 for Pedestrians and Cyclists.", + "bbox": [ + 214, + 763, + 784, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "MonoTTA", + "bbox": [ + 658, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/21e234440e204e27ec0c18cc793bbd57c303929d334ce98b0f5e2193c398f2d8.jpg", + "image_caption": [ + "Performance on Daytime to Night and Night to Daytime scenarios of nuScenes", + "Fig. 5: Comparison with baselines on Daytime $\\rightarrow$ Night and Night $\\rightarrow$ Daytime of nuScenes, regarding Mean $AP_{3D|R_{40}}$ . The **bold** number indicates the best result." + ], + "image_footnote": [], + "bbox": [ + 271, + 157, + 728, + 292 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.1 Comparisons with Previous Methods", + "text_level": 1, + "bbox": [ + 214, + 354, + 568, + 371 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We first compare our MonoTTA with previous methods in severity level 1 of KITTI-C. The results are reported in Table 1, which gives the following observations: 1) Due to distribution shifts, directly applying the pre-trained model to the test data (i.e., source-only) suffers severe performance degradation in all categories. 2) Existing TTA methods are able to mitigate the negative effect of distribution shifts for Mono 3Det to some degree. However, they only achieve suboptimal performance since they tend to increase the scores of all positive detections, containing severe noise. 3) MonoTTA consistently outperforms all compared methods over all categories within various base models in terms of mean $AP_{3D|R_{40}}$ . Specifically, MonoTTA achieves the best or comparable performance in all categories under all corruptions, attaining a large performance gain over TENT and EATA (e.g., improving an average $AP_{3D|R_{40}}$ about 5.1 and 7.4 of the Car category based on MonoGround).", + "bbox": [ + 212, + 380, + 787, + 579 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.2 More Severe Corruption and Real Scenario", + "text_level": 1, + "bbox": [ + 214, + 601, + 619, + 617 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "On the one hand, to fully validate the effectiveness of our MonoTTA, we visualize experimental results under more severe corruption conditions at severity level 2 as shown in Fig. 4, which clearly gives additional observations: 1) With the escalation of severity level, the pre-trained models suffer a larger performance decline within various corruptions, enlarging the difficulty of TTA. 2) The performance improvements of existing TTA methods become relatively limited, particularly in Pedestrian and Cyclist classes. 3) Even if the tasks are more challenging, MonoTTA still stably obtains the best average performance within all corruptions since $\\mathcal{L}_{Nreg}$ plays an important role in alleviating distribution shifts for certain extreme scenarios, i.e., only low-score objects exist.", + "bbox": [ + 212, + 627, + 787, + 777 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "On the other hand, we further validate different methods within real scenarios as shown in Fig. 5. The experimental results also give the following observations: 1) Under real corruptions, the pre-trained model still suffers severe performance degradation due to the data distribution shifts. 2) TENT tends to increase the", + "bbox": [ + 214, + 779, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Hongbin Lin, Yifan Zhang and et al.", + "bbox": [ + 271, + 114, + 514, + 128 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/bd5c73fe7e83b74d2fa3050e86c7f7fed709f5718e2f98020d5b4cc7b4c25e3a.jpg", + "table_caption": [ + "Table 2: Comparison with baselines based on the instance-level method (i.e., batch size is 1) on the KITTI-C validation set, regarding Mean $AP_{3D|R_{40}}$ and the severity levels 1 and 2. The bold number indicates the best result." + ], + "table_footnote": [], + "table_body": "
MethodLevel 1Level 2
CarPedes.CyclistAvg.CarPedes.CyclistAvg.
MonoNeRD19.845.962.279.3613.023.831.616.15
BN adaptation [33]30.738.853.8114.4626.476.942.9112.11
TENT [36]35.729.994.7516.8231.857.813.4014.35
EATA [26]34.6010.044.2016.2830.667.863.2813.93
MonoTTA (Ours)37.4010.394.3517.3833.998.253.3315.19
", + "bbox": [ + 330, + 198, + 671, + 250 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/58fae23c2e680d7613aa885f32347b57be1fe215d277b1b17e8cd08a98225fa2.jpg", + "table_caption": [ + "Table 3: Based on MonoGround [29], we conduct ablation studies of $\\mathcal{L}_{AO}$ and $\\mathcal{L}_{Nreg}$ on the KITTI-C validation set, regarding $AP_{3D|R_{40}}$ ." + ], + "table_footnote": [], + "table_body": "
Backbone\\( {\\mathcal{L}}_{AO} \\)\\( {\\mathcal{L}}_{Nreg} \\)Level 1Level 2
CarPedes.CyclistAvg.CarPedes.CyclistAvg.
✓9.674.421.425.175.722.430.953.03
✓✓27.9410.175.1014.4021.277.793.8210.96
✓✓23.268.604.6012.1518.936.813.559.77
✓✓✓32.6411.596.1516.7925.968.784.1212.85
", + "bbox": [ + 328, + 301, + 671, + 348 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "confidence of all positive detections and thus overfits to noise, i.e., failing to handle the extremely challenging task $\\mathrm{N} \\rightarrow \\mathrm{D}$ . 3) EATA still achieves sub-optimal performance while our MonoTTA brings sufficient average performance improvement on both MonoFlex (6.23 mAP) and MonoGround (8.26 mAP), maintaining the best performance in real scenarios. Detailed results are in Appendix $D$ .", + "bbox": [ + 212, + 366, + 787, + 444 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.3 Application to Instance-Level Inference Method", + "text_level": 1, + "bbox": [ + 214, + 465, + 658, + 479 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "In this section, we seek to investigate whether the proposed MonoTTA can be used to effectively enhance Mono 3Det methods which only process a single image sequentially, i.e., $B = 1$ . To be specific, it may be crucial for Mono 3Det methods to make immediate decisions based on the most recent scene (image) in real-world scenarios like autonomous driving. To this end, it is essential for TTA methods devised for Mono 3Det to allow a single image as input and then conduct model adaptation. To validate it, we integrate MonoTTA into the SOTA Mono 3Det method namely MonoNeRD [40] which accepts one image each time at the test phase. As shown in Table 2, MonoTTA achieves the best or comparable performance across all categories at both severity levels 1 and 2, illustrating the applicability of our method to boost these approaches for handling OOD test data. Detailed results can be also found in Appendix $D$ .", + "bbox": [ + 212, + 489, + 787, + 672 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.4 Ablation Studies and Quantitative Results", + "text_level": 1, + "bbox": [ + 214, + 693, + 612, + 709 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Ablation studies. To examine the effectiveness of the losses in MonoTTA, we show the results of the models optimized by different losses. As shown in Table 3, introducing $\\mathcal{L}_{AO}$ or $\\mathcal{L}_{Nreg}$ enhances the model performance compared to directly applying the pre-trained model (i.e., source only). On the one hand, such a result verifies that our strategy is able to alleviate the score decline issue in test OOD data with high-score objects. On the other hand, introducing $\\mathcal{L}_{Nreg}$ could obtain a higher average $AP_{3D|R_{40}}$ value, which also verifies that the negative regularization term is able to enhance the pre-trained model even only", + "bbox": [ + 212, + 718, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "MonoTTA", + "bbox": [ + 658, + 114, + 730, + 127 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/7dee429f5322d7f05be1b976335d70ae3b10fcc09f1d687224b5877ee2f9e237.jpg", + "image_caption": [ + "Fig.6: Qualitative results of baselines and the proposed MonoTTA based on MonoGround [29]. We visualize the results on KITTI-C validation set, where predicted cars, pedestrians and cyclists are in lime green, sky blue and yellow, respectively." + ], + "image_footnote": [], + "bbox": [ + 272, + 143, + 730, + 359 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "with low-score objects. When combining the losses (i.e. $\\mathcal{L}_{AO}$ , $\\mathcal{L}_{Nreg}$ ) together, we obtain the best performance.", + "bbox": [ + 212, + 426, + 782, + 457 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Quantitative Results. We provide visualizations within Snow, Fog and Frost based on Monoground as shown in Fig. 6. It is evident that the source-only setting suffers severe omissions, while BN adaptation, TENT and EATA alleviate distribution shifts to some degree and give more detections. As for our MonoTTA, it can produce superior detections even in severe conditions, including fewer omissions and accurate detections as highlighted by red circles.", + "bbox": [ + 212, + 458, + 784, + 547 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 569, + 359, + 584 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we propose a monocular test-time adaptation method to improve the pre-trained model on the shifted test data for monocular 3D object detection. Specifically, our method consists of two strategies: 1) Reliability-driven adaptation. To discover more potential objects, we devise a self-adaptive strategy to identify reliable objects for adaptive model adaptation. 2) Noise-guard adaptation. To avoid overfitting to noise and trivial solutions, we devise the negative regularization term to mitigate the negative effects of noisy detections and alleviate distribution shifts. Experiments on KITTI-C and nuScenes datasets demonstrate the effectiveness of MonoTTA in handling fully test-time adaptation for monocular 3D object detection.", + "bbox": [ + 212, + 598, + 787, + 748 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Future directions. 1) Our work focuses on 2D images, while future studies could explore 3D information in handling distribution shifts. 2) This work explores TTA by assuming one OOD distribution at a time, where the forgetting issue is not severe as the source model weights are recoverable. Exploring scenarios with dynamically OOD distributions offers a compelling future direction, where the forgetting issue would become more severe.", + "bbox": [ + 212, + 750, + 787, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Hongbin Lin, Yifan Zhang and et al.", + "bbox": [ + 271, + 114, + 514, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work was supported by NSFC with Grant No. 62293482, by the Basic Research Project No. HZQB-KCZYZ-2021067 of Hetao Shenzhen HK S&T Cooperation Zone, by Shenzhen General Program No. JCYJ20220530143600001, by Shenzhen-Hong Kong Joint Funding No. SGDX20211123112401002, by the Shenzhen Outstanding Talents Training Fund 202002, by Guangdong Research Project No. 2017ZT07X152 and No. 2019CX01X104, by the Guangdong Provincial Key Laboratory of Future Networks of Intelligence (Grant No. 2022B1212010001), by the Guangdong Provincial Key Laboratory of Big Data Computing, CHUK-Shenzhen, by the NSFC 61931024&12326610, by the Key Area R&D Program of Guangdong Province with grant No. 2018B030338001, by the Shenzhen Key Laboratory of Big Data and Artificial Intelligence (Grant No. ZDSYS201707251409055), and by Tencent & Huawei Open Fund.", + "bbox": [ + 212, + 176, + 789, + 357 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 380, + 321, + 396 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuscenes: A multimodal dataset for autonomous driving. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11621-11631 (2020)", + "2. Chen, X., Kundu, K., Zhang, Z., Ma, H., Fidler, S., Urtasun, R.: Monocular 3d object detection for autonomous driving. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2147-2156 (2016)", + "3. Chen, X., Kundu, K., Zhu, Y., Berneshawi, A.G., Ma, H., Fidler, S., Urtasun, R.: 3d object proposals for accurate object class detection. Advances in neural information processing systems 28 (2015)", + "4. Chen, Y., Tai, L., Sun, K., Li, M.: Monopair: Monocular 3d object detection using pairwise spatial relationships. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12093-12102 (2020)", + "5. Chen, Y., Liu, J., Zhang, X., Qi, X., Jia, J.: Voxelnext: Fully sparse voxelnet for 3d object detection and tracking. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21674-21683 (2023)", + "6. Ding, M., Huo, Y., Yi, H., Wang, Z., Shi, J., Lu, Z., Luo, P.: Learning depth-guided convolutions for monocular 3d object detection. In: Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition workshops. pp. 1000-1001 (2020)", + "7. Duan, K., Bai, S., Xie, L., Qi, H., Huang, Q., Tian, Q.: Centernet: Keypoint triplets for object detection. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6569-6578 (2019)", + "8. Fleuret, F., et al.: Test time adaptation through perturbation robustness. In: NeurIPS 2021 Workshop on Distribution Shifts: Connecting Methods and Applications (2021)", + "9. Geiger, A., Lenz, P., Urtasun, R.: Are we ready for autonomous driving? the kitti vision benchmark suite. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3354-3361. IEEE (2012)", + "0. Hegde, D., Kilic, V., Sindagi, V., Cooper, A.B., Foster, M., Patel, V.M.: Source-free unsupervised domain adaptation for 3d object detection in adverse weather." + ], + "bbox": [ + 225, + 410, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "MonoTTA", + "bbox": [ + 658, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 6973-6980. IEEE (2023)", + "11. Hendrycks, D., Dietterich, T.: Benchmarking neural network robustness to common corruptions and perturbations. In: International Conference on Learning Representations (2018)", + "12. Kim, J., Hwang, I., Kim, Y.M.: Ev-tta: Test-time adaptation for event-based object recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17745-17754 (2022)", + "13. Kim, Y., Yim, J., Yun, J., Kim, J.: Nlnl: Negative learning for noisy labels. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 101-110 (2019)", + "14. Kim, Y., Yun, J., Shon, H., Kim, J.: Joint negative and positive learning for noisy labels. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9442-9451 (2021)", + "15. Kumar, A., Brazil, G., Liu, X.: Groomed-nms: Grouped mathematically differentiable nms for monocular 3d object detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8973-8983 (2021)", + "16. Li, P., Chen, X., Shen, S.: Stereo r-cnn based 3d object detection for autonomous driving. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7644-7652 (2019)", + "17. Liang, J., Hu, D., Feng, J.: Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation. In: International conference on machine learning. pp. 6028-6039. PMLR (2020)", + "18. Lin, H., Zhang, Y., Qiu, Z., Niu, S., Gan, C., Liu, Y., Tan, M.: Prototype-guided continual adaptation for class-incremental unsupervised domain adaptation. In: European Conference on Computer Vision. pp. 351-368 (2022)", + "19. Liu, Y., Kothari, P., Van Delft, B., Bellot-Gurlet, B., Mordan, T., Alahi, A.: Ttt++: When does self-supervised test-time training fail or thrive? In: Advances in Neural Information Processing Systems. vol. 34, pp. 21808-21820 (2021)", + "20. Liu, Z., Wu, Z., Tóth, R.: Smoke: Single-stage monocular 3d object detection via keypoint estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. pp. 996-997 (2020)", + "21. Liu, Z., Tang, H., Amini, A., Yang, X., Mao, H., Rus, D.L., Han, S.: Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. In: 2023 IEEE international conference on robotics and automation (ICRA). pp. 2774-2781. IEEE (2023)", + "22. Luo, Y., Zheng, C., Yan, X., Kun, T., Zheng, C., Cui, S., Li, Z.: Latr: 3d lane detection from monocular images with transformer. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7941-7952 (2023)", + "23. Ma, X., Liu, S., Xia, Z., Zhang, H., Zeng, X., Ouyang, W.: Rethinking pseudo-lidar representation. In: European Conference on Computer Vision (2020)", + "24. Mirza, M.J., Soneira, P.J., Lin, W., Kozinski, M., Possegger, H., Bischof, H.: Actmad: Activation matching to align distributions for test-time-training. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 24152-24161 (2023)", + "25. Nado, Z., Padhy, S., Sculley, D., D'Amour, A., Lakshminarayanan, B., Snoek, J.: Evaluating prediction-time batch normalization for robustness under covariate shift. arXiv preprint arXiv:2006.10963 (2020)", + "26. Niu, S., Wu, J., Zhang, Y., Chen, Y., Zheng, S., Zhao, P., Tan, M.: Efficient test-time model adaptation without forgetting. In: The International Conference on Machine Learning (2022)" + ], + "bbox": [ + 217, + 146, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Hongbin Lin, Yifan Zhang and et al.", + "bbox": [ + 271, + 114, + 514, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "27. Niu, S., Wu, J., Zhang, Y., Wen, Z., Chen, Y., Zhao, P., Tan, M.: Towards stable test-time adaptation in dynamic wild world. In: International Conference on Learning Representations (2023)", + "28. Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., et al.: Pytorch: An imperative style, high-performance deep learning library. In: Advances in neural information processing systems. vol. 32 (2019)", + "29. Qin, Z., Li, X.: Monoground: Detecting monocular 3d objects from the ground. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3793-3802 (2022)", + "30. Qiu, Z., Zhang, Y., Lin, H., et al.: Source-free domain adaptation via avatar prototype generation and adaptation. In: International Joint Conference on Artificial Intelligence (2021)", + "31. Reading, C., Harakeh, A., Chae, J., Waslander, S.L.: Categorical depth distribution network for monocular 3d object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8555-8564 (2021)", + "32. Saltori, C., Lathuilière, S., Sebe, N., Ricci, E., Galasso, F.: Sf-uda 3d: Source-free unsupervised domain adaptation for lidar-based 3d object detection. In: 2020 International Conference on 3D Vision (3DV). pp. 771–780. IEEE (2020)", + "33. Schneider, S., Rusak, E., Eck, L., Bringmann, O., Brendel, W., Bethge, M.: Improving robustness against common corruptions by covariate shift adaptation. In: Advances in neural information processing systems. pp. 11539-11551 (2020)", + "34. Sun, Y., Wang, X., Liu, Z., Miller, J., Efros, A., Hardt, M.: Test-time training with self-supervision for generalization under distribution shifts. In: International conference on machine learning. pp. 9229-9248. PMLR (2020)", + "35. Veksler, O.: Test time adaptation with regularized loss for weakly supervised salient object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7360-7369 (2023)", + "36. Wang, D., Shelhamer, E., Liu, S., Olshausen, B., Darrell, T.: Tent: Fully test-time adaptation by entropy minimization. In: The International Conference on Machine Learning (2021)", + "37. Wang, Y., Chao, W.L., Garg, D., Hariharan, B., Campbell, M., Weinberger, K.Q.: Pseudo-lidar from visual depth estimation: Bridging the gap in 3d object detection for autonomous driving. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8445-8453 (2019)", + "38. Wu, H., Wen, C., Shi, S., Li, X., Wang, C.: Virtual sparse convolution for multimodal 3d object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21653-21662 (2023)", + "39. Xu, B., Chen, Z.: Multi-level fusion based 3d object detection from monocular images. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2345-2353 (2018)", + "40. Xu, J., Peng, L., Cheng, H., Li, H., Qian, W., Li, K., Wang, W., Cai, D.: Mononerd: Nerf-like representations for monocular 3d object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6814-6824 (2023)", + "41. Yang, J., Shi, S., Wang, Z., Li, H., Qi, X.: St3d: Self-training for unsupervised domain adaptation on 3d object detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10368-10378 (2021)", + "42. Ye, X., Shu, M., Li, H., Shi, Y., Li, Y., Wang, G., Tan, X., Ding, E.: Rope3d: The roadside perception dataset for autonomous driving and monocular 3d object detection task. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21341-21350 (2022)" + ], + "bbox": [ + 212, + 146, + 787, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "MonoTTA", + "bbox": [ + 658, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "43. Zhang, M., Levine, S., Finn, C.: Memo: Test time robustness via adaptation and augmentation. In: Advances in Neural Information Processing Systems. vol. 35, pp. 38629-38642 (2022)", + "44. Zhang, Y., Hooi, B., Hong, L., Feng, J.: Self-supervised aggregation of diverse experts for test-agnostic long-tailed recognition. In: Advances in Neural Information Processing Systems. vol. 35, pp. 34077-34090 (2022)", + "45. Zhang, Y., Hooi, B., Hu, D., Liang, J., Feng, J.: Unleashing the power of contrastive self-supervised visual models via contrast-regularized fine-tuning. In: Advances in Neural Information Processing Systems. vol. 34, pp. 29848-29860 (2021)", + "46. Zhang, Y., Kang, B., Hooi, B., Yan, S., Feng, J.: Deep long-tailed learning: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence 45(9), 10795-10816 (2023)", + "47. Zhang, Y., Wei, Y., Wu, Q., Zhao, P., Niu, S., Huang, J., Tan, M.: Collaborative unsupervised domain adaptation for medical image diagnosis. IEEE Transactions on Image Processing 29, 7834-7844 (2020)", + "48. Zhang, Y., Lu, J., Zhou, J.: Objects are different: Flexible monocular 3d object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3289-3298 (2021)", + "49. Zou, Z., Ye, X., Du, L., Cheng, X., Tan, X., Zhang, L., Feng, J., Xue, X., Ding, E.: The devil is in the task: Exploiting reciprocal appearance-localization features for monocular 3d object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2713-2722 (2021)" + ], + "bbox": [ + 215, + 146, + 785, + 452 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Hongbin Lin, Yifan Zhang and et al.", + "bbox": [ + 271, + 114, + 514, + 128 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/MonoTTA_ Fully Test-Time Adaptation for Monocular 3D Object Detection/cfa6dde5-bde3-4186-b95b-8bbf6c77c339_model.json b/2024/MonoTTA_ Fully Test-Time Adaptation for Monocular 3D Object Detection/cfa6dde5-bde3-4186-b95b-8bbf6c77c339_model.json new file mode 100644 index 0000000000000000000000000000000000000000..db0549a2a872f7cb6e01fbe70ccb690916d29ff7 --- /dev/null +++ b/2024/MonoTTA_ Fully Test-Time Adaptation for Monocular 3D Object Detection/cfa6dde5-bde3-4186-b95b-8bbf6c77c339_model.json @@ -0,0 +1,2293 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.257, + 0.141, + 0.75, + 0.187 + ], + "angle": 0, + "content": "MonoTTA: Fully Test-Time Adaptation for Monocular 3D Object Detection" + }, + { + "type": "text", + "bbox": [ + 0.293, + 0.212, + 0.71, + 0.244 + ], + "angle": 0, + "content": "Hongbin Lin\\(^{1,2*}\\), Yifan Zhang\\(^{3,4*}\\), Shuaicheng Niu\\(^{5*}\\), Shuguang Cui\\(^{2,1}\\), Zhen Li\\(^{2,1\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.254, + 0.744, + 0.299 + ], + "angle": 0, + "content": "\\(^{1}\\)FNii-Shenzhen, \\(^{2}\\) SSE, CUHK-Shenzhen, \\(^{3}\\) NUS, \\(^{4}\\) Skywork AI, \\(^{5}\\) NTU {hongbinlin@link.,shuguangcui@,lizhen@}cuhk.edu.cn, yifan.zhang@u.nus.edu, shuaicheng.niu@ntu.edu.sg" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.324, + 0.744, + 0.699 + ], + "angle": 0, + "content": "Abstract. Monocular 3D object detection (Mono 3Det) aims to identify 3D objects from a single RGB image. However, existing methods often assume training and test data follow the same distribution, which may not hold in real-world test scenarios. To address the out-of-distribution (OOD) problems, we explore a new adaptation paradigm for Mono 3Det, termed Fully Test-time Adaptation which aims to adapt a well-trained model to unlabeled test data by handling potential data distribution shifts at test time. However, applying this paradigm in Mono 3Det poses significant challenges due to OOD test data causing a remarkable decline in object detection scores. This decline conflicts with the pre-defined score thresholds of existing detection methods, leading to severe object omissions (i.e., rare positive detections and many false negatives). Consequently, the limited positive detection and plenty of noisy predictions cause test-time adaptation to fail in Mono 3Det. To handle this problem, we propose a novel Monocular Test-Time Adaptation (MonoTTA) method, based on two new strategies. 1) Reliability-driven adaptation: we empirically find that high-score objects are still reliable and the optimization of high-score objects can enhance confidence across all detections. Thus, we devise a self-adaptive strategy to identify reliable objects for model adaptation, which discovers potential objects and alleviates omissions. 2) Noise-guard adaptation: since high-score objects may be scarce, we develop a negative regularization term to exploit the numerous low-score objects via negative learning, preventing overfitting to noise and trivial solutions. Experimental results show that MonoTTA brings significant performance gains for Mono 3Det models in OOD test scenarios, approximately \\(190\\%\\) gains by average on KITTI and \\(198\\%\\) gains on nuScenes. The source code is now available at Hongbin98/MonoTTA." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.711, + 0.725, + 0.726 + ], + "angle": 0, + "content": "Keywords: Test-time Adaptation \\(\\cdot\\) Monocular 3D Object Detection" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.747, + 0.377, + 0.763 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.774, + 0.787, + 0.806 + ], + "angle": 0, + "content": "Three-dimensional (3D) Object Detection is a significant computer vision task, with the objective of identifying objects and determining their spatial and di" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.811, + 0.428, + 0.826 + ], + "angle": 0, + "content": "* Authors contributed equally." + }, + { + "type": "page_footnote", + "bbox": [ + 0.223, + 0.827, + 0.393, + 0.84 + ], + "angle": 0, + "content": "† Corresponding authors." + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.811, + 0.428, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.516, + 0.129 + ], + "angle": 0, + "content": "Hongbin Lin, Yifan Zhang and et al." + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.145, + 0.784, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.31, + 0.788, + 0.421 + ], + "angle": 0, + "content": "Fig. 1: An illustration of the generalizability issue of Mono 3Det models. Compared with in-distribution (In-dis) scenarios (e.g., sunny), the detection scores within out-of-distribution (OOD) test data suffer severe degradation when the well-trained model (MonoFlex [48]) is directly applied to test scenarios affected by common natural disruptions, like weather changes (e.g., snow and fog). Since existing Mono 3Det methods mainly adopt a pre-defined score threshold (e.g., 0.2) for object detection, it leads to severe omissions and unreliable detections, thereby suffering serious performance degradation. Note that test images are the same but under different weather conditions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.444, + 0.788, + 0.732 + ], + "angle": 0, + "content": "mensional attributes through diverse sensor inputs [3,5,16,37,38]. To reduce the cost of sensors, there is an increasing trend towards implementing autonomous driving systems via Monocular 3D Object Detection (Mono 3Det) [2,42], where only one single RGB image and the camera calibration information are given. Even if this practical task is challenging, Mono 3Det methods have achieved promising results across various tasks and datasets [4,22,29,40,48]. Behind the success, a common presupposition is assuming that test images have the same distribution as the training images. However, this assumption could be possibly invalidated in many real-world scenarios due to prevalent natural corruptions such as weather changes, diminished sharpness, and other factors that introduce noise and contribute to uncalibrated cameras. In such circumstances, the well-trained model often suffers substantial performance degradation as a consequence of the data distributional shifts between the training images and the unlabeled test images. As shown in Fig. 1, the model performance degrades from \\(46.2\\mathrm{mAP}\\) in in-distribution data to \\(0.3\\mathrm{mAP}\\) in Snow and \\(7.2\\mathrm{mAP}\\) in Fog. Considering the widespread application of Mono 3Det in autonomous driving, the severe performance degradation due to out-of-distribution (OOD) test data may lead to unexpected traffic accidents and pose serious safety risks. Therefore, it is crucial to deal with the OOD generalization problem for Mono 3Det." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.842 + ], + "angle": 0, + "content": "In addressing the OOD challenges specifically in test scenarios, one paradigm that has emerged as highly promising and gaining traction is Test-Time Adaptation (TTA), which seeks to tackle data distribution shifts by adapting a well-trained model to unlabeled test images in real time. Test-time training (TTT) [34] represents an initial approach of TTA in classification tasks, by adjusting the well-trained model to predict rotations through additional model training, while its computation demands at the adaptation stage are prohibitive" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.66, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "MonoTTA" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.787, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.254 + ], + "angle": 0, + "content": "in Mono 3Det applications, particularly in autonomous driving. To enhance efficiency, Tent [36] and EATA [26] have been developed for Fully Test-Time Adaptation (Fully TTA) where only unlabeled test images and a well-trained model are provided. Besides, Ev-TTA [12] and SOD [35] devise TTA methods to handle the event-based object recognition and weakly supervised salient object detection, respectively. Considering the constraints on time of Mono 3Det, we explore the fully TTA paradigm which seeks to deal with OOD test data in real time." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.259, + 0.79, + 0.486 + ], + "angle": 0, + "content": "To investigate this paradigm for Mono 3Det, we dig into the detection outcomes for objects within test scenarios with variations or corruptions that are commonly caused by weather or cameras. Specifically, we directly apply the well-trained model to the validation set of KITTI which has been artificially injected with four distinct types of weather-related corruptions, namely Brightness, Frost, Fog, and Snow. Subsequently, we plot their distributions of detection scores (c.f. Fig. 1). It is observed that the detection scores of test objects tend to markedly decline as well as the high-score objects are scarce in the extreme scenario (Snow) when the well-trained model is directly applied to the scenarios with corruptions. This phenomenon indicates that: 1) The pre-trained Mono 3Det model struggles to discriminate between objects and the background within OOD test data, presenting as quantities of omissions and unconfident detections. 2) Directly applying existing fully TTA methods to Mono 3Det could only get suboptimal performance since they struggle to optimize the model without enough high-score (positive) detections, especially in certain extreme scenarios." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.493, + 0.79, + 0.844 + ], + "angle": 0, + "content": "To handle it in Mono 3Det, we propose a Monocular Test-Time Adaptation (MonoTTA) method, consisting of the reliability-driven adaptation and noise-guard adaptation strategies: 1) Reliability-driven adaptation. Specifically, data distribution shifts lead to omissions and noisy detections while our empirical analysis suggests that high detection score objects are still reliable (c.f. Fig. 3 (a)). Moreover, even if we only optimize the model via high-score objects (\\(e.g., \\geq 0.5\\)), both the numbers of low-score and high-score objects increase (c.f. Fig. 3 (b)). These investigations motivate us that exploiting high-score objects rather than all objects for model adaptation would be a more reliable way to alleviate data distribution shifts and discover potential objects. Hence, we develop a self-adaptive strategy for the identification of reliable high-score objects in test images and devise the adaptive optimization loss \\(\\mathcal{L}_{AO}\\) to exploit the reliable subset for model adaptation, alleviating the detection score decline issue of OOD test data and digging out more potential objects. 2) Noise-guard adaptation. In addition, data distribution shifts may also result in a scarcity of high-score objects, i.e., the majority of objects presenting low scores as the 'Snow' scenario in Fig. 3 (a). To this end, we develop a negative regularization term to make rational use of the numerous low-scoring objects in the Negative Learning manner [13]. On the one hand, the negative regularization term \\(\\mathcal{L}_{Nreg}\\) allows the model to conduct adaptation via numerous noisy low-scoring objects. Thus, the model can achieve more high-score objects after alleviating distribution shifts. On the other hand, this term also prevents the model from overfitting to noise and trivial solutions, i.e., assigning all classes of one object with high scores." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.516, + 0.129 + ], + "angle": 0, + "content": "Hongbin Lin, Yifan Zhang and et al." + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.147, + 0.594, + 0.162 + ], + "angle": 0, + "content": "We summarize the main contributions as follows:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.17, + 0.783, + 0.242 + ], + "angle": 0, + "content": "- To the best of our knowledge, we are the first to explore Fully Test-Time Adaptation to address OOD generalization problems for Mono 3Det. We show that the explored novel paradigm can bring significant improvements to Mono 3Det models in OOD test scenarios, e.g., \\(137\\%\\) and \\(244\\%\\) average performance gains across 13 types of OOD shifts on KITTI." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.245, + 0.783, + 0.318 + ], + "angle": 0, + "content": "- Our empirical investigation reveals an important insight that high-score objects maintain their reliability amidst various corruptions, while optimizing these high-score objects significantly boosts model confidence across all detections. This motivates the first effective test-time adaptation method (i.e., our MonoTTA) in Mono 3Det." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.32, + 0.783, + 0.394 + ], + "angle": 0, + "content": "- Extensive experiments on 13 types of corruptions of KITTI and 2 real scenarios (daytime \\(\\leftrightarrow\\) night) of nuScenes demonstrate the effectiveness of our MonoTTA in boosting existing Mono 3Det methods [29,48] to handle test-time OOD problems. Even for instance-level methods [40], MonoTTA also maintains sufficient improvement, which further confirms its applicability." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.17, + 0.783, + 0.394 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.417, + 0.388, + 0.433 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.448, + 0.785, + 0.717 + ], + "angle": 0, + "content": "We first review the literature on Monocular 3D Object Detection, and then discuss Source-free Domain Adaptation and Test-Time Adaptation methods. More discussions on Unsupervised Domain Adaptation [18,41,47] are in Appendix A. Monocular 3D Object Detection aims to perceive 3D objects from a single 2D image. Existing Mono 3Det methods could be divided into two groups according to the use of extra information. On the one hand, some existing methods leverage extra pre-trained depth estimation modules [6,39,49] to solve one of the most difficult problems in Mono 3Det, i.e., depth estimation from a single image. Other methods utilize LiDAR information, e.g., generating pseudo-LiDAR [23,31,37]. It is worth noting that Monoground [29] proposes to introduce the ground plane as prior information, and MonoNeRD [40] proposes to utilize scene geometric clues to enhance the detector's performance in the implicit reconstruction manner. On the other hand, some Mono 3Det methods try to detect 3D objects without extra data. For example, SMOKE [20] proposes to detect 3D objects as the key points estimation task. Then, Monoflex [48] improves this idea by providing a flexible definition of object centers, which unifies the centers of regular and truncated objects. GrooMeD-NMS [15] proposes a grouped mathematically differentiable Non-Maximal Suppression for Mono 3Det." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.784, + 0.84 + ], + "angle": 0, + "content": "Source-free Domain Adaptation (SFDA) aims to adapt the pre-trained source model to an unlabeled target domain without using the source data due to privacy issues [17, 30]. SF-UDA\\(^3\\)D [32] first explores the SFDA framework to adapt the PointRCNN 3D detector to target domains, which consists of pseudolabeling, reversible scale-transformations and motion coherency. Recently, the authors [10] seek to exploit the source model more reliably and propose an uncertainty-aware teacher-student framework to filter incorrect pseudo labels during model adaptation, alleviating the negative impact of label noise." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.66, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "MonoTTA" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.192 + ], + "angle": 0, + "content": "Nonetheless, SFDA assumes all target data to be known in advance and makes predictions after multiple epochs of optimization, which may not be viable for real-time applications due to computational or time constraints." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.193, + 0.785, + 0.403 + ], + "angle": 0, + "content": "Test-Time Adaptation (TTA) seeks to improve model performance on test data via model adaptation through test samples even if data shifts exist. Early TTA methods [19, 34] endeavor to conduct additional model optimization on training data by self-supervised objectives, and then adapt the well-trained model to the test data via self-supervised objectives. However, in Mono 3Det applications like autonomous driving, the computation demands of such methods are prohibitive. To solve this, Fully Test-Time Adaptation methods are developed to adapt the well-trained model, where only unlabeled test images are available. Specifically, certain methods [25, 27, 33] tackle data distribution shifts by adapting the batch normalization layer statistics, while others alleviate this issue either by the entropy minimization of test data [8, 36] or maximizing the prediction consistency of different augmentations [43-46]. As for object detection tasks, Ev-TTA [12] and SOD [35] try to handle the event-based object recognition and weakly supervised salient object detection offline, respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.405, + 0.785, + 0.465 + ], + "angle": 0, + "content": "However, existing fully TTA methods struggle to optimize the model and solve distribution shifts in Mono 3Det due to numerous false negative detections. To the best of our knowledge, our MonoTTA stands as the first fully TTA method that handles distribution shifts for Mono 3Det models in real time." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.491, + 0.579, + 0.509 + ], + "angle": 0, + "content": "3 Monocular Test-Time Adaptation" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.525, + 0.424, + 0.54 + ], + "angle": 0, + "content": "3.1 Problem Statement" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.553, + 0.785, + 0.719 + ], + "angle": 0, + "content": "Without loss of generality, we denote the pre-trained (or well-trained) model as \\( f_{\\Theta_0}(\\mathbf{x}) \\), which is achieved via training on labeled training images \\( \\{(\\mathbf{x}_i^s,\\mathbf{y}_i^s)\\}_{i = 1}^N \\). The training images follow the training distribution \\( P(\\mathbf{x}) \\) (i.e., \\( \\mathbf{x}^s\\sim P(\\mathbf{x}) \\)). Here, \\( \\Theta_0 \\) represents the parameters of the pre-trained model and \\( N \\) is the number of training data. During the training stage, the model is optimized to fit (or overfit) the training data. Then, at the test stage, the model will be able to perform well if the unlabeled test images \\( \\mathcal{D}_t = \\{\\mathbf{x}_i\\}_{i = 1}^{N_t} \\) follows the identical data distribution, i.e., \\( \\mathbf{x}\\sim P(\\mathbf{x}) \\) where \\( N_{t} \\) is the total number of test images. However, in real applications, it is possible for the pre-trained model to encounter Out-Of-Distribution (OOD) test samples due to prevalent natural corruptions, namely distribution shifts, i.e., \\( \\mathbf{x}\\sim Q(\\mathbf{x}) \\) and \\( P(\\mathbf{x})\\neq Q(\\mathbf{x}) \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.787, + 0.84 + ], + "angle": 0, + "content": "To address this issue, fully test-time adaptation [36] seeks to tackle distribution shifts by adapting the pre-trained model \\( f_{\\Theta_0}(\\mathbf{x}) \\) to unlabeled test images \\( \\{\\mathbf{x}_i\\}_{i=1}^{N_t} \\) in real time. To achieve this goal, existing methods typically endeavor to update the model through the minimization of unsupervised objectives defined on test samples by \\( \\min_{\\hat{\\Theta}} \\mathcal{L}(\\mathbf{x}; \\Theta) \\), where \\( \\mathbf{x} \\sim Q(\\mathbf{x}) \\) and \\( \\hat{\\Theta} \\subseteq \\Theta \\). Here, \\( \\hat{\\Theta} \\) denotes the subset of model parameters that should be updated (i.e., batch normalization layers following existing methods [26,36]). Most existing fully TTA methods focus on classification tasks, heavily relying on sufficient positive predictions for" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.516, + 0.129 + ], + "angle": 0, + "content": "Hongbin Lin, Yifan Zhang and et al." + }, + { + "type": "image", + "bbox": [ + 0.237, + 0.145, + 0.775, + 0.374 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.385, + 0.789, + 0.498 + ], + "angle": 0, + "content": "Fig. 2: An illustration of our MonoTTA. During the test phase, only the pre-trained model \\( f_{\\Theta_0}(\\mathbf{x}) \\) and unlabeled test images \\( \\{\\mathbf{x}_i\\}_{i=1}^{N_t} \\) are given. To conduct model adaptation, we initialize the model \\( f_{\\Theta}(\\mathbf{x}) \\) by \\( \\Theta_0 \\) and only update the parameters of batch normalization layers. When a batch of test images arrives, we first compute test object scores and refine the adaptive threshold \\( \\alpha \\) to select the reliable high-score objects, thereby optimizing \\( \\Theta \\) via the adaptive optimization loss \\( \\mathcal{L}_{AO} \\). Meanwhile, we devise a negative regularization term \\( \\mathcal{L}_{Nreg} \\) to facilitate the model to avoid overfitting to noise and trivial solutions. Here, Ped. and Cyc. represent Pedestrian and Cyclist in KITTI." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.518, + 0.789, + 0.655 + ], + "angle": 0, + "content": "model adaptation. Nevertheless, there is a significant difference between conventional classification tasks and Mono 3Det. As previously indicated, the detection scores of test images \\(\\mathbf{x}\\) derived from \\(f_{\\Theta_0}(\\mathbf{x})\\) are prone to markedly decrease in the presence of corruptions as shown in Fig. 1, leading to severe omissions (numerous false negatives) in Mono 3Det. In such circumstances, the scarcity of positive detections presents a significant challenge for model adaptation to test distributions while adapting the model with unreliable low-score detections may significantly introduce the noise. Therefore, existing fully TTA methods tend to fail in the OOD generalization problems of Mono 3Det." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.678, + 0.391, + 0.692 + ], + "angle": 0, + "content": "3.2 Overall Scheme" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.789, + 0.795 + ], + "angle": 0, + "content": "After thoroughly examining the characteristics and challenges of Mono 3Det, we introduce a Monocular Test-Time Adaptation (MonoTTA) method to address the OOD problems for Mono 3Det models, which seeks to solve the object score declining issue within unlabeled OOD test data. As shown in Fig. 2, MonoTTA consists of two strategies: 1) Reliability-driven adaptation and 2) Noise-guard adaptation. We first briefly introduce the two strategies below." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.789, + 0.842 + ], + "angle": 0, + "content": "First, we develop a reliability-driven adaptation strategy (c.f. Section 3.3) to conduct reliable model adaptation for OOD test data based on dependable test objects. Our empirical investigations inspire us to exploit those relatively reliable" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.66, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "MonoTTA" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.126 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.258, + 0.145, + 0.48, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.373, + 0.281, + 0.387, + 0.291 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.487, + 0.149, + 0.739, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.625, + 0.28, + 0.641, + 0.29 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.294, + 0.788, + 0.392 + ], + "angle": 0, + "content": "Fig. 3: Based on MonoGround [29], we conduct two empirical studies (Car, KITTI), with the 3D IoU threshold of 0.5. (a) We visualize the accuracy of the objects across varied scoring ranges, which shows that the accuracy of objects with high scores remains relatively stable even in the presence of diverse corruptions (Ideal means in-distribution scenarios). (b) We visualize the number of low & high-score objects before and after optimization. Although only high-score objects are optimized, the model treats low-score objects with more confidence." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.412, + 0.785, + 0.472 + ], + "angle": 0, + "content": "test objects for alleviating distribution shifts, thereby discovering more potential objects. To this end, MonoTTA excludes unreliable test objects involving an adaptive threshold \\(\\alpha\\) for any unlabeled test data. Subsequently, the model is optimized by the adaptive optimization loss \\(\\mathcal{L}_{AO}\\) via the selected reliable objects." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.473, + 0.788, + 0.608 + ], + "angle": 0, + "content": "Second, we tend to utilize plenty of low-score objects to adapt the model in an indirect manner instead of directly optimizing the model since low-score objects are noisy. Hence, we devise a noise-guard adaptation strategy (c.f. Section 3.4) to prevent the model from overfitting to noisy predictions and falling into trivial solutions. Specifically, we randomly choose one of the negative classes of low-score objects and minimize the scores (e.g., score 0.03 of [0.10, 0.12, 0.03] c.f. Fig. 2) after simply filtering out extremely low-score detections. Even though the positive class is noisy (i.e., score 0.12), this term is capable of optimizing the model indirectly, i.e., learn to deny the negative category of the object." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.609, + 0.641, + 0.624 + ], + "angle": 0, + "content": "Overall, the training scheme of MonoTTA is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.402, + 0.634, + 0.787, + 0.66 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\hat {\\Theta}} \\mathcal {L} _ {A O} (\\hat {\\Theta}) + \\lambda \\mathcal {L} _ {N r e g} (\\hat {\\Theta}), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.671, + 0.788, + 0.703 + ], + "angle": 0, + "content": "where \\(\\lambda\\) is the balance hyper-parameter. The pseudo-code of MonoTTA is summarized in Algorithm 1." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.724, + 0.51, + 0.74 + ], + "angle": 0, + "content": "3.3 Reliability-Driven Adaptation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.785, + 0.795 + ], + "angle": 0, + "content": "To identify dependable test objects and conduct test-time model adaptation, we propose a reliability-driven adaptation strategy that consists of two components: 1) Reliable object identification and 2) Adaptive model optimization." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Reliable object identification. When unlabeled test images arrive, it is difficult for the pre-trained model to get accurate detections on OOD test data due to the decline in detection scores. To resolve this, we dig into the detection" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.516, + 0.129 + ], + "angle": 0, + "content": "Hongbin Lin, Yifan Zhang and et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.36 + ], + "angle": 0, + "content": "accuracy of the pre-trained model in test scenarios with various corruptions. Specifically, based on MonoGround [29], we visualize the detection precision for the Car category within the KITTI dataset across varied scoring ranges, with the 3D Intersection over Union (IoU) threshold of 0.5. As shown in Fig. 3 (a), we find that high-score objects are more reliable and relatively stable even in the presence of diverse corruptions. Following this, we propose to select reliable high-score objects to conduct model adaptation via an adaptive threshold \\(\\alpha\\). Specifically, we exploit the model \\(f_{\\Theta}(\\cdot)\\) (initialized by \\(\\Theta_0\\)) to infer a batch of test images \\(\\{\\mathbf{x}_b\\}_{b=1}^B\\) and obtain the heatmap \\(\\mathbf{h}_i\\) of each image \\(\\mathbf{x}_i\\) by \\(\\mathbf{h}_i = f_{\\Theta}(\\mathbf{x}_i)\\), where \\(i\\) ranges from 0 to \\(B-1\\) and \\(B\\) denotes the batch size. Then, we could achieve object score maps \\(\\mathbf{s}_i \\in \\mathbb{R}^{B \\times N_m}\\) like the peaks in \\(\\mathbf{h}_i\\) [7] after normalization, where \\(K\\) and \\(N_m\\) denotes the number of classes and the maximum of detected objects, respectively. With the score map \\(\\mathbf{s}_i\\), we update the adaptive threshold \\(\\alpha_t\\) at the iteration \\(t\\) in the exponential moving average manner by:" + }, + { + "type": "equation", + "bbox": [ + 0.368, + 0.377, + 0.786, + 0.412 + ], + "angle": 0, + "content": "\\[\n\\alpha_ {t} = \\left\\{ \\begin{array}{l l} \\gamma , & \\text {i f} t = 1 \\\\ \\beta \\bar {m} _ {t} + (1 - \\beta) \\alpha_ {t - 1}, & \\text {i f} t > 1 \\end{array} \\right., \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.378, + 0.436, + 0.786, + 0.48 + ], + "angle": 0, + "content": "\\[\n\\bar {m} _ {t} = \\frac {1}{B} \\sum_ {i = 1} ^ {B} \\frac {\\sum_ {j = 1} ^ {N _ {m}} s _ {i j} \\cdot \\mathbb {I} \\left(s _ {i j} \\geq \\gamma\\right)}{\\sum_ {j = 1} ^ {N _ {m}} \\mathbb {I} \\left(s _ {i j} \\geq \\gamma\\right)}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.492, + 0.788, + 0.583 + ], + "angle": 0, + "content": "In Eqn. (2), \\(\\bar{m}_t\\) denotes the average score of all detected objects in a single batch of \\(B\\) test samples at iteration \\(t\\), while \\(\\beta \\in [0,1]\\) is a decay coefficient. As for Eqn. (3), \\(s_{ij} \\in (0,1)\\) denotes the score of the \\(j\\)-th object in the \\(i\\)-th test image while \\(\\mathbb{I}(\\cdot)\\) is the indicator function. Note that \\(\\gamma \\in \\mathbb{R}^1\\) is a pre-defined object detection threshold adopted from existing methods at their original inference stage as well as \\(N_m\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.585, + 0.788, + 0.691 + ], + "angle": 0, + "content": "Adaptive model optimization. As shown in Fig. 3 (b), the optimization of high-score objects can also enhance the confidence of the model for relatively low-score objects. It motivates us that exploiting high-score objects rather than all objects for model adaptation would be a more reliable way to learn from OOD test data. Therefore, with the adaptive threshold \\(\\alpha_{t}\\), we select the reliable subset of high scores from \\(\\mathbf{s}_i\\) and calculate the adaptive optimization loss \\(\\mathcal{L}_{AO}\\) to adapt the model by:" + }, + { + "type": "equation", + "bbox": [ + 0.357, + 0.706, + 0.787, + 0.75 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {A O} = - \\frac {1}{B} \\sum_ {i = 1} ^ {B} \\sum_ {j = 1} ^ {N _ {m}} \\log \\left(s _ {i j} \\cdot \\mathbb {I} \\left(s _ {i j} \\geq \\alpha_ {t}\\right)\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.842 + ], + "angle": 0, + "content": "The adaptive optimization loss \\(\\mathcal{L}_{AO}\\) alleviates the potential data distribution shifts in OOD test data by allowing the model to confidently identify high-score test objects. As we mentioned, it solves the score decline issue of OOD test data in a more reliable way rather than directly optimizing all test objects, thereby avoiding overfitting to noise and discovering more potential test objects." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.66, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "MonoTTA" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.147, + 0.611, + 0.162 + ], + "angle": 0, + "content": "Algorithm 1 The pipeline of the proposed MonoTTA" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.167, + 0.786, + 0.196 + ], + "angle": 0, + "content": "Require: Unlabeled test data \\(\\mathcal{D}_t = \\{\\mathbf{x}_i\\}_{i=1}^{N_t}\\); Pre-trained model \\(f_{\\Theta_0}(\\mathbf{x})\\); Batch size \\(B\\); Parameters \\(\\lambda, \\beta, \\eta\\)." + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.196, + 0.491, + 0.21 + ], + "angle": 0, + "content": "1: for a batch images \\(\\{\\mathbf{x}_b\\}_{b = 1}^B\\) in \\(\\mathcal{D}_t\\) do" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.21, + 0.606, + 0.223 + ], + "angle": 0, + "content": "2: Update the adaptive threshold \\(\\alpha\\) based on Eqn. (2);" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.223, + 0.663, + 0.238 + ], + "angle": 0, + "content": "3: Calculate adaptive optimization loss \\(\\mathcal{L}_{AO}\\) based on Eqn. (4);" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.238, + 0.687, + 0.252 + ], + "angle": 0, + "content": "4: Calculate negative regularization term \\(\\mathcal{L}_{Nreg}\\) based on Eqn. (6);" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.252, + 0.483, + 0.266 + ], + "angle": 0, + "content": "5: Update \\(\\hat{\\Theta}\\) by optimizing Eqn. (1)" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.266, + 0.301, + 0.278 + ], + "angle": 0, + "content": "6: end for" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.278, + 0.517, + 0.293 + ], + "angle": 0, + "content": "7: return Detection Results for all \\(\\mathbf{x} \\in \\mathcal{D}_t\\)." + }, + { + "type": "list", + "bbox": [ + 0.223, + 0.196, + 0.687, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.326, + 0.466, + 0.34 + ], + "angle": 0, + "content": "3.4 Noise-Guard Adaptation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.35, + 0.788, + 0.579 + ], + "angle": 0, + "content": "Through the optimization of adaptive optimization loss \\(\\mathcal{L}_{AO}\\), the model is refined to yield more confident detection outcomes. Nevertheless, high-scoring objects may be scarce due to the distribution shifts, making the adaptation procedure difficult. Meanwhile, exclusive reliance on \\(\\mathcal{L}_{AO}\\) for adaptation may result in trivial solutions, whereby the model indiscriminately assigns high scores to all categories. Previous studies [13, 14] indicate that deep neural networks could learn from noisy pseudo labels in classification tasks through negative learning. Thus, MonoTTA proposes to learn from noisy low-score objects in a negative learning manner for Mono 3Det. Specifically, we denote \\(\\hat{\\mathbf{s}}_i\\in \\mathbb{R}^{B\\times N_m\\times K}\\) as the multi-class score map for the test image \\(\\mathbf{x}_i\\), i.e. the multi-class score map \\(\\hat{\\mathbf{s}}_i\\) contains not only the highest score of objects but also the scores for other classes. Here \\(\\mathbf{s}_i = \\arg \\max_k\\hat{\\mathbf{s}}_{ik}\\) where \\(k\\) is the class index. Employing a simple constant threshold \\(\\eta\\), we filter out extremely low-score objects and then randomly select a negative class \\(\\bar{k}\\) for each object. Subsequently, we compute the regularization loss \\(e_k\\) for each class \\(k\\) with relatively low object scores \\(s_{ij}\\in [\\eta ,\\alpha_t)\\) by:" + }, + { + "type": "equation", + "bbox": [ + 0.351, + 0.588, + 0.786, + 0.632 + ], + "angle": 0, + "content": "\\[\ne _ {k} = - \\sum_ {i = 1} ^ {B} \\sum_ {j = 1} ^ {N _ {m}} \\bar {y} _ {i j} \\log \\left(1 - s _ {i j \\bar {k}} \\cdot \\mathbb {I} (\\bar {k} = k)\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.642, + 0.787, + 0.673 + ], + "angle": 0, + "content": "where \\(\\bar{y}_{ij} = 1 - s_{ij\\bar{k}}\\) is a constant weight. We further define \\(n_k\\) as the frequency of negative scores corresponding to class \\(k\\) in the test batch and balance \\(e_k\\) by:" + }, + { + "type": "equation", + "bbox": [ + 0.44, + 0.683, + 0.786, + 0.725 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {N r e g} = \\sum_ {k = 1} ^ {K} \\frac {e _ {k}}{n _ {k}}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.843 + ], + "angle": 0, + "content": "As we mentioned before, detections with low and intermediate scores tend to be more noisy (c.f. Fig. 3 (a)), which is unreliable for direct model adaptation. To this end, we introduce the regularization term \\(\\mathcal{L}_{Nreg}\\) to leverage noisy low-score detections and improve the model for assimilating potentially accurate information (c.f. Term 2 in Fig. 2). Moreover, this term also prevents the model from trivial solutions, i.e., indiscriminately assigning high scores to all classes of a single object (e.g., TENT [36]). In addition, high-score objects may be absent" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.516, + 0.129 + ], + "angle": 0, + "content": "Hongbin Lin, Yifan Zhang and et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.784, + 0.174 + ], + "angle": 0, + "content": "Table 1: Comparison with baselines on the KITTI-C validation set, severity level 1 regarding Mean \\(AP_{3D|R_{40}}\\). The bold number indicates the best result." + }, + { + "type": "table", + "bbox": [ + 0.242, + 0.18, + 0.763, + 0.289 + ], + "angle": 0, + "content": "
Car, IoU @ 0.7, 0.5, 0.5
MethodNoiseBlurWeatherDigitalAvg.
Gauss.ShotImpul.Defoc.GlassMotionSnowFrostFogBrit.Contr.PixelSat.
Monoflex [48]3.847.485.312.593.7311.050.237.777.5724.876.9228.1631.4610.84
• BN adaptation [33]13.5821.9318.7815.878.5924.325.4221.4524.6331.8030.5841.0430.7122.21
• TENT [36]17.8027.0923.1821.6611.9028.756.8426.5830.7835.6534.7241.7135.9126.35
• EATA [26]16.6726.4225.0722.5413.2327.737.8726.5831.1035.3935.2841.4036.7226.62
• ActMAD [24]12.2621.4920.4711.107.5323.795.5719.1126.5426.2423.9238.5133.7520.79
• MonoTTA (Ours)21.1528.6526.6425.9119.2631.4812.4330.2433.7536.8436.8341.9738.1329.48
MonoGround [29]2.404.103.313.712.678.130.225.544.5925.374.0033.5728.089.67
• BN adaptation [33]13.4923.5219.6916.337.6123.997.9820.7124.0031.3429.0343.0632.9922.60
• TENT [36]17.9429.6019.9023.4513.9029.3910.3226.6533.3535.9636.3943.3537.7927.54
• EATA [26]16.0326.0818.0820.2812.4027.379.2223.7929.4933.6532.5843.6136.0025.28
• ActMAD [24]13.6522.3920.5614.998.1721.099.1318.7719.5228.3425.3242.6331.5821.24
• MonoTTA (Ours)26.1333.1128.6030.3825.4832.4418.7232.6037.7537.8739.5743.6737.9832.64
" + }, + { + "type": "table", + "bbox": [ + 0.242, + 0.289, + 0.762, + 0.394 + ], + "angle": 0, + "content": "
Pedestrian, IoU @ 0.5, 0.25
MethodNoiseBlurWeatherDigitalAvg.
Gauss.ShotImpul.Defoc.GlassMotionSnowFrostFogBrit.Contr.PixelSat.
Monoflex0.191.620.323.728.476.220.004.272.259.192.081.839.113.79
• BN adaptation [33]6.218.209.207.835.357.522.896.479.249.129.9312.739.768.03
• TENT [36]6.027.969.577.756.068.632.636.719.9110.2610.5512.3310.278.36
• EATA [26]6.057.969.747.936.068.012.486.249.949.7010.0212.4110.128.21
• ActMAD [24]2.414.274.681.461.626.481.174.254.846.426.198.297.074.55
• MonotTA (Ours)6.548.419.397.637.128.993.257.6410.2610.5510.0613.2810.668.75
MonoGround0.610.930.738.417.578.030.003.191.3913.821.833.707.264.42
• BN adaptation [33]5.097.087.777.636.639.452.246.728.409.7211.0616.7012.438.53
• TENT [36]7.2710.1010.028.538.3011.033.548.739.1111.7412.1217.7015.0310.25
• EATA [26]5.928.058.588.127.5910.953.317.9110.1011.0210.9517.3014.349.55
• ActMAD [24]3.965.756.485.615.737.141.585.375.747.278.1013.129.276.55
• MonotTA (Ours)8.5811.1811.799.229.4013.204.839.9514.4612.8513.2517.1314.8511.59
" + }, + { + "type": "table", + "bbox": [ + 0.242, + 0.394, + 0.762, + 0.499 + ], + "angle": 0, + "content": "
Cyclist, IoU @ 0.5, 0.25, 0.25
MethodNoiseBlurWeatherDigitalAvg.
Gauss.ShotImpul.Defoc.GlassMotionSnowFrostFogBrit.Contr.PixelSat.
Monoflex0.281.640.470.594.973.600.007.423.8113.073.793.808.393.99
• BN adaptation [33]2.396.264.365.786.769.091.708.539.1612.9111.2610.5511.027.67
• TENT [36]2.727.945.636.277.209.491.078.9410.9612.7512.729.6411.288.20
• EATA [26]2.337.465.467.197.237.511.248.6010.1812.8611.0210.8910.307.87
• ActMAD [24]2.225.324.362.291.575.640.485.686.4610.569.319.977.805.51
• MonotTA (Ours)3.017.245.987.006.099.511.4510.6311.2212.8511.8511.5910.818.40
MonoGround0.120.460.480.330.722.030.000.560.355.550.522.085.241.42
• BN adaptation [33]1.763.582.083.613.055.410.573.824.476.305.6011.028.874.63
• TENT [36]1.794.853.003.363.496.050.494.436.207.196.5010.439.235.15
• EATA [26]1.894.202.414.133.175.730.384.066.276.416.1410.937.724.88
• ActMAD [24]1.142.841.533.33.083.150.322.743.476.365.759.696.953.87
• MonotTA (Ours)3.935.784.555.434.706.090.694.667.537.697.7411.719.436.15
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.519, + 0.788, + 0.626 + ], + "angle": 0, + "content": "in certain extreme scenarios. For instance, the Snow scenario (c.f. Fig. 3 (a)) lacks objects with scores exceeding 0.4 when the pre-trained model is directly applied. Under such a circumstance, \\(\\mathcal{L}_{Nreg}\\) plays a more important role in model adaptation since it can alleviate distribution shifts even only low-score objects, i.e., deny the negative category. In other words, \\(\\mathcal{L}_{Nreg}\\) enables the model to alleviate distribution shifts and achieve more relatively high-score objects, thereby laying a crucial foundation for \\(\\mathcal{L}_{AO}\\) in extremely challenging scenarios." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.653, + 0.377, + 0.671 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.688, + 0.787, + 0.749 + ], + "angle": 0, + "content": "We conduct experiments based on KITTI [9] and nuScenes [1]. The results presented in this manuscript represent the average value across three difficulty levels, i.e., Easy, Moderate, Hard. Note that we provide more results of higher severity levels of KITTI and more detailed results in Appendix \\(D\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Datasets. For KITTI, we adopt the protocol established by Monoflex [48] to split the training images into the training set (3712) and validation set (3769) to perform model training and adaptation, respectively. Following the official KITTI evaluation criteria, we evaluate detection results on three levels of difficulty. Then, we construct the KITTI-C dataset through the incorporation of 13 distinct types of data corruptions [11] to the validation set and each corruption" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.66, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "MonoTTA" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image_caption", + "bbox": [ + 0.331, + 0.144, + 0.678, + 0.155 + ], + "angle": 0, + "content": "Average performance of 13 corruptions on KITTI-C Validation set, severity level 2" + }, + { + "type": "image", + "bbox": [ + 0.274, + 0.157, + 0.727, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.297, + 0.784, + 0.325 + ], + "angle": 0, + "content": "Fig. 4: We visualize the comparison with baselines on the KITTI-C validation set, severity level 2 regarding Mean \\(AP_{3D|R_{40}}\\). The bold number indicates the best result." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.348, + 0.784, + 0.379 + ], + "angle": 0, + "content": "has 5 severity levels. The model is trained on the original training set and then tested on the KITTI-C validation set within one of the corruptions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.388, + 0.785, + 0.493 + ], + "angle": 0, + "content": "As for nuScenes, we adopt the front-view images and construct the Daytime and Night scenarios via their scene descriptions following [21]. Specifically, there exist \\(24.7\\mathrm{k} / 5.4\\mathrm{k}\\) train/test images in Daytime while \\(3.3\\mathrm{k} / 0.6\\mathrm{k}\\) train/test images in Night. Based on these splits, we construct two real-world adaptation tasks, i.e., Daytime \\(\\rightarrow\\) Night and Night \\(\\rightarrow\\) Daytime. For simplification, we transfer the nuScenes dataset into the KITTI format and only consider the Car category. More details of data construction are provided in Appendix \\(\\pmb{B}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.503, + 0.785, + 0.624 + ], + "angle": 0, + "content": "Implementation details. We implement our method and other baselines in PyTorch [28]. In MonoTTA, we conduct model adaptation based on the public pre-trained weights and the parameter settings provided by their authors [29, 40, 48]. Besides, we employ the Stochastic Gradient Descent (SGD) optimizer with a half learning rate of the initial rate used in base training over different methods, a momentum of 0.9 and a batch size of 16 for KITTI, 4 for nuScenes. Parameters \\(\\lambda\\), \\(\\beta\\), \\(\\eta\\) are assigned default values of 1, 0.1, and 0.05, respectively. More training details of MonoTTA are provided in Appendix \\(C\\)." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.634, + 0.785, + 0.755 + ], + "angle": 0, + "content": "Compared methods. Based on three typical or state-of-the-art (SOTA) Mono 3Det methods [29, 40, 48], we fully compare MonoTTA with following methods: 1) source-only, i.e., directly apply the pre-trained model to the test data within corruptions; 2) BN adaptation [33] updates batch normalization statistics via target data; 3) TENT [36] minimizes the entropy loss of test data; 4) EATA [26] identifies reliable samples to update the model by entropy loss minimization. Here we compare MonoTTA with its variant Efficient Test-time Adaptation. 5) ActMAD [24] matches activation statistics for the distribution alignment." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.765, + 0.785, + 0.84 + ], + "angle": 0, + "content": "Evaluation protocols. In order to fully evaluate the proposed method, we report our experimental results in the Average Precision (AP) for 3D bounding boxes, denoted as \\( AP_{3D|R_{40}} \\). The results present the mean values across three levels of difficulty and the Intersection over Union (IoU) thresholds are set to 0.7, 0.5, 0.5 for Cars and 0.5, 0.25, 0.25 for Pedestrians and Cyclists." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.516, + 0.129 + ], + "angle": 0, + "content": "Hongbin Lin, Yifan Zhang and et al." + }, + { + "type": "image_caption", + "bbox": [ + 0.334, + 0.145, + 0.672, + 0.156 + ], + "angle": 0, + "content": "Performance on Daytime to Night and Night to Daytime scenarios of nuScenes" + }, + { + "type": "image", + "bbox": [ + 0.272, + 0.158, + 0.729, + 0.294 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.305, + 0.788, + 0.336 + ], + "angle": 0, + "content": "Fig. 5: Comparison with baselines on Daytime \\(\\rightarrow\\) Night and Night \\(\\rightarrow\\) Daytime of nuScenes, regarding Mean \\(AP_{3D|R_{40}}\\). The **bold** number indicates the best result." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.355, + 0.569, + 0.372 + ], + "angle": 0, + "content": "4.1 Comparisons with Previous Methods" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.381, + 0.789, + 0.58 + ], + "angle": 0, + "content": "We first compare our MonoTTA with previous methods in severity level 1 of KITTI-C. The results are reported in Table 1, which gives the following observations: 1) Due to distribution shifts, directly applying the pre-trained model to the test data (i.e., source-only) suffers severe performance degradation in all categories. 2) Existing TTA methods are able to mitigate the negative effect of distribution shifts for Mono 3Det to some degree. However, they only achieve suboptimal performance since they tend to increase the scores of all positive detections, containing severe noise. 3) MonoTTA consistently outperforms all compared methods over all categories within various base models in terms of mean \\( AP_{3D|R_{40}} \\). Specifically, MonoTTA achieves the best or comparable performance in all categories under all corruptions, attaining a large performance gain over TENT and EATA (e.g., improving an average \\( AP_{3D|R_{40}} \\) about 5.1 and 7.4 of the Car category based on MonoGround)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.602, + 0.62, + 0.618 + ], + "angle": 0, + "content": "4.2 More Severe Corruption and Real Scenario" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.628, + 0.788, + 0.779 + ], + "angle": 0, + "content": "On the one hand, to fully validate the effectiveness of our MonoTTA, we visualize experimental results under more severe corruption conditions at severity level 2 as shown in Fig. 4, which clearly gives additional observations: 1) With the escalation of severity level, the pre-trained models suffer a larger performance decline within various corruptions, enlarging the difficulty of TTA. 2) The performance improvements of existing TTA methods become relatively limited, particularly in Pedestrian and Cyclist classes. 3) Even if the tasks are more challenging, MonoTTA still stably obtains the best average performance within all corruptions since \\(\\mathcal{L}_{Nreg}\\) plays an important role in alleviating distribution shifts for certain extreme scenarios, i.e., only low-score objects exist." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "On the other hand, we further validate different methods within real scenarios as shown in Fig. 5. The experimental results also give the following observations: 1) Under real corruptions, the pre-trained model still suffers severe performance degradation due to the data distribution shifts. 2) TENT tends to increase the" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.66, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "MonoTTA" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.187 + ], + "angle": 0, + "content": "Table 2: Comparison with baselines based on the instance-level method (i.e., batch size is 1) on the KITTI-C validation set, regarding Mean \\(AP_{3D|R_{40}}\\) and the severity levels 1 and 2. The bold number indicates the best result." + }, + { + "type": "table", + "bbox": [ + 0.331, + 0.199, + 0.673, + 0.251 + ], + "angle": 0, + "content": "
MethodLevel 1Level 2
CarPedes.CyclistAvg.CarPedes.CyclistAvg.
MonoNeRD19.845.962.279.3613.023.831.616.15
BN adaptation [33]30.738.853.8114.4626.476.942.9112.11
TENT [36]35.729.994.7516.8231.857.813.4014.35
EATA [26]34.6010.044.2016.2830.667.863.2813.93
MonoTTA (Ours)37.4010.394.3517.3833.998.253.3315.19
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.263, + 0.784, + 0.293 + ], + "angle": 0, + "content": "Table 3: Based on MonoGround [29], we conduct ablation studies of \\(\\mathcal{L}_{AO}\\) and \\(\\mathcal{L}_{Nreg}\\) on the KITTI-C validation set, regarding \\(AP_{3D|R_{40}}\\)." + }, + { + "type": "table", + "bbox": [ + 0.33, + 0.303, + 0.673, + 0.349 + ], + "angle": 0, + "content": "
Backbone\\( {\\mathcal{L}}_{AO} \\)\\( {\\mathcal{L}}_{Nreg} \\)Level 1Level 2
CarPedes.CyclistAvg.CarPedes.CyclistAvg.
✓9.674.421.425.175.722.430.953.03
✓✓27.9410.175.1014.4021.277.793.8210.96
✓✓23.268.604.6012.1518.936.813.559.77
✓✓✓32.6411.596.1516.7925.968.784.1212.85
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.367, + 0.788, + 0.445 + ], + "angle": 0, + "content": "confidence of all positive detections and thus overfits to noise, i.e., failing to handle the extremely challenging task \\(\\mathrm{N} \\rightarrow \\mathrm{D}\\). 3) EATA still achieves sub-optimal performance while our MonoTTA brings sufficient average performance improvement on both MonoFlex (6.23 mAP) and MonoGround (8.26 mAP), maintaining the best performance in real scenarios. Detailed results are in Appendix \\(D\\)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.466, + 0.659, + 0.481 + ], + "angle": 0, + "content": "4.3 Application to Instance-Level Inference Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.49, + 0.789, + 0.673 + ], + "angle": 0, + "content": "In this section, we seek to investigate whether the proposed MonoTTA can be used to effectively enhance Mono 3Det methods which only process a single image sequentially, i.e., \\( B = 1 \\). To be specific, it may be crucial for Mono 3Det methods to make immediate decisions based on the most recent scene (image) in real-world scenarios like autonomous driving. To this end, it is essential for TTA methods devised for Mono 3Det to allow a single image as input and then conduct model adaptation. To validate it, we integrate MonoTTA into the SOTA Mono 3Det method namely MonoNeRD [40] which accepts one image each time at the test phase. As shown in Table 2, MonoTTA achieves the best or comparable performance across all categories at both severity levels 1 and 2, illustrating the applicability of our method to boost these approaches for handling OOD test data. Detailed results can be also found in Appendix \\( D \\)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.694, + 0.613, + 0.71 + ], + "angle": 0, + "content": "4.4 Ablation Studies and Quantitative Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.719, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Ablation studies. To examine the effectiveness of the losses in MonoTTA, we show the results of the models optimized by different losses. As shown in Table 3, introducing \\(\\mathcal{L}_{AO}\\) or \\(\\mathcal{L}_{Nreg}\\) enhances the model performance compared to directly applying the pre-trained model (i.e., source only). On the one hand, such a result verifies that our strategy is able to alleviate the score decline issue in test OOD data with high-score objects. On the other hand, introducing \\(\\mathcal{L}_{Nreg}\\) could obtain a higher average \\(AP_{3D|R_{40}}\\) value, which also verifies that the negative regularization term is able to enhance the pre-trained model even only" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.516, + 0.129 + ], + "angle": 0, + "content": "Hongbin Lin, Yifan Zhang and et al." + }, + { + "type": "image", + "bbox": [ + 0.274, + 0.144, + 0.731, + 0.361 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.372, + 0.788, + 0.415 + ], + "angle": 0, + "content": "Fig.6: Qualitative results of baselines and the proposed MonoTTA based on MonoGround [29]. We visualize the results on KITTI-C validation set, where predicted cars, pedestrians and cyclists are in lime green, sky blue and yellow, respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.428, + 0.784, + 0.458 + ], + "angle": 0, + "content": "with low-score objects. When combining the losses (i.e. \\(\\mathcal{L}_{AO}\\), \\(\\mathcal{L}_{Nreg}\\)) together, we obtain the best performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.459, + 0.785, + 0.549 + ], + "angle": 0, + "content": "Quantitative Results. We provide visualizations within Snow, Fog and Frost based on Monoground as shown in Fig. 6. It is evident that the source-only setting suffers severe omissions, while BN adaptation, TENT and EATA alleviate distribution shifts to some degree and give more detections. As for our MonoTTA, it can produce superior detections even in severe conditions, including fewer omissions and accurate detections as highlighted by red circles." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.57, + 0.36, + 0.585 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.788, + 0.749 + ], + "angle": 0, + "content": "In this paper, we propose a monocular test-time adaptation method to improve the pre-trained model on the shifted test data for monocular 3D object detection. Specifically, our method consists of two strategies: 1) Reliability-driven adaptation. To discover more potential objects, we devise a self-adaptive strategy to identify reliable objects for adaptive model adaptation. 2) Noise-guard adaptation. To avoid overfitting to noise and trivial solutions, we devise the negative regularization term to mitigate the negative effects of noisy detections and alleviate distribution shifts. Experiments on KITTI-C and nuScenes datasets demonstrate the effectiveness of MonoTTA in handling fully test-time adaptation for monocular 3D object detection." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Future directions. 1) Our work focuses on 2D images, while future studies could explore 3D information in handling distribution shifts. 2) This work explores TTA by assuming one OOD distribution at a time, where the forgetting issue is not severe as the source model weights are recoverable. Exploring scenarios with dynamically OOD distributions offers a compelling future direction, where the forgetting issue would become more severe." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.66, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "MonoTTA" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.177, + 0.79, + 0.358 + ], + "angle": 0, + "content": "This work was supported by NSFC with Grant No. 62293482, by the Basic Research Project No. HZQB-KCZYZ-2021067 of Hetao Shenzhen HK S&T Cooperation Zone, by Shenzhen General Program No. JCYJ20220530143600001, by Shenzhen-Hong Kong Joint Funding No. SGDX20211123112401002, by the Shenzhen Outstanding Talents Training Fund 202002, by Guangdong Research Project No. 2017ZT07X152 and No. 2019CX01X104, by the Guangdong Provincial Key Laboratory of Future Networks of Intelligence (Grant No. 2022B1212010001), by the Guangdong Provincial Key Laboratory of Big Data Computing, CHUK-Shenzhen, by the NSFC 61931024&12326610, by the Key Area R&D Program of Guangdong Province with grant No. 2018B030338001, by the Shenzhen Key Laboratory of Big Data and Artificial Intelligence (Grant No. ZDSYS201707251409055), and by Tencent & Huawei Open Fund." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.381, + 0.323, + 0.397 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.411, + 0.787, + 0.468 + ], + "angle": 0, + "content": "1. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuscenes: A multimodal dataset for autonomous driving. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11621-11631 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.468, + 0.787, + 0.509 + ], + "angle": 0, + "content": "2. Chen, X., Kundu, K., Zhang, Z., Ma, H., Fidler, S., Urtasun, R.: Monocular 3d object detection for autonomous driving. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2147-2156 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.509, + 0.787, + 0.55 + ], + "angle": 0, + "content": "3. Chen, X., Kundu, K., Zhu, Y., Berneshawi, A.G., Ma, H., Fidler, S., Urtasun, R.: 3d object proposals for accurate object class detection. Advances in neural information processing systems 28 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.55, + 0.787, + 0.592 + ], + "angle": 0, + "content": "4. Chen, Y., Tai, L., Sun, K., Li, M.: Monopair: Monocular 3d object detection using pairwise spatial relationships. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12093-12102 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.592, + 0.787, + 0.633 + ], + "angle": 0, + "content": "5. Chen, Y., Liu, J., Zhang, X., Qi, X., Jia, J.: Voxelnext: Fully sparse voxelnet for 3d object detection and tracking. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21674-21683 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.633, + 0.787, + 0.688 + ], + "angle": 0, + "content": "6. Ding, M., Huo, Y., Yi, H., Wang, Z., Shi, J., Lu, Z., Luo, P.: Learning depth-guided convolutions for monocular 3d object detection. In: Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition workshops. pp. 1000-1001 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.689, + 0.787, + 0.731 + ], + "angle": 0, + "content": "7. Duan, K., Bai, S., Xie, L., Qi, H., Huang, Q., Tian, Q.: Centernet: Keypoint triplets for object detection. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6569-6578 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.731, + 0.787, + 0.772 + ], + "angle": 0, + "content": "8. Fleuret, F., et al.: Test time adaptation through perturbation robustness. In: NeurIPS 2021 Workshop on Distribution Shifts: Connecting Methods and Applications (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.772, + 0.787, + 0.813 + ], + "angle": 0, + "content": "9. Geiger, A., Lenz, P., Urtasun, R.: Are we ready for autonomous driving? the kitti vision benchmark suite. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3354-3361. IEEE (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.813, + 0.787, + 0.84 + ], + "angle": 0, + "content": "0. Hegde, D., Kilic, V., Sindagi, V., Cooper, A.B., Foster, M., Patel, V.M.: Source-free unsupervised domain adaptation for 3d object detection in adverse weather." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.411, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.516, + 0.129 + ], + "angle": 0, + "content": "Hongbin Lin, Yifan Zhang and et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.241, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 6973-6980. IEEE (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.175, + 0.785, + 0.215 + ], + "angle": 0, + "content": "11. Hendrycks, D., Dietterich, T.: Benchmarking neural network robustness to common corruptions and perturbations. In: International Conference on Learning Representations (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.216, + 0.785, + 0.257 + ], + "angle": 0, + "content": "12. Kim, J., Hwang, I., Kim, Y.M.: Ev-tta: Test-time adaptation for event-based object recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17745-17754 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.257, + 0.785, + 0.297 + ], + "angle": 0, + "content": "13. Kim, Y., Yim, J., Yun, J., Kim, J.: Nlnl: Negative learning for noisy labels. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 101-110 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.297, + 0.785, + 0.338 + ], + "angle": 0, + "content": "14. Kim, Y., Yun, J., Shon, H., Kim, J.: Joint negative and positive learning for noisy labels. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9442-9451 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.338, + 0.785, + 0.379 + ], + "angle": 0, + "content": "15. Kumar, A., Brazil, G., Liu, X.: Groomed-nms: Grouped mathematically differentiable nms for monocular 3d object detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8973-8983 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.379, + 0.785, + 0.419 + ], + "angle": 0, + "content": "16. Li, P., Chen, X., Shen, S.: Stereo r-cnn based 3d object detection for autonomous driving. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7644-7652 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.42, + 0.785, + 0.46 + ], + "angle": 0, + "content": "17. Liang, J., Hu, D., Feng, J.: Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation. In: International conference on machine learning. pp. 6028-6039. PMLR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.46, + 0.785, + 0.5 + ], + "angle": 0, + "content": "18. Lin, H., Zhang, Y., Qiu, Z., Niu, S., Gan, C., Liu, Y., Tan, M.: Prototype-guided continual adaptation for class-incremental unsupervised domain adaptation. In: European Conference on Computer Vision. pp. 351-368 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.501, + 0.785, + 0.542 + ], + "angle": 0, + "content": "19. Liu, Y., Kothari, P., Van Delft, B., Bellot-Gurlet, B., Mordan, T., Alahi, A.: Ttt++: When does self-supervised test-time training fail or thrive? In: Advances in Neural Information Processing Systems. vol. 34, pp. 21808-21820 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.542, + 0.785, + 0.582 + ], + "angle": 0, + "content": "20. Liu, Z., Wu, Z., Tóth, R.: Smoke: Single-stage monocular 3d object detection via keypoint estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. pp. 996-997 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.582, + 0.785, + 0.636 + ], + "angle": 0, + "content": "21. Liu, Z., Tang, H., Amini, A., Yang, X., Mao, H., Rus, D.L., Han, S.: Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. In: 2023 IEEE international conference on robotics and automation (ICRA). pp. 2774-2781. IEEE (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.636, + 0.785, + 0.678 + ], + "angle": 0, + "content": "22. Luo, Y., Zheng, C., Yan, X., Kun, T., Zheng, C., Cui, S., Li, Z.: Latr: 3d lane detection from monocular images with transformer. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7941-7952 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.678, + 0.785, + 0.704 + ], + "angle": 0, + "content": "23. Ma, X., Liu, S., Xia, Z., Zhang, H., Zeng, X., Ouyang, W.: Rethinking pseudo-lidar representation. In: European Conference on Computer Vision (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.704, + 0.785, + 0.759 + ], + "angle": 0, + "content": "24. Mirza, M.J., Soneira, P.J., Lin, W., Kozinski, M., Possegger, H., Bischof, H.: Actmad: Activation matching to align distributions for test-time-training. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 24152-24161 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.759, + 0.785, + 0.8 + ], + "angle": 0, + "content": "25. Nado, Z., Padhy, S., Sculley, D., D'Amour, A., Lakshminarayanan, B., Snoek, J.: Evaluating prediction-time batch normalization for robustness under covariate shift. arXiv preprint arXiv:2006.10963 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "26. Niu, S., Wu, J., Zhang, Y., Chen, Y., Zheng, S., Zhao, P., Tan, M.: Efficient test-time model adaptation without forgetting. In: The International Conference on Machine Learning (2022)" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.66, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "MonoTTA" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "27. Niu, S., Wu, J., Zhang, Y., Wen, Z., Chen, Y., Zhao, P., Tan, M.: Towards stable test-time adaptation in dynamic wild world. In: International Conference on Learning Representations (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.189, + 0.788, + 0.244 + ], + "angle": 0, + "content": "28. Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., et al.: Pytorch: An imperative style, high-performance deep learning library. In: Advances in neural information processing systems. vol. 32 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.244, + 0.788, + 0.284 + ], + "angle": 0, + "content": "29. Qin, Z., Li, X.: Monoground: Detecting monocular 3d objects from the ground. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3793-3802 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.284, + 0.788, + 0.325 + ], + "angle": 0, + "content": "30. Qiu, Z., Zhang, Y., Lin, H., et al.: Source-free domain adaptation via avatar prototype generation and adaptation. In: International Joint Conference on Artificial Intelligence (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.325, + 0.788, + 0.365 + ], + "angle": 0, + "content": "31. Reading, C., Harakeh, A., Chae, J., Waslander, S.L.: Categorical depth distribution network for monocular 3d object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8555-8564 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.365, + 0.788, + 0.406 + ], + "angle": 0, + "content": "32. Saltori, C., Lathuilière, S., Sebe, N., Ricci, E., Galasso, F.: Sf-uda 3d: Source-free unsupervised domain adaptation for lidar-based 3d object detection. In: 2020 International Conference on 3D Vision (3DV). pp. 771–780. IEEE (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.406, + 0.788, + 0.447 + ], + "angle": 0, + "content": "33. Schneider, S., Rusak, E., Eck, L., Bringmann, O., Brendel, W., Bethge, M.: Improving robustness against common corruptions by covariate shift adaptation. In: Advances in neural information processing systems. pp. 11539-11551 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.447, + 0.788, + 0.488 + ], + "angle": 0, + "content": "34. Sun, Y., Wang, X., Liu, Z., Miller, J., Efros, A., Hardt, M.: Test-time training with self-supervision for generalization under distribution shifts. In: International conference on machine learning. pp. 9229-9248. PMLR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.488, + 0.788, + 0.528 + ], + "angle": 0, + "content": "35. Veksler, O.: Test time adaptation with regularized loss for weakly supervised salient object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7360-7369 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.528, + 0.788, + 0.569 + ], + "angle": 0, + "content": "36. Wang, D., Shelhamer, E., Liu, S., Olshausen, B., Darrell, T.: Tent: Fully test-time adaptation by entropy minimization. In: The International Conference on Machine Learning (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.569, + 0.788, + 0.623 + ], + "angle": 0, + "content": "37. Wang, Y., Chao, W.L., Garg, D., Hariharan, B., Campbell, M., Weinberger, K.Q.: Pseudo-lidar from visual depth estimation: Bridging the gap in 3d object detection for autonomous driving. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8445-8453 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.623, + 0.788, + 0.664 + ], + "angle": 0, + "content": "38. Wu, H., Wen, C., Shi, S., Li, X., Wang, C.: Virtual sparse convolution for multimodal 3d object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21653-21662 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.664, + 0.788, + 0.704 + ], + "angle": 0, + "content": "39. Xu, B., Chen, Z.: Multi-level fusion based 3d object detection from monocular images. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2345-2353 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.704, + 0.788, + 0.745 + ], + "angle": 0, + "content": "40. Xu, J., Peng, L., Cheng, H., Li, H., Qian, W., Li, K., Wang, W., Cai, D.: Mononerd: Nerf-like representations for monocular 3d object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6814-6824 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.745, + 0.788, + 0.786 + ], + "angle": 0, + "content": "41. Yang, J., Shi, S., Wang, Z., Li, H., Qi, X.: St3d: Self-training for unsupervised domain adaptation on 3d object detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10368-10378 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.786, + 0.788, + 0.84 + ], + "angle": 0, + "content": "42. Ye, X., Shu, M., Li, H., Shi, Y., Li, Y., Wang, G., Tan, X., Ding, E.: Rope3d: The roadside perception dataset for autonomous driving and monocular 3d object detection task. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21341-21350 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.516, + 0.129 + ], + "angle": 0, + "content": "Hongbin Lin, Yifan Zhang and et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "43. Zhang, M., Levine, S., Finn, C.: Memo: Test time robustness via adaptation and augmentation. In: Advances in Neural Information Processing Systems. vol. 35, pp. 38629-38642 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.787, + 0.232 + ], + "angle": 0, + "content": "44. Zhang, Y., Hooi, B., Hong, L., Feng, J.: Self-supervised aggregation of diverse experts for test-agnostic long-tailed recognition. In: Advances in Neural Information Processing Systems. vol. 35, pp. 34077-34090 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.787, + 0.273 + ], + "angle": 0, + "content": "45. Zhang, Y., Hooi, B., Hu, D., Liang, J., Feng, J.: Unleashing the power of contrastive self-supervised visual models via contrast-regularized fine-tuning. In: Advances in Neural Information Processing Systems. vol. 34, pp. 29848-29860 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.787, + 0.314 + ], + "angle": 0, + "content": "46. Zhang, Y., Kang, B., Hooi, B., Yan, S., Feng, J.: Deep long-tailed learning: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence 45(9), 10795-10816 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.787, + 0.357 + ], + "angle": 0, + "content": "47. Zhang, Y., Wei, Y., Wu, Q., Zhao, P., Niu, S., Huang, J., Tan, M.: Collaborative unsupervised domain adaptation for medical image diagnosis. IEEE Transactions on Image Processing 29, 7834-7844 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.357, + 0.787, + 0.398 + ], + "angle": 0, + "content": "48. Zhang, Y., Lu, J., Zhou, J.: Objects are different: Flexible monocular 3d object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3289-3298 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.787, + 0.453 + ], + "angle": 0, + "content": "49. Zou, Z., Ye, X., Du, L., Cheng, X., Tan, X., Zhang, L., Feng, J., Xue, X., Ding, E.: The devil is in the task: Exploiting reciprocal appearance-localization features for monocular 3d object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2713-2722 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.453 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/MonoTTA_ Fully Test-Time Adaptation for Monocular 3D Object Detection/cfa6dde5-bde3-4186-b95b-8bbf6c77c339_origin.pdf b/2024/MonoTTA_ Fully Test-Time Adaptation for Monocular 3D Object Detection/cfa6dde5-bde3-4186-b95b-8bbf6c77c339_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..5371a57ba73edfb62f43456aada19ae8ffa50952 --- /dev/null +++ b/2024/MonoTTA_ Fully Test-Time Adaptation for Monocular 3D Object Detection/cfa6dde5-bde3-4186-b95b-8bbf6c77c339_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dc94df8f5c91cea17eb3ea82484b77abb58e0226858df0903a392b3b0936e0c8 +size 1579683 diff --git a/2024/MonoTTA_ Fully Test-Time Adaptation for Monocular 3D Object Detection/full.md b/2024/MonoTTA_ Fully Test-Time Adaptation for Monocular 3D Object Detection/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d78f951dff8ea34478e2454a4add089ad881f16a --- /dev/null +++ b/2024/MonoTTA_ Fully Test-Time Adaptation for Monocular 3D Object Detection/full.md @@ -0,0 +1,271 @@ +# MonoTTA: Fully Test-Time Adaptation for Monocular 3D Object Detection + +Hongbin Lin $^{1,2*}$ , Yifan Zhang $^{3,4*}$ , Shuaicheng Niu $^{5*}$ , Shuguang Cui $^{2,1}$ , Zhen Li $^{2,1\dagger}$ + +$^{1}$ FNii-Shenzhen, $^{2}$ SSE, CUHK-Shenzhen, $^{3}$ NUS, $^{4}$ Skywork AI, $^{5}$ NTU {hongbinlin@link.,shuguangcui@,lizhen@}cuhk.edu.cn, yifan.zhang@u.nus.edu, shuaicheng.niu@ntu.edu.sg + +Abstract. Monocular 3D object detection (Mono 3Det) aims to identify 3D objects from a single RGB image. However, existing methods often assume training and test data follow the same distribution, which may not hold in real-world test scenarios. To address the out-of-distribution (OOD) problems, we explore a new adaptation paradigm for Mono 3Det, termed Fully Test-time Adaptation which aims to adapt a well-trained model to unlabeled test data by handling potential data distribution shifts at test time. However, applying this paradigm in Mono 3Det poses significant challenges due to OOD test data causing a remarkable decline in object detection scores. This decline conflicts with the pre-defined score thresholds of existing detection methods, leading to severe object omissions (i.e., rare positive detections and many false negatives). Consequently, the limited positive detection and plenty of noisy predictions cause test-time adaptation to fail in Mono 3Det. To handle this problem, we propose a novel Monocular Test-Time Adaptation (MonoTTA) method, based on two new strategies. 1) Reliability-driven adaptation: we empirically find that high-score objects are still reliable and the optimization of high-score objects can enhance confidence across all detections. Thus, we devise a self-adaptive strategy to identify reliable objects for model adaptation, which discovers potential objects and alleviates omissions. 2) Noise-guard adaptation: since high-score objects may be scarce, we develop a negative regularization term to exploit the numerous low-score objects via negative learning, preventing overfitting to noise and trivial solutions. Experimental results show that MonoTTA brings significant performance gains for Mono 3Det models in OOD test scenarios, approximately $190\%$ gains by average on KITTI and $198\%$ gains on nuScenes. The source code is now available at Hongbin98/MonoTTA. + +Keywords: Test-time Adaptation $\cdot$ Monocular 3D Object Detection + +# 1 Introduction + +Three-dimensional (3D) Object Detection is a significant computer vision task, with the objective of identifying objects and determining their spatial and di + +![](images/8eb188ab9bc574c6820cd170f345d2d622008005abc8b11024dcb6d4b96ad52a.jpg) +Fig. 1: An illustration of the generalizability issue of Mono 3Det models. Compared with in-distribution (In-dis) scenarios (e.g., sunny), the detection scores within out-of-distribution (OOD) test data suffer severe degradation when the well-trained model (MonoFlex [48]) is directly applied to test scenarios affected by common natural disruptions, like weather changes (e.g., snow and fog). Since existing Mono 3Det methods mainly adopt a pre-defined score threshold (e.g., 0.2) for object detection, it leads to severe omissions and unreliable detections, thereby suffering serious performance degradation. Note that test images are the same but under different weather conditions. + +mensional attributes through diverse sensor inputs [3,5,16,37,38]. To reduce the cost of sensors, there is an increasing trend towards implementing autonomous driving systems via Monocular 3D Object Detection (Mono 3Det) [2,42], where only one single RGB image and the camera calibration information are given. Even if this practical task is challenging, Mono 3Det methods have achieved promising results across various tasks and datasets [4,22,29,40,48]. Behind the success, a common presupposition is assuming that test images have the same distribution as the training images. However, this assumption could be possibly invalidated in many real-world scenarios due to prevalent natural corruptions such as weather changes, diminished sharpness, and other factors that introduce noise and contribute to uncalibrated cameras. In such circumstances, the well-trained model often suffers substantial performance degradation as a consequence of the data distributional shifts between the training images and the unlabeled test images. As shown in Fig. 1, the model performance degrades from $46.2\mathrm{mAP}$ in in-distribution data to $0.3\mathrm{mAP}$ in Snow and $7.2\mathrm{mAP}$ in Fog. Considering the widespread application of Mono 3Det in autonomous driving, the severe performance degradation due to out-of-distribution (OOD) test data may lead to unexpected traffic accidents and pose serious safety risks. Therefore, it is crucial to deal with the OOD generalization problem for Mono 3Det. + +In addressing the OOD challenges specifically in test scenarios, one paradigm that has emerged as highly promising and gaining traction is Test-Time Adaptation (TTA), which seeks to tackle data distribution shifts by adapting a well-trained model to unlabeled test images in real time. Test-time training (TTT) [34] represents an initial approach of TTA in classification tasks, by adjusting the well-trained model to predict rotations through additional model training, while its computation demands at the adaptation stage are prohibitive + +in Mono 3Det applications, particularly in autonomous driving. To enhance efficiency, Tent [36] and EATA [26] have been developed for Fully Test-Time Adaptation (Fully TTA) where only unlabeled test images and a well-trained model are provided. Besides, Ev-TTA [12] and SOD [35] devise TTA methods to handle the event-based object recognition and weakly supervised salient object detection, respectively. Considering the constraints on time of Mono 3Det, we explore the fully TTA paradigm which seeks to deal with OOD test data in real time. + +To investigate this paradigm for Mono 3Det, we dig into the detection outcomes for objects within test scenarios with variations or corruptions that are commonly caused by weather or cameras. Specifically, we directly apply the well-trained model to the validation set of KITTI which has been artificially injected with four distinct types of weather-related corruptions, namely Brightness, Frost, Fog, and Snow. Subsequently, we plot their distributions of detection scores (c.f. Fig. 1). It is observed that the detection scores of test objects tend to markedly decline as well as the high-score objects are scarce in the extreme scenario (Snow) when the well-trained model is directly applied to the scenarios with corruptions. This phenomenon indicates that: 1) The pre-trained Mono 3Det model struggles to discriminate between objects and the background within OOD test data, presenting as quantities of omissions and unconfident detections. 2) Directly applying existing fully TTA methods to Mono 3Det could only get suboptimal performance since they struggle to optimize the model without enough high-score (positive) detections, especially in certain extreme scenarios. + +To handle it in Mono 3Det, we propose a Monocular Test-Time Adaptation (MonoTTA) method, consisting of the reliability-driven adaptation and noise-guard adaptation strategies: 1) Reliability-driven adaptation. Specifically, data distribution shifts lead to omissions and noisy detections while our empirical analysis suggests that high detection score objects are still reliable (c.f. Fig. 3 (a)). Moreover, even if we only optimize the model via high-score objects ( $e.g., \geq 0.5$ ), both the numbers of low-score and high-score objects increase (c.f. Fig. 3 (b)). These investigations motivate us that exploiting high-score objects rather than all objects for model adaptation would be a more reliable way to alleviate data distribution shifts and discover potential objects. Hence, we develop a self-adaptive strategy for the identification of reliable high-score objects in test images and devise the adaptive optimization loss $\mathcal{L}_{AO}$ to exploit the reliable subset for model adaptation, alleviating the detection score decline issue of OOD test data and digging out more potential objects. 2) Noise-guard adaptation. In addition, data distribution shifts may also result in a scarcity of high-score objects, i.e., the majority of objects presenting low scores as the 'Snow' scenario in Fig. 3 (a). To this end, we develop a negative regularization term to make rational use of the numerous low-scoring objects in the Negative Learning manner [13]. On the one hand, the negative regularization term $\mathcal{L}_{Nreg}$ allows the model to conduct adaptation via numerous noisy low-scoring objects. Thus, the model can achieve more high-score objects after alleviating distribution shifts. On the other hand, this term also prevents the model from overfitting to noise and trivial solutions, i.e., assigning all classes of one object with high scores. + +We summarize the main contributions as follows: + +- To the best of our knowledge, we are the first to explore Fully Test-Time Adaptation to address OOD generalization problems for Mono 3Det. We show that the explored novel paradigm can bring significant improvements to Mono 3Det models in OOD test scenarios, e.g., $137\%$ and $244\%$ average performance gains across 13 types of OOD shifts on KITTI. +- Our empirical investigation reveals an important insight that high-score objects maintain their reliability amidst various corruptions, while optimizing these high-score objects significantly boosts model confidence across all detections. This motivates the first effective test-time adaptation method (i.e., our MonoTTA) in Mono 3Det. +- Extensive experiments on 13 types of corruptions of KITTI and 2 real scenarios (daytime $\leftrightarrow$ night) of nuScenes demonstrate the effectiveness of our MonoTTA in boosting existing Mono 3Det methods [29,48] to handle test-time OOD problems. Even for instance-level methods [40], MonoTTA also maintains sufficient improvement, which further confirms its applicability. + +# 2 Related Work + +We first review the literature on Monocular 3D Object Detection, and then discuss Source-free Domain Adaptation and Test-Time Adaptation methods. More discussions on Unsupervised Domain Adaptation [18,41,47] are in Appendix A. Monocular 3D Object Detection aims to perceive 3D objects from a single 2D image. Existing Mono 3Det methods could be divided into two groups according to the use of extra information. On the one hand, some existing methods leverage extra pre-trained depth estimation modules [6,39,49] to solve one of the most difficult problems in Mono 3Det, i.e., depth estimation from a single image. Other methods utilize LiDAR information, e.g., generating pseudo-LiDAR [23,31,37]. It is worth noting that Monoground [29] proposes to introduce the ground plane as prior information, and MonoNeRD [40] proposes to utilize scene geometric clues to enhance the detector's performance in the implicit reconstruction manner. On the other hand, some Mono 3Det methods try to detect 3D objects without extra data. For example, SMOKE [20] proposes to detect 3D objects as the key points estimation task. Then, Monoflex [48] improves this idea by providing a flexible definition of object centers, which unifies the centers of regular and truncated objects. GrooMeD-NMS [15] proposes a grouped mathematically differentiable Non-Maximal Suppression for Mono 3Det. + +Source-free Domain Adaptation (SFDA) aims to adapt the pre-trained source model to an unlabeled target domain without using the source data due to privacy issues [17, 30]. SF-UDA $^3$ D [32] first explores the SFDA framework to adapt the PointRCNN 3D detector to target domains, which consists of pseudolabeling, reversible scale-transformations and motion coherency. Recently, the authors [10] seek to exploit the source model more reliably and propose an uncertainty-aware teacher-student framework to filter incorrect pseudo labels during model adaptation, alleviating the negative impact of label noise. + +Nonetheless, SFDA assumes all target data to be known in advance and makes predictions after multiple epochs of optimization, which may not be viable for real-time applications due to computational or time constraints. + +Test-Time Adaptation (TTA) seeks to improve model performance on test data via model adaptation through test samples even if data shifts exist. Early TTA methods [19, 34] endeavor to conduct additional model optimization on training data by self-supervised objectives, and then adapt the well-trained model to the test data via self-supervised objectives. However, in Mono 3Det applications like autonomous driving, the computation demands of such methods are prohibitive. To solve this, Fully Test-Time Adaptation methods are developed to adapt the well-trained model, where only unlabeled test images are available. Specifically, certain methods [25, 27, 33] tackle data distribution shifts by adapting the batch normalization layer statistics, while others alleviate this issue either by the entropy minimization of test data [8, 36] or maximizing the prediction consistency of different augmentations [43-46]. As for object detection tasks, Ev-TTA [12] and SOD [35] try to handle the event-based object recognition and weakly supervised salient object detection offline, respectively. + +However, existing fully TTA methods struggle to optimize the model and solve distribution shifts in Mono 3Det due to numerous false negative detections. To the best of our knowledge, our MonoTTA stands as the first fully TTA method that handles distribution shifts for Mono 3Det models in real time. + +# 3 Monocular Test-Time Adaptation + +# 3.1 Problem Statement + +Without loss of generality, we denote the pre-trained (or well-trained) model as $f_{\Theta_0}(\mathbf{x})$ , which is achieved via training on labeled training images $\{(\mathbf{x}_i^s,\mathbf{y}_i^s)\}_{i = 1}^N$ . The training images follow the training distribution $P(\mathbf{x})$ (i.e., $\mathbf{x}^s\sim P(\mathbf{x})$ ). Here, $\Theta_0$ represents the parameters of the pre-trained model and $N$ is the number of training data. During the training stage, the model is optimized to fit (or overfit) the training data. Then, at the test stage, the model will be able to perform well if the unlabeled test images $\mathcal{D}_t = \{\mathbf{x}_i\}_{i = 1}^{N_t}$ follows the identical data distribution, i.e., $\mathbf{x}\sim P(\mathbf{x})$ where $N_{t}$ is the total number of test images. However, in real applications, it is possible for the pre-trained model to encounter Out-Of-Distribution (OOD) test samples due to prevalent natural corruptions, namely distribution shifts, i.e., $\mathbf{x}\sim Q(\mathbf{x})$ and $P(\mathbf{x})\neq Q(\mathbf{x})$ . + +To address this issue, fully test-time adaptation [36] seeks to tackle distribution shifts by adapting the pre-trained model $f_{\Theta_0}(\mathbf{x})$ to unlabeled test images $\{\mathbf{x}_i\}_{i=1}^{N_t}$ in real time. To achieve this goal, existing methods typically endeavor to update the model through the minimization of unsupervised objectives defined on test samples by $\min_{\hat{\Theta}} \mathcal{L}(\mathbf{x}; \Theta)$ , where $\mathbf{x} \sim Q(\mathbf{x})$ and $\hat{\Theta} \subseteq \Theta$ . Here, $\hat{\Theta}$ denotes the subset of model parameters that should be updated (i.e., batch normalization layers following existing methods [26,36]). Most existing fully TTA methods focus on classification tasks, heavily relying on sufficient positive predictions for + +![](images/7ba916103de92e85b6fa56879c52dc92c2b51841c794c47ecc76865a941152c8.jpg) +Fig. 2: An illustration of our MonoTTA. During the test phase, only the pre-trained model $f_{\Theta_0}(\mathbf{x})$ and unlabeled test images $\{\mathbf{x}_i\}_{i=1}^{N_t}$ are given. To conduct model adaptation, we initialize the model $f_{\Theta}(\mathbf{x})$ by $\Theta_0$ and only update the parameters of batch normalization layers. When a batch of test images arrives, we first compute test object scores and refine the adaptive threshold $\alpha$ to select the reliable high-score objects, thereby optimizing $\Theta$ via the adaptive optimization loss $\mathcal{L}_{AO}$ . Meanwhile, we devise a negative regularization term $\mathcal{L}_{Nreg}$ to facilitate the model to avoid overfitting to noise and trivial solutions. Here, Ped. and Cyc. represent Pedestrian and Cyclist in KITTI. + +model adaptation. Nevertheless, there is a significant difference between conventional classification tasks and Mono 3Det. As previously indicated, the detection scores of test images $\mathbf{x}$ derived from $f_{\Theta_0}(\mathbf{x})$ are prone to markedly decrease in the presence of corruptions as shown in Fig. 1, leading to severe omissions (numerous false negatives) in Mono 3Det. In such circumstances, the scarcity of positive detections presents a significant challenge for model adaptation to test distributions while adapting the model with unreliable low-score detections may significantly introduce the noise. Therefore, existing fully TTA methods tend to fail in the OOD generalization problems of Mono 3Det. + +# 3.2 Overall Scheme + +After thoroughly examining the characteristics and challenges of Mono 3Det, we introduce a Monocular Test-Time Adaptation (MonoTTA) method to address the OOD problems for Mono 3Det models, which seeks to solve the object score declining issue within unlabeled OOD test data. As shown in Fig. 2, MonoTTA consists of two strategies: 1) Reliability-driven adaptation and 2) Noise-guard adaptation. We first briefly introduce the two strategies below. + +First, we develop a reliability-driven adaptation strategy (c.f. Section 3.3) to conduct reliable model adaptation for OOD test data based on dependable test objects. Our empirical investigations inspire us to exploit those relatively reliable + +![](images/3acbb989c5da1a458cc9645f9fc38d02e6b56bbe28819edb4829bbe02c59e89f.jpg) +(a) + +![](images/b71877ed35d254d833d907c9acdf1fabec7900d73d66895bba8a8d1b23eeaa23.jpg) +(b) +Fig. 3: Based on MonoGround [29], we conduct two empirical studies (Car, KITTI), with the 3D IoU threshold of 0.5. (a) We visualize the accuracy of the objects across varied scoring ranges, which shows that the accuracy of objects with high scores remains relatively stable even in the presence of diverse corruptions (Ideal means in-distribution scenarios). (b) We visualize the number of low & high-score objects before and after optimization. Although only high-score objects are optimized, the model treats low-score objects with more confidence. + +test objects for alleviating distribution shifts, thereby discovering more potential objects. To this end, MonoTTA excludes unreliable test objects involving an adaptive threshold $\alpha$ for any unlabeled test data. Subsequently, the model is optimized by the adaptive optimization loss $\mathcal{L}_{AO}$ via the selected reliable objects. + +Second, we tend to utilize plenty of low-score objects to adapt the model in an indirect manner instead of directly optimizing the model since low-score objects are noisy. Hence, we devise a noise-guard adaptation strategy (c.f. Section 3.4) to prevent the model from overfitting to noisy predictions and falling into trivial solutions. Specifically, we randomly choose one of the negative classes of low-score objects and minimize the scores (e.g., score 0.03 of [0.10, 0.12, 0.03] c.f. Fig. 2) after simply filtering out extremely low-score detections. Even though the positive class is noisy (i.e., score 0.12), this term is capable of optimizing the model indirectly, i.e., learn to deny the negative category of the object. + +Overall, the training scheme of MonoTTA is as follows: + +$$ +\min _ {\hat {\Theta}} \mathcal {L} _ {A O} (\hat {\Theta}) + \lambda \mathcal {L} _ {N r e g} (\hat {\Theta}), \tag {1} +$$ + +where $\lambda$ is the balance hyper-parameter. The pseudo-code of MonoTTA is summarized in Algorithm 1. + +# 3.3 Reliability-Driven Adaptation + +To identify dependable test objects and conduct test-time model adaptation, we propose a reliability-driven adaptation strategy that consists of two components: 1) Reliable object identification and 2) Adaptive model optimization. + +Reliable object identification. When unlabeled test images arrive, it is difficult for the pre-trained model to get accurate detections on OOD test data due to the decline in detection scores. To resolve this, we dig into the detection + +accuracy of the pre-trained model in test scenarios with various corruptions. Specifically, based on MonoGround [29], we visualize the detection precision for the Car category within the KITTI dataset across varied scoring ranges, with the 3D Intersection over Union (IoU) threshold of 0.5. As shown in Fig. 3 (a), we find that high-score objects are more reliable and relatively stable even in the presence of diverse corruptions. Following this, we propose to select reliable high-score objects to conduct model adaptation via an adaptive threshold $\alpha$ . Specifically, we exploit the model $f_{\Theta}(\cdot)$ (initialized by $\Theta_0$ ) to infer a batch of test images $\{\mathbf{x}_b\}_{b=1}^B$ and obtain the heatmap $\mathbf{h}_i$ of each image $\mathbf{x}_i$ by $\mathbf{h}_i = f_{\Theta}(\mathbf{x}_i)$ , where $i$ ranges from 0 to $B-1$ and $B$ denotes the batch size. Then, we could achieve object score maps $\mathbf{s}_i \in \mathbb{R}^{B \times N_m}$ like the peaks in $\mathbf{h}_i$ [7] after normalization, where $K$ and $N_m$ denotes the number of classes and the maximum of detected objects, respectively. With the score map $\mathbf{s}_i$ , we update the adaptive threshold $\alpha_t$ at the iteration $t$ in the exponential moving average manner by: + +$$ +\alpha_ {t} = \left\{ \begin{array}{l l} \gamma , & \text {i f} t = 1 \\ \beta \bar {m} _ {t} + (1 - \beta) \alpha_ {t - 1}, & \text {i f} t > 1 \end{array} \right., \tag {2} +$$ + +$$ +\bar {m} _ {t} = \frac {1}{B} \sum_ {i = 1} ^ {B} \frac {\sum_ {j = 1} ^ {N _ {m}} s _ {i j} \cdot \mathbb {I} \left(s _ {i j} \geq \gamma\right)}{\sum_ {j = 1} ^ {N _ {m}} \mathbb {I} \left(s _ {i j} \geq \gamma\right)}. \tag {3} +$$ + +In Eqn. (2), $\bar{m}_t$ denotes the average score of all detected objects in a single batch of $B$ test samples at iteration $t$ , while $\beta \in [0,1]$ is a decay coefficient. As for Eqn. (3), $s_{ij} \in (0,1)$ denotes the score of the $j$ -th object in the $i$ -th test image while $\mathbb{I}(\cdot)$ is the indicator function. Note that $\gamma \in \mathbb{R}^1$ is a pre-defined object detection threshold adopted from existing methods at their original inference stage as well as $N_m$ . + +Adaptive model optimization. As shown in Fig. 3 (b), the optimization of high-score objects can also enhance the confidence of the model for relatively low-score objects. It motivates us that exploiting high-score objects rather than all objects for model adaptation would be a more reliable way to learn from OOD test data. Therefore, with the adaptive threshold $\alpha_{t}$ , we select the reliable subset of high scores from $\mathbf{s}_i$ and calculate the adaptive optimization loss $\mathcal{L}_{AO}$ to adapt the model by: + +$$ +\mathcal {L} _ {A O} = - \frac {1}{B} \sum_ {i = 1} ^ {B} \sum_ {j = 1} ^ {N _ {m}} \log \left(s _ {i j} \cdot \mathbb {I} \left(s _ {i j} \geq \alpha_ {t}\right)\right), \tag {4} +$$ + +The adaptive optimization loss $\mathcal{L}_{AO}$ alleviates the potential data distribution shifts in OOD test data by allowing the model to confidently identify high-score test objects. As we mentioned, it solves the score decline issue of OOD test data in a more reliable way rather than directly optimizing all test objects, thereby avoiding overfitting to noise and discovering more potential test objects. + +# Algorithm 1 The pipeline of the proposed MonoTTA + +Require: Unlabeled test data $\mathcal{D}_t = \{\mathbf{x}_i\}_{i=1}^{N_t}$ ; Pre-trained model $f_{\Theta_0}(\mathbf{x})$ ; Batch size $B$ ; Parameters $\lambda, \beta, \eta$ . + +1: for a batch images $\{\mathbf{x}_b\}_{b = 1}^B$ in $\mathcal{D}_t$ do +2: Update the adaptive threshold $\alpha$ based on Eqn. (2); +3: Calculate adaptive optimization loss $\mathcal{L}_{AO}$ based on Eqn. (4); +4: Calculate negative regularization term $\mathcal{L}_{Nreg}$ based on Eqn. (6); +5: Update $\hat{\Theta}$ by optimizing Eqn. (1) +6: end for +7: return Detection Results for all $\mathbf{x} \in \mathcal{D}_t$ . + +# 3.4 Noise-Guard Adaptation + +Through the optimization of adaptive optimization loss $\mathcal{L}_{AO}$ , the model is refined to yield more confident detection outcomes. Nevertheless, high-scoring objects may be scarce due to the distribution shifts, making the adaptation procedure difficult. Meanwhile, exclusive reliance on $\mathcal{L}_{AO}$ for adaptation may result in trivial solutions, whereby the model indiscriminately assigns high scores to all categories. Previous studies [13, 14] indicate that deep neural networks could learn from noisy pseudo labels in classification tasks through negative learning. Thus, MonoTTA proposes to learn from noisy low-score objects in a negative learning manner for Mono 3Det. Specifically, we denote $\hat{\mathbf{s}}_i\in \mathbb{R}^{B\times N_m\times K}$ as the multi-class score map for the test image $\mathbf{x}_i$ , i.e. the multi-class score map $\hat{\mathbf{s}}_i$ contains not only the highest score of objects but also the scores for other classes. Here $\mathbf{s}_i = \arg \max_k\hat{\mathbf{s}}_{ik}$ where $k$ is the class index. Employing a simple constant threshold $\eta$ , we filter out extremely low-score objects and then randomly select a negative class $\bar{k}$ for each object. Subsequently, we compute the regularization loss $e_k$ for each class $k$ with relatively low object scores $s_{ij}\in [\eta ,\alpha_t)$ by: + +$$ +e _ {k} = - \sum_ {i = 1} ^ {B} \sum_ {j = 1} ^ {N _ {m}} \bar {y} _ {i j} \log \left(1 - s _ {i j \bar {k}} \cdot \mathbb {I} (\bar {k} = k)\right), \tag {5} +$$ + +where $\bar{y}_{ij} = 1 - s_{ij\bar{k}}$ is a constant weight. We further define $n_k$ as the frequency of negative scores corresponding to class $k$ in the test batch and balance $e_k$ by: + +$$ +\mathcal {L} _ {N r e g} = \sum_ {k = 1} ^ {K} \frac {e _ {k}}{n _ {k}}. \tag {6} +$$ + +As we mentioned before, detections with low and intermediate scores tend to be more noisy (c.f. Fig. 3 (a)), which is unreliable for direct model adaptation. To this end, we introduce the regularization term $\mathcal{L}_{Nreg}$ to leverage noisy low-score detections and improve the model for assimilating potentially accurate information (c.f. Term 2 in Fig. 2). Moreover, this term also prevents the model from trivial solutions, i.e., indiscriminately assigning high scores to all classes of a single object (e.g., TENT [36]). In addition, high-score objects may be absent + +Table 1: Comparison with baselines on the KITTI-C validation set, severity level 1 regarding Mean $AP_{3D|R_{40}}$ . The bold number indicates the best result. + +
Car, IoU @ 0.7, 0.5, 0.5
MethodNoiseBlurWeatherDigitalAvg.
Gauss.ShotImpul.Defoc.GlassMotionSnowFrostFogBrit.Contr.PixelSat.
Monoflex [48]3.847.485.312.593.7311.050.237.777.5724.876.9228.1631.4610.84
• BN adaptation [33]13.5821.9318.7815.878.5924.325.4221.4524.6331.8030.5841.0430.7122.21
• TENT [36]17.8027.0923.1821.6611.9028.756.8426.5830.7835.6534.7241.7135.9126.35
• EATA [26]16.6726.4225.0722.5413.2327.737.8726.5831.1035.3935.2841.4036.7226.62
• ActMAD [24]12.2621.4920.4711.107.5323.795.5719.1126.5426.2423.9238.5133.7520.79
• MonoTTA (Ours)21.1528.6526.6425.9119.2631.4812.4330.2433.7536.8436.8341.9738.1329.48
MonoGround [29]2.404.103.313.712.678.130.225.544.5925.374.0033.5728.089.67
• BN adaptation [33]13.4923.5219.6916.337.6123.997.9820.7124.0031.3429.0343.0632.9922.60
• TENT [36]17.9429.6019.9023.4513.9029.3910.3226.6533.3535.9636.3943.3537.7927.54
• EATA [26]16.0326.0818.0820.2812.4027.379.2223.7929.4933.6532.5843.6136.0025.28
• ActMAD [24]13.6522.3920.5614.998.1721.099.1318.7719.5228.3425.3242.6331.5821.24
• MonoTTA (Ours)26.1333.1128.6030.3825.4832.4418.7232.6037.7537.8739.5743.6737.9832.64
+ +
Pedestrian, IoU @ 0.5, 0.25
MethodNoiseBlurWeatherDigitalAvg.
Gauss.ShotImpul.Defoc.GlassMotionSnowFrostFogBrit.Contr.PixelSat.
Monoflex0.191.620.323.728.476.220.004.272.259.192.081.839.113.79
• BN adaptation [33]6.218.209.207.835.357.522.896.479.249.129.9312.739.768.03
• TENT [36]6.027.969.577.756.068.632.636.719.9110.2610.5512.3310.278.36
• EATA [26]6.057.969.747.936.068.012.486.249.949.7010.0212.4110.128.21
• ActMAD [24]2.414.274.681.461.626.481.174.254.846.426.198.297.074.55
• MonotTA (Ours)6.548.419.397.637.128.993.257.6410.2610.5510.0613.2810.668.75
MonoGround0.610.930.738.417.578.030.003.191.3913.821.833.707.264.42
• BN adaptation [33]5.097.087.777.636.639.452.246.728.409.7211.0616.7012.438.53
• TENT [36]7.2710.1010.028.538.3011.033.548.739.1111.7412.1217.7015.0310.25
• EATA [26]5.928.058.588.127.5910.953.317.9110.1011.0210.9517.3014.349.55
• ActMAD [24]3.965.756.485.615.737.141.585.375.747.278.1013.129.276.55
• MonotTA (Ours)8.5811.1811.799.229.4013.204.839.9514.4612.8513.2517.1314.8511.59
+ +
Cyclist, IoU @ 0.5, 0.25, 0.25
MethodNoiseBlurWeatherDigitalAvg.
Gauss.ShotImpul.Defoc.GlassMotionSnowFrostFogBrit.Contr.PixelSat.
Monoflex0.281.640.470.594.973.600.007.423.8113.073.793.808.393.99
• BN adaptation [33]2.396.264.365.786.769.091.708.539.1612.9111.2610.5511.027.67
• TENT [36]2.727.945.636.277.209.491.078.9410.9612.7512.729.6411.288.20
• EATA [26]2.337.465.467.197.237.511.248.6010.1812.8611.0210.8910.307.87
• ActMAD [24]2.225.324.362.291.575.640.485.686.4610.569.319.977.805.51
• MonotTA (Ours)3.017.245.987.006.099.511.4510.6311.2212.8511.8511.5910.818.40
MonoGround0.120.460.480.330.722.030.000.560.355.550.522.085.241.42
• BN adaptation [33]1.763.582.083.613.055.410.573.824.476.305.6011.028.874.63
• TENT [36]1.794.853.003.363.496.050.494.436.207.196.5010.439.235.15
• EATA [26]1.894.202.414.133.175.730.384.066.276.416.1410.937.724.88
• ActMAD [24]1.142.841.533.33.083.150.322.743.476.365.759.696.953.87
• MonotTA (Ours)3.935.784.555.434.706.090.694.667.537.697.7411.719.436.15
+ +in certain extreme scenarios. For instance, the Snow scenario (c.f. Fig. 3 (a)) lacks objects with scores exceeding 0.4 when the pre-trained model is directly applied. Under such a circumstance, $\mathcal{L}_{Nreg}$ plays a more important role in model adaptation since it can alleviate distribution shifts even only low-score objects, i.e., deny the negative category. In other words, $\mathcal{L}_{Nreg}$ enables the model to alleviate distribution shifts and achieve more relatively high-score objects, thereby laying a crucial foundation for $\mathcal{L}_{AO}$ in extremely challenging scenarios. + +# 4 Experiments + +We conduct experiments based on KITTI [9] and nuScenes [1]. The results presented in this manuscript represent the average value across three difficulty levels, i.e., Easy, Moderate, Hard. Note that we provide more results of higher severity levels of KITTI and more detailed results in Appendix $D$ . + +Datasets. For KITTI, we adopt the protocol established by Monoflex [48] to split the training images into the training set (3712) and validation set (3769) to perform model training and adaptation, respectively. Following the official KITTI evaluation criteria, we evaluate detection results on three levels of difficulty. Then, we construct the KITTI-C dataset through the incorporation of 13 distinct types of data corruptions [11] to the validation set and each corruption + +![](images/1016ad6fb21f14f8f303f665c45098a31a4d72442d89e7b833a4fc5e03d381d3.jpg) +Average performance of 13 corruptions on KITTI-C Validation set, severity level 2 +Fig. 4: We visualize the comparison with baselines on the KITTI-C validation set, severity level 2 regarding Mean $AP_{3D|R_{40}}$ . The bold number indicates the best result. + +has 5 severity levels. The model is trained on the original training set and then tested on the KITTI-C validation set within one of the corruptions. + +As for nuScenes, we adopt the front-view images and construct the Daytime and Night scenarios via their scene descriptions following [21]. Specifically, there exist $24.7\mathrm{k} / 5.4\mathrm{k}$ train/test images in Daytime while $3.3\mathrm{k} / 0.6\mathrm{k}$ train/test images in Night. Based on these splits, we construct two real-world adaptation tasks, i.e., Daytime $\rightarrow$ Night and Night $\rightarrow$ Daytime. For simplification, we transfer the nuScenes dataset into the KITTI format and only consider the Car category. More details of data construction are provided in Appendix $\pmb{B}$ . + +Implementation details. We implement our method and other baselines in PyTorch [28]. In MonoTTA, we conduct model adaptation based on the public pre-trained weights and the parameter settings provided by their authors [29, 40, 48]. Besides, we employ the Stochastic Gradient Descent (SGD) optimizer with a half learning rate of the initial rate used in base training over different methods, a momentum of 0.9 and a batch size of 16 for KITTI, 4 for nuScenes. Parameters $\lambda$ , $\beta$ , $\eta$ are assigned default values of 1, 0.1, and 0.05, respectively. More training details of MonoTTA are provided in Appendix $C$ . + +Compared methods. Based on three typical or state-of-the-art (SOTA) Mono 3Det methods [29, 40, 48], we fully compare MonoTTA with following methods: 1) source-only, i.e., directly apply the pre-trained model to the test data within corruptions; 2) BN adaptation [33] updates batch normalization statistics via target data; 3) TENT [36] minimizes the entropy loss of test data; 4) EATA [26] identifies reliable samples to update the model by entropy loss minimization. Here we compare MonoTTA with its variant Efficient Test-time Adaptation. 5) ActMAD [24] matches activation statistics for the distribution alignment. + +Evaluation protocols. In order to fully evaluate the proposed method, we report our experimental results in the Average Precision (AP) for 3D bounding boxes, denoted as $AP_{3D|R_{40}}$ . The results present the mean values across three levels of difficulty and the Intersection over Union (IoU) thresholds are set to 0.7, 0.5, 0.5 for Cars and 0.5, 0.25, 0.25 for Pedestrians and Cyclists. + +![](images/21e234440e204e27ec0c18cc793bbd57c303929d334ce98b0f5e2193c398f2d8.jpg) +Performance on Daytime to Night and Night to Daytime scenarios of nuScenes +Fig. 5: Comparison with baselines on Daytime $\rightarrow$ Night and Night $\rightarrow$ Daytime of nuScenes, regarding Mean $AP_{3D|R_{40}}$ . The **bold** number indicates the best result. + +# 4.1 Comparisons with Previous Methods + +We first compare our MonoTTA with previous methods in severity level 1 of KITTI-C. The results are reported in Table 1, which gives the following observations: 1) Due to distribution shifts, directly applying the pre-trained model to the test data (i.e., source-only) suffers severe performance degradation in all categories. 2) Existing TTA methods are able to mitigate the negative effect of distribution shifts for Mono 3Det to some degree. However, they only achieve suboptimal performance since they tend to increase the scores of all positive detections, containing severe noise. 3) MonoTTA consistently outperforms all compared methods over all categories within various base models in terms of mean $AP_{3D|R_{40}}$ . Specifically, MonoTTA achieves the best or comparable performance in all categories under all corruptions, attaining a large performance gain over TENT and EATA (e.g., improving an average $AP_{3D|R_{40}}$ about 5.1 and 7.4 of the Car category based on MonoGround). + +# 4.2 More Severe Corruption and Real Scenario + +On the one hand, to fully validate the effectiveness of our MonoTTA, we visualize experimental results under more severe corruption conditions at severity level 2 as shown in Fig. 4, which clearly gives additional observations: 1) With the escalation of severity level, the pre-trained models suffer a larger performance decline within various corruptions, enlarging the difficulty of TTA. 2) The performance improvements of existing TTA methods become relatively limited, particularly in Pedestrian and Cyclist classes. 3) Even if the tasks are more challenging, MonoTTA still stably obtains the best average performance within all corruptions since $\mathcal{L}_{Nreg}$ plays an important role in alleviating distribution shifts for certain extreme scenarios, i.e., only low-score objects exist. + +On the other hand, we further validate different methods within real scenarios as shown in Fig. 5. The experimental results also give the following observations: 1) Under real corruptions, the pre-trained model still suffers severe performance degradation due to the data distribution shifts. 2) TENT tends to increase the + +Table 2: Comparison with baselines based on the instance-level method (i.e., batch size is 1) on the KITTI-C validation set, regarding Mean $AP_{3D|R_{40}}$ and the severity levels 1 and 2. The bold number indicates the best result. + +
MethodLevel 1Level 2
CarPedes.CyclistAvg.CarPedes.CyclistAvg.
MonoNeRD19.845.962.279.3613.023.831.616.15
BN adaptation [33]30.738.853.8114.4626.476.942.9112.11
TENT [36]35.729.994.7516.8231.857.813.4014.35
EATA [26]34.6010.044.2016.2830.667.863.2813.93
MonoTTA (Ours)37.4010.394.3517.3833.998.253.3315.19
+ +Table 3: Based on MonoGround [29], we conduct ablation studies of $\mathcal{L}_{AO}$ and $\mathcal{L}_{Nreg}$ on the KITTI-C validation set, regarding $AP_{3D|R_{40}}$ . + +
Backbone\( {\mathcal{L}}_{AO} \)\( {\mathcal{L}}_{Nreg} \)Level 1Level 2
CarPedes.CyclistAvg.CarPedes.CyclistAvg.
✓9.674.421.425.175.722.430.953.03
✓✓27.9410.175.1014.4021.277.793.8210.96
✓✓23.268.604.6012.1518.936.813.559.77
✓✓✓32.6411.596.1516.7925.968.784.1212.85
+ +confidence of all positive detections and thus overfits to noise, i.e., failing to handle the extremely challenging task $\mathrm{N} \rightarrow \mathrm{D}$ . 3) EATA still achieves sub-optimal performance while our MonoTTA brings sufficient average performance improvement on both MonoFlex (6.23 mAP) and MonoGround (8.26 mAP), maintaining the best performance in real scenarios. Detailed results are in Appendix $D$ . + +# 4.3 Application to Instance-Level Inference Method + +In this section, we seek to investigate whether the proposed MonoTTA can be used to effectively enhance Mono 3Det methods which only process a single image sequentially, i.e., $B = 1$ . To be specific, it may be crucial for Mono 3Det methods to make immediate decisions based on the most recent scene (image) in real-world scenarios like autonomous driving. To this end, it is essential for TTA methods devised for Mono 3Det to allow a single image as input and then conduct model adaptation. To validate it, we integrate MonoTTA into the SOTA Mono 3Det method namely MonoNeRD [40] which accepts one image each time at the test phase. As shown in Table 2, MonoTTA achieves the best or comparable performance across all categories at both severity levels 1 and 2, illustrating the applicability of our method to boost these approaches for handling OOD test data. Detailed results can be also found in Appendix $D$ . + +# 4.4 Ablation Studies and Quantitative Results + +Ablation studies. To examine the effectiveness of the losses in MonoTTA, we show the results of the models optimized by different losses. As shown in Table 3, introducing $\mathcal{L}_{AO}$ or $\mathcal{L}_{Nreg}$ enhances the model performance compared to directly applying the pre-trained model (i.e., source only). On the one hand, such a result verifies that our strategy is able to alleviate the score decline issue in test OOD data with high-score objects. On the other hand, introducing $\mathcal{L}_{Nreg}$ could obtain a higher average $AP_{3D|R_{40}}$ value, which also verifies that the negative regularization term is able to enhance the pre-trained model even only + +![](images/7dee429f5322d7f05be1b976335d70ae3b10fcc09f1d687224b5877ee2f9e237.jpg) +Fig.6: Qualitative results of baselines and the proposed MonoTTA based on MonoGround [29]. We visualize the results on KITTI-C validation set, where predicted cars, pedestrians and cyclists are in lime green, sky blue and yellow, respectively. + +with low-score objects. When combining the losses (i.e. $\mathcal{L}_{AO}$ , $\mathcal{L}_{Nreg}$ ) together, we obtain the best performance. + +Quantitative Results. We provide visualizations within Snow, Fog and Frost based on Monoground as shown in Fig. 6. It is evident that the source-only setting suffers severe omissions, while BN adaptation, TENT and EATA alleviate distribution shifts to some degree and give more detections. As for our MonoTTA, it can produce superior detections even in severe conditions, including fewer omissions and accurate detections as highlighted by red circles. + +# 5 Conclusion + +In this paper, we propose a monocular test-time adaptation method to improve the pre-trained model on the shifted test data for monocular 3D object detection. Specifically, our method consists of two strategies: 1) Reliability-driven adaptation. To discover more potential objects, we devise a self-adaptive strategy to identify reliable objects for adaptive model adaptation. 2) Noise-guard adaptation. To avoid overfitting to noise and trivial solutions, we devise the negative regularization term to mitigate the negative effects of noisy detections and alleviate distribution shifts. Experiments on KITTI-C and nuScenes datasets demonstrate the effectiveness of MonoTTA in handling fully test-time adaptation for monocular 3D object detection. + +Future directions. 1) Our work focuses on 2D images, while future studies could explore 3D information in handling distribution shifts. 2) This work explores TTA by assuming one OOD distribution at a time, where the forgetting issue is not severe as the source model weights are recoverable. Exploring scenarios with dynamically OOD distributions offers a compelling future direction, where the forgetting issue would become more severe. + +# Acknowledgements + +This work was supported by NSFC with Grant No. 62293482, by the Basic Research Project No. HZQB-KCZYZ-2021067 of Hetao Shenzhen HK S&T Cooperation Zone, by Shenzhen General Program No. JCYJ20220530143600001, by Shenzhen-Hong Kong Joint Funding No. SGDX20211123112401002, by the Shenzhen Outstanding Talents Training Fund 202002, by Guangdong Research Project No. 2017ZT07X152 and No. 2019CX01X104, by the Guangdong Provincial Key Laboratory of Future Networks of Intelligence (Grant No. 2022B1212010001), by the Guangdong Provincial Key Laboratory of Big Data Computing, CHUK-Shenzhen, by the NSFC 61931024&12326610, by the Key Area R&D Program of Guangdong Province with grant No. 2018B030338001, by the Shenzhen Key Laboratory of Big Data and Artificial Intelligence (Grant No. ZDSYS201707251409055), and by Tencent & Huawei Open Fund. + +# References + +1. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuscenes: A multimodal dataset for autonomous driving. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11621-11631 (2020) +2. Chen, X., Kundu, K., Zhang, Z., Ma, H., Fidler, S., Urtasun, R.: Monocular 3d object detection for autonomous driving. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2147-2156 (2016) +3. Chen, X., Kundu, K., Zhu, Y., Berneshawi, A.G., Ma, H., Fidler, S., Urtasun, R.: 3d object proposals for accurate object class detection. Advances in neural information processing systems 28 (2015) +4. Chen, Y., Tai, L., Sun, K., Li, M.: Monopair: Monocular 3d object detection using pairwise spatial relationships. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12093-12102 (2020) +5. Chen, Y., Liu, J., Zhang, X., Qi, X., Jia, J.: Voxelnext: Fully sparse voxelnet for 3d object detection and tracking. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21674-21683 (2023) +6. Ding, M., Huo, Y., Yi, H., Wang, Z., Shi, J., Lu, Z., Luo, P.: Learning depth-guided convolutions for monocular 3d object detection. In: Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition workshops. pp. 1000-1001 (2020) +7. Duan, K., Bai, S., Xie, L., Qi, H., Huang, Q., Tian, Q.: Centernet: Keypoint triplets for object detection. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6569-6578 (2019) +8. Fleuret, F., et al.: Test time adaptation through perturbation robustness. In: NeurIPS 2021 Workshop on Distribution Shifts: Connecting Methods and Applications (2021) +9. Geiger, A., Lenz, P., Urtasun, R.: Are we ready for autonomous driving? the kitti vision benchmark suite. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3354-3361. IEEE (2012) +0. Hegde, D., Kilic, V., Sindagi, V., Cooper, A.B., Foster, M., Patel, V.M.: Source-free unsupervised domain adaptation for 3d object detection in adverse weather. + +In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 6973-6980. IEEE (2023) +11. Hendrycks, D., Dietterich, T.: Benchmarking neural network robustness to common corruptions and perturbations. In: International Conference on Learning Representations (2018) +12. Kim, J., Hwang, I., Kim, Y.M.: Ev-tta: Test-time adaptation for event-based object recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17745-17754 (2022) +13. Kim, Y., Yim, J., Yun, J., Kim, J.: Nlnl: Negative learning for noisy labels. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 101-110 (2019) +14. Kim, Y., Yun, J., Shon, H., Kim, J.: Joint negative and positive learning for noisy labels. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9442-9451 (2021) +15. Kumar, A., Brazil, G., Liu, X.: Groomed-nms: Grouped mathematically differentiable nms for monocular 3d object detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8973-8983 (2021) +16. Li, P., Chen, X., Shen, S.: Stereo r-cnn based 3d object detection for autonomous driving. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7644-7652 (2019) +17. Liang, J., Hu, D., Feng, J.: Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation. In: International conference on machine learning. pp. 6028-6039. PMLR (2020) +18. Lin, H., Zhang, Y., Qiu, Z., Niu, S., Gan, C., Liu, Y., Tan, M.: Prototype-guided continual adaptation for class-incremental unsupervised domain adaptation. In: European Conference on Computer Vision. pp. 351-368 (2022) +19. Liu, Y., Kothari, P., Van Delft, B., Bellot-Gurlet, B., Mordan, T., Alahi, A.: Ttt++: When does self-supervised test-time training fail or thrive? In: Advances in Neural Information Processing Systems. vol. 34, pp. 21808-21820 (2021) +20. Liu, Z., Wu, Z., Tóth, R.: Smoke: Single-stage monocular 3d object detection via keypoint estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. pp. 996-997 (2020) +21. Liu, Z., Tang, H., Amini, A., Yang, X., Mao, H., Rus, D.L., Han, S.: Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. In: 2023 IEEE international conference on robotics and automation (ICRA). pp. 2774-2781. IEEE (2023) +22. Luo, Y., Zheng, C., Yan, X., Kun, T., Zheng, C., Cui, S., Li, Z.: Latr: 3d lane detection from monocular images with transformer. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7941-7952 (2023) +23. Ma, X., Liu, S., Xia, Z., Zhang, H., Zeng, X., Ouyang, W.: Rethinking pseudo-lidar representation. In: European Conference on Computer Vision (2020) +24. Mirza, M.J., Soneira, P.J., Lin, W., Kozinski, M., Possegger, H., Bischof, H.: Actmad: Activation matching to align distributions for test-time-training. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 24152-24161 (2023) +25. Nado, Z., Padhy, S., Sculley, D., D'Amour, A., Lakshminarayanan, B., Snoek, J.: Evaluating prediction-time batch normalization for robustness under covariate shift. arXiv preprint arXiv:2006.10963 (2020) +26. Niu, S., Wu, J., Zhang, Y., Chen, Y., Zheng, S., Zhao, P., Tan, M.: Efficient test-time model adaptation without forgetting. In: The International Conference on Machine Learning (2022) + +27. Niu, S., Wu, J., Zhang, Y., Wen, Z., Chen, Y., Zhao, P., Tan, M.: Towards stable test-time adaptation in dynamic wild world. In: International Conference on Learning Representations (2023) +28. Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., et al.: Pytorch: An imperative style, high-performance deep learning library. In: Advances in neural information processing systems. vol. 32 (2019) +29. Qin, Z., Li, X.: Monoground: Detecting monocular 3d objects from the ground. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3793-3802 (2022) +30. Qiu, Z., Zhang, Y., Lin, H., et al.: Source-free domain adaptation via avatar prototype generation and adaptation. In: International Joint Conference on Artificial Intelligence (2021) +31. Reading, C., Harakeh, A., Chae, J., Waslander, S.L.: Categorical depth distribution network for monocular 3d object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8555-8564 (2021) +32. Saltori, C., Lathuilière, S., Sebe, N., Ricci, E., Galasso, F.: Sf-uda 3d: Source-free unsupervised domain adaptation for lidar-based 3d object detection. In: 2020 International Conference on 3D Vision (3DV). pp. 771–780. IEEE (2020) +33. Schneider, S., Rusak, E., Eck, L., Bringmann, O., Brendel, W., Bethge, M.: Improving robustness against common corruptions by covariate shift adaptation. In: Advances in neural information processing systems. pp. 11539-11551 (2020) +34. Sun, Y., Wang, X., Liu, Z., Miller, J., Efros, A., Hardt, M.: Test-time training with self-supervision for generalization under distribution shifts. In: International conference on machine learning. pp. 9229-9248. PMLR (2020) +35. Veksler, O.: Test time adaptation with regularized loss for weakly supervised salient object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7360-7369 (2023) +36. Wang, D., Shelhamer, E., Liu, S., Olshausen, B., Darrell, T.: Tent: Fully test-time adaptation by entropy minimization. In: The International Conference on Machine Learning (2021) +37. Wang, Y., Chao, W.L., Garg, D., Hariharan, B., Campbell, M., Weinberger, K.Q.: Pseudo-lidar from visual depth estimation: Bridging the gap in 3d object detection for autonomous driving. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8445-8453 (2019) +38. Wu, H., Wen, C., Shi, S., Li, X., Wang, C.: Virtual sparse convolution for multimodal 3d object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21653-21662 (2023) +39. Xu, B., Chen, Z.: Multi-level fusion based 3d object detection from monocular images. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2345-2353 (2018) +40. Xu, J., Peng, L., Cheng, H., Li, H., Qian, W., Li, K., Wang, W., Cai, D.: Mononerd: Nerf-like representations for monocular 3d object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6814-6824 (2023) +41. Yang, J., Shi, S., Wang, Z., Li, H., Qi, X.: St3d: Self-training for unsupervised domain adaptation on 3d object detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10368-10378 (2021) +42. Ye, X., Shu, M., Li, H., Shi, Y., Li, Y., Wang, G., Tan, X., Ding, E.: Rope3d: The roadside perception dataset for autonomous driving and monocular 3d object detection task. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21341-21350 (2022) + +43. Zhang, M., Levine, S., Finn, C.: Memo: Test time robustness via adaptation and augmentation. In: Advances in Neural Information Processing Systems. vol. 35, pp. 38629-38642 (2022) +44. Zhang, Y., Hooi, B., Hong, L., Feng, J.: Self-supervised aggregation of diverse experts for test-agnostic long-tailed recognition. In: Advances in Neural Information Processing Systems. vol. 35, pp. 34077-34090 (2022) +45. Zhang, Y., Hooi, B., Hu, D., Liang, J., Feng, J.: Unleashing the power of contrastive self-supervised visual models via contrast-regularized fine-tuning. In: Advances in Neural Information Processing Systems. vol. 34, pp. 29848-29860 (2021) +46. Zhang, Y., Kang, B., Hooi, B., Yan, S., Feng, J.: Deep long-tailed learning: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence 45(9), 10795-10816 (2023) +47. Zhang, Y., Wei, Y., Wu, Q., Zhao, P., Niu, S., Huang, J., Tan, M.: Collaborative unsupervised domain adaptation for medical image diagnosis. IEEE Transactions on Image Processing 29, 7834-7844 (2020) +48. Zhang, Y., Lu, J., Zhou, J.: Objects are different: Flexible monocular 3d object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3289-3298 (2021) +49. Zou, Z., Ye, X., Du, L., Cheng, X., Tan, X., Zhang, L., Feng, J., Xue, X., Ding, E.: The devil is in the task: Exploiting reciprocal appearance-localization features for monocular 3d object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2713-2722 (2021) \ No newline at end of file diff --git a/2024/MonoTTA_ Fully Test-Time Adaptation for Monocular 3D Object Detection/images.zip b/2024/MonoTTA_ Fully Test-Time Adaptation for Monocular 3D Object Detection/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..4ea8e3cb5c5fbb2f6e7c0bdf5537c17823268824 --- /dev/null +++ b/2024/MonoTTA_ Fully Test-Time Adaptation for Monocular 3D Object Detection/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:852fa5ab2da8f6ead4b8949fd580dbd1767ff504f560e973a58ca04e21177342 +size 564881 diff --git a/2024/MonoTTA_ Fully Test-Time Adaptation for Monocular 3D Object Detection/layout.json b/2024/MonoTTA_ Fully Test-Time Adaptation for Monocular 3D Object Detection/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..c2019c00e873481817034d8f6b88e298b7f312fc --- /dev/null +++ b/2024/MonoTTA_ Fully Test-Time Adaptation for Monocular 3D Object Detection/layout.json @@ -0,0 +1,9917 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 157, + 111, + 459, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 111, + 459, + 148 + ], + "spans": [ + { + "bbox": [ + 157, + 111, + 459, + 148 + ], + "type": "text", + "content": "MonoTTA: Fully Test-Time Adaptation for Monocular 3D Object Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 179, + 167, + 434, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 167, + 434, + 193 + ], + "spans": [ + { + "bbox": [ + 179, + 167, + 434, + 193 + ], + "type": "text", + "content": "Hongbin Lin" + }, + { + "bbox": [ + 179, + 167, + 434, + 193 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 179, + 167, + 434, + 193 + ], + "type": "text", + "content": ", Yifan Zhang" + }, + { + "bbox": [ + 179, + 167, + 434, + 193 + ], + "type": "inline_equation", + "content": "^{3,4*}" + }, + { + "bbox": [ + 179, + 167, + 434, + 193 + ], + "type": "text", + "content": ", Shuaicheng Niu" + }, + { + "bbox": [ + 179, + 167, + 434, + 193 + ], + "type": "inline_equation", + "content": "^{5*}" + }, + { + "bbox": [ + 179, + 167, + 434, + 193 + ], + "type": "text", + "content": ", Shuguang Cui" + }, + { + "bbox": [ + 179, + 167, + 434, + 193 + ], + "type": "inline_equation", + "content": "^{2,1}" + }, + { + "bbox": [ + 179, + 167, + 434, + 193 + ], + "type": "text", + "content": ", Zhen Li" + }, + { + "bbox": [ + 179, + 167, + 434, + 193 + ], + "type": "inline_equation", + "content": "^{2,1\\dagger}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 159, + 201, + 455, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 201, + 455, + 236 + ], + "spans": [ + { + "bbox": [ + 159, + 201, + 455, + 236 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 159, + 201, + 455, + 236 + ], + "type": "text", + "content": "FNii-Shenzhen, " + }, + { + "bbox": [ + 159, + 201, + 455, + 236 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 159, + 201, + 455, + 236 + ], + "type": "text", + "content": " SSE, CUHK-Shenzhen, " + }, + { + "bbox": [ + 159, + 201, + 455, + 236 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 159, + 201, + 455, + 236 + ], + "type": "text", + "content": " NUS, " + }, + { + "bbox": [ + 159, + 201, + 455, + 236 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 159, + 201, + 455, + 236 + ], + "type": "text", + "content": " Skywork AI, " + }, + { + "bbox": [ + 159, + 201, + 455, + 236 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 159, + 201, + 455, + 236 + ], + "type": "text", + "content": " NTU {hongbinlin@link.,shuguangcui@,lizhen@}cuhk.edu.cn, yifan.zhang@u.nus.edu, shuaicheng.niu@ntu.edu.sg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 256, + 455, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 256, + 455, + 553 + ], + "spans": [ + { + "bbox": [ + 160, + 256, + 455, + 553 + ], + "type": "text", + "content": "Abstract. Monocular 3D object detection (Mono 3Det) aims to identify 3D objects from a single RGB image. However, existing methods often assume training and test data follow the same distribution, which may not hold in real-world test scenarios. To address the out-of-distribution (OOD) problems, we explore a new adaptation paradigm for Mono 3Det, termed Fully Test-time Adaptation which aims to adapt a well-trained model to unlabeled test data by handling potential data distribution shifts at test time. However, applying this paradigm in Mono 3Det poses significant challenges due to OOD test data causing a remarkable decline in object detection scores. This decline conflicts with the pre-defined score thresholds of existing detection methods, leading to severe object omissions (i.e., rare positive detections and many false negatives). Consequently, the limited positive detection and plenty of noisy predictions cause test-time adaptation to fail in Mono 3Det. To handle this problem, we propose a novel Monocular Test-Time Adaptation (MonoTTA) method, based on two new strategies. 1) Reliability-driven adaptation: we empirically find that high-score objects are still reliable and the optimization of high-score objects can enhance confidence across all detections. Thus, we devise a self-adaptive strategy to identify reliable objects for model adaptation, which discovers potential objects and alleviates omissions. 2) Noise-guard adaptation: since high-score objects may be scarce, we develop a negative regularization term to exploit the numerous low-score objects via negative learning, preventing overfitting to noise and trivial solutions. Experimental results show that MonoTTA brings significant performance gains for Mono 3Det models in OOD test scenarios, approximately " + }, + { + "bbox": [ + 160, + 256, + 455, + 553 + ], + "type": "inline_equation", + "content": "190\\%" + }, + { + "bbox": [ + 160, + 256, + 455, + 553 + ], + "type": "text", + "content": " gains by average on KITTI and " + }, + { + "bbox": [ + 160, + 256, + 455, + 553 + ], + "type": "inline_equation", + "content": "198\\%" + }, + { + "bbox": [ + 160, + 256, + 455, + 553 + ], + "type": "text", + "content": " gains on nuScenes. The source code is now available at Hongbin98/MonoTTA." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 563, + 443, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 563, + 443, + 574 + ], + "spans": [ + { + "bbox": [ + 160, + 563, + 443, + 574 + ], + "type": "text", + "content": "Keywords: Test-time Adaptation " + }, + { + "bbox": [ + 160, + 563, + 443, + 574 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 563, + 443, + 574 + ], + "type": "text", + "content": " Monocular 3D Object Detection" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 591, + 230, + 604 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 591, + 230, + 604 + ], + "spans": [ + { + "bbox": [ + 132, + 591, + 230, + 604 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 613, + 481, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 613, + 481, + 638 + ], + "spans": [ + { + "bbox": [ + 130, + 613, + 481, + 638 + ], + "type": "text", + "content": "Three-dimensional (3D) Object Detection is a significant computer vision task, with the objective of identifying objects and determining their spatial and di" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 642, + 261, + 654 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 642, + 261, + 654 + ], + "spans": [ + { + "bbox": [ + 133, + 642, + 261, + 654 + ], + "type": "text", + "content": "* Authors contributed equally." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 136, + 654, + 240, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 654, + 240, + 665 + ], + "spans": [ + { + "bbox": [ + 136, + 654, + 240, + 665 + ], + "type": "text", + "content": "† Corresponding authors." + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 114, + 479, + 235 + ], + "blocks": [ + { + "bbox": [ + 136, + 114, + 479, + 235 + ], + "lines": [ + { + "bbox": [ + 136, + 114, + 479, + 235 + ], + "spans": [ + { + "bbox": [ + 136, + 114, + 479, + 235 + ], + "type": "image", + "image_path": "8eb188ab9bc574c6820cd170f345d2d622008005abc8b11024dcb6d4b96ad52a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 245, + 482, + 333 + ], + "lines": [ + { + "bbox": [ + 130, + 245, + 482, + 333 + ], + "spans": [ + { + "bbox": [ + 130, + 245, + 482, + 333 + ], + "type": "text", + "content": "Fig. 1: An illustration of the generalizability issue of Mono 3Det models. Compared with in-distribution (In-dis) scenarios (e.g., sunny), the detection scores within out-of-distribution (OOD) test data suffer severe degradation when the well-trained model (MonoFlex [48]) is directly applied to test scenarios affected by common natural disruptions, like weather changes (e.g., snow and fog). Since existing Mono 3Det methods mainly adopt a pre-defined score threshold (e.g., 0.2) for object detection, it leads to severe omissions and unreliable detections, thereby suffering serious performance degradation. Note that test images are the same but under different weather conditions." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 351, + 482, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 351, + 482, + 579 + ], + "spans": [ + { + "bbox": [ + 130, + 351, + 482, + 579 + ], + "type": "text", + "content": "mensional attributes through diverse sensor inputs [3,5,16,37,38]. To reduce the cost of sensors, there is an increasing trend towards implementing autonomous driving systems via Monocular 3D Object Detection (Mono 3Det) [2,42], where only one single RGB image and the camera calibration information are given. Even if this practical task is challenging, Mono 3Det methods have achieved promising results across various tasks and datasets [4,22,29,40,48]. Behind the success, a common presupposition is assuming that test images have the same distribution as the training images. However, this assumption could be possibly invalidated in many real-world scenarios due to prevalent natural corruptions such as weather changes, diminished sharpness, and other factors that introduce noise and contribute to uncalibrated cameras. In such circumstances, the well-trained model often suffers substantial performance degradation as a consequence of the data distributional shifts between the training images and the unlabeled test images. As shown in Fig. 1, the model performance degrades from " + }, + { + "bbox": [ + 130, + 351, + 482, + 579 + ], + "type": "inline_equation", + "content": "46.2\\mathrm{mAP}" + }, + { + "bbox": [ + 130, + 351, + 482, + 579 + ], + "type": "text", + "content": " in in-distribution data to " + }, + { + "bbox": [ + 130, + 351, + 482, + 579 + ], + "type": "inline_equation", + "content": "0.3\\mathrm{mAP}" + }, + { + "bbox": [ + 130, + 351, + 482, + 579 + ], + "type": "text", + "content": " in Snow and " + }, + { + "bbox": [ + 130, + 351, + 482, + 579 + ], + "type": "inline_equation", + "content": "7.2\\mathrm{mAP}" + }, + { + "bbox": [ + 130, + 351, + 482, + 579 + ], + "type": "text", + "content": " in Fog. Considering the widespread application of Mono 3Det in autonomous driving, the severe performance degradation due to out-of-distribution (OOD) test data may lead to unexpected traffic accidents and pose serious safety risks. Therefore, it is crucial to deal with the OOD generalization problem for Mono 3Det." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "In addressing the OOD challenges specifically in test scenarios, one paradigm that has emerged as highly promising and gaining traction is Test-Time Adaptation (TTA), which seeks to tackle data distribution shifts by adapting a well-trained model to unlabeled test images in real time. Test-time training (TTT) [34] represents an initial approach of TTA in classification tasks, by adjusting the well-trained model to predict rotations through additional model training, while its computation demands at the adaptation stage are prohibitive" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 315, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 315, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 315, + 102 + ], + "type": "text", + "content": "Hongbin Lin, Yifan Zhang and et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "content": "in Mono 3Det applications, particularly in autonomous driving. To enhance efficiency, Tent [36] and EATA [26] have been developed for Fully Test-Time Adaptation (Fully TTA) where only unlabeled test images and a well-trained model are provided. Besides, Ev-TTA [12] and SOD [35] devise TTA methods to handle the event-based object recognition and weakly supervised salient object detection, respectively. Considering the constraints on time of Mono 3Det, we explore the fully TTA paradigm which seeks to deal with OOD test data in real time." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 205, + 483, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 205, + 483, + 384 + ], + "spans": [ + { + "bbox": [ + 130, + 205, + 483, + 384 + ], + "type": "text", + "content": "To investigate this paradigm for Mono 3Det, we dig into the detection outcomes for objects within test scenarios with variations or corruptions that are commonly caused by weather or cameras. Specifically, we directly apply the well-trained model to the validation set of KITTI which has been artificially injected with four distinct types of weather-related corruptions, namely Brightness, Frost, Fog, and Snow. Subsequently, we plot their distributions of detection scores (c.f. Fig. 1). It is observed that the detection scores of test objects tend to markedly decline as well as the high-score objects are scarce in the extreme scenario (Snow) when the well-trained model is directly applied to the scenarios with corruptions. This phenomenon indicates that: 1) The pre-trained Mono 3Det model struggles to discriminate between objects and the background within OOD test data, presenting as quantities of omissions and unconfident detections. 2) Directly applying existing fully TTA methods to Mono 3Det could only get suboptimal performance since they struggle to optimize the model without enough high-score (positive) detections, especially in certain extreme scenarios." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 390, + 483, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 390, + 483, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 390, + 483, + 668 + ], + "type": "text", + "content": "To handle it in Mono 3Det, we propose a Monocular Test-Time Adaptation (MonoTTA) method, consisting of the reliability-driven adaptation and noise-guard adaptation strategies: 1) Reliability-driven adaptation. Specifically, data distribution shifts lead to omissions and noisy detections while our empirical analysis suggests that high detection score objects are still reliable (c.f. Fig. 3 (a)). Moreover, even if we only optimize the model via high-score objects (" + }, + { + "bbox": [ + 130, + 390, + 483, + 668 + ], + "type": "inline_equation", + "content": "e.g., \\geq 0.5" + }, + { + "bbox": [ + 130, + 390, + 483, + 668 + ], + "type": "text", + "content": "), both the numbers of low-score and high-score objects increase (c.f. Fig. 3 (b)). These investigations motivate us that exploiting high-score objects rather than all objects for model adaptation would be a more reliable way to alleviate data distribution shifts and discover potential objects. Hence, we develop a self-adaptive strategy for the identification of reliable high-score objects in test images and devise the adaptive optimization loss " + }, + { + "bbox": [ + 130, + 390, + 483, + 668 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{AO}" + }, + { + "bbox": [ + 130, + 390, + 483, + 668 + ], + "type": "text", + "content": " to exploit the reliable subset for model adaptation, alleviating the detection score decline issue of OOD test data and digging out more potential objects. 2) Noise-guard adaptation. In addition, data distribution shifts may also result in a scarcity of high-score objects, i.e., the majority of objects presenting low scores as the 'Snow' scenario in Fig. 3 (a). To this end, we develop a negative regularization term to make rational use of the numerous low-scoring objects in the Negative Learning manner [13]. On the one hand, the negative regularization term " + }, + { + "bbox": [ + 130, + 390, + 483, + 668 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{Nreg}" + }, + { + "bbox": [ + 130, + 390, + 483, + 668 + ], + "type": "text", + "content": " allows the model to conduct adaptation via numerous noisy low-scoring objects. Thus, the model can achieve more high-score objects after alleviating distribution shifts. On the other hand, this term also prevents the model from overfitting to noise and trivial solutions, i.e., assigning all classes of one object with high scores." + } + ] + } + ], + "index": 4 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 403, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 403, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 403, + 91, + 447, + 101 + ], + "type": "text", + "content": "MonoTTA" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 147, + 116, + 363, + 128 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 116, + 363, + 128 + ], + "spans": [ + { + "bbox": [ + 147, + 116, + 363, + 128 + ], + "type": "text", + "content": "We summarize the main contributions as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 134, + 479, + 312 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 138, + 134, + 479, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 134, + 479, + 191 + ], + "spans": [ + { + "bbox": [ + 138, + 134, + 479, + 191 + ], + "type": "text", + "content": "- To the best of our knowledge, we are the first to explore Fully Test-Time Adaptation to address OOD generalization problems for Mono 3Det. We show that the explored novel paradigm can bring significant improvements to Mono 3Det models in OOD test scenarios, e.g., " + }, + { + "bbox": [ + 138, + 134, + 479, + 191 + ], + "type": "inline_equation", + "content": "137\\%" + }, + { + "bbox": [ + 138, + 134, + 479, + 191 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 138, + 134, + 479, + 191 + ], + "type": "inline_equation", + "content": "244\\%" + }, + { + "bbox": [ + 138, + 134, + 479, + 191 + ], + "type": "text", + "content": " average performance gains across 13 types of OOD shifts on KITTI." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 194, + 479, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 194, + 479, + 251 + ], + "spans": [ + { + "bbox": [ + 138, + 194, + 479, + 251 + ], + "type": "text", + "content": "- Our empirical investigation reveals an important insight that high-score objects maintain their reliability amidst various corruptions, while optimizing these high-score objects significantly boosts model confidence across all detections. This motivates the first effective test-time adaptation method (i.e., our MonoTTA) in Mono 3Det." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 253, + 479, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 253, + 479, + 312 + ], + "spans": [ + { + "bbox": [ + 138, + 253, + 479, + 312 + ], + "type": "text", + "content": "- Extensive experiments on 13 types of corruptions of KITTI and 2 real scenarios (daytime " + }, + { + "bbox": [ + 138, + 253, + 479, + 312 + ], + "type": "inline_equation", + "content": "\\leftrightarrow" + }, + { + "bbox": [ + 138, + 253, + 479, + 312 + ], + "type": "text", + "content": " night) of nuScenes demonstrate the effectiveness of our MonoTTA in boosting existing Mono 3Det methods [29,48] to handle test-time OOD problems. Even for instance-level methods [40], MonoTTA also maintains sufficient improvement, which further confirms its applicability." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 330, + 237, + 342 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 330, + 237, + 342 + ], + "spans": [ + { + "bbox": [ + 132, + 330, + 237, + 342 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 354, + 480, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 354, + 480, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 480, + 567 + ], + "type": "text", + "content": "We first review the literature on Monocular 3D Object Detection, and then discuss Source-free Domain Adaptation and Test-Time Adaptation methods. More discussions on Unsupervised Domain Adaptation [18,41,47] are in Appendix A. Monocular 3D Object Detection aims to perceive 3D objects from a single 2D image. Existing Mono 3Det methods could be divided into two groups according to the use of extra information. On the one hand, some existing methods leverage extra pre-trained depth estimation modules [6,39,49] to solve one of the most difficult problems in Mono 3Det, i.e., depth estimation from a single image. Other methods utilize LiDAR information, e.g., generating pseudo-LiDAR [23,31,37]. It is worth noting that Monoground [29] proposes to introduce the ground plane as prior information, and MonoNeRD [40] proposes to utilize scene geometric clues to enhance the detector's performance in the implicit reconstruction manner. On the other hand, some Mono 3Det methods try to detect 3D objects without extra data. For example, SMOKE [20] proposes to detect 3D objects as the key points estimation task. Then, Monoflex [48] improves this idea by providing a flexible definition of object centers, which unifies the centers of regular and truncated objects. GrooMeD-NMS [15] proposes a grouped mathematically differentiable Non-Maximal Suppression for Mono 3Det." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 570, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 479, + 665 + ], + "type": "text", + "content": "Source-free Domain Adaptation (SFDA) aims to adapt the pre-trained source model to an unlabeled target domain without using the source data due to privacy issues [17, 30]. SF-UDA" + }, + { + "bbox": [ + 130, + 570, + 479, + 665 + ], + "type": "inline_equation", + "content": "^3" + }, + { + "bbox": [ + 130, + 570, + 479, + 665 + ], + "type": "text", + "content": "D [32] first explores the SFDA framework to adapt the PointRCNN 3D detector to target domains, which consists of pseudolabeling, reversible scale-transformations and motion coherency. Recently, the authors [10] seek to exploit the source model more reliably and propose an uncertainty-aware teacher-student framework to filter incorrect pseudo labels during model adaptation, alleviating the negative impact of label noise." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 315, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 315, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 315, + 102 + ], + "type": "text", + "content": "Hongbin Lin, Yifan Zhang and et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "Nonetheless, SFDA assumes all target data to be known in advance and makes predictions after multiple epochs of optimization, which may not be viable for real-time applications due to computational or time constraints." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 480, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 480, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 480, + 319 + ], + "type": "text", + "content": "Test-Time Adaptation (TTA) seeks to improve model performance on test data via model adaptation through test samples even if data shifts exist. Early TTA methods [19, 34] endeavor to conduct additional model optimization on training data by self-supervised objectives, and then adapt the well-trained model to the test data via self-supervised objectives. However, in Mono 3Det applications like autonomous driving, the computation demands of such methods are prohibitive. To solve this, Fully Test-Time Adaptation methods are developed to adapt the well-trained model, where only unlabeled test images are available. Specifically, certain methods [25, 27, 33] tackle data distribution shifts by adapting the batch normalization layer statistics, while others alleviate this issue either by the entropy minimization of test data [8, 36] or maximizing the prediction consistency of different augmentations [43-46]. As for object detection tasks, Ev-TTA [12] and SOD [35] try to handle the event-based object recognition and weakly supervised salient object detection offline, respectively." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 320, + 480, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 320, + 480, + 368 + ], + "spans": [ + { + "bbox": [ + 130, + 320, + 480, + 368 + ], + "type": "text", + "content": "However, existing fully TTA methods struggle to optimize the model and solve distribution shifts in Mono 3Det due to numerous false negative detections. To the best of our knowledge, our MonoTTA stands as the first fully TTA method that handles distribution shifts for Mono 3Det models in real time." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 388, + 354, + 403 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 388, + 354, + 403 + ], + "spans": [ + { + "bbox": [ + 131, + 388, + 354, + 403 + ], + "type": "text", + "content": "3 Monocular Test-Time Adaptation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 415, + 259, + 427 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 415, + 259, + 427 + ], + "spans": [ + { + "bbox": [ + 131, + 415, + 259, + 427 + ], + "type": "text", + "content": "3.1 Problem Statement" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 437, + 480, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 480, + 569 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 480, + 569 + ], + "type": "text", + "content": "Without loss of generality, we denote the pre-trained (or well-trained) model as " + }, + { + "bbox": [ + 130, + 437, + 480, + 569 + ], + "type": "inline_equation", + "content": "f_{\\Theta_0}(\\mathbf{x})" + }, + { + "bbox": [ + 130, + 437, + 480, + 569 + ], + "type": "text", + "content": ", which is achieved via training on labeled training images " + }, + { + "bbox": [ + 130, + 437, + 480, + 569 + ], + "type": "inline_equation", + "content": "\\{(\\mathbf{x}_i^s,\\mathbf{y}_i^s)\\}_{i = 1}^N" + }, + { + "bbox": [ + 130, + 437, + 480, + 569 + ], + "type": "text", + "content": ". The training images follow the training distribution " + }, + { + "bbox": [ + 130, + 437, + 480, + 569 + ], + "type": "inline_equation", + "content": "P(\\mathbf{x})" + }, + { + "bbox": [ + 130, + 437, + 480, + 569 + ], + "type": "text", + "content": " (i.e., " + }, + { + "bbox": [ + 130, + 437, + 480, + 569 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^s\\sim P(\\mathbf{x})" + }, + { + "bbox": [ + 130, + 437, + 480, + 569 + ], + "type": "text", + "content": "). Here, " + }, + { + "bbox": [ + 130, + 437, + 480, + 569 + ], + "type": "inline_equation", + "content": "\\Theta_0" + }, + { + "bbox": [ + 130, + 437, + 480, + 569 + ], + "type": "text", + "content": " represents the parameters of the pre-trained model and " + }, + { + "bbox": [ + 130, + 437, + 480, + 569 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 437, + 480, + 569 + ], + "type": "text", + "content": " is the number of training data. During the training stage, the model is optimized to fit (or overfit) the training data. Then, at the test stage, the model will be able to perform well if the unlabeled test images " + }, + { + "bbox": [ + 130, + 437, + 480, + 569 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_t = \\{\\mathbf{x}_i\\}_{i = 1}^{N_t}" + }, + { + "bbox": [ + 130, + 437, + 480, + 569 + ], + "type": "text", + "content": " follows the identical data distribution, i.e., " + }, + { + "bbox": [ + 130, + 437, + 480, + 569 + ], + "type": "inline_equation", + "content": "\\mathbf{x}\\sim P(\\mathbf{x})" + }, + { + "bbox": [ + 130, + 437, + 480, + 569 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 130, + 437, + 480, + 569 + ], + "type": "inline_equation", + "content": "N_{t}" + }, + { + "bbox": [ + 130, + 437, + 480, + 569 + ], + "type": "text", + "content": " is the total number of test images. However, in real applications, it is possible for the pre-trained model to encounter Out-Of-Distribution (OOD) test samples due to prevalent natural corruptions, namely distribution shifts, i.e., " + }, + { + "bbox": [ + 130, + 437, + 480, + 569 + ], + "type": "inline_equation", + "content": "\\mathbf{x}\\sim Q(\\mathbf{x})" + }, + { + "bbox": [ + 130, + 437, + 480, + 569 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 437, + 480, + 569 + ], + "type": "inline_equation", + "content": "P(\\mathbf{x})\\neq Q(\\mathbf{x})" + }, + { + "bbox": [ + 130, + 437, + 480, + 569 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 570, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 481, + 665 + ], + "type": "text", + "content": "To address this issue, fully test-time adaptation [36] seeks to tackle distribution shifts by adapting the pre-trained model " + }, + { + "bbox": [ + 130, + 570, + 481, + 665 + ], + "type": "inline_equation", + "content": "f_{\\Theta_0}(\\mathbf{x})" + }, + { + "bbox": [ + 130, + 570, + 481, + 665 + ], + "type": "text", + "content": " to unlabeled test images " + }, + { + "bbox": [ + 130, + 570, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{x}_i\\}_{i=1}^{N_t}" + }, + { + "bbox": [ + 130, + 570, + 481, + 665 + ], + "type": "text", + "content": " in real time. To achieve this goal, existing methods typically endeavor to update the model through the minimization of unsupervised objectives defined on test samples by " + }, + { + "bbox": [ + 130, + 570, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\min_{\\hat{\\Theta}} \\mathcal{L}(\\mathbf{x}; \\Theta)" + }, + { + "bbox": [ + 130, + 570, + 481, + 665 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 570, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\sim Q(\\mathbf{x})" + }, + { + "bbox": [ + 130, + 570, + 481, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 570, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\hat{\\Theta} \\subseteq \\Theta" + }, + { + "bbox": [ + 130, + 570, + 481, + 665 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 130, + 570, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\hat{\\Theta}" + }, + { + "bbox": [ + 130, + 570, + 481, + 665 + ], + "type": "text", + "content": " denotes the subset of model parameters that should be updated (i.e., batch normalization layers following existing methods [26,36]). Most existing fully TTA methods focus on classification tasks, heavily relying on sufficient positive predictions for" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 403, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 403, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 403, + 91, + 447, + 101 + ], + "type": "text", + "content": "MonoTTA" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 145, + 114, + 474, + 296 + ], + "blocks": [ + { + "bbox": [ + 145, + 114, + 474, + 296 + ], + "lines": [ + { + "bbox": [ + 145, + 114, + 474, + 296 + ], + "spans": [ + { + "bbox": [ + 145, + 114, + 474, + 296 + ], + "type": "image", + "image_path": "7ba916103de92e85b6fa56879c52dc92c2b51841c794c47ecc76865a941152c8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 304, + 482, + 394 + ], + "lines": [ + { + "bbox": [ + 130, + 304, + 482, + 394 + ], + "spans": [ + { + "bbox": [ + 130, + 304, + 482, + 394 + ], + "type": "text", + "content": "Fig. 2: An illustration of our MonoTTA. During the test phase, only the pre-trained model " + }, + { + "bbox": [ + 130, + 304, + 482, + 394 + ], + "type": "inline_equation", + "content": "f_{\\Theta_0}(\\mathbf{x})" + }, + { + "bbox": [ + 130, + 304, + 482, + 394 + ], + "type": "text", + "content": " and unlabeled test images " + }, + { + "bbox": [ + 130, + 304, + 482, + 394 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{x}_i\\}_{i=1}^{N_t}" + }, + { + "bbox": [ + 130, + 304, + 482, + 394 + ], + "type": "text", + "content": " are given. To conduct model adaptation, we initialize the model " + }, + { + "bbox": [ + 130, + 304, + 482, + 394 + ], + "type": "inline_equation", + "content": "f_{\\Theta}(\\mathbf{x})" + }, + { + "bbox": [ + 130, + 304, + 482, + 394 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 130, + 304, + 482, + 394 + ], + "type": "inline_equation", + "content": "\\Theta_0" + }, + { + "bbox": [ + 130, + 304, + 482, + 394 + ], + "type": "text", + "content": " and only update the parameters of batch normalization layers. When a batch of test images arrives, we first compute test object scores and refine the adaptive threshold " + }, + { + "bbox": [ + 130, + 304, + 482, + 394 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 304, + 482, + 394 + ], + "type": "text", + "content": " to select the reliable high-score objects, thereby optimizing " + }, + { + "bbox": [ + 130, + 304, + 482, + 394 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 130, + 304, + 482, + 394 + ], + "type": "text", + "content": " via the adaptive optimization loss " + }, + { + "bbox": [ + 130, + 304, + 482, + 394 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{AO}" + }, + { + "bbox": [ + 130, + 304, + 482, + 394 + ], + "type": "text", + "content": ". Meanwhile, we devise a negative regularization term " + }, + { + "bbox": [ + 130, + 304, + 482, + 394 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{Nreg}" + }, + { + "bbox": [ + 130, + 304, + 482, + 394 + ], + "type": "text", + "content": " to facilitate the model to avoid overfitting to noise and trivial solutions. Here, Ped. and Cyc. represent Pedestrian and Cyclist in KITTI." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 410, + 482, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 410, + 482, + 518 + ], + "spans": [ + { + "bbox": [ + 130, + 410, + 482, + 518 + ], + "type": "text", + "content": "model adaptation. Nevertheless, there is a significant difference between conventional classification tasks and Mono 3Det. As previously indicated, the detection scores of test images " + }, + { + "bbox": [ + 130, + 410, + 482, + 518 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 130, + 410, + 482, + 518 + ], + "type": "text", + "content": " derived from " + }, + { + "bbox": [ + 130, + 410, + 482, + 518 + ], + "type": "inline_equation", + "content": "f_{\\Theta_0}(\\mathbf{x})" + }, + { + "bbox": [ + 130, + 410, + 482, + 518 + ], + "type": "text", + "content": " are prone to markedly decrease in the presence of corruptions as shown in Fig. 1, leading to severe omissions (numerous false negatives) in Mono 3Det. In such circumstances, the scarcity of positive detections presents a significant challenge for model adaptation to test distributions while adapting the model with unreliable low-score detections may significantly introduce the noise. Therefore, existing fully TTA methods tend to fail in the OOD generalization problems of Mono 3Det." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 536, + 239, + 548 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 536, + 239, + 548 + ], + "spans": [ + { + "bbox": [ + 132, + 536, + 239, + 548 + ], + "type": "text", + "content": "3.2 Overall Scheme" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 557, + 482, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 629 + ], + "type": "text", + "content": "After thoroughly examining the characteristics and challenges of Mono 3Det, we introduce a Monocular Test-Time Adaptation (MonoTTA) method to address the OOD problems for Mono 3Det models, which seeks to solve the object score declining issue within unlabeled OOD test data. As shown in Fig. 2, MonoTTA consists of two strategies: 1) Reliability-driven adaptation and 2) Noise-guard adaptation. We first briefly introduce the two strategies below." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "content": "First, we develop a reliability-driven adaptation strategy (c.f. Section 3.3) to conduct reliable model adaptation for OOD test data based on dependable test objects. Our empirical investigations inspire us to exploit those relatively reliable" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 315, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 315, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 315, + 102 + ], + "type": "text", + "content": "Hongbin Lin, Yifan Zhang and et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 157, + 114, + 293, + 220 + ], + "blocks": [ + { + "bbox": [ + 157, + 114, + 293, + 220 + ], + "lines": [ + { + "bbox": [ + 157, + 114, + 293, + 220 + ], + "spans": [ + { + "bbox": [ + 157, + 114, + 293, + 220 + ], + "type": "image", + "image_path": "3acbb989c5da1a458cc9645f9fc38d02e6b56bbe28819edb4829bbe02c59e89f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 228, + 222, + 236, + 230 + ], + "lines": [ + { + "bbox": [ + 228, + 222, + 236, + 230 + ], + "spans": [ + { + "bbox": [ + 228, + 222, + 236, + 230 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 298, + 118, + 452, + 220 + ], + "blocks": [ + { + "bbox": [ + 298, + 118, + 452, + 220 + ], + "lines": [ + { + "bbox": [ + 298, + 118, + 452, + 220 + ], + "spans": [ + { + "bbox": [ + 298, + 118, + 452, + 220 + ], + "type": "image", + "image_path": "b71877ed35d254d833d907c9acdf1fabec7900d73d66895bba8a8d1b23eeaa23.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 382, + 221, + 392, + 229 + ], + "lines": [ + { + "bbox": [ + 382, + 221, + 392, + 229 + ], + "spans": [ + { + "bbox": [ + 382, + 221, + 392, + 229 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 232, + 482, + 310 + ], + "lines": [ + { + "bbox": [ + 130, + 232, + 482, + 310 + ], + "spans": [ + { + "bbox": [ + 130, + 232, + 482, + 310 + ], + "type": "text", + "content": "Fig. 3: Based on MonoGround [29], we conduct two empirical studies (Car, KITTI), with the 3D IoU threshold of 0.5. (a) We visualize the accuracy of the objects across varied scoring ranges, which shows that the accuracy of objects with high scores remains relatively stable even in the presence of diverse corruptions (Ideal means in-distribution scenarios). (b) We visualize the number of low & high-score objects before and after optimization. Although only high-score objects are optimized, the model treats low-score objects with more confidence." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 326, + 480, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 326, + 480, + 373 + ], + "spans": [ + { + "bbox": [ + 130, + 326, + 480, + 373 + ], + "type": "text", + "content": "test objects for alleviating distribution shifts, thereby discovering more potential objects. To this end, MonoTTA excludes unreliable test objects involving an adaptive threshold " + }, + { + "bbox": [ + 130, + 326, + 480, + 373 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 326, + 480, + 373 + ], + "type": "text", + "content": " for any unlabeled test data. Subsequently, the model is optimized by the adaptive optimization loss " + }, + { + "bbox": [ + 130, + 326, + 480, + 373 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{AO}" + }, + { + "bbox": [ + 130, + 326, + 480, + 373 + ], + "type": "text", + "content": " via the selected reliable objects." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 374, + 482, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 374, + 482, + 481 + ], + "spans": [ + { + "bbox": [ + 130, + 374, + 482, + 481 + ], + "type": "text", + "content": "Second, we tend to utilize plenty of low-score objects to adapt the model in an indirect manner instead of directly optimizing the model since low-score objects are noisy. Hence, we devise a noise-guard adaptation strategy (c.f. Section 3.4) to prevent the model from overfitting to noisy predictions and falling into trivial solutions. Specifically, we randomly choose one of the negative classes of low-score objects and minimize the scores (e.g., score 0.03 of [0.10, 0.12, 0.03] c.f. Fig. 2) after simply filtering out extremely low-score detections. Even though the positive class is noisy (i.e., score 0.12), this term is capable of optimizing the model indirectly, i.e., learn to deny the negative category of the object." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 146, + 482, + 392, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 482, + 392, + 494 + ], + "spans": [ + { + "bbox": [ + 146, + 482, + 392, + 494 + ], + "type": "text", + "content": "Overall, the training scheme of MonoTTA is as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 246, + 502, + 481, + 522 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 246, + 502, + 481, + 522 + ], + "spans": [ + { + "bbox": [ + 246, + 502, + 481, + 522 + ], + "type": "interline_equation", + "content": "\\min _ {\\hat {\\Theta}} \\mathcal {L} _ {A O} (\\hat {\\Theta}) + \\lambda \\mathcal {L} _ {N r e g} (\\hat {\\Theta}), \\tag {1}", + "image_path": "5766cac4354dbdbe53404c58cec01f4f91cc57c62a791225269e3fb1e82b16db.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 531, + 482, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 531, + 482, + 556 + ], + "spans": [ + { + "bbox": [ + 130, + 531, + 482, + 556 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 531, + 482, + 556 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 531, + 482, + 556 + ], + "type": "text", + "content": " is the balance hyper-parameter. The pseudo-code of MonoTTA is summarized in Algorithm 1." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 573, + 312, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 573, + 312, + 586 + ], + "spans": [ + { + "bbox": [ + 131, + 573, + 312, + 586 + ], + "type": "text", + "content": "3.3 Reliability-Driven Adaptation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 594, + 480, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 480, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 480, + 629 + ], + "type": "text", + "content": "To identify dependable test objects and conduct test-time model adaptation, we propose a reliability-driven adaptation strategy that consists of two components: 1) Reliable object identification and 2) Adaptive model optimization." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "content": "Reliable object identification. When unlabeled test images arrive, it is difficult for the pre-trained model to get accurate detections on OOD test data due to the decline in detection scores. To resolve this, we dig into the detection" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "type": "text", + "content": "MonoTTA" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 285 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 285 + ], + "type": "text", + "content": "accuracy of the pre-trained model in test scenarios with various corruptions. Specifically, based on MonoGround [29], we visualize the detection precision for the Car category within the KITTI dataset across varied scoring ranges, with the 3D Intersection over Union (IoU) threshold of 0.5. As shown in Fig. 3 (a), we find that high-score objects are more reliable and relatively stable even in the presence of diverse corruptions. Following this, we propose to select reliable high-score objects to conduct model adaptation via an adaptive threshold " + }, + { + "bbox": [ + 130, + 116, + 482, + 285 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 116, + 482, + 285 + ], + "type": "text", + "content": ". Specifically, we exploit the model " + }, + { + "bbox": [ + 130, + 116, + 482, + 285 + ], + "type": "inline_equation", + "content": "f_{\\Theta}(\\cdot)" + }, + { + "bbox": [ + 130, + 116, + 482, + 285 + ], + "type": "text", + "content": " (initialized by " + }, + { + "bbox": [ + 130, + 116, + 482, + 285 + ], + "type": "inline_equation", + "content": "\\Theta_0" + }, + { + "bbox": [ + 130, + 116, + 482, + 285 + ], + "type": "text", + "content": ") to infer a batch of test images " + }, + { + "bbox": [ + 130, + 116, + 482, + 285 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{x}_b\\}_{b=1}^B" + }, + { + "bbox": [ + 130, + 116, + 482, + 285 + ], + "type": "text", + "content": " and obtain the heatmap " + }, + { + "bbox": [ + 130, + 116, + 482, + 285 + ], + "type": "inline_equation", + "content": "\\mathbf{h}_i" + }, + { + "bbox": [ + 130, + 116, + 482, + 285 + ], + "type": "text", + "content": " of each image " + }, + { + "bbox": [ + 130, + 116, + 482, + 285 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 130, + 116, + 482, + 285 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 130, + 116, + 482, + 285 + ], + "type": "inline_equation", + "content": "\\mathbf{h}_i = f_{\\Theta}(\\mathbf{x}_i)" + }, + { + "bbox": [ + 130, + 116, + 482, + 285 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 116, + 482, + 285 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 116, + 482, + 285 + ], + "type": "text", + "content": " ranges from 0 to " + }, + { + "bbox": [ + 130, + 116, + 482, + 285 + ], + "type": "inline_equation", + "content": "B-1" + }, + { + "bbox": [ + 130, + 116, + 482, + 285 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 482, + 285 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 130, + 116, + 482, + 285 + ], + "type": "text", + "content": " denotes the batch size. Then, we could achieve object score maps " + }, + { + "bbox": [ + 130, + 116, + 482, + 285 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_i \\in \\mathbb{R}^{B \\times N_m}" + }, + { + "bbox": [ + 130, + 116, + 482, + 285 + ], + "type": "text", + "content": " like the peaks in " + }, + { + "bbox": [ + 130, + 116, + 482, + 285 + ], + "type": "inline_equation", + "content": "\\mathbf{h}_i" + }, + { + "bbox": [ + 130, + 116, + 482, + 285 + ], + "type": "text", + "content": " [7] after normalization, where " + }, + { + "bbox": [ + 130, + 116, + 482, + 285 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 116, + 482, + 285 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 482, + 285 + ], + "type": "inline_equation", + "content": "N_m" + }, + { + "bbox": [ + 130, + 116, + 482, + 285 + ], + "type": "text", + "content": " denotes the number of classes and the maximum of detected objects, respectively. With the score map " + }, + { + "bbox": [ + 130, + 116, + 482, + 285 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_i" + }, + { + "bbox": [ + 130, + 116, + 482, + 285 + ], + "type": "text", + "content": ", we update the adaptive threshold " + }, + { + "bbox": [ + 130, + 116, + 482, + 285 + ], + "type": "inline_equation", + "content": "\\alpha_t" + }, + { + "bbox": [ + 130, + 116, + 482, + 285 + ], + "type": "text", + "content": " at the iteration " + }, + { + "bbox": [ + 130, + 116, + 482, + 285 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 116, + 482, + 285 + ], + "type": "text", + "content": " in the exponential moving average manner by:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 225, + 298, + 481, + 326 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 298, + 481, + 326 + ], + "spans": [ + { + "bbox": [ + 225, + 298, + 481, + 326 + ], + "type": "interline_equation", + "content": "\\alpha_ {t} = \\left\\{ \\begin{array}{l l} \\gamma , & \\text {i f} t = 1 \\\\ \\beta \\bar {m} _ {t} + (1 - \\beta) \\alpha_ {t - 1}, & \\text {i f} t > 1 \\end{array} \\right., \\tag {2}", + "image_path": "c328ab84755ee1042c0de4ae96acf91dc661d4315375a2e27bbba28aefc34efa.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 231, + 345, + 481, + 380 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 345, + 481, + 380 + ], + "spans": [ + { + "bbox": [ + 231, + 345, + 481, + 380 + ], + "type": "interline_equation", + "content": "\\bar {m} _ {t} = \\frac {1}{B} \\sum_ {i = 1} ^ {B} \\frac {\\sum_ {j = 1} ^ {N _ {m}} s _ {i j} \\cdot \\mathbb {I} \\left(s _ {i j} \\geq \\gamma\\right)}{\\sum_ {j = 1} ^ {N _ {m}} \\mathbb {I} \\left(s _ {i j} \\geq \\gamma\\right)}. \\tag {3}", + "image_path": "cede7efb103cd1ecf30ca18e07dd663db2cf3bbe87664de832ba487bc08ba875.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 389, + 482, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 389, + 482, + 461 + ], + "spans": [ + { + "bbox": [ + 130, + 389, + 482, + 461 + ], + "type": "text", + "content": "In Eqn. (2), " + }, + { + "bbox": [ + 130, + 389, + 482, + 461 + ], + "type": "inline_equation", + "content": "\\bar{m}_t" + }, + { + "bbox": [ + 130, + 389, + 482, + 461 + ], + "type": "text", + "content": " denotes the average score of all detected objects in a single batch of " + }, + { + "bbox": [ + 130, + 389, + 482, + 461 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 130, + 389, + 482, + 461 + ], + "type": "text", + "content": " test samples at iteration " + }, + { + "bbox": [ + 130, + 389, + 482, + 461 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 389, + 482, + 461 + ], + "type": "text", + "content": ", while " + }, + { + "bbox": [ + 130, + 389, + 482, + 461 + ], + "type": "inline_equation", + "content": "\\beta \\in [0,1]" + }, + { + "bbox": [ + 130, + 389, + 482, + 461 + ], + "type": "text", + "content": " is a decay coefficient. As for Eqn. (3), " + }, + { + "bbox": [ + 130, + 389, + 482, + 461 + ], + "type": "inline_equation", + "content": "s_{ij} \\in (0,1)" + }, + { + "bbox": [ + 130, + 389, + 482, + 461 + ], + "type": "text", + "content": " denotes the score of the " + }, + { + "bbox": [ + 130, + 389, + 482, + 461 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 130, + 389, + 482, + 461 + ], + "type": "text", + "content": "-th object in the " + }, + { + "bbox": [ + 130, + 389, + 482, + 461 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 389, + 482, + 461 + ], + "type": "text", + "content": "-th test image while " + }, + { + "bbox": [ + 130, + 389, + 482, + 461 + ], + "type": "inline_equation", + "content": "\\mathbb{I}(\\cdot)" + }, + { + "bbox": [ + 130, + 389, + 482, + 461 + ], + "type": "text", + "content": " is the indicator function. Note that " + }, + { + "bbox": [ + 130, + 389, + 482, + 461 + ], + "type": "inline_equation", + "content": "\\gamma \\in \\mathbb{R}^1" + }, + { + "bbox": [ + 130, + 389, + 482, + 461 + ], + "type": "text", + "content": " is a pre-defined object detection threshold adopted from existing methods at their original inference stage as well as " + }, + { + "bbox": [ + 130, + 389, + 482, + 461 + ], + "type": "inline_equation", + "content": "N_m" + }, + { + "bbox": [ + 130, + 389, + 482, + 461 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 463, + 482, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 463, + 482, + 547 + ], + "spans": [ + { + "bbox": [ + 130, + 463, + 482, + 547 + ], + "type": "text", + "content": "Adaptive model optimization. As shown in Fig. 3 (b), the optimization of high-score objects can also enhance the confidence of the model for relatively low-score objects. It motivates us that exploiting high-score objects rather than all objects for model adaptation would be a more reliable way to learn from OOD test data. Therefore, with the adaptive threshold " + }, + { + "bbox": [ + 130, + 463, + 482, + 547 + ], + "type": "inline_equation", + "content": "\\alpha_{t}" + }, + { + "bbox": [ + 130, + 463, + 482, + 547 + ], + "type": "text", + "content": ", we select the reliable subset of high scores from " + }, + { + "bbox": [ + 130, + 463, + 482, + 547 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_i" + }, + { + "bbox": [ + 130, + 463, + 482, + 547 + ], + "type": "text", + "content": " and calculate the adaptive optimization loss " + }, + { + "bbox": [ + 130, + 463, + 482, + 547 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{AO}" + }, + { + "bbox": [ + 130, + 463, + 482, + 547 + ], + "type": "text", + "content": " to adapt the model by:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 218, + 559, + 481, + 594 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 218, + 559, + 481, + 594 + ], + "spans": [ + { + "bbox": [ + 218, + 559, + 481, + 594 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {A O} = - \\frac {1}{B} \\sum_ {i = 1} ^ {B} \\sum_ {j = 1} ^ {N _ {m}} \\log \\left(s _ {i j} \\cdot \\mathbb {I} \\left(s _ {i j} \\geq \\alpha_ {t}\\right)\\right), \\tag {4}", + "image_path": "44dfc8a372b1b84e38bd3073b4634f66b2271aec4aa0de1beb9226f97ccc57ca.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "The adaptive optimization loss " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{AO}" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": " alleviates the potential data distribution shifts in OOD test data by allowing the model to confidently identify high-score test objects. As we mentioned, it solves the score decline issue of OOD test data in a more reliable way rather than directly optimizing all test objects, thereby avoiding overfitting to noise and discovering more potential test objects." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 315, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 315, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 315, + 102 + ], + "type": "text", + "content": "Hongbin Lin, Yifan Zhang and et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 116, + 373, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 373, + 128 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 373, + 128 + ], + "type": "text", + "content": "Algorithm 1 The pipeline of the proposed MonoTTA" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 132, + 481, + 155 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 132, + 481, + 155 + ], + "spans": [ + { + "bbox": [ + 132, + 132, + 481, + 155 + ], + "type": "text", + "content": "Require: Unlabeled test data " + }, + { + "bbox": [ + 132, + 132, + 481, + 155 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_t = \\{\\mathbf{x}_i\\}_{i=1}^{N_t}" + }, + { + "bbox": [ + 132, + 132, + 481, + 155 + ], + "type": "text", + "content": "; Pre-trained model " + }, + { + "bbox": [ + 132, + 132, + 481, + 155 + ], + "type": "inline_equation", + "content": "f_{\\Theta_0}(\\mathbf{x})" + }, + { + "bbox": [ + 132, + 132, + 481, + 155 + ], + "type": "text", + "content": "; Batch size " + }, + { + "bbox": [ + 132, + 132, + 481, + 155 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 132, + 132, + 481, + 155 + ], + "type": "text", + "content": "; Parameters " + }, + { + "bbox": [ + 132, + 132, + 481, + 155 + ], + "type": "inline_equation", + "content": "\\lambda, \\beta, \\eta" + }, + { + "bbox": [ + 132, + 132, + 481, + 155 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 136, + 155, + 420, + 232 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 136, + 155, + 300, + 166 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 155, + 300, + 166 + ], + "spans": [ + { + "bbox": [ + 136, + 155, + 300, + 166 + ], + "type": "text", + "content": "1: for a batch images " + }, + { + "bbox": [ + 136, + 155, + 300, + 166 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{x}_b\\}_{b = 1}^B" + }, + { + "bbox": [ + 136, + 155, + 300, + 166 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 136, + 155, + 300, + 166 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_t" + }, + { + "bbox": [ + 136, + 155, + 300, + 166 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 136, + 166, + 370, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 166, + 370, + 176 + ], + "spans": [ + { + "bbox": [ + 136, + 166, + 370, + 176 + ], + "type": "text", + "content": "2: Update the adaptive threshold " + }, + { + "bbox": [ + 136, + 166, + 370, + 176 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 136, + 166, + 370, + 176 + ], + "type": "text", + "content": " based on Eqn. (2);" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 136, + 176, + 405, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 176, + 405, + 188 + ], + "spans": [ + { + "bbox": [ + 136, + 176, + 405, + 188 + ], + "type": "text", + "content": "3: Calculate adaptive optimization loss " + }, + { + "bbox": [ + 136, + 176, + 405, + 188 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{AO}" + }, + { + "bbox": [ + 136, + 176, + 405, + 188 + ], + "type": "text", + "content": " based on Eqn. (4);" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 136, + 188, + 420, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 188, + 420, + 199 + ], + "spans": [ + { + "bbox": [ + 136, + 188, + 420, + 199 + ], + "type": "text", + "content": "4: Calculate negative regularization term " + }, + { + "bbox": [ + 136, + 188, + 420, + 199 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{Nreg}" + }, + { + "bbox": [ + 136, + 188, + 420, + 199 + ], + "type": "text", + "content": " based on Eqn. (6);" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 136, + 199, + 295, + 210 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 199, + 295, + 210 + ], + "spans": [ + { + "bbox": [ + 136, + 199, + 295, + 210 + ], + "type": "text", + "content": "5: Update " + }, + { + "bbox": [ + 136, + 199, + 295, + 210 + ], + "type": "inline_equation", + "content": "\\hat{\\Theta}" + }, + { + "bbox": [ + 136, + 199, + 295, + 210 + ], + "type": "text", + "content": " by optimizing Eqn. (1)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 136, + 210, + 184, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 210, + 184, + 220 + ], + "spans": [ + { + "bbox": [ + 136, + 210, + 184, + 220 + ], + "type": "text", + "content": "6: end for" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 136, + 220, + 316, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 220, + 316, + 232 + ], + "spans": [ + { + "bbox": [ + 136, + 220, + 316, + 232 + ], + "type": "text", + "content": "7: return Detection Results for all " + }, + { + "bbox": [ + 136, + 220, + 316, + 232 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\in \\mathcal{D}_t" + }, + { + "bbox": [ + 136, + 220, + 316, + 232 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 258, + 285, + 269 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 258, + 285, + 269 + ], + "spans": [ + { + "bbox": [ + 132, + 258, + 285, + 269 + ], + "type": "text", + "content": "3.4 Noise-Guard Adaptation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 277, + 482, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 277, + 482, + 458 + ], + "spans": [ + { + "bbox": [ + 130, + 277, + 482, + 458 + ], + "type": "text", + "content": "Through the optimization of adaptive optimization loss " + }, + { + "bbox": [ + 130, + 277, + 482, + 458 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{AO}" + }, + { + "bbox": [ + 130, + 277, + 482, + 458 + ], + "type": "text", + "content": ", the model is refined to yield more confident detection outcomes. Nevertheless, high-scoring objects may be scarce due to the distribution shifts, making the adaptation procedure difficult. Meanwhile, exclusive reliance on " + }, + { + "bbox": [ + 130, + 277, + 482, + 458 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{AO}" + }, + { + "bbox": [ + 130, + 277, + 482, + 458 + ], + "type": "text", + "content": " for adaptation may result in trivial solutions, whereby the model indiscriminately assigns high scores to all categories. Previous studies [13, 14] indicate that deep neural networks could learn from noisy pseudo labels in classification tasks through negative learning. Thus, MonoTTA proposes to learn from noisy low-score objects in a negative learning manner for Mono 3Det. Specifically, we denote " + }, + { + "bbox": [ + 130, + 277, + 482, + 458 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{s}}_i\\in \\mathbb{R}^{B\\times N_m\\times K}" + }, + { + "bbox": [ + 130, + 277, + 482, + 458 + ], + "type": "text", + "content": " as the multi-class score map for the test image " + }, + { + "bbox": [ + 130, + 277, + 482, + 458 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 130, + 277, + 482, + 458 + ], + "type": "text", + "content": ", i.e. the multi-class score map " + }, + { + "bbox": [ + 130, + 277, + 482, + 458 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{s}}_i" + }, + { + "bbox": [ + 130, + 277, + 482, + 458 + ], + "type": "text", + "content": " contains not only the highest score of objects but also the scores for other classes. Here " + }, + { + "bbox": [ + 130, + 277, + 482, + 458 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_i = \\arg \\max_k\\hat{\\mathbf{s}}_{ik}" + }, + { + "bbox": [ + 130, + 277, + 482, + 458 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 130, + 277, + 482, + 458 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 277, + 482, + 458 + ], + "type": "text", + "content": " is the class index. Employing a simple constant threshold " + }, + { + "bbox": [ + 130, + 277, + 482, + 458 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 130, + 277, + 482, + 458 + ], + "type": "text", + "content": ", we filter out extremely low-score objects and then randomly select a negative class " + }, + { + "bbox": [ + 130, + 277, + 482, + 458 + ], + "type": "inline_equation", + "content": "\\bar{k}" + }, + { + "bbox": [ + 130, + 277, + 482, + 458 + ], + "type": "text", + "content": " for each object. Subsequently, we compute the regularization loss " + }, + { + "bbox": [ + 130, + 277, + 482, + 458 + ], + "type": "inline_equation", + "content": "e_k" + }, + { + "bbox": [ + 130, + 277, + 482, + 458 + ], + "type": "text", + "content": " for each class " + }, + { + "bbox": [ + 130, + 277, + 482, + 458 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 277, + 482, + 458 + ], + "type": "text", + "content": " with relatively low object scores " + }, + { + "bbox": [ + 130, + 277, + 482, + 458 + ], + "type": "inline_equation", + "content": "s_{ij}\\in [\\eta ,\\alpha_t)" + }, + { + "bbox": [ + 130, + 277, + 482, + 458 + ], + "type": "text", + "content": " by:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 214, + 465, + 481, + 500 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 214, + 465, + 481, + 500 + ], + "spans": [ + { + "bbox": [ + 214, + 465, + 481, + 500 + ], + "type": "interline_equation", + "content": "e _ {k} = - \\sum_ {i = 1} ^ {B} \\sum_ {j = 1} ^ {N _ {m}} \\bar {y} _ {i j} \\log \\left(1 - s _ {i j \\bar {k}} \\cdot \\mathbb {I} (\\bar {k} = k)\\right), \\tag {5}", + "image_path": "51b274f95e6454178ddb59baaad76d750d2e790d94c472ceee101aa9c3212b59.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 508, + 481, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 508, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 508, + 481, + 533 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 508, + 481, + 533 + ], + "type": "inline_equation", + "content": "\\bar{y}_{ij} = 1 - s_{ij\\bar{k}}" + }, + { + "bbox": [ + 130, + 508, + 481, + 533 + ], + "type": "text", + "content": " is a constant weight. We further define " + }, + { + "bbox": [ + 130, + 508, + 481, + 533 + ], + "type": "inline_equation", + "content": "n_k" + }, + { + "bbox": [ + 130, + 508, + 481, + 533 + ], + "type": "text", + "content": " as the frequency of negative scores corresponding to class " + }, + { + "bbox": [ + 130, + 508, + 481, + 533 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 508, + 481, + 533 + ], + "type": "text", + "content": " in the test batch and balance " + }, + { + "bbox": [ + 130, + 508, + 481, + 533 + ], + "type": "inline_equation", + "content": "e_k" + }, + { + "bbox": [ + 130, + 508, + 481, + 533 + ], + "type": "text", + "content": " by:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 269, + 540, + 481, + 574 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 269, + 540, + 481, + 574 + ], + "spans": [ + { + "bbox": [ + 269, + 540, + 481, + 574 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {N r e g} = \\sum_ {k = 1} ^ {K} \\frac {e _ {k}}{n _ {k}}. \\tag {6}", + "image_path": "2e5cdacb6919fcff3f02e10219157764e69486c9e11dae582002a14f313a5e9f.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 582, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 667 + ], + "type": "text", + "content": "As we mentioned before, detections with low and intermediate scores tend to be more noisy (c.f. Fig. 3 (a)), which is unreliable for direct model adaptation. To this end, we introduce the regularization term " + }, + { + "bbox": [ + 130, + 582, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{Nreg}" + }, + { + "bbox": [ + 130, + 582, + 482, + 667 + ], + "type": "text", + "content": " to leverage noisy low-score detections and improve the model for assimilating potentially accurate information (c.f. Term 2 in Fig. 2). Moreover, this term also prevents the model from trivial solutions, i.e., indiscriminately assigning high scores to all classes of a single object (e.g., TENT [36]). In addition, high-score objects may be absent" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "type": "text", + "content": "MonoTTA" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 148, + 142, + 466, + 228 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "type": "text", + "content": "Table 1: Comparison with baselines on the KITTI-C validation set, severity level 1 regarding Mean " + }, + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "type": "inline_equation", + "content": "AP_{3D|R_{40}}" + }, + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "type": "text", + "content": ". The bold number indicates the best result." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 148, + 142, + 466, + 228 + ], + "lines": [ + { + "bbox": [ + 148, + 142, + 466, + 228 + ], + "spans": [ + { + "bbox": [ + 148, + 142, + 466, + 228 + ], + "type": "table", + "html": "
Car, IoU @ 0.7, 0.5, 0.5
MethodNoiseBlurWeatherDigitalAvg.
Gauss.ShotImpul.Defoc.GlassMotionSnowFrostFogBrit.Contr.PixelSat.
Monoflex [48]3.847.485.312.593.7311.050.237.777.5724.876.9228.1631.4610.84
• BN adaptation [33]13.5821.9318.7815.878.5924.325.4221.4524.6331.8030.5841.0430.7122.21
• TENT [36]17.8027.0923.1821.6611.9028.756.8426.5830.7835.6534.7241.7135.9126.35
• EATA [26]16.6726.4225.0722.5413.2327.737.8726.5831.1035.3935.2841.4036.7226.62
• ActMAD [24]12.2621.4920.4711.107.5323.795.5719.1126.5426.2423.9238.5133.7520.79
• MonoTTA (Ours)21.1528.6526.6425.9119.2631.4812.4330.2433.7536.8436.8341.9738.1329.48
MonoGround [29]2.404.103.313.712.678.130.225.544.5925.374.0033.5728.089.67
• BN adaptation [33]13.4923.5219.6916.337.6123.997.9820.7124.0031.3429.0343.0632.9922.60
• TENT [36]17.9429.6019.9023.4513.9029.3910.3226.6533.3535.9636.3943.3537.7927.54
• EATA [26]16.0326.0818.0820.2812.4027.379.2223.7929.4933.6532.5843.6136.0025.28
• ActMAD [24]13.6522.3920.5614.998.1721.099.1318.7719.5228.3425.3242.6331.5821.24
• MonoTTA (Ours)26.1333.1128.6030.3825.4832.4418.7232.6037.7537.8739.5743.6737.9832.64
", + "image_path": "63d80a18755b299d0077d6e63c29c38fd292bb69547e45ed5e25d1416a9e0bdb.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 148, + 228, + 466, + 312 + ], + "blocks": [ + { + "bbox": [ + 148, + 228, + 466, + 312 + ], + "lines": [ + { + "bbox": [ + 148, + 228, + 466, + 312 + ], + "spans": [ + { + "bbox": [ + 148, + 228, + 466, + 312 + ], + "type": "table", + "html": "
Pedestrian, IoU @ 0.5, 0.25
MethodNoiseBlurWeatherDigitalAvg.
Gauss.ShotImpul.Defoc.GlassMotionSnowFrostFogBrit.Contr.PixelSat.
Monoflex0.191.620.323.728.476.220.004.272.259.192.081.839.113.79
• BN adaptation [33]6.218.209.207.835.357.522.896.479.249.129.9312.739.768.03
• TENT [36]6.027.969.577.756.068.632.636.719.9110.2610.5512.3310.278.36
• EATA [26]6.057.969.747.936.068.012.486.249.949.7010.0212.4110.128.21
• ActMAD [24]2.414.274.681.461.626.481.174.254.846.426.198.297.074.55
• MonotTA (Ours)6.548.419.397.637.128.993.257.6410.2610.5510.0613.2810.668.75
MonoGround0.610.930.738.417.578.030.003.191.3913.821.833.707.264.42
• BN adaptation [33]5.097.087.777.636.639.452.246.728.409.7211.0616.7012.438.53
• TENT [36]7.2710.1010.028.538.3011.033.548.739.1111.7412.1217.7015.0310.25
• EATA [26]5.928.058.588.127.5910.953.317.9110.1011.0210.9517.3014.349.55
• ActMAD [24]3.965.756.485.615.737.141.585.375.747.278.1013.129.276.55
• MonotTA (Ours)8.5811.1811.799.229.4013.204.839.9514.4612.8513.2517.1314.8511.59
", + "image_path": "4e086364b9fcd8756e6d9347724bdf5e142b88d84507bf9576eb6564ab9951ad.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 148, + 312, + 466, + 395 + ], + "blocks": [ + { + "bbox": [ + 148, + 312, + 466, + 395 + ], + "lines": [ + { + "bbox": [ + 148, + 312, + 466, + 395 + ], + "spans": [ + { + "bbox": [ + 148, + 312, + 466, + 395 + ], + "type": "table", + "html": "
Cyclist, IoU @ 0.5, 0.25, 0.25
MethodNoiseBlurWeatherDigitalAvg.
Gauss.ShotImpul.Defoc.GlassMotionSnowFrostFogBrit.Contr.PixelSat.
Monoflex0.281.640.470.594.973.600.007.423.8113.073.793.808.393.99
• BN adaptation [33]2.396.264.365.786.769.091.708.539.1612.9111.2610.5511.027.67
• TENT [36]2.727.945.636.277.209.491.078.9410.9612.7512.729.6411.288.20
• EATA [26]2.337.465.467.197.237.511.248.6010.1812.8611.0210.8910.307.87
• ActMAD [24]2.225.324.362.291.575.640.485.686.4610.569.319.977.805.51
• MonotTA (Ours)3.017.245.987.006.099.511.4510.6311.2212.8511.8511.5910.818.40
MonoGround0.120.460.480.330.722.030.000.560.355.550.522.085.241.42
• BN adaptation [33]1.763.582.083.613.055.410.573.824.476.305.6011.028.874.63
• TENT [36]1.794.853.003.363.496.050.494.436.207.196.5010.439.235.15
• EATA [26]1.894.202.414.133.175.730.384.066.276.416.1410.937.724.88
• ActMAD [24]1.142.841.533.33.083.150.322.743.476.365.759.696.953.87
• MonotTA (Ours)3.935.784.555.434.706.090.694.667.537.697.7411.719.436.15
", + "image_path": "44412cd6db4e954e0783727b5a869eb8bf6953821f7a71c2f6828db9bd1d6114.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 411, + 482, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 411, + 482, + 495 + ], + "spans": [ + { + "bbox": [ + 130, + 411, + 482, + 495 + ], + "type": "text", + "content": "in certain extreme scenarios. For instance, the Snow scenario (c.f. Fig. 3 (a)) lacks objects with scores exceeding 0.4 when the pre-trained model is directly applied. Under such a circumstance, " + }, + { + "bbox": [ + 130, + 411, + 482, + 495 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{Nreg}" + }, + { + "bbox": [ + 130, + 411, + 482, + 495 + ], + "type": "text", + "content": " plays a more important role in model adaptation since it can alleviate distribution shifts even only low-score objects, i.e., deny the negative category. In other words, " + }, + { + "bbox": [ + 130, + 411, + 482, + 495 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{Nreg}" + }, + { + "bbox": [ + 130, + 411, + 482, + 495 + ], + "type": "text", + "content": " enables the model to alleviate distribution shifts and achieve more relatively high-score objects, thereby laying a crucial foundation for " + }, + { + "bbox": [ + 130, + 411, + 482, + 495 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{AO}" + }, + { + "bbox": [ + 130, + 411, + 482, + 495 + ], + "type": "text", + "content": " in extremely challenging scenarios." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 517, + 230, + 531 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 517, + 230, + 531 + ], + "spans": [ + { + "bbox": [ + 132, + 517, + 230, + 531 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 544, + 481, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 544, + 481, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 544, + 481, + 593 + ], + "type": "text", + "content": "We conduct experiments based on KITTI [9] and nuScenes [1]. The results presented in this manuscript represent the average value across three difficulty levels, i.e., Easy, Moderate, Hard. Note that we provide more results of higher severity levels of KITTI and more detailed results in Appendix " + }, + { + "bbox": [ + 130, + 544, + 481, + 593 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 130, + 544, + 481, + 593 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "Datasets. For KITTI, we adopt the protocol established by Monoflex [48] to split the training images into the training set (3712) and validation set (3769) to perform model training and adaptation, respectively. Following the official KITTI evaluation criteria, we evaluate detection results on three levels of difficulty. Then, we construct the KITTI-C dataset through the incorporation of 13 distinct types of data corruptions [11] to the validation set and each corruption" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 315, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 315, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 315, + 102 + ], + "type": "text", + "content": "Hongbin Lin, Yifan Zhang and et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 167, + 124, + 444, + 224 + ], + "blocks": [ + { + "bbox": [ + 202, + 114, + 414, + 122 + ], + "lines": [ + { + "bbox": [ + 202, + 114, + 414, + 122 + ], + "spans": [ + { + "bbox": [ + 202, + 114, + 414, + 122 + ], + "type": "text", + "content": "Average performance of 13 corruptions on KITTI-C Validation set, severity level 2" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 167, + 124, + 444, + 224 + ], + "lines": [ + { + "bbox": [ + 167, + 124, + 444, + 224 + ], + "spans": [ + { + "bbox": [ + 167, + 124, + 444, + 224 + ], + "type": "image", + "image_path": "1016ad6fb21f14f8f303f665c45098a31a4d72442d89e7b833a4fc5e03d381d3.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 235, + 479, + 257 + ], + "lines": [ + { + "bbox": [ + 131, + 235, + 479, + 257 + ], + "spans": [ + { + "bbox": [ + 131, + 235, + 479, + 257 + ], + "type": "text", + "content": "Fig. 4: We visualize the comparison with baselines on the KITTI-C validation set, severity level 2 regarding Mean " + }, + { + "bbox": [ + 131, + 235, + 479, + 257 + ], + "type": "inline_equation", + "content": "AP_{3D|R_{40}}" + }, + { + "bbox": [ + 131, + 235, + 479, + 257 + ], + "type": "text", + "content": ". The bold number indicates the best result." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 275, + 479, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 275, + 479, + 300 + ], + "spans": [ + { + "bbox": [ + 131, + 275, + 479, + 300 + ], + "type": "text", + "content": "has 5 severity levels. The model is trained on the original training set and then tested on the KITTI-C validation set within one of the corruptions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 307, + 480, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 307, + 480, + 390 + ], + "spans": [ + { + "bbox": [ + 130, + 307, + 480, + 390 + ], + "type": "text", + "content": "As for nuScenes, we adopt the front-view images and construct the Daytime and Night scenarios via their scene descriptions following [21]. Specifically, there exist " + }, + { + "bbox": [ + 130, + 307, + 480, + 390 + ], + "type": "inline_equation", + "content": "24.7\\mathrm{k} / 5.4\\mathrm{k}" + }, + { + "bbox": [ + 130, + 307, + 480, + 390 + ], + "type": "text", + "content": " train/test images in Daytime while " + }, + { + "bbox": [ + 130, + 307, + 480, + 390 + ], + "type": "inline_equation", + "content": "3.3\\mathrm{k} / 0.6\\mathrm{k}" + }, + { + "bbox": [ + 130, + 307, + 480, + 390 + ], + "type": "text", + "content": " train/test images in Night. Based on these splits, we construct two real-world adaptation tasks, i.e., Daytime " + }, + { + "bbox": [ + 130, + 307, + 480, + 390 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 130, + 307, + 480, + 390 + ], + "type": "text", + "content": " Night and Night " + }, + { + "bbox": [ + 130, + 307, + 480, + 390 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 130, + 307, + 480, + 390 + ], + "type": "text", + "content": " Daytime. For simplification, we transfer the nuScenes dataset into the KITTI format and only consider the Car category. More details of data construction are provided in Appendix " + }, + { + "bbox": [ + 130, + 307, + 480, + 390 + ], + "type": "inline_equation", + "content": "\\pmb{B}" + }, + { + "bbox": [ + 130, + 307, + 480, + 390 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 398, + 480, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 398, + 480, + 494 + ], + "spans": [ + { + "bbox": [ + 130, + 398, + 480, + 494 + ], + "type": "text", + "content": "Implementation details. We implement our method and other baselines in PyTorch [28]. In MonoTTA, we conduct model adaptation based on the public pre-trained weights and the parameter settings provided by their authors [29, 40, 48]. Besides, we employ the Stochastic Gradient Descent (SGD) optimizer with a half learning rate of the initial rate used in base training over different methods, a momentum of 0.9 and a batch size of 16 for KITTI, 4 for nuScenes. Parameters " + }, + { + "bbox": [ + 130, + 398, + 480, + 494 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 398, + 480, + 494 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 398, + 480, + 494 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 130, + 398, + 480, + 494 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 398, + 480, + 494 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 130, + 398, + 480, + 494 + ], + "type": "text", + "content": " are assigned default values of 1, 0.1, and 0.05, respectively. More training details of MonoTTA are provided in Appendix " + }, + { + "bbox": [ + 130, + 398, + 480, + 494 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 398, + 480, + 494 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 502, + 480, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 502, + 480, + 597 + ], + "spans": [ + { + "bbox": [ + 131, + 502, + 480, + 597 + ], + "type": "text", + "content": "Compared methods. Based on three typical or state-of-the-art (SOTA) Mono 3Det methods [29, 40, 48], we fully compare MonoTTA with following methods: 1) source-only, i.e., directly apply the pre-trained model to the test data within corruptions; 2) BN adaptation [33] updates batch normalization statistics via target data; 3) TENT [36] minimizes the entropy loss of test data; 4) EATA [26] identifies reliable samples to update the model by entropy loss minimization. Here we compare MonoTTA with its variant Efficient Test-time Adaptation. 5) ActMAD [24] matches activation statistics for the distribution alignment." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 605, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 605, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 605, + 480, + 665 + ], + "type": "text", + "content": "Evaluation protocols. In order to fully evaluate the proposed method, we report our experimental results in the Average Precision (AP) for 3D bounding boxes, denoted as " + }, + { + "bbox": [ + 131, + 605, + 480, + 665 + ], + "type": "inline_equation", + "content": "AP_{3D|R_{40}}" + }, + { + "bbox": [ + 131, + 605, + 480, + 665 + ], + "type": "text", + "content": ". The results present the mean values across three levels of difficulty and the Intersection over Union (IoU) thresholds are set to 0.7, 0.5, 0.5 for Cars and 0.5, 0.25, 0.25 for Pedestrians and Cyclists." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "type": "text", + "content": "MonoTTA" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 166, + 125, + 446, + 232 + ], + "blocks": [ + { + "bbox": [ + 204, + 114, + 411, + 123 + ], + "lines": [ + { + "bbox": [ + 204, + 114, + 411, + 123 + ], + "spans": [ + { + "bbox": [ + 204, + 114, + 411, + 123 + ], + "type": "text", + "content": "Performance on Daytime to Night and Night to Daytime scenarios of nuScenes" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 166, + 125, + 446, + 232 + ], + "lines": [ + { + "bbox": [ + 166, + 125, + 446, + 232 + ], + "spans": [ + { + "bbox": [ + 166, + 125, + 446, + 232 + ], + "type": "image", + "image_path": "21e234440e204e27ec0c18cc793bbd57c303929d334ce98b0f5e2193c398f2d8.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 241, + 482, + 266 + ], + "lines": [ + { + "bbox": [ + 131, + 241, + 482, + 266 + ], + "spans": [ + { + "bbox": [ + 131, + 241, + 482, + 266 + ], + "type": "text", + "content": "Fig. 5: Comparison with baselines on Daytime " + }, + { + "bbox": [ + 131, + 241, + 482, + 266 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 131, + 241, + 482, + 266 + ], + "type": "text", + "content": " Night and Night " + }, + { + "bbox": [ + 131, + 241, + 482, + 266 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 131, + 241, + 482, + 266 + ], + "type": "text", + "content": " Daytime of nuScenes, regarding Mean " + }, + { + "bbox": [ + 131, + 241, + 482, + 266 + ], + "type": "inline_equation", + "content": "AP_{3D|R_{40}}" + }, + { + "bbox": [ + 131, + 241, + 482, + 266 + ], + "type": "text", + "content": ". The **bold** number indicates the best result." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 281, + 348, + 294 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 281, + 348, + 294 + ], + "spans": [ + { + "bbox": [ + 131, + 281, + 348, + 294 + ], + "type": "text", + "content": "4.1 Comparisons with Previous Methods" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 301, + 482, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 301, + 482, + 459 + ], + "spans": [ + { + "bbox": [ + 130, + 301, + 482, + 459 + ], + "type": "text", + "content": "We first compare our MonoTTA with previous methods in severity level 1 of KITTI-C. The results are reported in Table 1, which gives the following observations: 1) Due to distribution shifts, directly applying the pre-trained model to the test data (i.e., source-only) suffers severe performance degradation in all categories. 2) Existing TTA methods are able to mitigate the negative effect of distribution shifts for Mono 3Det to some degree. However, they only achieve suboptimal performance since they tend to increase the scores of all positive detections, containing severe noise. 3) MonoTTA consistently outperforms all compared methods over all categories within various base models in terms of mean " + }, + { + "bbox": [ + 130, + 301, + 482, + 459 + ], + "type": "inline_equation", + "content": "AP_{3D|R_{40}}" + }, + { + "bbox": [ + 130, + 301, + 482, + 459 + ], + "type": "text", + "content": ". Specifically, MonoTTA achieves the best or comparable performance in all categories under all corruptions, attaining a large performance gain over TENT and EATA (e.g., improving an average " + }, + { + "bbox": [ + 130, + 301, + 482, + 459 + ], + "type": "inline_equation", + "content": "AP_{3D|R_{40}}" + }, + { + "bbox": [ + 130, + 301, + 482, + 459 + ], + "type": "text", + "content": " about 5.1 and 7.4 of the Car category based on MonoGround)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 476, + 379, + 489 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 476, + 379, + 489 + ], + "spans": [ + { + "bbox": [ + 131, + 476, + 379, + 489 + ], + "type": "text", + "content": "4.2 More Severe Corruption and Real Scenario" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 497, + 482, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 497, + 482, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 497, + 482, + 616 + ], + "type": "text", + "content": "On the one hand, to fully validate the effectiveness of our MonoTTA, we visualize experimental results under more severe corruption conditions at severity level 2 as shown in Fig. 4, which clearly gives additional observations: 1) With the escalation of severity level, the pre-trained models suffer a larger performance decline within various corruptions, enlarging the difficulty of TTA. 2) The performance improvements of existing TTA methods become relatively limited, particularly in Pedestrian and Cyclist classes. 3) Even if the tasks are more challenging, MonoTTA still stably obtains the best average performance within all corruptions since " + }, + { + "bbox": [ + 130, + 497, + 482, + 616 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{Nreg}" + }, + { + "bbox": [ + 130, + 497, + 482, + 616 + ], + "type": "text", + "content": " plays an important role in alleviating distribution shifts for certain extreme scenarios, i.e., only low-score objects exist." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 617, + 482, + 666 + ], + "type": "text", + "content": "On the other hand, we further validate different methods within real scenarios as shown in Fig. 5. The experimental results also give the following observations: 1) Under real corruptions, the pre-trained model still suffers severe performance degradation due to the data distribution shifts. 2) TENT tends to increase the" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 315, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 315, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 315, + 102 + ], + "type": "text", + "content": "Hongbin Lin, Yifan Zhang and et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 202, + 157, + 411, + 198 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 2: Comparison with baselines based on the instance-level method (i.e., batch size is 1) on the KITTI-C validation set, regarding Mean " + }, + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "type": "inline_equation", + "content": "AP_{3D|R_{40}}" + }, + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "type": "text", + "content": " and the severity levels 1 and 2. The bold number indicates the best result." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 202, + 157, + 411, + 198 + ], + "lines": [ + { + "bbox": [ + 202, + 157, + 411, + 198 + ], + "spans": [ + { + "bbox": [ + 202, + 157, + 411, + 198 + ], + "type": "table", + "html": "
MethodLevel 1Level 2
CarPedes.CyclistAvg.CarPedes.CyclistAvg.
MonoNeRD19.845.962.279.3613.023.831.616.15
BN adaptation [33]30.738.853.8114.4626.476.942.9112.11
TENT [36]35.729.994.7516.8231.857.813.4014.35
EATA [26]34.6010.044.2016.2830.667.863.2813.93
MonoTTA (Ours)37.4010.394.3517.3833.998.253.3315.19
", + "image_path": "bd5c73fe7e83b74d2fa3050e86c7f7fed709f5718e2f98020d5b4cc7b4c25e3a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 201, + 239, + 411, + 276 + ], + "blocks": [ + { + "bbox": [ + 130, + 208, + 479, + 232 + ], + "lines": [ + { + "bbox": [ + 130, + 208, + 479, + 232 + ], + "spans": [ + { + "bbox": [ + 130, + 208, + 479, + 232 + ], + "type": "text", + "content": "Table 3: Based on MonoGround [29], we conduct ablation studies of " + }, + { + "bbox": [ + 130, + 208, + 479, + 232 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{AO}" + }, + { + "bbox": [ + 130, + 208, + 479, + 232 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 208, + 479, + 232 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{Nreg}" + }, + { + "bbox": [ + 130, + 208, + 479, + 232 + ], + "type": "text", + "content": " on the KITTI-C validation set, regarding " + }, + { + "bbox": [ + 130, + 208, + 479, + 232 + ], + "type": "inline_equation", + "content": "AP_{3D|R_{40}}" + }, + { + "bbox": [ + 130, + 208, + 479, + 232 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 201, + 239, + 411, + 276 + ], + "lines": [ + { + "bbox": [ + 201, + 239, + 411, + 276 + ], + "spans": [ + { + "bbox": [ + 201, + 239, + 411, + 276 + ], + "type": "table", + "html": "
Backbone\\( {\\mathcal{L}}_{AO} \\)\\( {\\mathcal{L}}_{Nreg} \\)Level 1Level 2
CarPedes.CyclistAvg.CarPedes.CyclistAvg.
✓9.674.421.425.175.722.430.953.03
✓✓27.9410.175.1014.4021.277.793.8210.96
✓✓23.268.604.6012.1518.936.813.559.77
✓✓✓32.6411.596.1516.7925.968.784.1212.85
", + "image_path": "58fae23c2e680d7613aa885f32347b57be1fe215d277b1b17e8cd08a98225fa2.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 290, + 482, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 290, + 482, + 352 + ], + "spans": [ + { + "bbox": [ + 130, + 290, + 482, + 352 + ], + "type": "text", + "content": "confidence of all positive detections and thus overfits to noise, i.e., failing to handle the extremely challenging task " + }, + { + "bbox": [ + 130, + 290, + 482, + 352 + ], + "type": "inline_equation", + "content": "\\mathrm{N} \\rightarrow \\mathrm{D}" + }, + { + "bbox": [ + 130, + 290, + 482, + 352 + ], + "type": "text", + "content": ". 3) EATA still achieves sub-optimal performance while our MonoTTA brings sufficient average performance improvement on both MonoFlex (6.23 mAP) and MonoGround (8.26 mAP), maintaining the best performance in real scenarios. Detailed results are in Appendix " + }, + { + "bbox": [ + 130, + 290, + 482, + 352 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 130, + 290, + 482, + 352 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 369, + 403, + 380 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 369, + 403, + 380 + ], + "spans": [ + { + "bbox": [ + 131, + 369, + 403, + 380 + ], + "type": "text", + "content": "4.3 Application to Instance-Level Inference Method" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 388, + 482, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 388, + 482, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 388, + 482, + 533 + ], + "type": "text", + "content": "In this section, we seek to investigate whether the proposed MonoTTA can be used to effectively enhance Mono 3Det methods which only process a single image sequentially, i.e., " + }, + { + "bbox": [ + 130, + 388, + 482, + 533 + ], + "type": "inline_equation", + "content": "B = 1" + }, + { + "bbox": [ + 130, + 388, + 482, + 533 + ], + "type": "text", + "content": ". To be specific, it may be crucial for Mono 3Det methods to make immediate decisions based on the most recent scene (image) in real-world scenarios like autonomous driving. To this end, it is essential for TTA methods devised for Mono 3Det to allow a single image as input and then conduct model adaptation. To validate it, we integrate MonoTTA into the SOTA Mono 3Det method namely MonoNeRD [40] which accepts one image each time at the test phase. As shown in Table 2, MonoTTA achieves the best or comparable performance across all categories at both severity levels 1 and 2, illustrating the applicability of our method to boost these approaches for handling OOD test data. Detailed results can be also found in Appendix " + }, + { + "bbox": [ + 130, + 388, + 482, + 533 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 130, + 388, + 482, + 533 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 549, + 375, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 549, + 375, + 562 + ], + "spans": [ + { + "bbox": [ + 131, + 549, + 375, + 562 + ], + "type": "text", + "content": "4.4 Ablation Studies and Quantitative Results" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "content": "Ablation studies. To examine the effectiveness of the losses in MonoTTA, we show the results of the models optimized by different losses. As shown in Table 3, introducing " + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{AO}" + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{Nreg}" + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "content": " enhances the model performance compared to directly applying the pre-trained model (i.e., source only). On the one hand, such a result verifies that our strategy is able to alleviate the score decline issue in test OOD data with high-score objects. On the other hand, introducing " + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{Nreg}" + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "content": " could obtain a higher average " + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "inline_equation", + "content": "AP_{3D|R_{40}}" + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "content": " value, which also verifies that the negative regularization term is able to enhance the pre-trained model even only" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 403, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 403, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 403, + 91, + 447, + 101 + ], + "type": "text", + "content": "MonoTTA" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 167, + 114, + 447, + 285 + ], + "blocks": [ + { + "bbox": [ + 167, + 114, + 447, + 285 + ], + "lines": [ + { + "bbox": [ + 167, + 114, + 447, + 285 + ], + "spans": [ + { + "bbox": [ + 167, + 114, + 447, + 285 + ], + "type": "image", + "image_path": "7dee429f5322d7f05be1b976335d70ae3b10fcc09f1d687224b5877ee2f9e237.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 294, + 482, + 328 + ], + "lines": [ + { + "bbox": [ + 130, + 294, + 482, + 328 + ], + "spans": [ + { + "bbox": [ + 130, + 294, + 482, + 328 + ], + "type": "text", + "content": "Fig.6: Qualitative results of baselines and the proposed MonoTTA based on MonoGround [29]. We visualize the results on KITTI-C validation set, where predicted cars, pedestrians and cyclists are in lime green, sky blue and yellow, respectively." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 338, + 479, + 362 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 338, + 479, + 362 + ], + "spans": [ + { + "bbox": [ + 130, + 338, + 479, + 362 + ], + "type": "text", + "content": "with low-score objects. When combining the losses (i.e. " + }, + { + "bbox": [ + 130, + 338, + 479, + 362 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{AO}" + }, + { + "bbox": [ + 130, + 338, + 479, + 362 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 338, + 479, + 362 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{Nreg}" + }, + { + "bbox": [ + 130, + 338, + 479, + 362 + ], + "type": "text", + "content": ") together, we obtain the best performance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 363, + 480, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 363, + 480, + 434 + ], + "spans": [ + { + "bbox": [ + 130, + 363, + 480, + 434 + ], + "type": "text", + "content": "Quantitative Results. We provide visualizations within Snow, Fog and Frost based on Monoground as shown in Fig. 6. It is evident that the source-only setting suffers severe omissions, while BN adaptation, TENT and EATA alleviate distribution shifts to some degree and give more detections. As for our MonoTTA, it can produce superior detections even in severe conditions, including fewer omissions and accurate detections as highlighted by red circles." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 451, + 220, + 463 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 451, + 220, + 463 + ], + "spans": [ + { + "bbox": [ + 132, + 451, + 220, + 463 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 474, + 482, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 482, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 482, + 593 + ], + "type": "text", + "content": "In this paper, we propose a monocular test-time adaptation method to improve the pre-trained model on the shifted test data for monocular 3D object detection. Specifically, our method consists of two strategies: 1) Reliability-driven adaptation. To discover more potential objects, we devise a self-adaptive strategy to identify reliable objects for adaptive model adaptation. 2) Noise-guard adaptation. To avoid overfitting to noise and trivial solutions, we devise the negative regularization term to mitigate the negative effects of noisy detections and alleviate distribution shifts. Experiments on KITTI-C and nuScenes datasets demonstrate the effectiveness of MonoTTA in handling fully test-time adaptation for monocular 3D object detection." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "content": "Future directions. 1) Our work focuses on 2D images, while future studies could explore 3D information in handling distribution shifts. 2) This work explores TTA by assuming one OOD distribution at a time, where the forgetting issue is not severe as the source model weights are recoverable. Exploring scenarios with dynamically OOD distributions offers a compelling future direction, where the forgetting issue would become more severe." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 315, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 315, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 315, + 102 + ], + "type": "text", + "content": "Hongbin Lin, Yifan Zhang and et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 483, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 483, + 283 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 483, + 283 + ], + "type": "text", + "content": "This work was supported by NSFC with Grant No. 62293482, by the Basic Research Project No. HZQB-KCZYZ-2021067 of Hetao Shenzhen HK S&T Cooperation Zone, by Shenzhen General Program No. JCYJ20220530143600001, by Shenzhen-Hong Kong Joint Funding No. SGDX20211123112401002, by the Shenzhen Outstanding Talents Training Fund 202002, by Guangdong Research Project No. 2017ZT07X152 and No. 2019CX01X104, by the Guangdong Provincial Key Laboratory of Future Networks of Intelligence (Grant No. 2022B1212010001), by the Guangdong Provincial Key Laboratory of Big Data Computing, CHUK-Shenzhen, by the NSFC 61931024&12326610, by the Key Area R&D Program of Guangdong Province with grant No. 2018B030338001, by the Shenzhen Key Laboratory of Big Data and Artificial Intelligence (Grant No. ZDSYS201707251409055), and by Tencent & Huawei Open Fund." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 301, + 197, + 314 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 301, + 197, + 314 + ], + "spans": [ + { + "bbox": [ + 133, + 301, + 197, + 314 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 325, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 138, + 325, + 481, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 325, + 481, + 370 + ], + "spans": [ + { + "bbox": [ + 138, + 325, + 481, + 370 + ], + "type": "text", + "content": "1. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuscenes: A multimodal dataset for autonomous driving. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11621-11631 (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 370, + 481, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 370, + 481, + 403 + ], + "spans": [ + { + "bbox": [ + 138, + 370, + 481, + 403 + ], + "type": "text", + "content": "2. Chen, X., Kundu, K., Zhang, Z., Ma, H., Fidler, S., Urtasun, R.: Monocular 3d object detection for autonomous driving. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2147-2156 (2016)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 403, + 481, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 403, + 481, + 435 + ], + "spans": [ + { + "bbox": [ + 138, + 403, + 481, + 435 + ], + "type": "text", + "content": "3. Chen, X., Kundu, K., Zhu, Y., Berneshawi, A.G., Ma, H., Fidler, S., Urtasun, R.: 3d object proposals for accurate object class detection. Advances in neural information processing systems 28 (2015)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 435, + 481, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 435, + 481, + 468 + ], + "spans": [ + { + "bbox": [ + 138, + 435, + 481, + 468 + ], + "type": "text", + "content": "4. Chen, Y., Tai, L., Sun, K., Li, M.: Monopair: Monocular 3d object detection using pairwise spatial relationships. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12093-12102 (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 468, + 481, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 468, + 481, + 501 + ], + "spans": [ + { + "bbox": [ + 138, + 468, + 481, + 501 + ], + "type": "text", + "content": "5. Chen, Y., Liu, J., Zhang, X., Qi, X., Jia, J.: Voxelnext: Fully sparse voxelnet for 3d object detection and tracking. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21674-21683 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 501, + 481, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 501, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 138, + 501, + 481, + 544 + ], + "type": "text", + "content": "6. Ding, M., Huo, Y., Yi, H., Wang, Z., Shi, J., Lu, Z., Luo, P.: Learning depth-guided convolutions for monocular 3d object detection. In: Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition workshops. pp. 1000-1001 (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 545, + 481, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 545, + 481, + 578 + ], + "spans": [ + { + "bbox": [ + 138, + 545, + 481, + 578 + ], + "type": "text", + "content": "7. Duan, K., Bai, S., Xie, L., Qi, H., Huang, Q., Tian, Q.: Centernet: Keypoint triplets for object detection. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6569-6578 (2019)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 578, + 481, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 578, + 481, + 611 + ], + "spans": [ + { + "bbox": [ + 138, + 578, + 481, + 611 + ], + "type": "text", + "content": "8. Fleuret, F., et al.: Test time adaptation through perturbation robustness. In: NeurIPS 2021 Workshop on Distribution Shifts: Connecting Methods and Applications (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 611, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 611, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 138, + 611, + 481, + 643 + ], + "type": "text", + "content": "9. Geiger, A., Lenz, P., Urtasun, R.: Are we ready for autonomous driving? the kitti vision benchmark suite. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3354-3361. IEEE (2012)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 643, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 643, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 643, + 481, + 665 + ], + "type": "text", + "content": "0. Hegde, D., Kilic, V., Sindagi, V., Cooper, A.B., Foster, M., Patel, V.M.: Source-free unsupervised domain adaptation for 3d object detection in adverse weather." + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "type": "text", + "content": "MonoTTA" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 147, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 147, + 116, + 480, + 138 + ], + "type": "text", + "content": "In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 6973-6980. IEEE (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 138, + 480, + 170 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 138, + 480, + 170 + ], + "spans": [ + { + "bbox": [ + 133, + 138, + 480, + 170 + ], + "type": "text", + "content": "11. Hendrycks, D., Dietterich, T.: Benchmarking neural network robustness to common corruptions and perturbations. In: International Conference on Learning Representations (2018)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 171, + 480, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 171, + 480, + 203 + ], + "spans": [ + { + "bbox": [ + 133, + 171, + 480, + 203 + ], + "type": "text", + "content": "12. Kim, J., Hwang, I., Kim, Y.M.: Ev-tta: Test-time adaptation for event-based object recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17745-17754 (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 203, + 480, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 203, + 480, + 235 + ], + "spans": [ + { + "bbox": [ + 133, + 203, + 480, + 235 + ], + "type": "text", + "content": "13. Kim, Y., Yim, J., Yun, J., Kim, J.: Nlnl: Negative learning for noisy labels. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 101-110 (2019)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 235, + 480, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 235, + 480, + 267 + ], + "spans": [ + { + "bbox": [ + 133, + 235, + 480, + 267 + ], + "type": "text", + "content": "14. Kim, Y., Yun, J., Shon, H., Kim, J.: Joint negative and positive learning for noisy labels. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9442-9451 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 267, + 480, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 267, + 480, + 300 + ], + "spans": [ + { + "bbox": [ + 133, + 267, + 480, + 300 + ], + "type": "text", + "content": "15. Kumar, A., Brazil, G., Liu, X.: Groomed-nms: Grouped mathematically differentiable nms for monocular 3d object detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8973-8983 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 300, + 480, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 300, + 480, + 331 + ], + "spans": [ + { + "bbox": [ + 133, + 300, + 480, + 331 + ], + "type": "text", + "content": "16. Li, P., Chen, X., Shen, S.: Stereo r-cnn based 3d object detection for autonomous driving. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7644-7652 (2019)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 133, + 332, + 480, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 332, + 480, + 364 + ], + "spans": [ + { + "bbox": [ + 133, + 332, + 480, + 364 + ], + "type": "text", + "content": "17. Liang, J., Hu, D., Feng, J.: Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation. In: International conference on machine learning. pp. 6028-6039. PMLR (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 133, + 364, + 480, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 364, + 480, + 396 + ], + "spans": [ + { + "bbox": [ + 133, + 364, + 480, + 396 + ], + "type": "text", + "content": "18. Lin, H., Zhang, Y., Qiu, Z., Niu, S., Gan, C., Liu, Y., Tan, M.: Prototype-guided continual adaptation for class-incremental unsupervised domain adaptation. In: European Conference on Computer Vision. pp. 351-368 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 133, + 396, + 480, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 396, + 480, + 429 + ], + "spans": [ + { + "bbox": [ + 133, + 396, + 480, + 429 + ], + "type": "text", + "content": "19. Liu, Y., Kothari, P., Van Delft, B., Bellot-Gurlet, B., Mordan, T., Alahi, A.: Ttt++: When does self-supervised test-time training fail or thrive? In: Advances in Neural Information Processing Systems. vol. 34, pp. 21808-21820 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 133, + 429, + 480, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 429, + 480, + 460 + ], + "spans": [ + { + "bbox": [ + 133, + 429, + 480, + 460 + ], + "type": "text", + "content": "20. Liu, Z., Wu, Z., Tóth, R.: Smoke: Single-stage monocular 3d object detection via keypoint estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. pp. 996-997 (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 133, + 460, + 480, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 460, + 480, + 503 + ], + "spans": [ + { + "bbox": [ + 133, + 460, + 480, + 503 + ], + "type": "text", + "content": "21. Liu, Z., Tang, H., Amini, A., Yang, X., Mao, H., Rus, D.L., Han, S.: Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. In: 2023 IEEE international conference on robotics and automation (ICRA). pp. 2774-2781. IEEE (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 133, + 503, + 480, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 503, + 480, + 536 + ], + "spans": [ + { + "bbox": [ + 133, + 503, + 480, + 536 + ], + "type": "text", + "content": "22. Luo, Y., Zheng, C., Yan, X., Kun, T., Zheng, C., Cui, S., Li, Z.: Latr: 3d lane detection from monocular images with transformer. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7941-7952 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 536, + 480, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 536, + 480, + 557 + ], + "spans": [ + { + "bbox": [ + 133, + 536, + 480, + 557 + ], + "type": "text", + "content": "23. Ma, X., Liu, S., Xia, Z., Zhang, H., Zeng, X., Ouyang, W.: Rethinking pseudo-lidar representation. In: European Conference on Computer Vision (2020)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 133, + 557, + 480, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 557, + 480, + 601 + ], + "spans": [ + { + "bbox": [ + 133, + 557, + 480, + 601 + ], + "type": "text", + "content": "24. Mirza, M.J., Soneira, P.J., Lin, W., Kozinski, M., Possegger, H., Bischof, H.: Actmad: Activation matching to align distributions for test-time-training. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 24152-24161 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 133, + 601, + 480, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 601, + 480, + 633 + ], + "spans": [ + { + "bbox": [ + 133, + 601, + 480, + 633 + ], + "type": "text", + "content": "25. Nado, Z., Padhy, S., Sculley, D., D'Amour, A., Lakshminarayanan, B., Snoek, J.: Evaluating prediction-time batch normalization for robustness under covariate shift. arXiv preprint arXiv:2006.10963 (2020)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 133, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 633, + 480, + 665 + ], + "type": "text", + "content": "26. Niu, S., Wu, J., Zhang, Y., Chen, Y., Zheng, S., Zhao, P., Tan, M.: Efficient test-time model adaptation without forgetting. In: The International Conference on Machine Learning (2022)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 315, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 315, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 315, + 102 + ], + "type": "text", + "content": "Hongbin Lin, Yifan Zhang and et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "27. Niu, S., Wu, J., Zhang, Y., Wen, Z., Chen, Y., Zhao, P., Tan, M.: Towards stable test-time adaptation in dynamic wild world. In: International Conference on Learning Representations (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 149, + 482, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 149, + 482, + 193 + ], + "spans": [ + { + "bbox": [ + 130, + 149, + 482, + 193 + ], + "type": "text", + "content": "28. Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., et al.: Pytorch: An imperative style, high-performance deep learning library. In: Advances in neural information processing systems. vol. 32 (2019)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 193, + 482, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 193, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 193, + 482, + 224 + ], + "type": "text", + "content": "29. Qin, Z., Li, X.: Monoground: Detecting monocular 3d objects from the ground. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3793-3802 (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 224, + 482, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 224, + 482, + 257 + ], + "spans": [ + { + "bbox": [ + 130, + 224, + 482, + 257 + ], + "type": "text", + "content": "30. Qiu, Z., Zhang, Y., Lin, H., et al.: Source-free domain adaptation via avatar prototype generation and adaptation. In: International Joint Conference on Artificial Intelligence (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 257, + 482, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 257, + 482, + 289 + ], + "spans": [ + { + "bbox": [ + 130, + 257, + 482, + 289 + ], + "type": "text", + "content": "31. Reading, C., Harakeh, A., Chae, J., Waslander, S.L.: Categorical depth distribution network for monocular 3d object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8555-8564 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 289, + 482, + 321 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 289, + 482, + 321 + ], + "spans": [ + { + "bbox": [ + 130, + 289, + 482, + 321 + ], + "type": "text", + "content": "32. Saltori, C., Lathuilière, S., Sebe, N., Ricci, E., Galasso, F.: Sf-uda 3d: Source-free unsupervised domain adaptation for lidar-based 3d object detection. In: 2020 International Conference on 3D Vision (3DV). pp. 771–780. IEEE (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 321, + 482, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 321, + 482, + 354 + ], + "spans": [ + { + "bbox": [ + 130, + 321, + 482, + 354 + ], + "type": "text", + "content": "33. Schneider, S., Rusak, E., Eck, L., Bringmann, O., Brendel, W., Bethge, M.: Improving robustness against common corruptions by covariate shift adaptation. In: Advances in neural information processing systems. pp. 11539-11551 (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 354, + 482, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 354, + 482, + 386 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 482, + 386 + ], + "type": "text", + "content": "34. Sun, Y., Wang, X., Liu, Z., Miller, J., Efros, A., Hardt, M.: Test-time training with self-supervision for generalization under distribution shifts. In: International conference on machine learning. pp. 9229-9248. PMLR (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 386, + 482, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 386, + 482, + 418 + ], + "spans": [ + { + "bbox": [ + 130, + 386, + 482, + 418 + ], + "type": "text", + "content": "35. Veksler, O.: Test time adaptation with regularized loss for weakly supervised salient object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7360-7369 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 418, + 482, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 418, + 482, + 450 + ], + "spans": [ + { + "bbox": [ + 130, + 418, + 482, + 450 + ], + "type": "text", + "content": "36. Wang, D., Shelhamer, E., Liu, S., Olshausen, B., Darrell, T.: Tent: Fully test-time adaptation by entropy minimization. In: The International Conference on Machine Learning (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 450, + 482, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 450, + 482, + 493 + ], + "spans": [ + { + "bbox": [ + 130, + 450, + 482, + 493 + ], + "type": "text", + "content": "37. Wang, Y., Chao, W.L., Garg, D., Hariharan, B., Campbell, M., Weinberger, K.Q.: Pseudo-lidar from visual depth estimation: Bridging the gap in 3d object detection for autonomous driving. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8445-8453 (2019)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 493, + 482, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 493, + 482, + 525 + ], + "spans": [ + { + "bbox": [ + 130, + 493, + 482, + 525 + ], + "type": "text", + "content": "38. Wu, H., Wen, C., Shi, S., Li, X., Wang, C.: Virtual sparse convolution for multimodal 3d object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21653-21662 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 525, + 482, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 525, + 482, + 557 + ], + "spans": [ + { + "bbox": [ + 130, + 525, + 482, + 557 + ], + "type": "text", + "content": "39. Xu, B., Chen, Z.: Multi-level fusion based 3d object detection from monocular images. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2345-2353 (2018)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 557, + 482, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 590 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 590 + ], + "type": "text", + "content": "40. Xu, J., Peng, L., Cheng, H., Li, H., Qian, W., Li, K., Wang, W., Cai, D.: Mononerd: Nerf-like representations for monocular 3d object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6814-6824 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 590, + 482, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 590, + 482, + 622 + ], + "spans": [ + { + "bbox": [ + 130, + 590, + 482, + 622 + ], + "type": "text", + "content": "41. Yang, J., Shi, S., Wang, Z., Li, H., Qi, X.: St3d: Self-training for unsupervised domain adaptation on 3d object detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10368-10378 (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 622, + 482, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 622, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 622, + 482, + 665 + ], + "type": "text", + "content": "42. Ye, X., Shu, M., Li, H., Shi, Y., Li, Y., Wang, G., Tan, X., Ding, E.: Rope3d: The roadside perception dataset for autonomous driving and monocular 3d object detection task. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21341-21350 (2022)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "type": "text", + "content": "MonoTTA" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 358 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "text", + "content": "43. Zhang, M., Levine, S., Finn, C.: Memo: Test time robustness via adaptation and augmentation. In: Advances in Neural Information Processing Systems. vol. 35, pp. 38629-38642 (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 481, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 481, + 183 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 481, + 183 + ], + "type": "text", + "content": "44. Zhang, Y., Hooi, B., Hong, L., Feng, J.: Self-supervised aggregation of diverse experts for test-agnostic long-tailed recognition. In: Advances in Neural Information Processing Systems. vol. 35, pp. 34077-34090 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 481, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 481, + 216 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 481, + 216 + ], + "type": "text", + "content": "45. Zhang, Y., Hooi, B., Hu, D., Liang, J., Feng, J.: Unleashing the power of contrastive self-supervised visual models via contrast-regularized fine-tuning. In: Advances in Neural Information Processing Systems. vol. 34, pp. 29848-29860 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 216, + 481, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 481, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 481, + 248 + ], + "type": "text", + "content": "46. Zhang, Y., Kang, B., Hooi, B., Yan, S., Feng, J.: Deep long-tailed learning: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence 45(9), 10795-10816 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 249, + 481, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 481, + 282 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 481, + 282 + ], + "type": "text", + "content": "47. Zhang, Y., Wei, Y., Wu, Q., Zhao, P., Niu, S., Huang, J., Tan, M.: Collaborative unsupervised domain adaptation for medical image diagnosis. IEEE Transactions on Image Processing 29, 7834-7844 (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 282, + 481, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 282, + 481, + 315 + ], + "spans": [ + { + "bbox": [ + 132, + 282, + 481, + 315 + ], + "type": "text", + "content": "48. Zhang, Y., Lu, J., Zhou, J.: Objects are different: Flexible monocular 3d object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3289-3298 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 315, + 481, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 481, + 358 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 481, + 358 + ], + "type": "text", + "content": "49. Zou, Z., Ye, X., Du, L., Cheng, X., Tan, X., Zhang, L., Feng, J., Xue, X., Ding, E.: The devil is in the task: Exploiting reciprocal appearance-localization features for monocular 3d object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2713-2722 (2021)" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 315, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 315, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 315, + 102 + ], + "type": "text", + "content": "Hongbin Lin, Yifan Zhang and et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/MonoWAD_ Weather-Adaptive Diffusion Model for Robust Monocular 3D Object Detection/4d73977a-6d19-47be-84e6-b3aec99a372e_content_list.json b/2024/MonoWAD_ Weather-Adaptive Diffusion Model for Robust Monocular 3D Object Detection/4d73977a-6d19-47be-84e6-b3aec99a372e_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..769b30e9f59bac04fe9a30b5c5edaf5faf34466f --- /dev/null +++ b/2024/MonoWAD_ Weather-Adaptive Diffusion Model for Robust Monocular 3D Object Detection/4d73977a-6d19-47be-84e6-b3aec99a372e_content_list.json @@ -0,0 +1,1866 @@ +[ + { + "type": "text", + "text": "MonoWAD: Weather-Adaptive Diffusion Model for Robust Monocular 3D Object Detection", + "text_level": 1, + "bbox": [ + 228, + 140, + 774, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Youngmin Oh1, Hyung-Il Kim2, Seong Tae Kim1†, and Jung Uk Kim1†", + "bbox": [ + 222, + 210, + 781, + 228 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Kyung Hee University, Yong-in, South Korea {oym9104, st.kim, ju.kim}@khu.ac.kr 2 ETRI, Daejeon, South Korea hikim@etri.re.kr", + "bbox": [ + 343, + 238, + 658, + 295 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Monocular 3D object detection is an important challenging task in autonomous driving. Existing methods mainly focus on performing 3D detection in ideal weather conditions, characterized by scenarios with clear and optimal visibility. However, the challenge of autonomous driving requires the ability to handle changes in weather conditions, such as foggy weather, not just clear weather. We introduce MonoWAD, a novel weather-robust monocular 3D object detector with a weather-adaptive diffusion model. It contains two components: (1) the weather codebook to memorize the knowledge of the clear weather and generate a weather-reference feature for any input, and (2) the weather-adaptive diffusion model to enhance the feature representation of the input feature by incorporating a weather-reference feature. This serves as an attention role in indicating how much improvement is needed for the input feature according to the weather conditions. To achieve this goal, we introduce a weather-adaptive enhancement loss to enhance the feature representation under both clear and foggy weather conditions. Extensive experiments under various weather conditions demonstrate that MonoWAD achieves weather-robust monocular 3D object detection. The code and dataset are released at https://github.com/VisualAIKHU/MonoWAD.", + "bbox": [ + 261, + 332, + 743, + 597 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Monocular 3D Object Detection $\\cdot$ Weather-Adaptive Diffusion $\\cdot$ Weather Codebook", + "bbox": [ + 259, + 609, + 740, + 638 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 662, + 375, + 679 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Monocular 3D object detection aims to detect 3D objects using only a single camera [16, 18, 28, 46, 55, 57]. In contrast to LiDAR-based methods that rely on expensive LiDAR sensors for depth estimation [24, 44, 45, 56], and stereo-based methods that require synchronized stereo cameras, monocular 3D object detection only requires monocular images, offering the advantage of computational cost-effectiveness and requiring fewer resources. Due to this characteristic, the monocular 3D object detection is applied to a wide range of real-world applications, such as autonomous driving [12, 19, 53] and robotics [50].", + "bbox": [ + 212, + 694, + 787, + 816 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author", + "bbox": [ + 217, + 824, + 382, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4096afacdb8773efa89658e3f125d0abdc0e7e5a72b2515d361cce17f37609ca.jpg", + "image_caption": [ + "Fig. 1: Conceptual diagram of the proposed method (foggy example). (a) In the training phase, weather codebook learns the clear knowledge to transfer it to weather-adaptive diffusion model to enhance content related to the weather conditions. (b) By doing so, even with input images under various weather conditions (e.g., foggy images), monocular 3D object detection becomes adaptable to various weather scenarios." + ], + "image_footnote": [], + "bbox": [ + 210, + 148, + 792, + 255 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, existing monocular 3D object detectors mainly focus on ideal autonomous driving environments (i.e., clear weather). There are challenges in applying them to real-world scenarios with adverse weather conditions, such as fog and rain. Among these, fog poses the most significant challenge compared to other weather [14, 17]. This is due to the dense and diffuse nature of fog, which strongly scatters and absorbs light, leading to difficulties in object detection [2]. Since monocular 3D object detection relies solely on visual information from a single image, unlike LiDAR, it is crucial to design detectors to achieve enhanced performance in challenging visibility scenarios.", + "bbox": [ + 212, + 340, + 787, + 477 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose MonoWAD, a novel weather-robust monocular 3D object detector to address the aforementioned issues. As mentioned earlier, due to the inherent challenges posed by foggy weather among various adverse weather conditions, we focus primarily on clear and foggy weather (results for other weather conditions such as rainy and sunset are presented in Section 4). For weather-robust object detection, clear weather requires relatively modest improvements to enhance the visual representation of features. In contrast, significant enhancements in this feature representation are required for foggy weather. To address this, we consider two key aspects: (1) how to quantify the degree of improvement needed for the input image, and (2) how to guide the representation of the input image.", + "bbox": [ + 212, + 478, + 787, + 643 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the two key aspects, as shown in Fig. 1, our MonoWAD consists of a weather codebook and weather-adaptive diffusion model. First, we introduce the weather codebook to generate a weather-reference feature that contains knowledge about reference weather in a given scene. The reference weather acts as a guide, indicating the degree of weather improvement required. Since the clear weather contains a richer visual representation of objects, we adopt it as a reference weather. At this time, we devise a clear knowledge recalling (CKR) loss to guide the weather codebook to memorize information about clear weather and generate a weather-reference feature for any input (clear or foggy). As a result, our detector can understand where improvements are needed in the input features based on the weather-reference feature.", + "bbox": [ + 212, + 643, + 787, + 808 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Second, we propose a weather-adaptive diffusion model to effectively enhance feature representations in accordance with weather conditions. Given input fea-", + "bbox": [ + 212, + 809, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Y. Oh et al.", + "bbox": [ + 271, + 114, + 354, + 127 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ture (clear or foggy), the weather-adaptive diffusion model dynamically enhances the representation of the input feature based on the weather-reference feature. The weather-reference feature plays a role of attention, determining the extent to which the input features need improvement. At this time, we define the difference between clear and foggy weather (i.e., weather changes) as the fog distribution to adopt it as the noise for our diffusion model. With fog distribution, our weather-adaptive diffusion model can enhance the feature representation according to the weather conditions through multiple steps of reverse processes. To achieve this goal, we introduce a weather-adaptive enhancement (WAE) loss. As a result, our MonoWAD performs weather-robust detection by adaptively improving feature representation according to the weather conditions.", + "bbox": [ + 212, + 146, + 782, + 311 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To adaptively enhance the feature representation through the difference between weather conditions, we generate a new foggy KITTI dataset based on the KITTI dataset [12]. Comprehensive experimental results on several datasets [11, 12] show that our MonoWAD outperforms the existing state-of-the-art monocular 3D object detectors [16,28,37,40,55] under foggy weather, which is the most challenging weather condition. While our method primarily focuses on foggy, experiments conducted under various weather conditions (e.g., foggy, rainy, and sunset) have demonstrated its applicability to other weather scenarios.", + "bbox": [ + 212, + 313, + 784, + 431 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The main contributions of our paper can be summarized as follows:", + "bbox": [ + 238, + 433, + 725, + 448 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce a new weather-robust monocular 3D object detector, called MonoWAD, that is robust to various weather conditions.", + "- We design a weather codebook with clear knowledge recalling loss for learning about clear weather, providing reference information for enhancement.", + "- We propose weather-adaptive diffusion model with weather-adaptive enhancement loss to dynamically enhance the feature representation of the input images according to the weather conditions." + ], + "bbox": [ + 225, + 458, + 782, + 561 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 585, + 387, + 602 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Monocular 3D Object Detection", + "text_level": 1, + "bbox": [ + 215, + 618, + 527, + 633 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Monocular 3D object detection task can be categorized into two directions according to the type of data used in the training phase: (1) using only a monocular image and (2) incorporating additional data, such as depth along with a monocular image. The first category relies on the geometric relationship between 2D and 3D [3,7,25,26,28,31,34,40,46,55,57]. For example, Deep3Dbox [34] utilizes the geometric information of 2D bounding boxes to predict 3D bounding boxes. In [3], M3D-RPN was proposed to understand the 3D scene from the depth-aware convolution. MonoRCNN [46] predicted 3D bounding boxes through geometry-based distance decomposition, and MonoCon [25] learns mono context for 3D object detection. MonoDETR [55] introduces a depth-guided transformer that utilizes geometric depth cues without requiring additional data.", + "bbox": [ + 212, + 643, + 784, + 808 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Moreover, since monocular image contains limited information for estimating 3D object cues, monocular 3D object detectors have adopted the additional", + "bbox": [ + 214, + 810, + 784, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "MonoWAD", + "bbox": [ + 653, + 114, + 730, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "data for more robust detection [4,8,16,27,32,37,41,48]. For example, the depth-conditioned dynamic message propagation (DDMP) [48] was proposed to integrate prior depth information with the image context. CaDDN [41] was introduced to utilize the depth distribution of predicted categories for each pixel to project the context information onto 3D space, deriving 3D bounding boxes. MonoDTR [16] employs the transformer architecture to integrate depth features and context, thus estimating more accurate depth information.", + "bbox": [ + 212, + 146, + 787, + 252 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Despite recent progress, the existing monocular 3D object detectors mainly rely on the benchmark data collected under clear weather conditions. However, it is essential to account for challenging adverse weather conditions, such as fog, to more accurately reflect real-world scenarios. In this paper, we aim to introduce weather-robust 3D object detection by enhancing visual features with the proposed weather-adaptive diffusion model and weather codebook.", + "bbox": [ + 212, + 252, + 787, + 343 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Computer Vision Tasks for Foggy Weather", + "text_level": 1, + "bbox": [ + 214, + 354, + 614, + 371 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "There have been a lot of studies on improving performance in various weather conditions for real-world application of computer vision technology [2, 13, 20, 21, 33, 38, 43, 51, 54]. In particular, fog is considered one of the most critical issues due to its significant degradation of visual information compared to other weather conditions. To deal with the foggy weather conditions, the authors in [43] generated synthetic fog images (so-called Foggy Cityscapes) from clear weather images, and utilized them for training semantic segmentation and 2D object detection. Similar to Foggy Cityscapes, Martin et al. [13] naturally synthesizes fog into LIDAR to enhance the performance of LiDAR-based 3D object detectors in foggy weather. Bijelic et al. [2] use real dataset including foggy weather, for 2D detection with the multimodal fusion networks. Mai et al. [33] synthesize fog for LiDAR and stereo images and perform fusion-based 3D object detection using the SLS-Fusion network. Xin et al. [51] focused on 2D detection in foggy weather by applying domain adaptation. In this context, we focus on monocular 3D object detection in scenarios that rely solely on visual information in challenging foggy environments. To address this challenge, we propose a weather-robust diffusion model that dynamically improves features based on reference feature.", + "bbox": [ + 212, + 380, + 787, + 636 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.3 Diffusion Models", + "text_level": 1, + "bbox": [ + 214, + 648, + 401, + 662 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Recently, the diffusion model [15] has attracted considerable interest in computer vision due to its impressive progress in image generation [22, 29, 35, 42], as well as its potential application to other vision tasks such as segmentation [1] and image captioning [30]. Inspired by the remarkable generative ability, we design the diffusion model for robust monocular 3D object detection by considering a foggy effect (which is one of the challenging adverse weather conditions for monocular 3D object detection) as a form of noise in the model. That is, we propose a method in which visual features obscured by fog are progressively improved by training a diffusion model based on the forward/reverse diffusion process. In particular, we present an adaptive method that allows the diffusion model to control the degree of improvement by weather conditions.", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Y. Oh et al.", + "bbox": [ + 271, + 114, + 354, + 127 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/2105952e910218bb8044c128b12c704c3f65278f2fb9df10aeaedb76fc42a5ae.jpg", + "image_caption": [ + "Fig. 2: Overview of our MonoWAD in the inference phase. It mainly contains three parts: weather codebook, weather-adaptive diffusion model, and detection block. Through the weather codebook and weather-adaptive diffusion model, our method can maintain robustness against various weather conditions (i.e., clear or foggy)." + ], + "image_footnote": [], + "bbox": [ + 218, + 146, + 782, + 268 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Proposed Method", + "text_level": 1, + "bbox": [ + 215, + 347, + 426, + 364 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Fig. 2 shows the overall framework of the proposed MonoWAD in the inference phase. A backbone network receives an input image (clear image $I_{c}$ or foggy image $I_{f}$ ) to encode the corresponding input feature (input clear feature $x^{c}$ or input foggy feature $x^{f}$ ). It interacts with the weather codebook $\\mathcal{Z}$ to generate weather-reference feature $x^{r}$ , indicating the amount of enhancement for the given input feature. Subsequently, the weather-adaptive diffusion model attempts to enhance the input feature over $T$ timesteps to obtain an enhanced feature $\\tilde{x}^{c}$ or $\\tilde{x}^{f}$ . Finally, monocular 3D object detection is performed through the detection block. Note that, in the training phase, our MonoWAD use the clear feature $x^{c}$ to train our diffusion as well as the detection block.", + "bbox": [ + 212, + 383, + 787, + 532 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We address two key issues: (1) how to guide weather-reference feature $x^r$ to serve as a reference feature, and (2) how to guide the weather-adaptive diffusion model to effectively enhance feature representations based on weather conditions. Details are in the following subsections.", + "bbox": [ + 212, + 535, + 787, + 595 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Weather Codebook", + "text_level": 1, + "bbox": [ + 215, + 613, + 421, + 627 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In foggy weather conditions, the overall visual quality of the scene is generally poor, requiring a significant enhancement. Conversely, in clear weather, the amount of improvement is expected to be relatively minimal compared to foggy conditions. Therefore, inspired by [9,36], we devise a weather codebook $\\mathcal{Z}$ to provide the reference knowledge about the weather for appropriate enhancement based on the weather conditions. At this time, as the clear weather contains abundant visual representations, we use it as a reference weather knowledge.", + "bbox": [ + 212, + 642, + 787, + 747 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in Fig. 3, the reference knowledge embedding procedure involves receiving paired clear-foggy features during the training phase. The weather codebook $\\mathcal{Z}$ consists of $K$ learnable slots, denoted as $\\mathcal{Z} = \\{z_k\\}_{k=1}^K (z_k \\in \\mathbb{R}^{1 \\times c})$ , where $c$ represents the dimensionality of each slot. The paired clear feature $x^c$ and foggy feature $x^f$ pass through a convolution layer to generate $\\hat{x}^c \\in \\mathbb{R}^{h \\times w \\times c}$ and $\\hat{x}^f \\in \\mathbb{R}^{h \\times w \\times c}$ ( $w$ denotes width and $h$ indicates height). Each element of the", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "MonoWAD", + "bbox": [ + 653, + 114, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a461609d2f407e75e3eec09b04ff65dc201e192f796c9785c68517f768250a5b.jpg", + "image_caption": [ + "Fig. 3: Illustration of the proposed (a) weather-invariant guiding (WIG) loss and (b) clear knowledge embedding (CKE) loss. The clear knowledge recalling (CKR) loss, obtained from combining WIG and CKE, aims to memorize the knowledge of the clear weather and recall the same clear knowledge from the foggy weather." + ], + "image_footnote": [], + "bbox": [ + 212, + 146, + 787, + 342 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "feature denoted as $\\hat{x}_{ij}^{c}\\in \\mathbb{R}^{1\\times c}$ and $\\hat{x}_{ij}^{f}\\in \\mathbb{R}^{1\\times c}$ . Subsequently, we obtain weather-reference feature for clear weather, $x^{r(c)}\\in \\mathbb{R}^{h\\times w\\times c}$ , by conducting element-wise quantization process $\\mathbf{q}(\\cdot)$ , calculated as:", + "bbox": [ + 212, + 412, + 787, + 462 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nx ^ {r (c)} = \\mathbf {q} (\\hat {x} ^ {c}) := \\left(\\underset {z _ {k} \\in \\mathcal {Z}} {\\arg \\min } \\| \\hat {x} _ {i j} ^ {c} - z _ {k} \\|\\right). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 357, + 472, + 785, + 506 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Utilizing $x^{c}$ and $x^{r(c)}$ , we introduce a clear knowledge embedding (CKE) loss $\\mathcal{L}_{cke}$ to guide $x^{r(c)}$ to follow the representation of $x^{c}$ . To this end, we perform global average pooling (GAP) for $x^{c}$ and $x^{r(c)}$ with softmax, generating $s^{c}$ and $s^{r(c)}$ , respectively. Each element in the vector indicates the probability of the significance of each channel. With $s^{c}$ and $s^{r(c)}$ , we employ KL divergence $D_{KL}(\\cdot)$ for $\\mathcal{L}_{cke}$ to compare the probability distributions, formulated as:", + "bbox": [ + 214, + 518, + 787, + 611 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {c k e} = D _ {K L} \\left(s ^ {c} \\| s ^ {r (c)}\\right). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 416, + 619, + 785, + 640 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Through $\\mathcal{L}_{cke}$ , the weather codebook $\\mathcal{Z}$ can memorize the knowledge of the clear weather, allowing it to effectively reconstruct the clear weather knowledge.", + "bbox": [ + 212, + 650, + 785, + 680 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Additionally, as the paired clear-foggy images are identical except for weather conditions, the quantization process of the foggy feature $x^{f}$ with the weather codebook should generate an equivalent weather-reference feature. For obtaining weather-reference feature for foggy $x^{r(f)}$ , the element-wise quantization process is also conducted for $\\hat{x}^{f}$ and $\\mathcal{Z}$ :", + "bbox": [ + 212, + 681, + 787, + 755 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nx ^ {r (f)} = \\mathbf {q} (\\hat {x} ^ {f}) := \\left(\\underset {z _ {k} \\in \\mathcal {Z}} {\\arg \\min } \\| \\hat {x} _ {i j} ^ {f} - z _ {k} \\|\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 356, + 763, + 785, + 799 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Next, we introduce the weather-invariant guiding (WIG) loss $\\mathcal{L}_{wig}$ to guide $\\mathcal{Z}$ that the weather codebook recalls the same clear knowledge of clear feature for", + "bbox": [ + 214, + 809, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Y. Oh et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/30c2bd0b9fc9be3b440b6c3e119c54193994e8d55260f719d12000e097d5ad66.jpg", + "image_caption": [ + "Fig. 4: Training process of the weather-adaptive diffusion model, which consists of two processes: (1) Adding fog variant $\\epsilon_{n}$ from input clear feature $x^{c}$ (forward process) and (2) enhancing representation with weather-reference feature $x^{r}$ (reverse process)." + ], + "image_footnote": [], + "bbox": [ + 225, + 146, + 776, + 311 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the foggy feature, which can be represented as follows:", + "bbox": [ + 215, + 371, + 607, + 385 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {w i g} = \\left\\| x ^ {r (c)} - x ^ {r (f)} \\right\\| _ {2} ^ {2}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 410, + 397, + 785, + 426 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Finally, the clear knowledge recalling (CKR) loss $\\mathcal{L}_{ckr}$ is obtained by adding $\\mathcal{L}_{cke}$ and $\\mathcal{L}_{wig}$ , which is defined as:", + "bbox": [ + 215, + 436, + 784, + 467 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {c k r} = \\mathcal {L} _ {c k e} + \\mathcal {L} _ {w i g}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 426, + 479, + 784, + 494 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In the training phase, the weight parameters of embedding $K$ slots of weather codebook $\\mathcal{Z}$ are initialized randomly and they are updated through Eq. (5). In the inference phase, all parameters are fixed to recall clear weather, generating weather-reference features for any weather conditions.", + "bbox": [ + 212, + 505, + 784, + 566 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.2 Weather-Adaptive Diffusion Model", + "text_level": 1, + "bbox": [ + 215, + 588, + 550, + 603 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Through Section 3.1, we now know that the weather codebook outputs the weather-reference feature $x^{r(c)}$ for the clear weather and $x^{r(f)}$ for the foggy weather through Eq. (5). From now on, since our method can receive any input images (clear or foggy), we denote the weather-reference feature as $x^r$ .", + "bbox": [ + 212, + 613, + 784, + 672 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Fig. 4 shows the training process of the proposed weather-adaptive diffusion model. The key idea of the diffusion model [15,42] is to gradually enhance $x_{T}$ to $x_{0}$ with fixed Markov Chain of $T$ timesteps. To this end, the forward and reverse processes are conducted in the training phase, and only the reverse process is used for the inference phase. Motivated by [15,42], we construct the weather-adaptive diffusion model to enhance representation related to the weather conditions. To this end, unlike traditional diffusion methods [15,42] that adopt the Gaussian noise to the image or latent space, we adopt $\\mathcal{F} = x^{f} - x^{c}$ , called fog distribution, to guide our diffusion model to be aware of the foggy weather. Ideally, the information contained within $\\mathcal{F}$ should include the information of fog, as it represents the difference between the foggy scene and the same scene with clear weather.", + "bbox": [ + 212, + 674, + 785, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "MonoWAD", + "bbox": [ + 653, + 114, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "By doing so, our diffusion model learns the variation in weather by repeatedly adding and removing the fog. Note that, as we take clear feature $x^{c}$ for $x_{0}$ to make a reference input for our diffusion model, we newly denote $x_{0}$ as $x_{0}^{c}$ .", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "For the forward process at the $t$ -th timestep, $q(x_{t}^{c}|x_{t - 1}^{c})$ takes the previous feature $x_{t - 1}^{c}$ and the noise related to the fog (i.e., $\\mathcal{F}$ ) as inputs to generate $x_{t}^{c}$ . This procedure is repeated over $T$ timesteps, which can be represented as:", + "bbox": [ + 212, + 191, + 784, + 237 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nq (x _ {t} ^ {c} | x _ {t - 1} ^ {c}) = \\mathcal {F} (x _ {t} ^ {c}; \\sqrt {1 - \\beta_ {t}} x _ {t - 1} ^ {c}, \\beta_ {t} \\mathbf {I}), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 366, + 247, + 785, + 267 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nq \\left(x _ {1} ^ {c}, \\dots , x _ {T} ^ {c} \\mid x ^ {c}\\right) = \\prod_ {t = 1} ^ {T} q \\left(x _ {t} ^ {c} \\mid x _ {t - 1} ^ {c}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 385, + 277, + 785, + 316 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\beta_{t}$ denotes the variance schedule.", + "bbox": [ + 214, + 323, + 501, + 335 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Next, the reverse process at the $t$ -th timestep aims to estimate fog variant $\\epsilon_{n}$ using $x_{t}^{c}$ to enhance the feature representation of the foggy. To this end, we adopt conditional autoencoder $\\epsilon_{\\theta}(x_{t}^{c}, t, x^{r})$ that receives $x_{t}^{c}$ and reference feature $x^{r}$ from the weather codebook $\\mathcal{Z}$ . Specifically, $\\epsilon_{\\theta}$ estimates mean $\\mu_{\\theta}$ and variance $\\pmb{\\Sigma}_{\\theta}$ of the fog distribution at the $t$ -th timestep, denoted as $\\tilde{\\mathcal{F}}(\\cdot)$ . The reverse process is also repeated over $T$ timesteps, which can be represented as:", + "bbox": [ + 212, + 338, + 782, + 428 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\theta} \\left(x _ {t - 1} ^ {c} \\mid x _ {t} ^ {c}, x ^ {r}\\right) = \\tilde {\\mathcal {F}} \\left(x _ {t - 1} ^ {c}; \\boldsymbol {\\mu} _ {\\theta} \\left(x _ {t} ^ {c}, t, x ^ {r}\\right), \\boldsymbol {\\Sigma} _ {\\theta} \\left(x _ {t} ^ {c}, t\\right)\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 325, + 440, + 785, + 458 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\theta} \\left(x ^ {c}, \\dots , x _ {T} ^ {c}\\right) = p \\left(x _ {T} ^ {c}\\right) \\prod_ {t = 1} ^ {T} p _ {\\theta} \\left(x _ {t - 1} ^ {c} \\mid x _ {t} ^ {c}, x ^ {r}\\right), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 356, + 469, + 785, + 508 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $p_{\\theta}(x_{t - 1}^{c}|x_{t}^{c},x^{r})$ includes cross-attention layer in $\\epsilon_{\\theta}$", + "bbox": [ + 214, + 515, + 627, + 530 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The cross-attention layer receives the flattened clear feature $x_{t}^{c}$ and weather-reference feature $x^{r}$ , i.e., $\\bar{x}_{t}^{c}$ and $\\bar{x}^{r}$ , respectively. The similarity between $\\bar{x}^{r}$ and $\\bar{x}_{t}^{c}$ is calculated in the cross-attention layer to transfer the enhancement to $\\bar{x}_{t}^{c}$ , which can be formulated as:", + "bbox": [ + 212, + 530, + 782, + 589 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {A t t e n t i o n} (Q, K, V) = \\operatorname {s o f t m a x} \\left(\\frac {Q K ^ {T}}{\\sqrt {d}}\\right) V, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 352, + 598, + 785, + 633 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $Q = W_{i}^{q} \\cdot \\bar{x}_{t}^{c}, K = W_{i}^{k} \\cdot \\bar{x}^{r}, V = W_{i}^{v} \\cdot \\bar{x}^{r}$ with learnable parameters $W_{i}^{q}, W_{i}^{k}, W_{i}^{v}$ .", + "bbox": [ + 214, + 642, + 782, + 672 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To ensure that the estimated fog variant $\\epsilon_{\\theta}$ is similar to the fog variant $\\epsilon_{n}$ applied to make $x_{t}^{c}$ from the input clear feature $x^{c}$ , we propose a weather-adaptive enhancement loss $\\mathcal{L}_{wae}$ , which is formulated as:", + "bbox": [ + 212, + 674, + 782, + 719 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {w a e} = \\mathbb {E} _ {x ^ {c}, \\boldsymbol {\\epsilon} _ {n} \\sim \\mathcal {F}, t} \\left[ \\| \\boldsymbol {\\epsilon} _ {n} - \\boldsymbol {\\epsilon} _ {\\theta} \\left(x _ {t} ^ {c}, t, x ^ {r}\\right) \\| _ {2} ^ {2} \\right]. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 356, + 729, + 785, + 755 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "With $\\mathcal{L}_{wae}$ , $\\epsilon_{\\theta}$ can estimate the fog variant by leveraging the fog distribution as noise for our diffusion model through multiple forward/reverse processes. Additionally, the cross-attention layer within $\\epsilon_{\\theta}$ dynamically enhances the feature representation based on combining knowledge of the input feature (foggy or clear) and the weather-reference feature. Since our diffusion model learns the", + "bbox": [ + 212, + 763, + 785, + 839 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Y. Oh et al.", + "bbox": [ + 271, + 114, + 354, + 127 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "degree of improvement needed corresponding to weather conditions, it can improve the representation of any input, whether clear or foggy in inference phase. This leads our monocular 3D object detector can handle both clear and foggy input images, resulting in weather-robust detection.", + "bbox": [ + 212, + 146, + 782, + 205 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "3.3 Total Loss Function", + "text_level": 1, + "bbox": [ + 214, + 218, + 426, + 232 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The total loss function of our MonoWAD is represented as follows:", + "bbox": [ + 214, + 243, + 694, + 258 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {T o t a l}} = \\mathcal {L} _ {O D} + \\lambda_ {1} \\mathcal {L} _ {\\text {c k r}} + \\lambda_ {2} \\mathcal {L} _ {\\text {w a e}}, \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 380, + 272, + 784, + 287 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\lambda_{1}$ and $\\lambda_{2}$ denote balancing hyper-parameters. In our experiments, we set $\\lambda_{1} = \\lambda_{2} = 1$ . $\\mathcal{L}_{OD}$ is the detection loss for 3D object detection [16,55]. It includes loss functions of classification, regression, and depth loss, that are similar to prior works [8,16,31]. The overall weight parameters are updated through $\\mathcal{L}_{Total}$ .", + "bbox": [ + 214, + 299, + 782, + 359 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 214, + 382, + 375, + 400 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Dataset and Evaluation Metrics", + "text_level": 1, + "bbox": [ + 214, + 414, + 526, + 429 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Datasets. We utilize the KITTI 3D object detection dataset [12], the most widely adopted for 3D object detection. It contains 7,481 training images and 7,518 test images under clear weather conditions. Due to unavailable ground-truth annotations for test images and limited evaluation on the test server, we follow [6] by splitting 3,712 training images and 3,769 validation images. In addition, as our work requires paired images to learn about weather changes, we generate foggy images from all images in the KITTI dataset, called foggy KITTI, that emulate the foggy scene. Following the protocols of [2,43], we generate foggy images based on object distances using depth maps estimated by DORN [10]. Please refer to the supplementary materials for details.", + "bbox": [ + 212, + 439, + 784, + 589 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In addition, as our work is focused on robust monocular 3D object detection in various weather conditions, we further adopt the Virtual KITTI dataset [5,11], which contains photo-realistic synthetic images under various weather conditions (e.g., foggy, rainy, sunset). It is associated with the original real-world KITTI dataset, which has 3D annotations for each weather.", + "bbox": [ + 212, + 590, + 784, + 666 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Evaluation Metrics. We adopt average precision in both 3D detection $(\\mathrm{AP}_{3D})$ and bird-eye view detection $(\\mathrm{AP}_{BEV})$ under three difficulty levels ('Easy', 'Moderate', 'Hard') according to size, occlusion, and truncation. Following [47], we use 40 recall position metric $\\mathrm{AP}_{40}$ and report scores for the car category under the IoU threshold 0.7 for the KITTI dataset and 0.5 for the Virtual KITTI dataset.", + "bbox": [ + 214, + 672, + 782, + 750 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.2 Implementation Details", + "text_level": 1, + "bbox": [ + 214, + 768, + 455, + 784 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We use DLA-102 [52] as our backbone and adopt the transformer architecture of [16] for the detection block. We train MonoWAD on a single RTX 4090 GPU with a batch size of 4 over 120 epochs using Adam optimizer [23] (initial learning", + "bbox": [ + 212, + 794, + 782, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "MonoWAD", + "bbox": [ + 653, + 114, + 730, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/43d987fd65abcd4539352a30f7ddc7a32010e6856d7d49d74baa9d01a9cba868.jpg", + "table_caption": [ + "Table 1: Detection results of car category on KITTI validation set under foggy weather and clear weather conditions. The results of the state-of-the-art methods under foggy weather are obtained through our reproduction with the official source code. Bold/underlined fonts indicate the best/second-best results." + ], + "table_footnote": [], + "table_body": "
MethodFoggy (AP3D)Clear (AP3D)Average
EasyMod.HardEasyMod.HardEasyMod.Hard
GUPNet [28] (ICCV'21)2.742.192.1622.7616.4613.7212.759.337.94
DID-M3D [37] (ECCV'22)1.150.610.6422.9816.1214.0312.078.377.34
MonoGround [40] (CVPR'22)0.000.000.0625.2418.6915.5812.629.357.82
MonoDTR [16] (CVPR'22)16.8911.869.8724.5218.5715.5120.7115.2212.69
MonoDETR [55] (ICCV'23)7.405.744.5328.8420.6116.3818.1213.1810.46
MonoWAD (Ours)27.1719.5716.2129.1021.0817.7328.1420.3316.97
", + "bbox": [ + 220, + 203, + 782, + 306 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/34ef2b7aba8cdc8c396a824d68a1df5f4a92af2375954e1aaeb4d4e5c784a916.jpg", + "table_caption": [ + "Table 2: Detection results of car category on KITTI test set under foggy weather. Bold/underlined fonts indicate the best/second-best results." + ], + "table_footnote": [], + "table_body": "
MethodAP3DAPBEV
EasyMod.HardEasyMod.Hard
GUPNet [28] (ICCV'21)3.012.421.134.903.022.91
DID-M3D [37] (ECCV'22)3.102.392.195.343.012.91
MonoGround [40] (CVPR'22)0.140.200.220.230.380.39
MonoDTR [16] (CVPR'22)11.077.415.2615.7610.157.53
MonoDETR [55] (ICCV'23)9.335.544.0613.158.066.30
MonoWAD (Ours)19.7513.3211.0427.9519.0615.61
", + "bbox": [ + 259, + 349, + 736, + 455 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "rate of $10^{-4}$ ). For the weather codebook, we utilize embedding slot $K = 4096$ and set the dimension of each slot $D = 256$ . We set 4 heads for cross-attention of the weather-adaptive diffusion model, and the timesteps for the forward and reverse process of diffusion default to 15. The number of channels for the diffusion output features is $C = 256$ . As recent monocular 3D object detectors [16,28,37,40,55] have not been explored under adverse weather conditions, we implemented them with available official source code to faithfully reproduce them.", + "bbox": [ + 212, + 469, + 787, + 575 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.3 Comparison", + "text_level": 1, + "bbox": [ + 215, + 589, + 362, + 604 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Results on KITTI 3D Dataset. We compared MonoWAD with state-of-the-art monocular 3D object detectors [16,28,37,40,55] that do not use additional data (e.g., depth maps or LiDAR) during inference on the KITTI and foggy KITTI validation sets. Table 1 shows the $AP_{3D}$ results, and $AP_{BEV}$ results are in the supplementary materials. While recent methods have shown improved performance in clear weather, their performance drops in foggy weather, limiting their applicability in real-world applications (e.g., autonomous driving and robotics). In contrast, MonoWAD showed stable 3D detection performance under both foggy and clear weather. Since our weather codebook learns knowledge about clear weather and the weather-adaptive diffusion model can enhance the feature representation of the input images under both clear and foggy weather, MonoWAD shows a more weather-robust detection performance than that of the existing methods. Also, to explore the weather robustness of MonoWAD, we conducted experiments with mixed foggy and clear weather conditions at various ratios (please see the supplementary materials).", + "bbox": [ + 212, + 612, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Y. Oh et al.", + "bbox": [ + 271, + 114, + 354, + 127 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/8f1f5b8491bac96f0f5e3211761154bb7fe958c2abbea8ff8288f4ecb58b66f1.jpg", + "table_caption": [ + "Table 3: Detection results of car category on Virtual KITTI under foggy, rainy, sunset conditions. Bold/underlined fonts indicate the best/second-best results." + ], + "table_footnote": [], + "table_body": "
MethodFoggy (AP3D)Rainy (AP3D)Sunset (AP3D)
EasyMod.HardEasyMod.HardEasyMod.Hard
GUPNet [28] (ICCV'21)1.761.571.572.341.241.212.771.641.65
DID-M3D [37] (ECCV'22)0.910.390.390.400.130.130.340.100.10
MonoGround [40] (CVPR'22)0.290.300.255.492.822.777.684.244.20
MonoDTR [16] (CVPR'22)8.795.755.7211.736.256.749.865.425.42
MonoDETR [55] (ICCV'23)4.502.992.966.613.463.427.084.174.16
MonoWAD (Ours)13.338.568.5014.128.338.2413.387.897.80
", + "bbox": [ + 220, + 176, + 777, + 280 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/92ef456f169f950eafccc5b80a73a7c9bb1b71ae5ff26f4d5d6731618d513186.jpg", + "table_caption": [ + "Table 4: Effect of the proposed method on KITTI validation set for car category. WC denotes our weather codebook, WAD indicates our weather-adaptive diffusion model." + ], + "table_footnote": [], + "table_body": "
MethodWCWADFoggy (AP3D)Clear (AP3D)
EasyMod.HardEasyMod.Hard
Baseline--13.759.618.1022.6317.1614.28
MonoWAD (Ours)-✓25.6218.6615.5626.3419.1716.15
✓✓27.1719.5716.2129.1021.0817.73
", + "bbox": [ + 279, + 313, + 717, + 385 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/7ac33ed1df4e17379d7d4fd068da7ffec2ff38f541fb5f02817366a9c2141fb6.jpg", + "table_caption": [ + "Table 5: Detection results on KITTI validation set by changing diffusion timestep $T$ ." + ], + "table_footnote": [], + "table_body": "
Timestep TFoggy (AP3D)Clear (AP3D)
EasyMod.HardEasyMod.Hard
-13.759.618.1022.6317.1614.28
523.5717.9114.9626.0319.2116.01
1025.2818.4915.4226.7919.9016.78
1527.1719.5716.2129.1021.0817.73
2024.5418.2915.1024.8518.5415.34
", + "bbox": [ + 274, + 407, + 722, + 502 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We further compared 3D detection performances on the foggy KITTI test set (Table 2). Similar to Table 1, existing monocular 3D object detectors show lower performance under foggy weather. Through the results, our method shows the weather-robust monocular 3D performance under foggy and clear weather.", + "bbox": [ + 215, + 513, + 782, + 574 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Results on Virtual KITTI Dataset. We also conducted experiments on the Virtual KITTI dataset to validate the generalization ability of our method across various weather conditions (i.e., foggy, rainy, sunset). The results are shown in Table 3. MonoWAD also shows the highest performance in foggy weather. Moreover, it outperforms the existing method in the rainy and sunset conditions. The results demonstrate that MonoWAD is robust and insensitive to various weather conditions that can be faced in real-world scenarios, not just clear weather.", + "bbox": [ + 215, + 582, + 784, + 688 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.4 Ablation Studies", + "text_level": 1, + "bbox": [ + 215, + 707, + 398, + 722 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We conducted ablation studies to examine (1) the effect of each proposed component (i.e., weather codebook and weather-adaptive diffusion model) and (2) the effect of the weather-adaptive diffusion model by varying the timestep $T$ . These experiments were performed on the KITTI and foggy KITTI 3D validation sets.", + "bbox": [ + 215, + 729, + 782, + 789 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Effect of the Proposed Modules. The results regarding the effectiveness of the proposed weather codebook and weather-adaptive diffusion model are pre-", + "bbox": [ + 215, + 809, + 784, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "MonoWAD", + "bbox": [ + 653, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/c1a042afafb71c984ae10b776ed7bb8d251369f7e429fb46a9e0001f466d4e66.jpg", + "table_caption": [ + "Table 6: Performance comparison on KITTI validation set under foggy and clear weather conditions. We compare our MonoWAD with MonoDTR [16], which demonstrates superior performance in foggy weather using state-of-the-art dehazing methods." + ], + "table_footnote": [], + "table_body": "
MethodDehazing MethodFoggy (AP3D)Clear (AP3D)Average
EasyMod.HardEasyMod.HardEasyMod.Hard
MonoDTR(B) [16] (CVPR'22)-16.8911.869.8724.5218.5715.5120.7115.2212.69
B + RIDCP [49] (CVPR'23)Image-level17.2312.4110.4424.0217.8914.7820.6315.1512.61
B + DENet [39] (ACCV'22)Feature-level22.3517.4414.477.105.704.5314.7311.579.50
B + Yang et al. [51] (ACCV'22)Feature-level22.8715.2112.1717.9613.1010.6420.4214.1611.41
MonoWAD (Ours)-27.1719.5716.2129.1021.0817.7328.1420.3316.97
", + "bbox": [ + 220, + 188, + 777, + 282 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/f612a81c020046afd3345f974015b5172f1ffaa62f69e2d8a3c7124aaae1bd75.jpg", + "table_caption": [ + "Table 7: Comparison of diffusion models on KITTI validation set for car category: $\\mathcal{B}$ is baseline detection block (transformer encoder-decoder), CDM (Conditional Diffusion Model [42]), WC (weather codebook), and WAD (weather adaptive diffusion model)." + ], + "table_footnote": [], + "table_body": "
MethodFoggy (AP3D)Clear (AP3D)
EasyMod.HardEasyMod.Hard
B+DDPM [15]5.323.842.7718.3112.7110.20
B+CDM [42]2.742.102.0020.5014.5111.63
B+WC+CDM [42]17.5112.7410.4021.0515.1112.54
B+WAD25.6218.6615.5626.3419.1716.15
MonoWAD(B+WC+WAD)27.1719.5716.2129.1021.0817.73
", + "bbox": [ + 274, + 340, + 723, + 450 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "sented in Table 4. Since our weather-adaptive diffusion model is designed to enhance feature representation, it alone has shown significant performance improvement. When the weather codebook is additionally considered, the weather-adaptive diffusion model can leverage the knowledge of the weather-reference feature for clear weather, leading to enhanced performance. It makes MonoWAD can be robust to various weather conditions for monocular 3D object detection.", + "bbox": [ + 212, + 465, + 787, + 556 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Effect of Timestep $T$ . We also conduct experiments by varying timestep $T$ of the weather-adaptive diffusion model. Timestep is the number of steps for the forward and reverse process of the diffusion model. Table 5 indicates that the highest monocular 3D detection performance is achieved when $T = 15$ under both clear and foggy weather conditions. Our MonoWAD consistently outperforms the baseline and other existing methods across all timesteps.", + "bbox": [ + 212, + 566, + 787, + 657 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.5 Discussions", + "text_level": 1, + "bbox": [ + 215, + 672, + 356, + 686 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Comparison with Dehazing Methods. We investigate the weather-robustness of our method for monocular 3D object detection compared to the other monocular detectors using dehazing methods. To this end, we compared MonoWAD with MonoDTR [16], which shows the best performance in foggy weather through the application of state-of-the-art image-level and feature-level dehazing methods [39,49,51]. The results are shown in Table 6. Recent dehazing methods are primarily focused on specific weather conditions, such as foggy weather. While they have shown some improvement under foggy conditions, they exhibit reduced performances under clear weather conditions. In contrast, our MonoWAD shows robust performance across both foggy and clear weather conditions.", + "bbox": [ + 212, + 696, + 792, + 849 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Y. Oh et al.", + "bbox": [ + 271, + 114, + 354, + 127 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/1ce9cb7a8af6a016833dac592e5eb58cf51c6f754a3faf5c18414187a51fa69e.jpg", + "image_caption": [ + "Fig. 5: Comparison of 3D detection examples (green: ground-truth, red: predicted 3D bounding-box) between our MonoWAD and two detectors, MonoDTR [16] and MonoDETR [55], that show the most improved performances among existing methods." + ], + "image_footnote": [], + "bbox": [ + 223, + 146, + 781, + 290 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/4df94299132fc98cb15113316d4d2591056ee507d5585a7166dd4da1fed82b6a.jpg", + "image_caption": [ + "Fig. 6: 3D detection results on real-world images of various weather conditions." + ], + "image_footnote": [], + "bbox": [ + 212, + 345, + 789, + 594 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Effect of Weather-Adaptive Diffusion Model. Table 7 shows the effectiveness of MonoWAD with existing diffusion models [15,42] on the KITTI and foggy KITTI validation sets. Existing methods adopt Gaussian noise for forward and reverse processes, but they can not fully understand about weather. In contrast, our weather-adaptive diffusion understands weather variances, allowing our MonoWAD to surpass existing methods in clear and foggy weather.", + "bbox": [ + 212, + 622, + 787, + 714 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.6 Visualization Results", + "text_level": 1, + "bbox": [ + 215, + 737, + 434, + 752 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Results on KITTI Dataset. We visualize several 3D detection results on the KITTI 3D dataset, comparing MonoWAD with MonoDTR and MonoDETR, which exhibit the highest performances among existing methods under foggy condition (Fig. 5). Existing methods struggle to detect objects obscured by fog, indicating limitations in detecting only fully visible objects. In contrast, even", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "MonoWAD", + "bbox": [ + 653, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/9a7aeccd9d723fbb3756503070f2e04bc5fb3e1ddc1cc2d526c39a3329162339.jpg", + "image_caption": [ + "Fig. 7: t-SNE visualization results (Red: Clear, Blue: Foggy)." + ], + "image_footnote": [], + "bbox": [ + 251, + 143, + 401, + 239 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/a32b11d56c1326e5fd5cc8acac38664ea01b1127ef3278cbfcfff7f6b61d76a0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 426, + 143, + 575, + 239 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/2590033f084d124308f014d93a14eb5c6c55e66ce963acf9541378c89e17472c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 602, + 143, + 751, + 239 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "in dense fog, MonoWAD effectively detects both close and fog-obscured objects with the aid of the weather codebook and weather-adaptive diffusion model.", + "bbox": [ + 212, + 266, + 784, + 297 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Results on Real-World Images. We further visualize 3D detection results on real-world images from the Seeing Through Fog dataset [2] under various weather conditions (i.e., foggy, rainy, snowy). In Fig. 6, MonoWAD shows robust detection under diverse weather conditions. This demonstrates that our proposed method maintains weather-robustness even in real-world scenarios by dynamically enhancing the input scenarios.", + "bbox": [ + 212, + 304, + 782, + 396 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "t-SNE Visualization. We conducted t-SNE visualization to analyze feature representations of MonoDTR, MonoDETR, and our MonoWAD on the KITTI and foggy KITTI validation set. As depicted in Fig. 7, the existing methods exhibit distinct feature representations for foggy and clear weather conditions. In contrast, MonoWAD, leveraging weather-robust feature learning from the weather codebook and weather-adaptive diffusion model, demonstrates similar feature representations for both clear and foggy weather conditions.", + "bbox": [ + 212, + 402, + 784, + 508 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "4.7 Limitations", + "text_level": 1, + "bbox": [ + 214, + 522, + 356, + 536 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "The experimental results show the weather-robustness of our method. However, due to the iterative nature of the diffusion model, our method shows $144\\mathrm{ms}/$ image at timestep $T = 15$ (110ms/image at $T = 5$ ), slower than the latest work, MonoDETR (38ms/image). Moreover, our method dynamically enhances the representation of the input feature based on the weather-reference feature and weather difference which needs paired images. Thus, exploring a method to achieve faster processing speeds while maintaining weather-robust performance without paired images could be an interesting direction for our future work.", + "bbox": [ + 212, + 542, + 784, + 662 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 214, + 676, + 359, + 691 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We proposed MonoWAD, a novel weather-robust monocular 3D object detector to handle various weather conditions. Addressing challenges in applying existing monocular 3D object detectors to real-world scenarios with various weather, we design a weather codebook with clear knowledge recalling loss to memorize the knowledge of the clear weather and to generate a weather-reference feature from both clear and foggy features. Also, we design a weather-adaptive diffusion model with weather-adaptive enhancement loss to enhance feature representation according to the weather conditions. As a result, our MonoWAD can detect objects occluded by fog and perform well in clear weather.", + "bbox": [ + 212, + 703, + 785, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Y. Oh et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work was supported by the NRF grant funded by the Korea government (MSIT) (No. RS-2023-00252391), and by IITP grant funded by the Korea government (MSIT) (No. RS-2022-00155911: Artificial Intelligence Convergence Innovation Human Resources Development (Kyung Hee University), IITP-2023-RS-2023-00266615: Convergence Security Core Talent Training Business Support Program, No. 2022-0-00124: Development of Artificial Intelligence Technology for Self-Improving Competency-Aware Learning Capabilities).", + "bbox": [ + 212, + 176, + 787, + 285 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 306, + 321, + 321 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Baranchuk, D., Rubachev, I., Voynov, A., Khrulkov, V., Babenko, A.: Label-efficient semantic segmentation with diffusion models. In: Int. Conf. Learn. Represent. (2021)", + "2. Bijelic, M., Gruber, T., Mannan, F., Kraus, F., Ritter, W., Dietmayer, K., Heide, F.: Seeing through fog without seeing fog: Deep multimodal sensor fusion in unseen adverse weather. In: IEEE Conf. Comput. Vis. Pattern Recog. (2020)", + "3. Brazil, G., Liu, X.: M3d-rpn: Monocular 3d region proposal network for object detection. In: Int. Conf. Comput. Vis. (2019)", + "4. Brazil, G., Pons-Moll, G., Liu, X., Schiele, B.: Kinematic 3d object detection in monocular video. In: Eur. Conf. Comput. Vis. Springer (2020)", + "5. Cabon, Y., Murray, N., Humenberger, M.: Virtual kitti 2 (2020)", + "6. Chen, X., Kundu, K., Zhang, Z., Ma, H., Fidler, S., Urtasun, R.: Monocular 3d object detection for autonomous driving. In: IEEE Conf. Comput. Vis. Pattern Recog. (2016)", + "7. Chen, Y., Tai, L., Sun, K., Li, M.: Monopair: Monocular 3d object detection using pairwise spatial relationships. In: IEEE Conf. Comput. Vis. Pattern Recog. (2020)", + "8. Ding, M., Huo, Y., Yi, H., Wang, Z., Shi, J., Lu, Z., Luo, P.: Learning depth-guided convolutions for monocular 3d object detection. In: IEEE Conf. Comput. Vis. Pattern Recog. Worksh. (2020)", + "9. Esser, P., Rombach, R., Ommer, B.: Taming transformers for high-resolution image synthesis. In: IEEE Conf. Comput. Vis. Pattern Recog. (2021)", + "10. Fu, H., Gong, M., Wang, C., Batmanghelich, K., Tao, D.: Deep ordinal regression network for monocular depth estimation. In: IEEE Conf. Comput. Vis. Pattern Recog. (2018)", + "1. Gaidon, A., Wang, Q., Cabon, Y., Vig, E.: Virtual worlds as proxy for multi-object tracking analysis. In: IEEE Conf. Comput. Vis. Pattern Recog. (2016)", + "2. Geiger, A., Lenz, P., Urtasun, R.: Are we ready for autonomous driving? the kitti vision benchmark suite. In: IEEE Conf. Comput. Vis. Pattern Recog. (2012). https://doi.org/10.1109/CVPR.2012.6248074", + "3. Hahner, M., Sakaridis, C., Dai, D., Van Gool, L.: Fog simulation on real lidar point clouds for 3d object detection in adverse weather. In: Int. Conf. Comput. Vis. (2021)", + "4. Hamilton, B., Tefft, B., Arnold, L., Grabowski, J.: Hidden highways: Fog and traffic crashes on america's roads (november 2014). Montana 40 (2006)", + "5. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. In: Adv. Neural Inform. Process. Syst. vol. 33 (2020)" + ], + "bbox": [ + 225, + 339, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "MonoWAD", + "bbox": [ + 653, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "16. Huang, K.C., Wu, T.H., Su, H.T., Hsu, W.H.: Monodtr: Monocular 3d object detection with depth-aware transformer. In: IEEE Conf. Comput. Vis. Pattern Recog. (2022)", + "17. Juneja, A., Kumar, V., Singla, S.K.: A systematic review on foggy datasets: Applications and challenges. Archives of Computational Methods in Engineering 29(3) (2022)", + "18. Kim, J.U., Kim, H.I., Ro, Y.M.: Stereoscopic vision recalling memory for monocular 3d object detection. IEEE Trans. Image Process. 32 (2023). https://doi.org/10.1109/TIP.2023.3274479", + "19. Kim, J.U., Park, S., Ro, Y.M.: Robust small-scale pedestrian detection with cued recall via memory learning. In: Int. Conf. Comput. Vis. (2021)", + "20. Kim, J.U., Park, S., Ro, Y.M.: Uncertainty-guided cross-modal learning for robust multispectral pedestrian detection. IEEE Trans. Circuit Syst. Video Technol. 32(3) (2021)", + "21. Kim, J.U., Park, S., Ro, Y.M.: Towards versatile pedestrian detector with multisensory-matching and multispectral recalling memory. In: AAAI. vol. 36 (2022)", + "22. Kim, S.W., Brown, B., Yin, K., Kreis, K., Schwarz, K., Li, D., Rombach, R., Torralba, A., Fidler, S.: Neuralfield-ldm: Scene generation with hierarchical latent diffusion models. In: IEEE Conf. Comput. Vis. Pattern Recog. (2023)", + "23. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014)", + "24. Lang, A.H., Vora, S., Caesar, H., Zhou, L., Yang, J., Beijbom, O.: Pointpillars: Fast encoders for object detection from point clouds. In: IEEE Conf. Comput. Vis. Pattern Recog. (2019)", + "25. Liu, X., Xue, N., Wu, T.: Learning auxiliary monocular contexts helps monocular 3d object detection. In: AAAI. vol. 36 (2022)", + "26. Liu, Z., Wu, Z., Toth, R.: Smoke: Single-stage monocular 3d object detection via keypoint estimation. In: IEEE Conf. Comput. Vis. Pattern Recog. Worksh. (2020)", + "27. Liu, Z., Zhou, D., Lu, F., Fang, J., Zhang, L.: Autoshape: Real-time shape-aware monocular 3d object detection. In: Int. Conf. Comput. Vis. (2021)", + "28. Lu, Y., Ma, X., Yang, L., Zhang, T., Liu, Y., Chu, Q., Yan, J., Ouyang, W.: Geometry uncertainty projection network for monocular 3d object detection. In: Int. Conf. Comput. Vis. (2021)", + "29. Lugmayr, A., Danelljan, M., Romero, A., Yu, F., Timofte, R., Van Gool, L.: Repaint: Inpainting using denoising diffusion probabilistic models. In: IEEE Conf. Comput. Vis. Pattern Recog. (2022)", + "30. Luo, J., Li, Y., Pan, Y., Yao, T., Feng, J., Chao, H., Mei, T.: Semantic-conditional diffusion networks for image captioning. In: IEEE Conf. Comput. Vis. Pattern Recog. (2023)", + "31. Luo, S., Dai, H., Shao, L., Ding, Y.: M3dssd: Monocular 3d single stage object detector. In: IEEE Conf. Comput. Vis. Pattern Recog. (2021)", + "32. Ma, X., Liu, S., Xia, Z., Zhang, H., Zeng, X., Ouyang, W.: Rethinking pseudo-lidar representation. In: Eur. Conf. Comput. Vis. Springer (2020)", + "33. Mai, N.A.M., Duthon, P., Khoudour, L., Crouzil, A., Velastin, S.A.: 3d object detection with sls-fusion network in foggy weather conditions. Sensors 21(20) (2021)", + "34. Mousavian, A., Anguelov, D., Flynn, J., Kosecka, J.: 3d bounding box estimation using deep learning and geometry. In: IEEE Conf. Comput. Vis. Pattern Recog. (2017)" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Y. Oh et al.", + "bbox": [ + 271, + 114, + 352, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "35. Ni, H., Shi, C., Li, K., Huang, S.X., Min, M.R.: Conditional image-to-video generation with latent flow diffusion models. In: IEEE Conf. Comput. Vis. Pattern Recog. (2023)", + "36. van den Oord, A., Vinyals, O., Kavukcuoglu, K.: Neural discrete representation learning (2018)", + "37. Peng, L., Wu, X., Yang, Z., Liu, H., Cai, D.: Did-m3d: Decoupling instance depth for monocular 3d object detection. In: Eur. Conf. Comput. Vis. Springer (2022)", + "38. Pfeuffer, A., Dietmayer, K.: Robust semantic segmentation in adverse weather conditions by means of sensor data fusion. In: 2019 22th International Conference on Information Fusion (FUSION) (2019). https://doi.org/10.23919/FUSION43075.2019.9011192", + "39. Qin, Q., Chang, K., Huang, M., Li, G.: Denet: Detection-driven enhancement network for object detection under adverse weather conditions. In: Asian Conf. Comput. Vis. (2022)", + "40. Qin, Z., Li, X.: Monoground: Detecting monocular 3d objects from the ground. In: IEEE Conf. Comput. Vis. Pattern Recog. (2022)", + "41. Reading, C., Harakeh, A., Chae, J., Waslander, S.L.: Categorical depth distribution network for monocular 3d object detection. In: IEEE Conf. Comput. Vis. Pattern Recog. (2021)", + "42. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: IEEE Conf. Comput. Vis. Pattern Recog. (2022)", + "43. Sakaridis, C., Dai, D., Van Gool, L.: Semantic foggy scene understanding with synthetic data. Int. J. Comput. Vis. 126 (2018)", + "44. Shi, S., Guo, C., Jiang, L., Wang, Z., Shi, J., Wang, X., Li, H.: Pv-rcnn: Point-voxel feature set abstraction for 3d object detection. In: IEEE Conf. Comput. Vis. Pattern Recog. (2020)", + "45. Shi, S., Wang, X., Li, H.: Pointcnn: 3d object proposal generation and detection from point cloud. In: IEEE Conf. Comput. Vis. Pattern Recog. (2019)", + "46. Shi, X., Ye, Q., Chen, X., Chen, C., Chen, Z., Kim, T.K.: Geometry-based distance decomposition for monocular 3d object detection. In: Int. Conf. Comput. Vis. (2021)", + "47. Simonelli, A., Bulo, S.R., Porzi, L., Lopez-Antequera, M., Kontschieder, P.: Disentangling monocular 3d object detection. In: Int. Conf. Comput. Vis. (2019)", + "48. Wang, L., Du, L., Ye, X., Fu, Y., Guo, G., Xue, X., Feng, J., Zhang, L.: Depth-conditioned dynamic message propagation for monocular 3d object detection. In: IEEE Conf. Comput. Vis. Pattern Recog. (2021)", + "49. Wu, R.Q., Duan, Z.P., Guo, C.L., Chai, Z., Li, C.: Ridcp: Revitalizing real image dehazing via high-quality codebook priors. In: IEEE Conf. Comput. Vis. Pattern Recog. (2023)", + "50. Yang, S., Scherer, S.: Cubeslam: Monocular 3-d object slam. IEEE Transactions on Robotics 35(4) (2019). https://doi.org/10.1109/TR0.2019.2909168", + "51. Yang, X., Mi, M.B., Yuan, Y., Wang, X., Tan, R.T.: Object detection in foggy scenes by embedding depth and reconstruction into domain adaptation. In: Asian Conf. Comput. Vis. (2022)", + "52. Yu, F., Wang, D., Shelhamer, E., Darrell, T.: Deep layer aggregation. In: IEEE Conf. Comput. Vis. Pattern Recog. (2018)", + "53. Yurtsever, E., Lambert, J., Carballo, A., Takeda, K.: A survey of autonomous driving: Common practices and emerging technologies. IEEE Access 8 (2020). https://doi.org/10.1109/ACCESS.2020.2983149" + ], + "bbox": [ + 212, + 146, + 787, + 840 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "MonoWAD", + "bbox": [ + 653, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "54. Zhang, C., Wang, H., Cai, Y., Chen, L., Li, Y., Sotelo, M.A., Li, Z.: Robust-fusionnet: Deep multimodal sensor fusion for 3-d object detection under severe weather conditions. IEEE Transactions on Instrumentation and Measurement 71 (2022). https://doi.org/10.1109/TIM.2022.3191724", + "55. Zhang, R., Qiu, H., Wang, T., Guo, Z., Cui, Z., Qiao, Y., Li, H., Gao, P.: Monodetr: Depth-guided transformer for monocular 3d object detection. In: Int. Conf. Comput. Vis. (2023)", + "56. Zhou, Y., Tuzel, O.: Voxelnet: End-to-end learning for point cloud based 3d object detection. In: IEEE Conf. Comput. Vis. Pattern Recog. (2018)", + "57. Zhou, Y., He, Y., Zhu, H., Wang, C., Li, H., Jiang, Q.: Monocular 3d object detection: An extrinsic parameter free approach. In: IEEE Conf. Comput. Vis. Pattern Recog. (2021)" + ], + "bbox": [ + 215, + 146, + 785, + 314 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Y. Oh et al.", + "bbox": [ + 271, + 114, + 352, + 126 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/MonoWAD_ Weather-Adaptive Diffusion Model for Robust Monocular 3D Object Detection/4d73977a-6d19-47be-84e6-b3aec99a372e_model.json b/2024/MonoWAD_ Weather-Adaptive Diffusion Model for Robust Monocular 3D Object Detection/4d73977a-6d19-47be-84e6-b3aec99a372e_model.json new file mode 100644 index 0000000000000000000000000000000000000000..b3c303df2245a0615b04fa05e7b2360879721561 --- /dev/null +++ b/2024/MonoWAD_ Weather-Adaptive Diffusion Model for Robust Monocular 3D Object Detection/4d73977a-6d19-47be-84e6-b3aec99a372e_model.json @@ -0,0 +1,2546 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.23, + 0.141, + 0.776, + 0.187 + ], + "angle": 0, + "content": "MonoWAD: Weather-Adaptive Diffusion Model for Robust Monocular 3D Object Detection" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.212, + 0.782, + 0.229 + ], + "angle": 0, + "content": "Youngmin Oh1, Hyung-Il Kim2, Seong Tae Kim1†, and Jung Uk Kim1†" + }, + { + "type": "text", + "bbox": [ + 0.344, + 0.239, + 0.66, + 0.296 + ], + "angle": 0, + "content": "1 Kyung Hee University, Yong-in, South Korea {oym9104, st.kim, ju.kim}@khu.ac.kr 2 ETRI, Daejeon, South Korea hikim@etri.re.kr" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.333, + 0.744, + 0.598 + ], + "angle": 0, + "content": "Abstract. Monocular 3D object detection is an important challenging task in autonomous driving. Existing methods mainly focus on performing 3D detection in ideal weather conditions, characterized by scenarios with clear and optimal visibility. However, the challenge of autonomous driving requires the ability to handle changes in weather conditions, such as foggy weather, not just clear weather. We introduce MonoWAD, a novel weather-robust monocular 3D object detector with a weather-adaptive diffusion model. It contains two components: (1) the weather codebook to memorize the knowledge of the clear weather and generate a weather-reference feature for any input, and (2) the weather-adaptive diffusion model to enhance the feature representation of the input feature by incorporating a weather-reference feature. This serves as an attention role in indicating how much improvement is needed for the input feature according to the weather conditions. To achieve this goal, we introduce a weather-adaptive enhancement loss to enhance the feature representation under both clear and foggy weather conditions. Extensive experiments under various weather conditions demonstrate that MonoWAD achieves weather-robust monocular 3D object detection. The code and dataset are released at https://github.com/VisualAIKHU/MonoWAD." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.61, + 0.741, + 0.639 + ], + "angle": 0, + "content": "Keywords: Monocular 3D Object Detection \\(\\cdot\\) Weather-Adaptive Diffusion \\(\\cdot\\) Weather Codebook" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.664, + 0.377, + 0.68 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.695, + 0.788, + 0.817 + ], + "angle": 0, + "content": "Monocular 3D object detection aims to detect 3D objects using only a single camera [16, 18, 28, 46, 55, 57]. In contrast to LiDAR-based methods that rely on expensive LiDAR sensors for depth estimation [24, 44, 45, 56], and stereo-based methods that require synchronized stereo cameras, monocular 3D object detection only requires monocular images, offering the advantage of computational cost-effectiveness and requiring fewer resources. Due to this characteristic, the monocular 3D object detection is applied to a wide range of real-world applications, such as autonomous driving [12, 19, 53] and robotics [50]." + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.383, + 0.841 + ], + "angle": 0, + "content": "† Corresponding author" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.128 + ], + "angle": 0, + "content": "Y. Oh et al." + }, + { + "type": "image", + "bbox": [ + 0.212, + 0.149, + 0.794, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.256, + 0.788, + 0.327 + ], + "angle": 0, + "content": "Fig. 1: Conceptual diagram of the proposed method (foggy example). (a) In the training phase, weather codebook learns the clear knowledge to transfer it to weather-adaptive diffusion model to enhance content related to the weather conditions. (b) By doing so, even with input images under various weather conditions (e.g., foggy images), monocular 3D object detection becomes adaptable to various weather scenarios." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.342, + 0.789, + 0.478 + ], + "angle": 0, + "content": "However, existing monocular 3D object detectors mainly focus on ideal autonomous driving environments (i.e., clear weather). There are challenges in applying them to real-world scenarios with adverse weather conditions, such as fog and rain. Among these, fog poses the most significant challenge compared to other weather [14, 17]. This is due to the dense and diffuse nature of fog, which strongly scatters and absorbs light, leading to difficulties in object detection [2]. Since monocular 3D object detection relies solely on visual information from a single image, unlike LiDAR, it is crucial to design detectors to achieve enhanced performance in challenging visibility scenarios." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.479, + 0.789, + 0.644 + ], + "angle": 0, + "content": "In this paper, we propose MonoWAD, a novel weather-robust monocular 3D object detector to address the aforementioned issues. As mentioned earlier, due to the inherent challenges posed by foggy weather among various adverse weather conditions, we focus primarily on clear and foggy weather (results for other weather conditions such as rainy and sunset are presented in Section 4). For weather-robust object detection, clear weather requires relatively modest improvements to enhance the visual representation of features. In contrast, significant enhancements in this feature representation are required for foggy weather. To address this, we consider two key aspects: (1) how to quantify the degree of improvement needed for the input image, and (2) how to guide the representation of the input image." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.645, + 0.789, + 0.809 + ], + "angle": 0, + "content": "To address the two key aspects, as shown in Fig. 1, our MonoWAD consists of a weather codebook and weather-adaptive diffusion model. First, we introduce the weather codebook to generate a weather-reference feature that contains knowledge about reference weather in a given scene. The reference weather acts as a guide, indicating the degree of weather improvement required. Since the clear weather contains a richer visual representation of objects, we adopt it as a reference weather. At this time, we devise a clear knowledge recalling (CKR) loss to guide the weather codebook to memorize information about clear weather and generate a weather-reference feature for any input (clear or foggy). As a result, our detector can understand where improvements are needed in the input features based on the weather-reference feature." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Second, we propose a weather-adaptive diffusion model to effectively enhance feature representations in accordance with weather conditions. Given input fea-" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.655, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "MonoWAD" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.312 + ], + "angle": 0, + "content": "ture (clear or foggy), the weather-adaptive diffusion model dynamically enhances the representation of the input feature based on the weather-reference feature. The weather-reference feature plays a role of attention, determining the extent to which the input features need improvement. At this time, we define the difference between clear and foggy weather (i.e., weather changes) as the fog distribution to adopt it as the noise for our diffusion model. With fog distribution, our weather-adaptive diffusion model can enhance the feature representation according to the weather conditions through multiple steps of reverse processes. To achieve this goal, we introduce a weather-adaptive enhancement (WAE) loss. As a result, our MonoWAD performs weather-robust detection by adaptively improving feature representation according to the weather conditions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.314, + 0.785, + 0.433 + ], + "angle": 0, + "content": "To adaptively enhance the feature representation through the difference between weather conditions, we generate a new foggy KITTI dataset based on the KITTI dataset [12]. Comprehensive experimental results on several datasets [11, 12] show that our MonoWAD outperforms the existing state-of-the-art monocular 3D object detectors [16,28,37,40,55] under foggy weather, which is the most challenging weather condition. While our method primarily focuses on foggy, experiments conducted under various weather conditions (e.g., foggy, rainy, and sunset) have demonstrated its applicability to other weather scenarios." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.434, + 0.726, + 0.449 + ], + "angle": 0, + "content": "The main contributions of our paper can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.459, + 0.784, + 0.486 + ], + "angle": 0, + "content": "- We introduce a new weather-robust monocular 3D object detector, called MonoWAD, that is robust to various weather conditions." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.489, + 0.784, + 0.517 + ], + "angle": 0, + "content": "- We design a weather codebook with clear knowledge recalling loss for learning about clear weather, providing reference information for enhancement." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.52, + 0.784, + 0.562 + ], + "angle": 0, + "content": "- We propose weather-adaptive diffusion model with weather-adaptive enhancement loss to dynamically enhance the feature representation of the input images according to the weather conditions." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.459, + 0.784, + 0.562 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.587, + 0.388, + 0.603 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.619, + 0.529, + 0.634 + ], + "angle": 0, + "content": "2.1 Monocular 3D Object Detection" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.785, + 0.809 + ], + "angle": 0, + "content": "Monocular 3D object detection task can be categorized into two directions according to the type of data used in the training phase: (1) using only a monocular image and (2) incorporating additional data, such as depth along with a monocular image. The first category relies on the geometric relationship between 2D and 3D [3,7,25,26,28,31,34,40,46,55,57]. For example, Deep3Dbox [34] utilizes the geometric information of 2D bounding boxes to predict 3D bounding boxes. In [3], M3D-RPN was proposed to understand the 3D scene from the depth-aware convolution. MonoRCNN [46] predicted 3D bounding boxes through geometry-based distance decomposition, and MonoCon [25] learns mono context for 3D object detection. MonoDETR [55] introduces a depth-guided transformer that utilizes geometric depth cues without requiring additional data." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.811, + 0.785, + 0.84 + ], + "angle": 0, + "content": "Moreover, since monocular image contains limited information for estimating 3D object cues, monocular 3D object detectors have adopted the additional" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.128 + ], + "angle": 0, + "content": "Y. Oh et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.253 + ], + "angle": 0, + "content": "data for more robust detection [4,8,16,27,32,37,41,48]. For example, the depth-conditioned dynamic message propagation (DDMP) [48] was proposed to integrate prior depth information with the image context. CaDDN [41] was introduced to utilize the depth distribution of predicted categories for each pixel to project the context information onto 3D space, deriving 3D bounding boxes. MonoDTR [16] employs the transformer architecture to integrate depth features and context, thus estimating more accurate depth information." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.253, + 0.788, + 0.344 + ], + "angle": 0, + "content": "Despite recent progress, the existing monocular 3D object detectors mainly rely on the benchmark data collected under clear weather conditions. However, it is essential to account for challenging adverse weather conditions, such as fog, to more accurately reflect real-world scenarios. In this paper, we aim to introduce weather-robust 3D object detection by enhancing visual features with the proposed weather-adaptive diffusion model and weather codebook." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.356, + 0.616, + 0.372 + ], + "angle": 0, + "content": "2.2 Computer Vision Tasks for Foggy Weather" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.381, + 0.789, + 0.637 + ], + "angle": 0, + "content": "There have been a lot of studies on improving performance in various weather conditions for real-world application of computer vision technology [2, 13, 20, 21, 33, 38, 43, 51, 54]. In particular, fog is considered one of the most critical issues due to its significant degradation of visual information compared to other weather conditions. To deal with the foggy weather conditions, the authors in [43] generated synthetic fog images (so-called Foggy Cityscapes) from clear weather images, and utilized them for training semantic segmentation and 2D object detection. Similar to Foggy Cityscapes, Martin et al. [13] naturally synthesizes fog into LIDAR to enhance the performance of LiDAR-based 3D object detectors in foggy weather. Bijelic et al. [2] use real dataset including foggy weather, for 2D detection with the multimodal fusion networks. Mai et al. [33] synthesize fog for LiDAR and stereo images and perform fusion-based 3D object detection using the SLS-Fusion network. Xin et al. [51] focused on 2D detection in foggy weather by applying domain adaptation. In this context, we focus on monocular 3D object detection in scenarios that rely solely on visual information in challenging foggy environments. To address this challenge, we propose a weather-robust diffusion model that dynamically improves features based on reference feature." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.649, + 0.403, + 0.663 + ], + "angle": 0, + "content": "2.3 Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Recently, the diffusion model [15] has attracted considerable interest in computer vision due to its impressive progress in image generation [22, 29, 35, 42], as well as its potential application to other vision tasks such as segmentation [1] and image captioning [30]. Inspired by the remarkable generative ability, we design the diffusion model for robust monocular 3D object detection by considering a foggy effect (which is one of the challenging adverse weather conditions for monocular 3D object detection) as a form of noise in the model. That is, we propose a method in which visual features obscured by fog are progressively improved by training a diffusion model based on the forward/reverse diffusion process. In particular, we present an adaptive method that allows the diffusion model to control the degree of improvement by weather conditions." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.655, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "MonoWAD" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.147, + 0.784, + 0.269 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.277, + 0.788, + 0.334 + ], + "angle": 0, + "content": "Fig. 2: Overview of our MonoWAD in the inference phase. It mainly contains three parts: weather codebook, weather-adaptive diffusion model, and detection block. Through the weather codebook and weather-adaptive diffusion model, our method can maintain robustness against various weather conditions (i.e., clear or foggy)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.348, + 0.427, + 0.365 + ], + "angle": 0, + "content": "3 Proposed Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.384, + 0.788, + 0.534 + ], + "angle": 0, + "content": "Fig. 2 shows the overall framework of the proposed MonoWAD in the inference phase. A backbone network receives an input image (clear image \\( I_{c} \\) or foggy image \\( I_{f} \\)) to encode the corresponding input feature (input clear feature \\( x^{c} \\) or input foggy feature \\( x^{f} \\)). It interacts with the weather codebook \\( \\mathcal{Z} \\) to generate weather-reference feature \\( x^{r} \\), indicating the amount of enhancement for the given input feature. Subsequently, the weather-adaptive diffusion model attempts to enhance the input feature over \\( T \\) timesteps to obtain an enhanced feature \\( \\tilde{x}^{c} \\) or \\( \\tilde{x}^{f} \\). Finally, monocular 3D object detection is performed through the detection block. Note that, in the training phase, our MonoWAD use the clear feature \\( x^{c} \\) to train our diffusion as well as the detection block." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.536, + 0.788, + 0.596 + ], + "angle": 0, + "content": "We address two key issues: (1) how to guide weather-reference feature \\( x^r \\) to serve as a reference feature, and (2) how to guide the weather-adaptive diffusion model to effectively enhance feature representations based on weather conditions. Details are in the following subsections." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.614, + 0.422, + 0.628 + ], + "angle": 0, + "content": "3.1 Weather Codebook" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.643, + 0.788, + 0.748 + ], + "angle": 0, + "content": "In foggy weather conditions, the overall visual quality of the scene is generally poor, requiring a significant enhancement. Conversely, in clear weather, the amount of improvement is expected to be relatively minimal compared to foggy conditions. Therefore, inspired by [9,36], we devise a weather codebook \\(\\mathcal{Z}\\) to provide the reference knowledge about the weather for appropriate enhancement based on the weather conditions. At this time, as the clear weather contains abundant visual representations, we use it as a reference weather knowledge." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "As shown in Fig. 3, the reference knowledge embedding procedure involves receiving paired clear-foggy features during the training phase. The weather codebook \\(\\mathcal{Z}\\) consists of \\(K\\) learnable slots, denoted as \\(\\mathcal{Z} = \\{z_k\\}_{k=1}^K (z_k \\in \\mathbb{R}^{1 \\times c})\\), where \\(c\\) represents the dimensionality of each slot. The paired clear feature \\(x^c\\) and foggy feature \\(x^f\\) pass through a convolution layer to generate \\(\\hat{x}^c \\in \\mathbb{R}^{h \\times w \\times c}\\) and \\(\\hat{x}^f \\in \\mathbb{R}^{h \\times w \\times c}\\) (\\(w\\) denotes width and \\(h\\) indicates height). Each element of the" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.127 + ], + "angle": 0, + "content": "Y. Oh et al." + }, + { + "type": "image", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.343 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.346, + 0.788, + 0.403 + ], + "angle": 0, + "content": "Fig. 3: Illustration of the proposed (a) weather-invariant guiding (WIG) loss and (b) clear knowledge embedding (CKE) loss. The clear knowledge recalling (CKR) loss, obtained from combining WIG and CKE, aims to memorize the knowledge of the clear weather and recall the same clear knowledge from the foggy weather." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.413, + 0.788, + 0.463 + ], + "angle": 0, + "content": "feature denoted as \\(\\hat{x}_{ij}^{c}\\in \\mathbb{R}^{1\\times c}\\) and \\(\\hat{x}_{ij}^{f}\\in \\mathbb{R}^{1\\times c}\\). Subsequently, we obtain weather-reference feature for clear weather, \\(x^{r(c)}\\in \\mathbb{R}^{h\\times w\\times c}\\), by conducting element-wise quantization process \\(\\mathbf{q}(\\cdot)\\), calculated as:" + }, + { + "type": "equation", + "bbox": [ + 0.359, + 0.473, + 0.787, + 0.507 + ], + "angle": 0, + "content": "\\[\nx ^ {r (c)} = \\mathbf {q} (\\hat {x} ^ {c}) := \\left(\\underset {z _ {k} \\in \\mathcal {Z}} {\\arg \\min } \\| \\hat {x} _ {i j} ^ {c} - z _ {k} \\|\\right). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.52, + 0.788, + 0.612 + ], + "angle": 0, + "content": "Utilizing \\( x^{c} \\) and \\( x^{r(c)} \\), we introduce a clear knowledge embedding (CKE) loss \\( \\mathcal{L}_{cke} \\) to guide \\( x^{r(c)} \\) to follow the representation of \\( x^{c} \\). To this end, we perform global average pooling (GAP) for \\( x^{c} \\) and \\( x^{r(c)} \\) with softmax, generating \\( s^{c} \\) and \\( s^{r(c)} \\), respectively. Each element in the vector indicates the probability of the significance of each channel. With \\( s^{c} \\) and \\( s^{r(c)} \\), we employ KL divergence \\( D_{KL}(\\cdot) \\) for \\( \\mathcal{L}_{cke} \\) to compare the probability distributions, formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.418, + 0.621, + 0.787, + 0.641 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {c k e} = D _ {K L} \\left(s ^ {c} \\| s ^ {r (c)}\\right). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.651, + 0.787, + 0.681 + ], + "angle": 0, + "content": "Through \\(\\mathcal{L}_{cke}\\), the weather codebook \\(\\mathcal{Z}\\) can memorize the knowledge of the clear weather, allowing it to effectively reconstruct the clear weather knowledge." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.682, + 0.788, + 0.756 + ], + "angle": 0, + "content": "Additionally, as the paired clear-foggy images are identical except for weather conditions, the quantization process of the foggy feature \\( x^{f} \\) with the weather codebook should generate an equivalent weather-reference feature. For obtaining weather-reference feature for foggy \\( x^{r(f)} \\), the element-wise quantization process is also conducted for \\( \\hat{x}^{f} \\) and \\( \\mathcal{Z} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.357, + 0.765, + 0.787, + 0.8 + ], + "angle": 0, + "content": "\\[\nx ^ {r (f)} = \\mathbf {q} (\\hat {x} ^ {f}) := \\left(\\underset {z _ {k} \\in \\mathcal {Z}} {\\arg \\min } \\| \\hat {x} _ {i j} ^ {f} - z _ {k} \\|\\right). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Next, we introduce the weather-invariant guiding (WIG) loss \\(\\mathcal{L}_{wig}\\) to guide \\(\\mathcal{Z}\\) that the weather codebook recalls the same clear knowledge of clear feature for" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.655, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "MonoWAD" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.227, + 0.147, + 0.777, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.315, + 0.785, + 0.358 + ], + "angle": 0, + "content": "Fig. 4: Training process of the weather-adaptive diffusion model, which consists of two processes: (1) Adding fog variant \\(\\epsilon_{n}\\) from input clear feature \\(x^{c}\\) (forward process) and (2) enhancing representation with weather-reference feature \\(x^{r}\\) (reverse process)." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.372, + 0.609, + 0.386 + ], + "angle": 0, + "content": "the foggy feature, which can be represented as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.411, + 0.398, + 0.786, + 0.427 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {w i g} = \\left\\| x ^ {r (c)} - x ^ {r (f)} \\right\\| _ {2} ^ {2}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.437, + 0.785, + 0.468 + ], + "angle": 0, + "content": "Finally, the clear knowledge recalling (CKR) loss \\(\\mathcal{L}_{ckr}\\) is obtained by adding \\(\\mathcal{L}_{cke}\\) and \\(\\mathcal{L}_{wig}\\), which is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.428, + 0.48, + 0.785, + 0.496 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {c k r} = \\mathcal {L} _ {c k e} + \\mathcal {L} _ {w i g}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.506, + 0.785, + 0.567 + ], + "angle": 0, + "content": "In the training phase, the weight parameters of embedding \\( K \\) slots of weather codebook \\( \\mathcal{Z} \\) are initialized randomly and they are updated through Eq. (5). In the inference phase, all parameters are fixed to recall clear weather, generating weather-reference features for any weather conditions." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.589, + 0.551, + 0.604 + ], + "angle": 0, + "content": "3.2 Weather-Adaptive Diffusion Model" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.785, + 0.674 + ], + "angle": 0, + "content": "Through Section 3.1, we now know that the weather codebook outputs the weather-reference feature \\( x^{r(c)} \\) for the clear weather and \\( x^{r(f)} \\) for the foggy weather through Eq. (5). From now on, since our method can receive any input images (clear or foggy), we denote the weather-reference feature as \\( x^r \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.675, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Fig. 4 shows the training process of the proposed weather-adaptive diffusion model. The key idea of the diffusion model [15,42] is to gradually enhance \\( x_{T} \\) to \\( x_{0} \\) with fixed Markov Chain of \\( T \\) timesteps. To this end, the forward and reverse processes are conducted in the training phase, and only the reverse process is used for the inference phase. Motivated by [15,42], we construct the weather-adaptive diffusion model to enhance representation related to the weather conditions. To this end, unlike traditional diffusion methods [15,42] that adopt the Gaussian noise to the image or latent space, we adopt \\( \\mathcal{F} = x^{f} - x^{c} \\), called fog distribution, to guide our diffusion model to be aware of the foggy weather. Ideally, the information contained within \\( \\mathcal{F} \\) should include the information of fog, as it represents the difference between the foggy scene and the same scene with clear weather." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.128 + ], + "angle": 0, + "content": "Y. Oh et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.192 + ], + "angle": 0, + "content": "By doing so, our diffusion model learns the variation in weather by repeatedly adding and removing the fog. Note that, as we take clear feature \\( x^{c} \\) for \\( x_{0} \\) to make a reference input for our diffusion model, we newly denote \\( x_{0} \\) as \\( x_{0}^{c} \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.193, + 0.785, + 0.238 + ], + "angle": 0, + "content": "For the forward process at the \\(t\\)-th timestep, \\(q(x_{t}^{c}|x_{t - 1}^{c})\\) takes the previous feature \\(x_{t - 1}^{c}\\) and the noise related to the fog (i.e., \\(\\mathcal{F}\\)) as inputs to generate \\(x_{t}^{c}\\). This procedure is repeated over \\(T\\) timesteps, which can be represented as:" + }, + { + "type": "equation", + "bbox": [ + 0.367, + 0.248, + 0.786, + 0.268 + ], + "angle": 0, + "content": "\\[\nq (x _ {t} ^ {c} | x _ {t - 1} ^ {c}) = \\mathcal {F} (x _ {t} ^ {c}; \\sqrt {1 - \\beta_ {t}} x _ {t - 1} ^ {c}, \\beta_ {t} \\mathbf {I}), \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.387, + 0.279, + 0.786, + 0.318 + ], + "angle": 0, + "content": "\\[\nq \\left(x _ {1} ^ {c}, \\dots , x _ {T} ^ {c} \\mid x ^ {c}\\right) = \\prod_ {t = 1} ^ {T} q \\left(x _ {t} ^ {c} \\mid x _ {t - 1} ^ {c}\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.324, + 0.503, + 0.337 + ], + "angle": 0, + "content": "where \\(\\beta_{t}\\) denotes the variance schedule." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.339, + 0.784, + 0.429 + ], + "angle": 0, + "content": "Next, the reverse process at the \\(t\\)-th timestep aims to estimate fog variant \\(\\epsilon_{n}\\) using \\(x_{t}^{c}\\) to enhance the feature representation of the foggy. To this end, we adopt conditional autoencoder \\(\\epsilon_{\\theta}(x_{t}^{c}, t, x^{r})\\) that receives \\(x_{t}^{c}\\) and reference feature \\(x^{r}\\) from the weather codebook \\(\\mathcal{Z}\\). Specifically, \\(\\epsilon_{\\theta}\\) estimates mean \\(\\mu_{\\theta}\\) and variance \\(\\pmb{\\Sigma}_{\\theta}\\) of the fog distribution at the \\(t\\)-th timestep, denoted as \\(\\tilde{\\mathcal{F}}(\\cdot)\\). The reverse process is also repeated over \\(T\\) timesteps, which can be represented as:" + }, + { + "type": "equation", + "bbox": [ + 0.326, + 0.441, + 0.786, + 0.459 + ], + "angle": 0, + "content": "\\[\np _ {\\theta} \\left(x _ {t - 1} ^ {c} \\mid x _ {t} ^ {c}, x ^ {r}\\right) = \\tilde {\\mathcal {F}} \\left(x _ {t - 1} ^ {c}; \\boldsymbol {\\mu} _ {\\theta} \\left(x _ {t} ^ {c}, t, x ^ {r}\\right), \\boldsymbol {\\Sigma} _ {\\theta} \\left(x _ {t} ^ {c}, t\\right)\\right), \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.357, + 0.47, + 0.786, + 0.51 + ], + "angle": 0, + "content": "\\[\np _ {\\theta} \\left(x ^ {c}, \\dots , x _ {T} ^ {c}\\right) = p \\left(x _ {T} ^ {c}\\right) \\prod_ {t = 1} ^ {T} p _ {\\theta} \\left(x _ {t - 1} ^ {c} \\mid x _ {t} ^ {c}, x ^ {r}\\right), \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.516, + 0.628, + 0.531 + ], + "angle": 0, + "content": "where \\(p_{\\theta}(x_{t - 1}^{c}|x_{t}^{c},x^{r})\\) includes cross-attention layer in \\(\\epsilon_{\\theta}\\)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.531, + 0.784, + 0.59 + ], + "angle": 0, + "content": "The cross-attention layer receives the flattened clear feature \\( x_{t}^{c} \\) and weather-reference feature \\( x^{r} \\), i.e., \\( \\bar{x}_{t}^{c} \\) and \\( \\bar{x}^{r} \\), respectively. The similarity between \\( \\bar{x}^{r} \\) and \\( \\bar{x}_{t}^{c} \\) is calculated in the cross-attention layer to transfer the enhancement to \\( \\bar{x}_{t}^{c} \\), which can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.353, + 0.599, + 0.786, + 0.634 + ], + "angle": 0, + "content": "\\[\n\\operatorname {A t t e n t i o n} (Q, K, V) = \\operatorname {s o f t m a x} \\left(\\frac {Q K ^ {T}}{\\sqrt {d}}\\right) V, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.643, + 0.784, + 0.674 + ], + "angle": 0, + "content": "where \\( Q = W_{i}^{q} \\cdot \\bar{x}_{t}^{c}, K = W_{i}^{k} \\cdot \\bar{x}^{r}, V = W_{i}^{v} \\cdot \\bar{x}^{r} \\) with learnable parameters \\( W_{i}^{q}, W_{i}^{k}, W_{i}^{v} \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.675, + 0.784, + 0.72 + ], + "angle": 0, + "content": "To ensure that the estimated fog variant \\(\\epsilon_{\\theta}\\) is similar to the fog variant \\(\\epsilon_{n}\\) applied to make \\(x_{t}^{c}\\) from the input clear feature \\(x^{c}\\), we propose a weather-adaptive enhancement loss \\(\\mathcal{L}_{wae}\\), which is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.357, + 0.731, + 0.786, + 0.756 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {w a e} = \\mathbb {E} _ {x ^ {c}, \\boldsymbol {\\epsilon} _ {n} \\sim \\mathcal {F}, t} \\left[ \\| \\boldsymbol {\\epsilon} _ {n} - \\boldsymbol {\\epsilon} _ {\\theta} \\left(x _ {t} ^ {c}, t, x ^ {r}\\right) \\| _ {2} ^ {2} \\right]. \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.786, + 0.84 + ], + "angle": 0, + "content": "With \\(\\mathcal{L}_{wae}\\), \\(\\epsilon_{\\theta}\\) can estimate the fog variant by leveraging the fog distribution as noise for our diffusion model through multiple forward/reverse processes. Additionally, the cross-attention layer within \\(\\epsilon_{\\theta}\\) dynamically enhances the feature representation based on combining knowledge of the input feature (foggy or clear) and the weather-reference feature. Since our diffusion model learns the" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.654, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "MonoWAD" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.207 + ], + "angle": 0, + "content": "degree of improvement needed corresponding to weather conditions, it can improve the representation of any input, whether clear or foggy in inference phase. This leads our monocular 3D object detector can handle both clear and foggy input images, resulting in weather-robust detection." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.219, + 0.427, + 0.233 + ], + "angle": 0, + "content": "3.3 Total Loss Function" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.244, + 0.696, + 0.259 + ], + "angle": 0, + "content": "The total loss function of our MonoWAD is represented as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.381, + 0.273, + 0.785, + 0.288 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {T o t a l}} = \\mathcal {L} _ {O D} + \\lambda_ {1} \\mathcal {L} _ {\\text {c k r}} + \\lambda_ {2} \\mathcal {L} _ {\\text {w a e}}, \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.3, + 0.784, + 0.36 + ], + "angle": 0, + "content": "where \\(\\lambda_{1}\\) and \\(\\lambda_{2}\\) denote balancing hyper-parameters. In our experiments, we set \\(\\lambda_{1} = \\lambda_{2} = 1\\). \\(\\mathcal{L}_{OD}\\) is the detection loss for 3D object detection [16,55]. It includes loss functions of classification, regression, and depth loss, that are similar to prior works [8,16,31]. The overall weight parameters are updated through \\(\\mathcal{L}_{Total}\\)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.383, + 0.376, + 0.401 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.415, + 0.527, + 0.43 + ], + "angle": 0, + "content": "4.1 Dataset and Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.44, + 0.785, + 0.59 + ], + "angle": 0, + "content": "Datasets. We utilize the KITTI 3D object detection dataset [12], the most widely adopted for 3D object detection. It contains 7,481 training images and 7,518 test images under clear weather conditions. Due to unavailable ground-truth annotations for test images and limited evaluation on the test server, we follow [6] by splitting 3,712 training images and 3,769 validation images. In addition, as our work requires paired images to learn about weather changes, we generate foggy images from all images in the KITTI dataset, called foggy KITTI, that emulate the foggy scene. Following the protocols of [2,43], we generate foggy images based on object distances using depth maps estimated by DORN [10]. Please refer to the supplementary materials for details." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.591, + 0.785, + 0.667 + ], + "angle": 0, + "content": "In addition, as our work is focused on robust monocular 3D object detection in various weather conditions, we further adopt the Virtual KITTI dataset [5,11], which contains photo-realistic synthetic images under various weather conditions (e.g., foggy, rainy, sunset). It is associated with the original real-world KITTI dataset, which has 3D annotations for each weather." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.674, + 0.784, + 0.75 + ], + "angle": 0, + "content": "Evaluation Metrics. We adopt average precision in both 3D detection \\((\\mathrm{AP}_{3D})\\) and bird-eye view detection \\((\\mathrm{AP}_{BEV})\\) under three difficulty levels ('Easy', 'Moderate', 'Hard') according to size, occlusion, and truncation. Following [47], we use 40 recall position metric \\(\\mathrm{AP}_{40}\\) and report scores for the car category under the IoU threshold 0.7 for the KITTI dataset and 0.5 for the Virtual KITTI dataset." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.77, + 0.457, + 0.785 + ], + "angle": 0, + "content": "4.2 Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.784, + 0.841 + ], + "angle": 0, + "content": "We use DLA-102 [52] as our backbone and adopt the transformer architecture of [16] for the detection block. We train MonoWAD on a single RTX 4090 GPU with a batch size of 4 over 120 epochs using Adam optimizer [23] (initial learning" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.128 + ], + "angle": 0, + "content": "Y. Oh et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.201 + ], + "angle": 0, + "content": "Table 1: Detection results of car category on KITTI validation set under foggy weather and clear weather conditions. The results of the state-of-the-art methods under foggy weather are obtained through our reproduction with the official source code. Bold/underlined fonts indicate the best/second-best results." + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.204, + 0.784, + 0.307 + ], + "angle": 0, + "content": "
MethodFoggy (AP3D)Clear (AP3D)Average
EasyMod.HardEasyMod.HardEasyMod.Hard
GUPNet [28] (ICCV'21)2.742.192.1622.7616.4613.7212.759.337.94
DID-M3D [37] (ECCV'22)1.150.610.6422.9816.1214.0312.078.377.34
MonoGround [40] (CVPR'22)0.000.000.0625.2418.6915.5812.629.357.82
MonoDTR [16] (CVPR'22)16.8911.869.8724.5218.5715.5120.7115.2212.69
MonoDETR [55] (ICCV'23)7.405.744.5328.8420.6116.3818.1213.1810.46
MonoWAD (Ours)27.1719.5716.2129.1021.0817.7328.1420.3316.97
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.319, + 0.784, + 0.347 + ], + "angle": 0, + "content": "Table 2: Detection results of car category on KITTI test set under foggy weather. Bold/underlined fonts indicate the best/second-best results." + }, + { + "type": "table", + "bbox": [ + 0.261, + 0.351, + 0.738, + 0.457 + ], + "angle": 0, + "content": "
MethodAP3DAPBEV
EasyMod.HardEasyMod.Hard
GUPNet [28] (ICCV'21)3.012.421.134.903.022.91
DID-M3D [37] (ECCV'22)3.102.392.195.343.012.91
MonoGround [40] (CVPR'22)0.140.200.220.230.380.39
MonoDTR [16] (CVPR'22)11.077.415.2615.7610.157.53
MonoDETR [55] (ICCV'23)9.335.544.0613.158.066.30
MonoWAD (Ours)19.7513.3211.0427.9519.0615.61
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.47, + 0.788, + 0.577 + ], + "angle": 0, + "content": "rate of \\(10^{-4}\\)). For the weather codebook, we utilize embedding slot \\(K = 4096\\) and set the dimension of each slot \\(D = 256\\). We set 4 heads for cross-attention of the weather-adaptive diffusion model, and the timesteps for the forward and reverse process of diffusion default to 15. The number of channels for the diffusion output features is \\(C = 256\\). As recent monocular 3D object detectors [16,28,37,40,55] have not been explored under adverse weather conditions, we implemented them with available official source code to faithfully reproduce them." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.59, + 0.363, + 0.606 + ], + "angle": 0, + "content": "4.3 Comparison" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.613, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Results on KITTI 3D Dataset. We compared MonoWAD with state-of-the-art monocular 3D object detectors [16,28,37,40,55] that do not use additional data (e.g., depth maps or LiDAR) during inference on the KITTI and foggy KITTI validation sets. Table 1 shows the \\(AP_{3D}\\) results, and \\(AP_{BEV}\\) results are in the supplementary materials. While recent methods have shown improved performance in clear weather, their performance drops in foggy weather, limiting their applicability in real-world applications (e.g., autonomous driving and robotics). In contrast, MonoWAD showed stable 3D detection performance under both foggy and clear weather. Since our weather codebook learns knowledge about clear weather and the weather-adaptive diffusion model can enhance the feature representation of the input images under both clear and foggy weather, MonoWAD shows a more weather-robust detection performance than that of the existing methods. Also, to explore the weather robustness of MonoWAD, we conducted experiments with mixed foggy and clear weather conditions at various ratios (please see the supplementary materials)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.655, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "MonoWAD" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.145, + 0.784, + 0.173 + ], + "angle": 0, + "content": "Table 3: Detection results of car category on Virtual KITTI under foggy, rainy, sunset conditions. Bold/underlined fonts indicate the best/second-best results." + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.177, + 0.778, + 0.281 + ], + "angle": 0, + "content": "
MethodFoggy (AP3D)Rainy (AP3D)Sunset (AP3D)
EasyMod.HardEasyMod.HardEasyMod.Hard
GUPNet [28] (ICCV'21)1.761.571.572.341.241.212.771.641.65
DID-M3D [37] (ECCV'22)0.910.390.390.400.130.130.340.100.10
MonoGround [40] (CVPR'22)0.290.300.255.492.822.777.684.244.20
MonoDTR [16] (CVPR'22)8.795.755.7211.736.256.749.865.425.42
MonoDETR [55] (ICCV'23)4.502.992.966.613.463.427.084.174.16
MonoWAD (Ours)13.338.568.5014.128.338.2413.387.897.80
" + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.286, + 0.785, + 0.313 + ], + "angle": 0, + "content": "Table 4: Effect of the proposed method on KITTI validation set for car category. WC denotes our weather codebook, WAD indicates our weather-adaptive diffusion model." + }, + { + "type": "table", + "bbox": [ + 0.28, + 0.314, + 0.718, + 0.386 + ], + "angle": 0, + "content": "
MethodWCWADFoggy (AP3D)Clear (AP3D)
EasyMod.HardEasyMod.Hard
Baseline--13.759.618.1022.6317.1614.28
MonoWAD (Ours)-✓25.6218.6615.5626.3419.1716.15
✓✓27.1719.5716.2129.1021.0817.73
" + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.393, + 0.784, + 0.407 + ], + "angle": 0, + "content": "Table 5: Detection results on KITTI validation set by changing diffusion timestep \\( T \\) ." + }, + { + "type": "table", + "bbox": [ + 0.276, + 0.408, + 0.723, + 0.503 + ], + "angle": 0, + "content": "
Timestep TFoggy (AP3D)Clear (AP3D)
EasyMod.HardEasyMod.Hard
-13.759.618.1022.6317.1614.28
523.5717.9114.9626.0319.2116.01
1025.2818.4915.4226.7919.9016.78
1527.1719.5716.2129.1021.0817.73
2024.5418.2915.1024.8518.5415.34
" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.514, + 0.784, + 0.575 + ], + "angle": 0, + "content": "We further compared 3D detection performances on the foggy KITTI test set (Table 2). Similar to Table 1, existing monocular 3D object detectors show lower performance under foggy weather. Through the results, our method shows the weather-robust monocular 3D performance under foggy and clear weather." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.583, + 0.785, + 0.689 + ], + "angle": 0, + "content": "Results on Virtual KITTI Dataset. We also conducted experiments on the Virtual KITTI dataset to validate the generalization ability of our method across various weather conditions (i.e., foggy, rainy, sunset). The results are shown in Table 3. MonoWAD also shows the highest performance in foggy weather. Moreover, it outperforms the existing method in the rainy and sunset conditions. The results demonstrate that MonoWAD is robust and insensitive to various weather conditions that can be faced in real-world scenarios, not just clear weather." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.708, + 0.4, + 0.723 + ], + "angle": 0, + "content": "4.4 Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.73, + 0.784, + 0.79 + ], + "angle": 0, + "content": "We conducted ablation studies to examine (1) the effect of each proposed component (i.e., weather codebook and weather-adaptive diffusion model) and (2) the effect of the weather-adaptive diffusion model by varying the timestep \\( T \\). These experiments were performed on the KITTI and foggy KITTI 3D validation sets." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.81, + 0.785, + 0.84 + ], + "angle": 0, + "content": "Effect of the Proposed Modules. The results regarding the effectiveness of the proposed weather codebook and weather-adaptive diffusion model are pre-" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.128 + ], + "angle": 0, + "content": "Y. Oh et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.187 + ], + "angle": 0, + "content": "Table 6: Performance comparison on KITTI validation set under foggy and clear weather conditions. We compare our MonoWAD with MonoDTR [16], which demonstrates superior performance in foggy weather using state-of-the-art dehazing methods." + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.189, + 0.779, + 0.283 + ], + "angle": 0, + "content": "
MethodDehazing MethodFoggy (AP3D)Clear (AP3D)Average
EasyMod.HardEasyMod.HardEasyMod.Hard
MonoDTR(B) [16] (CVPR'22)-16.8911.869.8724.5218.5715.5120.7115.2212.69
B + RIDCP [49] (CVPR'23)Image-level17.2312.4110.4424.0217.8914.7820.6315.1512.61
B + DENet [39] (ACCV'22)Feature-level22.3517.4414.477.105.704.5314.7311.579.50
B + Yang et al. [51] (ACCV'22)Feature-level22.8715.2112.1717.9613.1010.6420.4214.1611.41
MonoWAD (Ours)-27.1719.5716.2129.1021.0817.7328.1420.3316.97
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.295, + 0.789, + 0.338 + ], + "angle": 0, + "content": "Table 7: Comparison of diffusion models on KITTI validation set for car category: \\(\\mathcal{B}\\) is baseline detection block (transformer encoder-decoder), CDM (Conditional Diffusion Model [42]), WC (weather codebook), and WAD (weather adaptive diffusion model)." + }, + { + "type": "table", + "bbox": [ + 0.276, + 0.341, + 0.724, + 0.451 + ], + "angle": 0, + "content": "
MethodFoggy (AP3D)Clear (AP3D)
EasyMod.HardEasyMod.Hard
B+DDPM [15]5.323.842.7718.3112.7110.20
B+CDM [42]2.742.102.0020.5014.5111.63
B+WC+CDM [42]17.5112.7410.4021.0515.1112.54
B+WAD25.6218.6615.5626.3419.1716.15
MonoWAD(B+WC+WAD)27.1719.5716.2129.1021.0817.73
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.466, + 0.788, + 0.557 + ], + "angle": 0, + "content": "sented in Table 4. Since our weather-adaptive diffusion model is designed to enhance feature representation, it alone has shown significant performance improvement. When the weather codebook is additionally considered, the weather-adaptive diffusion model can leverage the knowledge of the weather-reference feature for clear weather, leading to enhanced performance. It makes MonoWAD can be robust to various weather conditions for monocular 3D object detection." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.567, + 0.788, + 0.659 + ], + "angle": 0, + "content": "Effect of Timestep \\(T\\). We also conduct experiments by varying timestep \\(T\\) of the weather-adaptive diffusion model. Timestep is the number of steps for the forward and reverse process of the diffusion model. Table 5 indicates that the highest monocular 3D detection performance is achieved when \\(T = 15\\) under both clear and foggy weather conditions. Our MonoWAD consistently outperforms the baseline and other existing methods across all timesteps." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.673, + 0.357, + 0.687 + ], + "angle": 0, + "content": "4.5 Discussions" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.697, + 0.794, + 0.851 + ], + "angle": 0, + "content": "Comparison with Dehazing Methods. We investigate the weather-robustness of our method for monocular 3D object detection compared to the other monocular detectors using dehazing methods. To this end, we compared MonoWAD with MonoDTR [16], which shows the best performance in foggy weather through the application of state-of-the-art image-level and feature-level dehazing methods [39,49,51]. The results are shown in Table 6. Recent dehazing methods are primarily focused on specific weather conditions, such as foggy weather. While they have shown some improvement under foggy conditions, they exhibit reduced performances under clear weather conditions. In contrast, our MonoWAD shows robust performance across both foggy and clear weather conditions." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.655, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "MonoWAD" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.224, + 0.147, + 0.782, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.293, + 0.785, + 0.335 + ], + "angle": 0, + "content": "Fig. 5: Comparison of 3D detection examples (green: ground-truth, red: predicted 3D bounding-box) between our MonoWAD and two detectors, MonoDTR [16] and MonoDETR [55], that show the most improved performances among existing methods." + }, + { + "type": "image", + "bbox": [ + 0.214, + 0.346, + 0.79, + 0.595 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.236, + 0.597, + 0.766, + 0.612 + ], + "angle": 0, + "content": "Fig. 6: 3D detection results on real-world images of various weather conditions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.623, + 0.788, + 0.715 + ], + "angle": 0, + "content": "Effect of Weather-Adaptive Diffusion Model. Table 7 shows the effectiveness of MonoWAD with existing diffusion models [15,42] on the KITTI and foggy KITTI validation sets. Existing methods adopt Gaussian noise for forward and reverse processes, but they can not fully understand about weather. In contrast, our weather-adaptive diffusion understands weather variances, allowing our MonoWAD to surpass existing methods in clear and foggy weather." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.738, + 0.436, + 0.753 + ], + "angle": 0, + "content": "4.6 Visualization Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Results on KITTI Dataset. We visualize several 3D detection results on the KITTI 3D dataset, comparing MonoWAD with MonoDTR and MonoDETR, which exhibit the highest performances among existing methods under foggy condition (Fig. 5). Existing methods struggle to detect objects obscured by fog, indicating limitations in detecting only fully visible objects. In contrast, even" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.127 + ], + "angle": 0, + "content": "Y. Oh et al." + }, + { + "type": "image", + "bbox": [ + 0.252, + 0.144, + 0.402, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.427, + 0.144, + 0.576, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.604, + 0.144, + 0.753, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.295, + 0.242, + 0.706, + 0.256 + ], + "angle": 0, + "content": "Fig. 7: t-SNE visualization results (Red: Clear, Blue: Foggy)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.267, + 0.785, + 0.299 + ], + "angle": 0, + "content": "in dense fog, MonoWAD effectively detects both close and fog-obscured objects with the aid of the weather codebook and weather-adaptive diffusion model." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.305, + 0.784, + 0.397 + ], + "angle": 0, + "content": "Results on Real-World Images. We further visualize 3D detection results on real-world images from the Seeing Through Fog dataset [2] under various weather conditions (i.e., foggy, rainy, snowy). In Fig. 6, MonoWAD shows robust detection under diverse weather conditions. This demonstrates that our proposed method maintains weather-robustness even in real-world scenarios by dynamically enhancing the input scenarios." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.403, + 0.785, + 0.509 + ], + "angle": 0, + "content": "t-SNE Visualization. We conducted t-SNE visualization to analyze feature representations of MonoDTR, MonoDETR, and our MonoWAD on the KITTI and foggy KITTI validation set. As depicted in Fig. 7, the existing methods exhibit distinct feature representations for foggy and clear weather conditions. In contrast, MonoWAD, leveraging weather-robust feature learning from the weather codebook and weather-adaptive diffusion model, demonstrates similar feature representations for both clear and foggy weather conditions." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.523, + 0.357, + 0.537 + ], + "angle": 0, + "content": "4.7 Limitations" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.544, + 0.785, + 0.664 + ], + "angle": 0, + "content": "The experimental results show the weather-robustness of our method. However, due to the iterative nature of the diffusion model, our method shows \\(144\\mathrm{ms}/\\) image at timestep \\(T = 15\\) (110ms/image at \\(T = 5\\)), slower than the latest work, MonoDETR (38ms/image). Moreover, our method dynamically enhances the representation of the input feature based on the weather-reference feature and weather difference which needs paired images. Thus, exploring a method to achieve faster processing speeds while maintaining weather-robust performance without paired images could be an interesting direction for our future work." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.677, + 0.36, + 0.693 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.787, + 0.84 + ], + "angle": 0, + "content": "We proposed MonoWAD, a novel weather-robust monocular 3D object detector to handle various weather conditions. Addressing challenges in applying existing monocular 3D object detectors to real-world scenarios with various weather, we design a weather codebook with clear knowledge recalling loss to memorize the knowledge of the clear weather and to generate a weather-reference feature from both clear and foggy features. Also, we design a weather-adaptive diffusion model with weather-adaptive enhancement loss to enhance feature representation according to the weather conditions. As a result, our MonoWAD can detect objects occluded by fog and perform well in clear weather." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.655, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "MonoWAD" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.788, + 0.286 + ], + "angle": 0, + "content": "This work was supported by the NRF grant funded by the Korea government (MSIT) (No. RS-2023-00252391), and by IITP grant funded by the Korea government (MSIT) (No. RS-2022-00155911: Artificial Intelligence Convergence Innovation Human Resources Development (Kyung Hee University), IITP-2023-RS-2023-00266615: Convergence Security Core Talent Training Business Support Program, No. 2022-0-00124: Development of Artificial Intelligence Technology for Self-Improving Competency-Aware Learning Capabilities)." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.308, + 0.323, + 0.323 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.34, + 0.785, + 0.38 + ], + "angle": 0, + "content": "1. Baranchuk, D., Rubachev, I., Voynov, A., Khrulkov, V., Babenko, A.: Label-efficient semantic segmentation with diffusion models. In: Int. Conf. Learn. Represent. (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.382, + 0.785, + 0.422 + ], + "angle": 0, + "content": "2. Bijelic, M., Gruber, T., Mannan, F., Kraus, F., Ritter, W., Dietmayer, K., Heide, F.: Seeing through fog without seeing fog: Deep multimodal sensor fusion in unseen adverse weather. In: IEEE Conf. Comput. Vis. Pattern Recog. (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.424, + 0.785, + 0.45 + ], + "angle": 0, + "content": "3. Brazil, G., Liu, X.: M3d-rpn: Monocular 3d region proposal network for object detection. In: Int. Conf. Comput. Vis. (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.452, + 0.785, + 0.478 + ], + "angle": 0, + "content": "4. Brazil, G., Pons-Moll, G., Liu, X., Schiele, B.: Kinematic 3d object detection in monocular video. In: Eur. Conf. Comput. Vis. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.48, + 0.671, + 0.492 + ], + "angle": 0, + "content": "5. Cabon, Y., Murray, N., Humenberger, M.: Virtual kitti 2 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.494, + 0.785, + 0.533 + ], + "angle": 0, + "content": "6. Chen, X., Kundu, K., Zhang, Z., Ma, H., Fidler, S., Urtasun, R.: Monocular 3d object detection for autonomous driving. In: IEEE Conf. Comput. Vis. Pattern Recog. (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.535, + 0.785, + 0.562 + ], + "angle": 0, + "content": "7. Chen, Y., Tai, L., Sun, K., Li, M.: Monopair: Monocular 3d object detection using pairwise spatial relationships. In: IEEE Conf. Comput. Vis. Pattern Recog. (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.564, + 0.785, + 0.603 + ], + "angle": 0, + "content": "8. Ding, M., Huo, Y., Yi, H., Wang, Z., Shi, J., Lu, Z., Luo, P.: Learning depth-guided convolutions for monocular 3d object detection. In: IEEE Conf. Comput. Vis. Pattern Recog. Worksh. (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.605, + 0.785, + 0.631 + ], + "angle": 0, + "content": "9. Esser, P., Rombach, R., Ommer, B.: Taming transformers for high-resolution image synthesis. In: IEEE Conf. Comput. Vis. Pattern Recog. (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.633, + 0.785, + 0.673 + ], + "angle": 0, + "content": "10. Fu, H., Gong, M., Wang, C., Batmanghelich, K., Tao, D.: Deep ordinal regression network for monocular depth estimation. In: IEEE Conf. Comput. Vis. Pattern Recog. (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.675, + 0.785, + 0.701 + ], + "angle": 0, + "content": "1. Gaidon, A., Wang, Q., Cabon, Y., Vig, E.: Virtual worlds as proxy for multi-object tracking analysis. In: IEEE Conf. Comput. Vis. Pattern Recog. (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.703, + 0.785, + 0.741 + ], + "angle": 0, + "content": "2. Geiger, A., Lenz, P., Urtasun, R.: Are we ready for autonomous driving? the kitti vision benchmark suite. In: IEEE Conf. Comput. Vis. Pattern Recog. (2012). https://doi.org/10.1109/CVPR.2012.6248074" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.743, + 0.785, + 0.784 + ], + "angle": 0, + "content": "3. Hahner, M., Sakaridis, C., Dai, D., Van Gool, L.: Fog simulation on real lidar point clouds for 3d object detection in adverse weather. In: Int. Conf. Comput. Vis. (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.786, + 0.785, + 0.812 + ], + "angle": 0, + "content": "4. Hamilton, B., Tefft, B., Arnold, L., Grabowski, J.: Hidden highways: Fog and traffic crashes on america's roads (november 2014). Montana 40 (2006)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "5. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. In: Adv. Neural Inform. Process. Syst. vol. 33 (2020)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.34, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.353, + 0.127 + ], + "angle": 0, + "content": "Y. Oh et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "16. Huang, K.C., Wu, T.H., Su, H.T., Hsu, W.H.: Monodtr: Monocular 3d object detection with depth-aware transformer. In: IEEE Conf. Comput. Vis. Pattern Recog. (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.785, + 0.232 + ], + "angle": 0, + "content": "17. Juneja, A., Kumar, V., Singla, S.K.: A systematic review on foggy datasets: Applications and challenges. Archives of Computational Methods in Engineering 29(3) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.233, + 0.785, + 0.273 + ], + "angle": 0, + "content": "18. Kim, J.U., Kim, H.I., Ro, Y.M.: Stereoscopic vision recalling memory for monocular 3d object detection. IEEE Trans. Image Process. 32 (2023). https://doi.org/10.1109/TIP.2023.3274479" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.274, + 0.785, + 0.303 + ], + "angle": 0, + "content": "19. Kim, J.U., Park, S., Ro, Y.M.: Robust small-scale pedestrian detection with cued recall via memory learning. In: Int. Conf. Comput. Vis. (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.304, + 0.785, + 0.345 + ], + "angle": 0, + "content": "20. Kim, J.U., Park, S., Ro, Y.M.: Uncertainty-guided cross-modal learning for robust multispectral pedestrian detection. IEEE Trans. Circuit Syst. Video Technol. 32(3) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.346, + 0.785, + 0.387 + ], + "angle": 0, + "content": "21. Kim, J.U., Park, S., Ro, Y.M.: Towards versatile pedestrian detector with multisensory-matching and multispectral recalling memory. In: AAAI. vol. 36 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.388, + 0.785, + 0.43 + ], + "angle": 0, + "content": "22. Kim, S.W., Brown, B., Yin, K., Kreis, K., Schwarz, K., Li, D., Rombach, R., Torralba, A., Fidler, S.: Neuralfield-ldm: Scene generation with hierarchical latent diffusion models. In: IEEE Conf. Comput. Vis. Pattern Recog. (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.431, + 0.785, + 0.458 + ], + "angle": 0, + "content": "23. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.459, + 0.785, + 0.5 + ], + "angle": 0, + "content": "24. Lang, A.H., Vora, S., Caesar, H., Zhou, L., Yang, J., Beijbom, O.: Pointpillars: Fast encoders for object detection from point clouds. In: IEEE Conf. Comput. Vis. Pattern Recog. (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.501, + 0.785, + 0.529 + ], + "angle": 0, + "content": "25. Liu, X., Xue, N., Wu, T.: Learning auxiliary monocular contexts helps monocular 3d object detection. In: AAAI. vol. 36 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.53, + 0.785, + 0.558 + ], + "angle": 0, + "content": "26. Liu, Z., Wu, Z., Toth, R.: Smoke: Single-stage monocular 3d object detection via keypoint estimation. In: IEEE Conf. Comput. Vis. Pattern Recog. Worksh. (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.559, + 0.785, + 0.586 + ], + "angle": 0, + "content": "27. Liu, Z., Zhou, D., Lu, F., Fang, J., Zhang, L.: Autoshape: Real-time shape-aware monocular 3d object detection. In: Int. Conf. Comput. Vis. (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.587, + 0.785, + 0.628 + ], + "angle": 0, + "content": "28. Lu, Y., Ma, X., Yang, L., Zhang, T., Liu, Y., Chu, Q., Yan, J., Ouyang, W.: Geometry uncertainty projection network for monocular 3d object detection. In: Int. Conf. Comput. Vis. (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.629, + 0.785, + 0.67 + ], + "angle": 0, + "content": "29. Lugmayr, A., Danelljan, M., Romero, A., Yu, F., Timofte, R., Van Gool, L.: Repaint: Inpainting using denoising diffusion probabilistic models. In: IEEE Conf. Comput. Vis. Pattern Recog. (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.671, + 0.785, + 0.712 + ], + "angle": 0, + "content": "30. Luo, J., Li, Y., Pan, Y., Yao, T., Feng, J., Chao, H., Mei, T.: Semantic-conditional diffusion networks for image captioning. In: IEEE Conf. Comput. Vis. Pattern Recog. (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.713, + 0.785, + 0.741 + ], + "angle": 0, + "content": "31. Luo, S., Dai, H., Shao, L., Ding, Y.: M3dssd: Monocular 3d single stage object detector. In: IEEE Conf. Comput. Vis. Pattern Recog. (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.742, + 0.785, + 0.77 + ], + "angle": 0, + "content": "32. Ma, X., Liu, S., Xia, Z., Zhang, H., Zeng, X., Ouyang, W.: Rethinking pseudo-lidar representation. In: Eur. Conf. Comput. Vis. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.771, + 0.785, + 0.798 + ], + "angle": 0, + "content": "33. Mai, N.A.M., Duthon, P., Khoudour, L., Crouzil, A., Velastin, S.A.: 3d object detection with sls-fusion network in foggy weather conditions. Sensors 21(20) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "34. Mousavian, A., Anguelov, D., Flynn, J., Kosecka, J.: 3d bounding box estimation using deep learning and geometry. In: IEEE Conf. Comput. Vis. Pattern Recog. (2017)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.655, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "MonoWAD" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "35. Ni, H., Shi, C., Li, K., Huang, S.X., Min, M.R.: Conditional image-to-video generation with latent flow diffusion models. In: IEEE Conf. Comput. Vis. Pattern Recog. (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.787, + 0.218 + ], + "angle": 0, + "content": "36. van den Oord, A., Vinyals, O., Kavukcuoglu, K.: Neural discrete representation learning (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.218, + 0.787, + 0.246 + ], + "angle": 0, + "content": "37. Peng, L., Wu, X., Yang, Z., Liu, H., Cai, D.: Did-m3d: Decoupling instance depth for monocular 3d object detection. In: Eur. Conf. Comput. Vis. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.246, + 0.787, + 0.299 + ], + "angle": 0, + "content": "38. Pfeuffer, A., Dietmayer, K.: Robust semantic segmentation in adverse weather conditions by means of sensor data fusion. In: 2019 22th International Conference on Information Fusion (FUSION) (2019). https://doi.org/10.23919/FUSION43075.2019.9011192" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.3, + 0.787, + 0.342 + ], + "angle": 0, + "content": "39. Qin, Q., Chang, K., Huang, M., Li, G.: Denet: Detection-driven enhancement network for object detection under adverse weather conditions. In: Asian Conf. Comput. Vis. (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.342, + 0.787, + 0.37 + ], + "angle": 0, + "content": "40. Qin, Z., Li, X.: Monoground: Detecting monocular 3d objects from the ground. In: IEEE Conf. Comput. Vis. Pattern Recog. (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.371, + 0.787, + 0.412 + ], + "angle": 0, + "content": "41. Reading, C., Harakeh, A., Chae, J., Waslander, S.L.: Categorical depth distribution network for monocular 3d object detection. In: IEEE Conf. Comput. Vis. Pattern Recog. (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.412, + 0.787, + 0.453 + ], + "angle": 0, + "content": "42. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: IEEE Conf. Comput. Vis. Pattern Recog. (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.453, + 0.787, + 0.479 + ], + "angle": 0, + "content": "43. Sakaridis, C., Dai, D., Van Gool, L.: Semantic foggy scene understanding with synthetic data. Int. J. Comput. Vis. 126 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.48, + 0.787, + 0.522 + ], + "angle": 0, + "content": "44. Shi, S., Guo, C., Jiang, L., Wang, Z., Shi, J., Wang, X., Li, H.: Pv-rcnn: Point-voxel feature set abstraction for 3d object detection. In: IEEE Conf. Comput. Vis. Pattern Recog. (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.522, + 0.787, + 0.55 + ], + "angle": 0, + "content": "45. Shi, S., Wang, X., Li, H.: Pointcnn: 3d object proposal generation and detection from point cloud. In: IEEE Conf. Comput. Vis. Pattern Recog. (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.55, + 0.787, + 0.591 + ], + "angle": 0, + "content": "46. Shi, X., Ye, Q., Chen, X., Chen, C., Chen, Z., Kim, T.K.: Geometry-based distance decomposition for monocular 3d object detection. In: Int. Conf. Comput. Vis. (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.591, + 0.787, + 0.619 + ], + "angle": 0, + "content": "47. Simonelli, A., Bulo, S.R., Porzi, L., Lopez-Antequera, M., Kontschieder, P.: Disentangling monocular 3d object detection. In: Int. Conf. Comput. Vis. (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.62, + 0.787, + 0.66 + ], + "angle": 0, + "content": "48. Wang, L., Du, L., Ye, X., Fu, Y., Guo, G., Xue, X., Feng, J., Zhang, L.: Depth-conditioned dynamic message propagation for monocular 3d object detection. In: IEEE Conf. Comput. Vis. Pattern Recog. (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.66, + 0.787, + 0.702 + ], + "angle": 0, + "content": "49. Wu, R.Q., Duan, Z.P., Guo, C.L., Chai, Z., Li, C.: Ridcp: Revitalizing real image dehazing via high-quality codebook priors. In: IEEE Conf. Comput. Vis. Pattern Recog. (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.702, + 0.787, + 0.73 + ], + "angle": 0, + "content": "50. Yang, S., Scherer, S.: Cubeslam: Monocular 3-d object slam. IEEE Transactions on Robotics 35(4) (2019). https://doi.org/10.1109/TR0.2019.2909168" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.73, + 0.787, + 0.771 + ], + "angle": 0, + "content": "51. Yang, X., Mi, M.B., Yuan, Y., Wang, X., Tan, R.T.: Object detection in foggy scenes by embedding depth and reconstruction into domain adaptation. In: Asian Conf. Comput. Vis. (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.771, + 0.787, + 0.799 + ], + "angle": 0, + "content": "52. Yu, F., Wang, D., Shelhamer, E., Darrell, T.: Deep layer aggregation. In: IEEE Conf. Comput. Vis. Pattern Recog. (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.799, + 0.787, + 0.841 + ], + "angle": 0, + "content": "53. Yurtsever, E., Lambert, J., Carballo, A., Takeda, K.: A survey of autonomous driving: Common practices and emerging technologies. IEEE Access 8 (2020). https://doi.org/10.1109/ACCESS.2020.2983149" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.353, + 0.127 + ], + "angle": 0, + "content": "Y. Oh et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.203 + ], + "angle": 0, + "content": "54. Zhang, C., Wang, H., Cai, Y., Chen, L., Li, Y., Sotelo, M.A., Li, Z.: Robust-fusionnet: Deep multimodal sensor fusion for 3-d object detection under severe weather conditions. IEEE Transactions on Instrumentation and Measurement 71 (2022). https://doi.org/10.1109/TIM.2022.3191724" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.204, + 0.787, + 0.245 + ], + "angle": 0, + "content": "55. Zhang, R., Qiu, H., Wang, T., Guo, Z., Cui, Z., Qiao, Y., Li, H., Gao, P.: Monodetr: Depth-guided transformer for monocular 3d object detection. In: Int. Conf. Comput. Vis. (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.246, + 0.787, + 0.273 + ], + "angle": 0, + "content": "56. Zhou, Y., Tuzel, O.: Voxelnet: End-to-end learning for point cloud based 3d object detection. In: IEEE Conf. Comput. Vis. Pattern Recog. (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.787, + 0.315 + ], + "angle": 0, + "content": "57. Zhou, Y., He, Y., Zhu, H., Wang, C., Li, H., Jiang, Q.: Monocular 3d object detection: An extrinsic parameter free approach. In: IEEE Conf. Comput. Vis. Pattern Recog. (2021)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.315 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/MonoWAD_ Weather-Adaptive Diffusion Model for Robust Monocular 3D Object Detection/4d73977a-6d19-47be-84e6-b3aec99a372e_origin.pdf b/2024/MonoWAD_ Weather-Adaptive Diffusion Model for Robust Monocular 3D Object Detection/4d73977a-6d19-47be-84e6-b3aec99a372e_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..fd89450868484b807e860851bc0c26f6f76d3746 --- /dev/null +++ b/2024/MonoWAD_ Weather-Adaptive Diffusion Model for Robust Monocular 3D Object Detection/4d73977a-6d19-47be-84e6-b3aec99a372e_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:10ee0c2027c15d1c0979b668e589c704a6c74cbf0d253e646523fc170187a059 +size 3334913 diff --git a/2024/MonoWAD_ Weather-Adaptive Diffusion Model for Robust Monocular 3D Object Detection/full.md b/2024/MonoWAD_ Weather-Adaptive Diffusion Model for Robust Monocular 3D Object Detection/full.md new file mode 100644 index 0000000000000000000000000000000000000000..3dccfc32b65cd0e9186fed979ddb88d85e2c1f06 --- /dev/null +++ b/2024/MonoWAD_ Weather-Adaptive Diffusion Model for Robust Monocular 3D Object Detection/full.md @@ -0,0 +1,337 @@ +# MonoWAD: Weather-Adaptive Diffusion Model for Robust Monocular 3D Object Detection + +Youngmin Oh1, Hyung-Il Kim2, Seong Tae Kim1†, and Jung Uk Kim1† + +1 Kyung Hee University, Yong-in, South Korea {oym9104, st.kim, ju.kim}@khu.ac.kr 2 ETRI, Daejeon, South Korea hikim@etri.re.kr + +Abstract. Monocular 3D object detection is an important challenging task in autonomous driving. Existing methods mainly focus on performing 3D detection in ideal weather conditions, characterized by scenarios with clear and optimal visibility. However, the challenge of autonomous driving requires the ability to handle changes in weather conditions, such as foggy weather, not just clear weather. We introduce MonoWAD, a novel weather-robust monocular 3D object detector with a weather-adaptive diffusion model. It contains two components: (1) the weather codebook to memorize the knowledge of the clear weather and generate a weather-reference feature for any input, and (2) the weather-adaptive diffusion model to enhance the feature representation of the input feature by incorporating a weather-reference feature. This serves as an attention role in indicating how much improvement is needed for the input feature according to the weather conditions. To achieve this goal, we introduce a weather-adaptive enhancement loss to enhance the feature representation under both clear and foggy weather conditions. Extensive experiments under various weather conditions demonstrate that MonoWAD achieves weather-robust monocular 3D object detection. The code and dataset are released at https://github.com/VisualAIKHU/MonoWAD. + +Keywords: Monocular 3D Object Detection $\cdot$ Weather-Adaptive Diffusion $\cdot$ Weather Codebook + +# 1 Introduction + +Monocular 3D object detection aims to detect 3D objects using only a single camera [16, 18, 28, 46, 55, 57]. In contrast to LiDAR-based methods that rely on expensive LiDAR sensors for depth estimation [24, 44, 45, 56], and stereo-based methods that require synchronized stereo cameras, monocular 3D object detection only requires monocular images, offering the advantage of computational cost-effectiveness and requiring fewer resources. Due to this characteristic, the monocular 3D object detection is applied to a wide range of real-world applications, such as autonomous driving [12, 19, 53] and robotics [50]. + +![](images/4096afacdb8773efa89658e3f125d0abdc0e7e5a72b2515d361cce17f37609ca.jpg) +Fig. 1: Conceptual diagram of the proposed method (foggy example). (a) In the training phase, weather codebook learns the clear knowledge to transfer it to weather-adaptive diffusion model to enhance content related to the weather conditions. (b) By doing so, even with input images under various weather conditions (e.g., foggy images), monocular 3D object detection becomes adaptable to various weather scenarios. + +However, existing monocular 3D object detectors mainly focus on ideal autonomous driving environments (i.e., clear weather). There are challenges in applying them to real-world scenarios with adverse weather conditions, such as fog and rain. Among these, fog poses the most significant challenge compared to other weather [14, 17]. This is due to the dense and diffuse nature of fog, which strongly scatters and absorbs light, leading to difficulties in object detection [2]. Since monocular 3D object detection relies solely on visual information from a single image, unlike LiDAR, it is crucial to design detectors to achieve enhanced performance in challenging visibility scenarios. + +In this paper, we propose MonoWAD, a novel weather-robust monocular 3D object detector to address the aforementioned issues. As mentioned earlier, due to the inherent challenges posed by foggy weather among various adverse weather conditions, we focus primarily on clear and foggy weather (results for other weather conditions such as rainy and sunset are presented in Section 4). For weather-robust object detection, clear weather requires relatively modest improvements to enhance the visual representation of features. In contrast, significant enhancements in this feature representation are required for foggy weather. To address this, we consider two key aspects: (1) how to quantify the degree of improvement needed for the input image, and (2) how to guide the representation of the input image. + +To address the two key aspects, as shown in Fig. 1, our MonoWAD consists of a weather codebook and weather-adaptive diffusion model. First, we introduce the weather codebook to generate a weather-reference feature that contains knowledge about reference weather in a given scene. The reference weather acts as a guide, indicating the degree of weather improvement required. Since the clear weather contains a richer visual representation of objects, we adopt it as a reference weather. At this time, we devise a clear knowledge recalling (CKR) loss to guide the weather codebook to memorize information about clear weather and generate a weather-reference feature for any input (clear or foggy). As a result, our detector can understand where improvements are needed in the input features based on the weather-reference feature. + +Second, we propose a weather-adaptive diffusion model to effectively enhance feature representations in accordance with weather conditions. Given input fea- + +ture (clear or foggy), the weather-adaptive diffusion model dynamically enhances the representation of the input feature based on the weather-reference feature. The weather-reference feature plays a role of attention, determining the extent to which the input features need improvement. At this time, we define the difference between clear and foggy weather (i.e., weather changes) as the fog distribution to adopt it as the noise for our diffusion model. With fog distribution, our weather-adaptive diffusion model can enhance the feature representation according to the weather conditions through multiple steps of reverse processes. To achieve this goal, we introduce a weather-adaptive enhancement (WAE) loss. As a result, our MonoWAD performs weather-robust detection by adaptively improving feature representation according to the weather conditions. + +To adaptively enhance the feature representation through the difference between weather conditions, we generate a new foggy KITTI dataset based on the KITTI dataset [12]. Comprehensive experimental results on several datasets [11, 12] show that our MonoWAD outperforms the existing state-of-the-art monocular 3D object detectors [16,28,37,40,55] under foggy weather, which is the most challenging weather condition. While our method primarily focuses on foggy, experiments conducted under various weather conditions (e.g., foggy, rainy, and sunset) have demonstrated its applicability to other weather scenarios. + +The main contributions of our paper can be summarized as follows: + +- We introduce a new weather-robust monocular 3D object detector, called MonoWAD, that is robust to various weather conditions. +- We design a weather codebook with clear knowledge recalling loss for learning about clear weather, providing reference information for enhancement. +- We propose weather-adaptive diffusion model with weather-adaptive enhancement loss to dynamically enhance the feature representation of the input images according to the weather conditions. + +# 2 Related Work + +# 2.1 Monocular 3D Object Detection + +Monocular 3D object detection task can be categorized into two directions according to the type of data used in the training phase: (1) using only a monocular image and (2) incorporating additional data, such as depth along with a monocular image. The first category relies on the geometric relationship between 2D and 3D [3,7,25,26,28,31,34,40,46,55,57]. For example, Deep3Dbox [34] utilizes the geometric information of 2D bounding boxes to predict 3D bounding boxes. In [3], M3D-RPN was proposed to understand the 3D scene from the depth-aware convolution. MonoRCNN [46] predicted 3D bounding boxes through geometry-based distance decomposition, and MonoCon [25] learns mono context for 3D object detection. MonoDETR [55] introduces a depth-guided transformer that utilizes geometric depth cues without requiring additional data. + +Moreover, since monocular image contains limited information for estimating 3D object cues, monocular 3D object detectors have adopted the additional + +data for more robust detection [4,8,16,27,32,37,41,48]. For example, the depth-conditioned dynamic message propagation (DDMP) [48] was proposed to integrate prior depth information with the image context. CaDDN [41] was introduced to utilize the depth distribution of predicted categories for each pixel to project the context information onto 3D space, deriving 3D bounding boxes. MonoDTR [16] employs the transformer architecture to integrate depth features and context, thus estimating more accurate depth information. + +Despite recent progress, the existing monocular 3D object detectors mainly rely on the benchmark data collected under clear weather conditions. However, it is essential to account for challenging adverse weather conditions, such as fog, to more accurately reflect real-world scenarios. In this paper, we aim to introduce weather-robust 3D object detection by enhancing visual features with the proposed weather-adaptive diffusion model and weather codebook. + +# 2.2 Computer Vision Tasks for Foggy Weather + +There have been a lot of studies on improving performance in various weather conditions for real-world application of computer vision technology [2, 13, 20, 21, 33, 38, 43, 51, 54]. In particular, fog is considered one of the most critical issues due to its significant degradation of visual information compared to other weather conditions. To deal with the foggy weather conditions, the authors in [43] generated synthetic fog images (so-called Foggy Cityscapes) from clear weather images, and utilized them for training semantic segmentation and 2D object detection. Similar to Foggy Cityscapes, Martin et al. [13] naturally synthesizes fog into LIDAR to enhance the performance of LiDAR-based 3D object detectors in foggy weather. Bijelic et al. [2] use real dataset including foggy weather, for 2D detection with the multimodal fusion networks. Mai et al. [33] synthesize fog for LiDAR and stereo images and perform fusion-based 3D object detection using the SLS-Fusion network. Xin et al. [51] focused on 2D detection in foggy weather by applying domain adaptation. In this context, we focus on monocular 3D object detection in scenarios that rely solely on visual information in challenging foggy environments. To address this challenge, we propose a weather-robust diffusion model that dynamically improves features based on reference feature. + +# 2.3 Diffusion Models + +Recently, the diffusion model [15] has attracted considerable interest in computer vision due to its impressive progress in image generation [22, 29, 35, 42], as well as its potential application to other vision tasks such as segmentation [1] and image captioning [30]. Inspired by the remarkable generative ability, we design the diffusion model for robust monocular 3D object detection by considering a foggy effect (which is one of the challenging adverse weather conditions for monocular 3D object detection) as a form of noise in the model. That is, we propose a method in which visual features obscured by fog are progressively improved by training a diffusion model based on the forward/reverse diffusion process. In particular, we present an adaptive method that allows the diffusion model to control the degree of improvement by weather conditions. + +![](images/2105952e910218bb8044c128b12c704c3f65278f2fb9df10aeaedb76fc42a5ae.jpg) +Fig. 2: Overview of our MonoWAD in the inference phase. It mainly contains three parts: weather codebook, weather-adaptive diffusion model, and detection block. Through the weather codebook and weather-adaptive diffusion model, our method can maintain robustness against various weather conditions (i.e., clear or foggy). + +# 3 Proposed Method + +Fig. 2 shows the overall framework of the proposed MonoWAD in the inference phase. A backbone network receives an input image (clear image $I_{c}$ or foggy image $I_{f}$ ) to encode the corresponding input feature (input clear feature $x^{c}$ or input foggy feature $x^{f}$ ). It interacts with the weather codebook $\mathcal{Z}$ to generate weather-reference feature $x^{r}$ , indicating the amount of enhancement for the given input feature. Subsequently, the weather-adaptive diffusion model attempts to enhance the input feature over $T$ timesteps to obtain an enhanced feature $\tilde{x}^{c}$ or $\tilde{x}^{f}$ . Finally, monocular 3D object detection is performed through the detection block. Note that, in the training phase, our MonoWAD use the clear feature $x^{c}$ to train our diffusion as well as the detection block. + +We address two key issues: (1) how to guide weather-reference feature $x^r$ to serve as a reference feature, and (2) how to guide the weather-adaptive diffusion model to effectively enhance feature representations based on weather conditions. Details are in the following subsections. + +# 3.1 Weather Codebook + +In foggy weather conditions, the overall visual quality of the scene is generally poor, requiring a significant enhancement. Conversely, in clear weather, the amount of improvement is expected to be relatively minimal compared to foggy conditions. Therefore, inspired by [9,36], we devise a weather codebook $\mathcal{Z}$ to provide the reference knowledge about the weather for appropriate enhancement based on the weather conditions. At this time, as the clear weather contains abundant visual representations, we use it as a reference weather knowledge. + +As shown in Fig. 3, the reference knowledge embedding procedure involves receiving paired clear-foggy features during the training phase. The weather codebook $\mathcal{Z}$ consists of $K$ learnable slots, denoted as $\mathcal{Z} = \{z_k\}_{k=1}^K (z_k \in \mathbb{R}^{1 \times c})$ , where $c$ represents the dimensionality of each slot. The paired clear feature $x^c$ and foggy feature $x^f$ pass through a convolution layer to generate $\hat{x}^c \in \mathbb{R}^{h \times w \times c}$ and $\hat{x}^f \in \mathbb{R}^{h \times w \times c}$ ( $w$ denotes width and $h$ indicates height). Each element of the + +![](images/a461609d2f407e75e3eec09b04ff65dc201e192f796c9785c68517f768250a5b.jpg) +Fig. 3: Illustration of the proposed (a) weather-invariant guiding (WIG) loss and (b) clear knowledge embedding (CKE) loss. The clear knowledge recalling (CKR) loss, obtained from combining WIG and CKE, aims to memorize the knowledge of the clear weather and recall the same clear knowledge from the foggy weather. + +feature denoted as $\hat{x}_{ij}^{c}\in \mathbb{R}^{1\times c}$ and $\hat{x}_{ij}^{f}\in \mathbb{R}^{1\times c}$ . Subsequently, we obtain weather-reference feature for clear weather, $x^{r(c)}\in \mathbb{R}^{h\times w\times c}$ , by conducting element-wise quantization process $\mathbf{q}(\cdot)$ , calculated as: + +$$ +x ^ {r (c)} = \mathbf {q} (\hat {x} ^ {c}) := \left(\underset {z _ {k} \in \mathcal {Z}} {\arg \min } \| \hat {x} _ {i j} ^ {c} - z _ {k} \|\right). \tag {1} +$$ + +Utilizing $x^{c}$ and $x^{r(c)}$ , we introduce a clear knowledge embedding (CKE) loss $\mathcal{L}_{cke}$ to guide $x^{r(c)}$ to follow the representation of $x^{c}$ . To this end, we perform global average pooling (GAP) for $x^{c}$ and $x^{r(c)}$ with softmax, generating $s^{c}$ and $s^{r(c)}$ , respectively. Each element in the vector indicates the probability of the significance of each channel. With $s^{c}$ and $s^{r(c)}$ , we employ KL divergence $D_{KL}(\cdot)$ for $\mathcal{L}_{cke}$ to compare the probability distributions, formulated as: + +$$ +\mathcal {L} _ {c k e} = D _ {K L} \left(s ^ {c} \| s ^ {r (c)}\right). \tag {2} +$$ + +Through $\mathcal{L}_{cke}$ , the weather codebook $\mathcal{Z}$ can memorize the knowledge of the clear weather, allowing it to effectively reconstruct the clear weather knowledge. + +Additionally, as the paired clear-foggy images are identical except for weather conditions, the quantization process of the foggy feature $x^{f}$ with the weather codebook should generate an equivalent weather-reference feature. For obtaining weather-reference feature for foggy $x^{r(f)}$ , the element-wise quantization process is also conducted for $\hat{x}^{f}$ and $\mathcal{Z}$ : + +$$ +x ^ {r (f)} = \mathbf {q} (\hat {x} ^ {f}) := \left(\underset {z _ {k} \in \mathcal {Z}} {\arg \min } \| \hat {x} _ {i j} ^ {f} - z _ {k} \|\right). \tag {3} +$$ + +Next, we introduce the weather-invariant guiding (WIG) loss $\mathcal{L}_{wig}$ to guide $\mathcal{Z}$ that the weather codebook recalls the same clear knowledge of clear feature for + +![](images/30c2bd0b9fc9be3b440b6c3e119c54193994e8d55260f719d12000e097d5ad66.jpg) +Fig. 4: Training process of the weather-adaptive diffusion model, which consists of two processes: (1) Adding fog variant $\epsilon_{n}$ from input clear feature $x^{c}$ (forward process) and (2) enhancing representation with weather-reference feature $x^{r}$ (reverse process). + +the foggy feature, which can be represented as follows: + +$$ +\mathcal {L} _ {w i g} = \left\| x ^ {r (c)} - x ^ {r (f)} \right\| _ {2} ^ {2}. \tag {4} +$$ + +Finally, the clear knowledge recalling (CKR) loss $\mathcal{L}_{ckr}$ is obtained by adding $\mathcal{L}_{cke}$ and $\mathcal{L}_{wig}$ , which is defined as: + +$$ +\mathcal {L} _ {c k r} = \mathcal {L} _ {c k e} + \mathcal {L} _ {w i g}. \tag {5} +$$ + +In the training phase, the weight parameters of embedding $K$ slots of weather codebook $\mathcal{Z}$ are initialized randomly and they are updated through Eq. (5). In the inference phase, all parameters are fixed to recall clear weather, generating weather-reference features for any weather conditions. + +# 3.2 Weather-Adaptive Diffusion Model + +Through Section 3.1, we now know that the weather codebook outputs the weather-reference feature $x^{r(c)}$ for the clear weather and $x^{r(f)}$ for the foggy weather through Eq. (5). From now on, since our method can receive any input images (clear or foggy), we denote the weather-reference feature as $x^r$ . + +Fig. 4 shows the training process of the proposed weather-adaptive diffusion model. The key idea of the diffusion model [15,42] is to gradually enhance $x_{T}$ to $x_{0}$ with fixed Markov Chain of $T$ timesteps. To this end, the forward and reverse processes are conducted in the training phase, and only the reverse process is used for the inference phase. Motivated by [15,42], we construct the weather-adaptive diffusion model to enhance representation related to the weather conditions. To this end, unlike traditional diffusion methods [15,42] that adopt the Gaussian noise to the image or latent space, we adopt $\mathcal{F} = x^{f} - x^{c}$ , called fog distribution, to guide our diffusion model to be aware of the foggy weather. Ideally, the information contained within $\mathcal{F}$ should include the information of fog, as it represents the difference between the foggy scene and the same scene with clear weather. + +By doing so, our diffusion model learns the variation in weather by repeatedly adding and removing the fog. Note that, as we take clear feature $x^{c}$ for $x_{0}$ to make a reference input for our diffusion model, we newly denote $x_{0}$ as $x_{0}^{c}$ . + +For the forward process at the $t$ -th timestep, $q(x_{t}^{c}|x_{t - 1}^{c})$ takes the previous feature $x_{t - 1}^{c}$ and the noise related to the fog (i.e., $\mathcal{F}$ ) as inputs to generate $x_{t}^{c}$ . This procedure is repeated over $T$ timesteps, which can be represented as: + +$$ +q (x _ {t} ^ {c} | x _ {t - 1} ^ {c}) = \mathcal {F} (x _ {t} ^ {c}; \sqrt {1 - \beta_ {t}} x _ {t - 1} ^ {c}, \beta_ {t} \mathbf {I}), \tag {6} +$$ + +$$ +q \left(x _ {1} ^ {c}, \dots , x _ {T} ^ {c} \mid x ^ {c}\right) = \prod_ {t = 1} ^ {T} q \left(x _ {t} ^ {c} \mid x _ {t - 1} ^ {c}\right), \tag {7} +$$ + +where $\beta_{t}$ denotes the variance schedule. + +Next, the reverse process at the $t$ -th timestep aims to estimate fog variant $\epsilon_{n}$ using $x_{t}^{c}$ to enhance the feature representation of the foggy. To this end, we adopt conditional autoencoder $\epsilon_{\theta}(x_{t}^{c}, t, x^{r})$ that receives $x_{t}^{c}$ and reference feature $x^{r}$ from the weather codebook $\mathcal{Z}$ . Specifically, $\epsilon_{\theta}$ estimates mean $\mu_{\theta}$ and variance $\pmb{\Sigma}_{\theta}$ of the fog distribution at the $t$ -th timestep, denoted as $\tilde{\mathcal{F}}(\cdot)$ . The reverse process is also repeated over $T$ timesteps, which can be represented as: + +$$ +p _ {\theta} \left(x _ {t - 1} ^ {c} \mid x _ {t} ^ {c}, x ^ {r}\right) = \tilde {\mathcal {F}} \left(x _ {t - 1} ^ {c}; \boldsymbol {\mu} _ {\theta} \left(x _ {t} ^ {c}, t, x ^ {r}\right), \boldsymbol {\Sigma} _ {\theta} \left(x _ {t} ^ {c}, t\right)\right), \tag {8} +$$ + +$$ +p _ {\theta} \left(x ^ {c}, \dots , x _ {T} ^ {c}\right) = p \left(x _ {T} ^ {c}\right) \prod_ {t = 1} ^ {T} p _ {\theta} \left(x _ {t - 1} ^ {c} \mid x _ {t} ^ {c}, x ^ {r}\right), \tag {9} +$$ + +where $p_{\theta}(x_{t - 1}^{c}|x_{t}^{c},x^{r})$ includes cross-attention layer in $\epsilon_{\theta}$ + +The cross-attention layer receives the flattened clear feature $x_{t}^{c}$ and weather-reference feature $x^{r}$ , i.e., $\bar{x}_{t}^{c}$ and $\bar{x}^{r}$ , respectively. The similarity between $\bar{x}^{r}$ and $\bar{x}_{t}^{c}$ is calculated in the cross-attention layer to transfer the enhancement to $\bar{x}_{t}^{c}$ , which can be formulated as: + +$$ +\operatorname {A t t e n t i o n} (Q, K, V) = \operatorname {s o f t m a x} \left(\frac {Q K ^ {T}}{\sqrt {d}}\right) V, \tag {10} +$$ + +where $Q = W_{i}^{q} \cdot \bar{x}_{t}^{c}, K = W_{i}^{k} \cdot \bar{x}^{r}, V = W_{i}^{v} \cdot \bar{x}^{r}$ with learnable parameters $W_{i}^{q}, W_{i}^{k}, W_{i}^{v}$ . + +To ensure that the estimated fog variant $\epsilon_{\theta}$ is similar to the fog variant $\epsilon_{n}$ applied to make $x_{t}^{c}$ from the input clear feature $x^{c}$ , we propose a weather-adaptive enhancement loss $\mathcal{L}_{wae}$ , which is formulated as: + +$$ +\mathcal {L} _ {w a e} = \mathbb {E} _ {x ^ {c}, \boldsymbol {\epsilon} _ {n} \sim \mathcal {F}, t} \left[ \| \boldsymbol {\epsilon} _ {n} - \boldsymbol {\epsilon} _ {\theta} \left(x _ {t} ^ {c}, t, x ^ {r}\right) \| _ {2} ^ {2} \right]. \tag {11} +$$ + +With $\mathcal{L}_{wae}$ , $\epsilon_{\theta}$ can estimate the fog variant by leveraging the fog distribution as noise for our diffusion model through multiple forward/reverse processes. Additionally, the cross-attention layer within $\epsilon_{\theta}$ dynamically enhances the feature representation based on combining knowledge of the input feature (foggy or clear) and the weather-reference feature. Since our diffusion model learns the + +degree of improvement needed corresponding to weather conditions, it can improve the representation of any input, whether clear or foggy in inference phase. This leads our monocular 3D object detector can handle both clear and foggy input images, resulting in weather-robust detection. + +# 3.3 Total Loss Function + +The total loss function of our MonoWAD is represented as follows: + +$$ +\mathcal {L} _ {\text {T o t a l}} = \mathcal {L} _ {O D} + \lambda_ {1} \mathcal {L} _ {\text {c k r}} + \lambda_ {2} \mathcal {L} _ {\text {w a e}}, \tag {12} +$$ + +where $\lambda_{1}$ and $\lambda_{2}$ denote balancing hyper-parameters. In our experiments, we set $\lambda_{1} = \lambda_{2} = 1$ . $\mathcal{L}_{OD}$ is the detection loss for 3D object detection [16,55]. It includes loss functions of classification, regression, and depth loss, that are similar to prior works [8,16,31]. The overall weight parameters are updated through $\mathcal{L}_{Total}$ . + +# 4 Experiments + +# 4.1 Dataset and Evaluation Metrics + +Datasets. We utilize the KITTI 3D object detection dataset [12], the most widely adopted for 3D object detection. It contains 7,481 training images and 7,518 test images under clear weather conditions. Due to unavailable ground-truth annotations for test images and limited evaluation on the test server, we follow [6] by splitting 3,712 training images and 3,769 validation images. In addition, as our work requires paired images to learn about weather changes, we generate foggy images from all images in the KITTI dataset, called foggy KITTI, that emulate the foggy scene. Following the protocols of [2,43], we generate foggy images based on object distances using depth maps estimated by DORN [10]. Please refer to the supplementary materials for details. + +In addition, as our work is focused on robust monocular 3D object detection in various weather conditions, we further adopt the Virtual KITTI dataset [5,11], which contains photo-realistic synthetic images under various weather conditions (e.g., foggy, rainy, sunset). It is associated with the original real-world KITTI dataset, which has 3D annotations for each weather. + +Evaluation Metrics. We adopt average precision in both 3D detection $(\mathrm{AP}_{3D})$ and bird-eye view detection $(\mathrm{AP}_{BEV})$ under three difficulty levels ('Easy', 'Moderate', 'Hard') according to size, occlusion, and truncation. Following [47], we use 40 recall position metric $\mathrm{AP}_{40}$ and report scores for the car category under the IoU threshold 0.7 for the KITTI dataset and 0.5 for the Virtual KITTI dataset. + +# 4.2 Implementation Details + +We use DLA-102 [52] as our backbone and adopt the transformer architecture of [16] for the detection block. We train MonoWAD on a single RTX 4090 GPU with a batch size of 4 over 120 epochs using Adam optimizer [23] (initial learning + +Table 1: Detection results of car category on KITTI validation set under foggy weather and clear weather conditions. The results of the state-of-the-art methods under foggy weather are obtained through our reproduction with the official source code. Bold/underlined fonts indicate the best/second-best results. + +
MethodFoggy (AP3D)Clear (AP3D)Average
EasyMod.HardEasyMod.HardEasyMod.Hard
GUPNet [28] (ICCV'21)2.742.192.1622.7616.4613.7212.759.337.94
DID-M3D [37] (ECCV'22)1.150.610.6422.9816.1214.0312.078.377.34
MonoGround [40] (CVPR'22)0.000.000.0625.2418.6915.5812.629.357.82
MonoDTR [16] (CVPR'22)16.8911.869.8724.5218.5715.5120.7115.2212.69
MonoDETR [55] (ICCV'23)7.405.744.5328.8420.6116.3818.1213.1810.46
MonoWAD (Ours)27.1719.5716.2129.1021.0817.7328.1420.3316.97
+ +Table 2: Detection results of car category on KITTI test set under foggy weather. Bold/underlined fonts indicate the best/second-best results. + +
MethodAP3DAPBEV
EasyMod.HardEasyMod.Hard
GUPNet [28] (ICCV'21)3.012.421.134.903.022.91
DID-M3D [37] (ECCV'22)3.102.392.195.343.012.91
MonoGround [40] (CVPR'22)0.140.200.220.230.380.39
MonoDTR [16] (CVPR'22)11.077.415.2615.7610.157.53
MonoDETR [55] (ICCV'23)9.335.544.0613.158.066.30
MonoWAD (Ours)19.7513.3211.0427.9519.0615.61
+ +rate of $10^{-4}$ ). For the weather codebook, we utilize embedding slot $K = 4096$ and set the dimension of each slot $D = 256$ . We set 4 heads for cross-attention of the weather-adaptive diffusion model, and the timesteps for the forward and reverse process of diffusion default to 15. The number of channels for the diffusion output features is $C = 256$ . As recent monocular 3D object detectors [16,28,37,40,55] have not been explored under adverse weather conditions, we implemented them with available official source code to faithfully reproduce them. + +# 4.3 Comparison + +Results on KITTI 3D Dataset. We compared MonoWAD with state-of-the-art monocular 3D object detectors [16,28,37,40,55] that do not use additional data (e.g., depth maps or LiDAR) during inference on the KITTI and foggy KITTI validation sets. Table 1 shows the $AP_{3D}$ results, and $AP_{BEV}$ results are in the supplementary materials. While recent methods have shown improved performance in clear weather, their performance drops in foggy weather, limiting their applicability in real-world applications (e.g., autonomous driving and robotics). In contrast, MonoWAD showed stable 3D detection performance under both foggy and clear weather. Since our weather codebook learns knowledge about clear weather and the weather-adaptive diffusion model can enhance the feature representation of the input images under both clear and foggy weather, MonoWAD shows a more weather-robust detection performance than that of the existing methods. Also, to explore the weather robustness of MonoWAD, we conducted experiments with mixed foggy and clear weather conditions at various ratios (please see the supplementary materials). + +Table 3: Detection results of car category on Virtual KITTI under foggy, rainy, sunset conditions. Bold/underlined fonts indicate the best/second-best results. + +
MethodFoggy (AP3D)Rainy (AP3D)Sunset (AP3D)
EasyMod.HardEasyMod.HardEasyMod.Hard
GUPNet [28] (ICCV'21)1.761.571.572.341.241.212.771.641.65
DID-M3D [37] (ECCV'22)0.910.390.390.400.130.130.340.100.10
MonoGround [40] (CVPR'22)0.290.300.255.492.822.777.684.244.20
MonoDTR [16] (CVPR'22)8.795.755.7211.736.256.749.865.425.42
MonoDETR [55] (ICCV'23)4.502.992.966.613.463.427.084.174.16
MonoWAD (Ours)13.338.568.5014.128.338.2413.387.897.80
+ +Table 4: Effect of the proposed method on KITTI validation set for car category. WC denotes our weather codebook, WAD indicates our weather-adaptive diffusion model. + +
MethodWCWADFoggy (AP3D)Clear (AP3D)
EasyMod.HardEasyMod.Hard
Baseline--13.759.618.1022.6317.1614.28
MonoWAD (Ours)-✓25.6218.6615.5626.3419.1716.15
✓✓27.1719.5716.2129.1021.0817.73
+ +Table 5: Detection results on KITTI validation set by changing diffusion timestep $T$ . + +
Timestep TFoggy (AP3D)Clear (AP3D)
EasyMod.HardEasyMod.Hard
-13.759.618.1022.6317.1614.28
523.5717.9114.9626.0319.2116.01
1025.2818.4915.4226.7919.9016.78
1527.1719.5716.2129.1021.0817.73
2024.5418.2915.1024.8518.5415.34
+ +We further compared 3D detection performances on the foggy KITTI test set (Table 2). Similar to Table 1, existing monocular 3D object detectors show lower performance under foggy weather. Through the results, our method shows the weather-robust monocular 3D performance under foggy and clear weather. + +Results on Virtual KITTI Dataset. We also conducted experiments on the Virtual KITTI dataset to validate the generalization ability of our method across various weather conditions (i.e., foggy, rainy, sunset). The results are shown in Table 3. MonoWAD also shows the highest performance in foggy weather. Moreover, it outperforms the existing method in the rainy and sunset conditions. The results demonstrate that MonoWAD is robust and insensitive to various weather conditions that can be faced in real-world scenarios, not just clear weather. + +# 4.4 Ablation Studies + +We conducted ablation studies to examine (1) the effect of each proposed component (i.e., weather codebook and weather-adaptive diffusion model) and (2) the effect of the weather-adaptive diffusion model by varying the timestep $T$ . These experiments were performed on the KITTI and foggy KITTI 3D validation sets. + +Effect of the Proposed Modules. The results regarding the effectiveness of the proposed weather codebook and weather-adaptive diffusion model are pre- + +Table 6: Performance comparison on KITTI validation set under foggy and clear weather conditions. We compare our MonoWAD with MonoDTR [16], which demonstrates superior performance in foggy weather using state-of-the-art dehazing methods. + +
MethodDehazing MethodFoggy (AP3D)Clear (AP3D)Average
EasyMod.HardEasyMod.HardEasyMod.Hard
MonoDTR(B) [16] (CVPR'22)-16.8911.869.8724.5218.5715.5120.7115.2212.69
B + RIDCP [49] (CVPR'23)Image-level17.2312.4110.4424.0217.8914.7820.6315.1512.61
B + DENet [39] (ACCV'22)Feature-level22.3517.4414.477.105.704.5314.7311.579.50
B + Yang et al. [51] (ACCV'22)Feature-level22.8715.2112.1717.9613.1010.6420.4214.1611.41
MonoWAD (Ours)-27.1719.5716.2129.1021.0817.7328.1420.3316.97
+ +Table 7: Comparison of diffusion models on KITTI validation set for car category: $\mathcal{B}$ is baseline detection block (transformer encoder-decoder), CDM (Conditional Diffusion Model [42]), WC (weather codebook), and WAD (weather adaptive diffusion model). + +
MethodFoggy (AP3D)Clear (AP3D)
EasyMod.HardEasyMod.Hard
B+DDPM [15]5.323.842.7718.3112.7110.20
B+CDM [42]2.742.102.0020.5014.5111.63
B+WC+CDM [42]17.5112.7410.4021.0515.1112.54
B+WAD25.6218.6615.5626.3419.1716.15
MonoWAD(B+WC+WAD)27.1719.5716.2129.1021.0817.73
+ +sented in Table 4. Since our weather-adaptive diffusion model is designed to enhance feature representation, it alone has shown significant performance improvement. When the weather codebook is additionally considered, the weather-adaptive diffusion model can leverage the knowledge of the weather-reference feature for clear weather, leading to enhanced performance. It makes MonoWAD can be robust to various weather conditions for monocular 3D object detection. + +Effect of Timestep $T$ . We also conduct experiments by varying timestep $T$ of the weather-adaptive diffusion model. Timestep is the number of steps for the forward and reverse process of the diffusion model. Table 5 indicates that the highest monocular 3D detection performance is achieved when $T = 15$ under both clear and foggy weather conditions. Our MonoWAD consistently outperforms the baseline and other existing methods across all timesteps. + +# 4.5 Discussions + +Comparison with Dehazing Methods. We investigate the weather-robustness of our method for monocular 3D object detection compared to the other monocular detectors using dehazing methods. To this end, we compared MonoWAD with MonoDTR [16], which shows the best performance in foggy weather through the application of state-of-the-art image-level and feature-level dehazing methods [39,49,51]. The results are shown in Table 6. Recent dehazing methods are primarily focused on specific weather conditions, such as foggy weather. While they have shown some improvement under foggy conditions, they exhibit reduced performances under clear weather conditions. In contrast, our MonoWAD shows robust performance across both foggy and clear weather conditions. + +![](images/1ce9cb7a8af6a016833dac592e5eb58cf51c6f754a3faf5c18414187a51fa69e.jpg) +Fig. 5: Comparison of 3D detection examples (green: ground-truth, red: predicted 3D bounding-box) between our MonoWAD and two detectors, MonoDTR [16] and MonoDETR [55], that show the most improved performances among existing methods. + +![](images/4df94299132fc98cb15113316d4d2591056ee507d5585a7166dd4da1fed82b6a.jpg) +Fig. 6: 3D detection results on real-world images of various weather conditions. + +Effect of Weather-Adaptive Diffusion Model. Table 7 shows the effectiveness of MonoWAD with existing diffusion models [15,42] on the KITTI and foggy KITTI validation sets. Existing methods adopt Gaussian noise for forward and reverse processes, but they can not fully understand about weather. In contrast, our weather-adaptive diffusion understands weather variances, allowing our MonoWAD to surpass existing methods in clear and foggy weather. + +# 4.6 Visualization Results + +Results on KITTI Dataset. We visualize several 3D detection results on the KITTI 3D dataset, comparing MonoWAD with MonoDTR and MonoDETR, which exhibit the highest performances among existing methods under foggy condition (Fig. 5). Existing methods struggle to detect objects obscured by fog, indicating limitations in detecting only fully visible objects. In contrast, even + +![](images/9a7aeccd9d723fbb3756503070f2e04bc5fb3e1ddc1cc2d526c39a3329162339.jpg) +Fig. 7: t-SNE visualization results (Red: Clear, Blue: Foggy). + +![](images/a32b11d56c1326e5fd5cc8acac38664ea01b1127ef3278cbfcfff7f6b61d76a0.jpg) + +![](images/2590033f084d124308f014d93a14eb5c6c55e66ce963acf9541378c89e17472c.jpg) + +in dense fog, MonoWAD effectively detects both close and fog-obscured objects with the aid of the weather codebook and weather-adaptive diffusion model. + +Results on Real-World Images. We further visualize 3D detection results on real-world images from the Seeing Through Fog dataset [2] under various weather conditions (i.e., foggy, rainy, snowy). In Fig. 6, MonoWAD shows robust detection under diverse weather conditions. This demonstrates that our proposed method maintains weather-robustness even in real-world scenarios by dynamically enhancing the input scenarios. + +t-SNE Visualization. We conducted t-SNE visualization to analyze feature representations of MonoDTR, MonoDETR, and our MonoWAD on the KITTI and foggy KITTI validation set. As depicted in Fig. 7, the existing methods exhibit distinct feature representations for foggy and clear weather conditions. In contrast, MonoWAD, leveraging weather-robust feature learning from the weather codebook and weather-adaptive diffusion model, demonstrates similar feature representations for both clear and foggy weather conditions. + +# 4.7 Limitations + +The experimental results show the weather-robustness of our method. However, due to the iterative nature of the diffusion model, our method shows $144\mathrm{ms}/$ image at timestep $T = 15$ (110ms/image at $T = 5$ ), slower than the latest work, MonoDETR (38ms/image). Moreover, our method dynamically enhances the representation of the input feature based on the weather-reference feature and weather difference which needs paired images. Thus, exploring a method to achieve faster processing speeds while maintaining weather-robust performance without paired images could be an interesting direction for our future work. + +# 5 Conclusion + +We proposed MonoWAD, a novel weather-robust monocular 3D object detector to handle various weather conditions. Addressing challenges in applying existing monocular 3D object detectors to real-world scenarios with various weather, we design a weather codebook with clear knowledge recalling loss to memorize the knowledge of the clear weather and to generate a weather-reference feature from both clear and foggy features. Also, we design a weather-adaptive diffusion model with weather-adaptive enhancement loss to enhance feature representation according to the weather conditions. As a result, our MonoWAD can detect objects occluded by fog and perform well in clear weather. + +# Acknowledgements + +This work was supported by the NRF grant funded by the Korea government (MSIT) (No. RS-2023-00252391), and by IITP grant funded by the Korea government (MSIT) (No. RS-2022-00155911: Artificial Intelligence Convergence Innovation Human Resources Development (Kyung Hee University), IITP-2023-RS-2023-00266615: Convergence Security Core Talent Training Business Support Program, No. 2022-0-00124: Development of Artificial Intelligence Technology for Self-Improving Competency-Aware Learning Capabilities). + +# References + +1. Baranchuk, D., Rubachev, I., Voynov, A., Khrulkov, V., Babenko, A.: Label-efficient semantic segmentation with diffusion models. In: Int. Conf. Learn. Represent. (2021) +2. Bijelic, M., Gruber, T., Mannan, F., Kraus, F., Ritter, W., Dietmayer, K., Heide, F.: Seeing through fog without seeing fog: Deep multimodal sensor fusion in unseen adverse weather. In: IEEE Conf. Comput. Vis. Pattern Recog. (2020) +3. Brazil, G., Liu, X.: M3d-rpn: Monocular 3d region proposal network for object detection. In: Int. Conf. Comput. Vis. (2019) +4. Brazil, G., Pons-Moll, G., Liu, X., Schiele, B.: Kinematic 3d object detection in monocular video. In: Eur. Conf. Comput. Vis. Springer (2020) +5. Cabon, Y., Murray, N., Humenberger, M.: Virtual kitti 2 (2020) +6. Chen, X., Kundu, K., Zhang, Z., Ma, H., Fidler, S., Urtasun, R.: Monocular 3d object detection for autonomous driving. In: IEEE Conf. Comput. Vis. Pattern Recog. (2016) +7. Chen, Y., Tai, L., Sun, K., Li, M.: Monopair: Monocular 3d object detection using pairwise spatial relationships. In: IEEE Conf. Comput. Vis. Pattern Recog. (2020) +8. Ding, M., Huo, Y., Yi, H., Wang, Z., Shi, J., Lu, Z., Luo, P.: Learning depth-guided convolutions for monocular 3d object detection. In: IEEE Conf. Comput. Vis. Pattern Recog. Worksh. (2020) +9. Esser, P., Rombach, R., Ommer, B.: Taming transformers for high-resolution image synthesis. In: IEEE Conf. Comput. Vis. Pattern Recog. (2021) +10. Fu, H., Gong, M., Wang, C., Batmanghelich, K., Tao, D.: Deep ordinal regression network for monocular depth estimation. In: IEEE Conf. Comput. Vis. Pattern Recog. (2018) +1. Gaidon, A., Wang, Q., Cabon, Y., Vig, E.: Virtual worlds as proxy for multi-object tracking analysis. In: IEEE Conf. Comput. Vis. Pattern Recog. (2016) +2. Geiger, A., Lenz, P., Urtasun, R.: Are we ready for autonomous driving? the kitti vision benchmark suite. In: IEEE Conf. Comput. Vis. Pattern Recog. (2012). https://doi.org/10.1109/CVPR.2012.6248074 +3. Hahner, M., Sakaridis, C., Dai, D., Van Gool, L.: Fog simulation on real lidar point clouds for 3d object detection in adverse weather. In: Int. Conf. Comput. Vis. (2021) +4. Hamilton, B., Tefft, B., Arnold, L., Grabowski, J.: Hidden highways: Fog and traffic crashes on america's roads (november 2014). Montana 40 (2006) +5. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. In: Adv. Neural Inform. Process. Syst. vol. 33 (2020) + +16. Huang, K.C., Wu, T.H., Su, H.T., Hsu, W.H.: Monodtr: Monocular 3d object detection with depth-aware transformer. In: IEEE Conf. Comput. Vis. Pattern Recog. (2022) +17. Juneja, A., Kumar, V., Singla, S.K.: A systematic review on foggy datasets: Applications and challenges. Archives of Computational Methods in Engineering 29(3) (2022) +18. Kim, J.U., Kim, H.I., Ro, Y.M.: Stereoscopic vision recalling memory for monocular 3d object detection. IEEE Trans. Image Process. 32 (2023). https://doi.org/10.1109/TIP.2023.3274479 +19. Kim, J.U., Park, S., Ro, Y.M.: Robust small-scale pedestrian detection with cued recall via memory learning. In: Int. Conf. Comput. Vis. (2021) +20. Kim, J.U., Park, S., Ro, Y.M.: Uncertainty-guided cross-modal learning for robust multispectral pedestrian detection. IEEE Trans. Circuit Syst. Video Technol. 32(3) (2021) +21. Kim, J.U., Park, S., Ro, Y.M.: Towards versatile pedestrian detector with multisensory-matching and multispectral recalling memory. In: AAAI. vol. 36 (2022) +22. Kim, S.W., Brown, B., Yin, K., Kreis, K., Schwarz, K., Li, D., Rombach, R., Torralba, A., Fidler, S.: Neuralfield-ldm: Scene generation with hierarchical latent diffusion models. In: IEEE Conf. Comput. Vis. Pattern Recog. (2023) +23. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014) +24. Lang, A.H., Vora, S., Caesar, H., Zhou, L., Yang, J., Beijbom, O.: Pointpillars: Fast encoders for object detection from point clouds. In: IEEE Conf. Comput. Vis. Pattern Recog. (2019) +25. Liu, X., Xue, N., Wu, T.: Learning auxiliary monocular contexts helps monocular 3d object detection. In: AAAI. vol. 36 (2022) +26. Liu, Z., Wu, Z., Toth, R.: Smoke: Single-stage monocular 3d object detection via keypoint estimation. In: IEEE Conf. Comput. Vis. Pattern Recog. Worksh. (2020) +27. Liu, Z., Zhou, D., Lu, F., Fang, J., Zhang, L.: Autoshape: Real-time shape-aware monocular 3d object detection. In: Int. Conf. Comput. Vis. (2021) +28. Lu, Y., Ma, X., Yang, L., Zhang, T., Liu, Y., Chu, Q., Yan, J., Ouyang, W.: Geometry uncertainty projection network for monocular 3d object detection. In: Int. Conf. Comput. Vis. (2021) +29. Lugmayr, A., Danelljan, M., Romero, A., Yu, F., Timofte, R., Van Gool, L.: Repaint: Inpainting using denoising diffusion probabilistic models. In: IEEE Conf. Comput. Vis. Pattern Recog. (2022) +30. Luo, J., Li, Y., Pan, Y., Yao, T., Feng, J., Chao, H., Mei, T.: Semantic-conditional diffusion networks for image captioning. In: IEEE Conf. Comput. Vis. Pattern Recog. (2023) +31. Luo, S., Dai, H., Shao, L., Ding, Y.: M3dssd: Monocular 3d single stage object detector. In: IEEE Conf. Comput. Vis. Pattern Recog. (2021) +32. Ma, X., Liu, S., Xia, Z., Zhang, H., Zeng, X., Ouyang, W.: Rethinking pseudo-lidar representation. In: Eur. Conf. Comput. Vis. Springer (2020) +33. Mai, N.A.M., Duthon, P., Khoudour, L., Crouzil, A., Velastin, S.A.: 3d object detection with sls-fusion network in foggy weather conditions. Sensors 21(20) (2021) +34. Mousavian, A., Anguelov, D., Flynn, J., Kosecka, J.: 3d bounding box estimation using deep learning and geometry. In: IEEE Conf. Comput. Vis. Pattern Recog. (2017) + +35. Ni, H., Shi, C., Li, K., Huang, S.X., Min, M.R.: Conditional image-to-video generation with latent flow diffusion models. In: IEEE Conf. Comput. Vis. Pattern Recog. (2023) +36. van den Oord, A., Vinyals, O., Kavukcuoglu, K.: Neural discrete representation learning (2018) +37. Peng, L., Wu, X., Yang, Z., Liu, H., Cai, D.: Did-m3d: Decoupling instance depth for monocular 3d object detection. In: Eur. Conf. Comput. Vis. Springer (2022) +38. Pfeuffer, A., Dietmayer, K.: Robust semantic segmentation in adverse weather conditions by means of sensor data fusion. In: 2019 22th International Conference on Information Fusion (FUSION) (2019). https://doi.org/10.23919/FUSION43075.2019.9011192 +39. Qin, Q., Chang, K., Huang, M., Li, G.: Denet: Detection-driven enhancement network for object detection under adverse weather conditions. In: Asian Conf. Comput. Vis. (2022) +40. Qin, Z., Li, X.: Monoground: Detecting monocular 3d objects from the ground. In: IEEE Conf. Comput. Vis. Pattern Recog. (2022) +41. Reading, C., Harakeh, A., Chae, J., Waslander, S.L.: Categorical depth distribution network for monocular 3d object detection. In: IEEE Conf. Comput. Vis. Pattern Recog. (2021) +42. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: IEEE Conf. Comput. Vis. Pattern Recog. (2022) +43. Sakaridis, C., Dai, D., Van Gool, L.: Semantic foggy scene understanding with synthetic data. Int. J. Comput. Vis. 126 (2018) +44. Shi, S., Guo, C., Jiang, L., Wang, Z., Shi, J., Wang, X., Li, H.: Pv-rcnn: Point-voxel feature set abstraction for 3d object detection. In: IEEE Conf. Comput. Vis. Pattern Recog. (2020) +45. Shi, S., Wang, X., Li, H.: Pointcnn: 3d object proposal generation and detection from point cloud. In: IEEE Conf. Comput. Vis. Pattern Recog. (2019) +46. Shi, X., Ye, Q., Chen, X., Chen, C., Chen, Z., Kim, T.K.: Geometry-based distance decomposition for monocular 3d object detection. In: Int. Conf. Comput. Vis. (2021) +47. Simonelli, A., Bulo, S.R., Porzi, L., Lopez-Antequera, M., Kontschieder, P.: Disentangling monocular 3d object detection. In: Int. Conf. Comput. Vis. (2019) +48. Wang, L., Du, L., Ye, X., Fu, Y., Guo, G., Xue, X., Feng, J., Zhang, L.: Depth-conditioned dynamic message propagation for monocular 3d object detection. In: IEEE Conf. Comput. Vis. Pattern Recog. (2021) +49. Wu, R.Q., Duan, Z.P., Guo, C.L., Chai, Z., Li, C.: Ridcp: Revitalizing real image dehazing via high-quality codebook priors. In: IEEE Conf. Comput. Vis. Pattern Recog. (2023) +50. Yang, S., Scherer, S.: Cubeslam: Monocular 3-d object slam. IEEE Transactions on Robotics 35(4) (2019). https://doi.org/10.1109/TR0.2019.2909168 +51. Yang, X., Mi, M.B., Yuan, Y., Wang, X., Tan, R.T.: Object detection in foggy scenes by embedding depth and reconstruction into domain adaptation. In: Asian Conf. Comput. Vis. (2022) +52. Yu, F., Wang, D., Shelhamer, E., Darrell, T.: Deep layer aggregation. In: IEEE Conf. Comput. Vis. Pattern Recog. (2018) +53. Yurtsever, E., Lambert, J., Carballo, A., Takeda, K.: A survey of autonomous driving: Common practices and emerging technologies. IEEE Access 8 (2020). https://doi.org/10.1109/ACCESS.2020.2983149 + +54. Zhang, C., Wang, H., Cai, Y., Chen, L., Li, Y., Sotelo, M.A., Li, Z.: Robust-fusionnet: Deep multimodal sensor fusion for 3-d object detection under severe weather conditions. IEEE Transactions on Instrumentation and Measurement 71 (2022). https://doi.org/10.1109/TIM.2022.3191724 +55. Zhang, R., Qiu, H., Wang, T., Guo, Z., Cui, Z., Qiao, Y., Li, H., Gao, P.: Monodetr: Depth-guided transformer for monocular 3d object detection. In: Int. Conf. Comput. Vis. (2023) +56. Zhou, Y., Tuzel, O.: Voxelnet: End-to-end learning for point cloud based 3d object detection. In: IEEE Conf. Comput. Vis. Pattern Recog. (2018) +57. Zhou, Y., He, Y., Zhu, H., Wang, C., Li, H., Jiang, Q.: Monocular 3d object detection: An extrinsic parameter free approach. In: IEEE Conf. Comput. Vis. Pattern Recog. (2021) \ No newline at end of file diff --git a/2024/MonoWAD_ Weather-Adaptive Diffusion Model for Robust Monocular 3D Object Detection/images.zip b/2024/MonoWAD_ Weather-Adaptive Diffusion Model for Robust Monocular 3D Object Detection/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..5b3dea89f5da5c7d1b694a6000ce69518fdfa06b --- /dev/null +++ b/2024/MonoWAD_ Weather-Adaptive Diffusion Model for Robust Monocular 3D Object Detection/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ba630e9abcd9904fa8bba5a36e2b70df6967262edb91ecd6ebb1f6556d9b4fd +size 713153 diff --git a/2024/MonoWAD_ Weather-Adaptive Diffusion Model for Robust Monocular 3D Object Detection/layout.json b/2024/MonoWAD_ Weather-Adaptive Diffusion Model for Robust Monocular 3D Object Detection/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..79165f7987e54edc2120b0fdbcf4a5df806a9de0 --- /dev/null +++ b/2024/MonoWAD_ Weather-Adaptive Diffusion Model for Robust Monocular 3D Object Detection/layout.json @@ -0,0 +1,10454 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 140, + 111, + 474, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 111, + 474, + 148 + ], + "spans": [ + { + "bbox": [ + 140, + 111, + 474, + 148 + ], + "type": "text", + "content": "MonoWAD: Weather-Adaptive Diffusion Model for Robust Monocular 3D Object Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 136, + 167, + 478, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 167, + 478, + 181 + ], + "spans": [ + { + "bbox": [ + 136, + 167, + 478, + 181 + ], + "type": "text", + "content": "Youngmin Oh1, Hyung-Il Kim2, Seong Tae Kim1†, and Jung Uk Kim1†" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 210, + 189, + 403, + 234 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 210, + 189, + 403, + 234 + ], + "spans": [ + { + "bbox": [ + 210, + 189, + 403, + 234 + ], + "type": "text", + "content": "1 Kyung Hee University, Yong-in, South Korea {oym9104, st.kim, ju.kim}@khu.ac.kr 2 ETRI, Daejeon, South Korea hikim@etri.re.kr" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 263, + 455, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 263, + 455, + 473 + ], + "spans": [ + { + "bbox": [ + 160, + 263, + 455, + 473 + ], + "type": "text", + "content": "Abstract. Monocular 3D object detection is an important challenging task in autonomous driving. Existing methods mainly focus on performing 3D detection in ideal weather conditions, characterized by scenarios with clear and optimal visibility. However, the challenge of autonomous driving requires the ability to handle changes in weather conditions, such as foggy weather, not just clear weather. We introduce MonoWAD, a novel weather-robust monocular 3D object detector with a weather-adaptive diffusion model. It contains two components: (1) the weather codebook to memorize the knowledge of the clear weather and generate a weather-reference feature for any input, and (2) the weather-adaptive diffusion model to enhance the feature representation of the input feature by incorporating a weather-reference feature. This serves as an attention role in indicating how much improvement is needed for the input feature according to the weather conditions. To achieve this goal, we introduce a weather-adaptive enhancement loss to enhance the feature representation under both clear and foggy weather conditions. Extensive experiments under various weather conditions demonstrate that MonoWAD achieves weather-robust monocular 3D object detection. The code and dataset are released at https://github.com/VisualAIKHU/MonoWAD." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 159, + 483, + 453, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 483, + 453, + 506 + ], + "spans": [ + { + "bbox": [ + 159, + 483, + 453, + 506 + ], + "type": "text", + "content": "Keywords: Monocular 3D Object Detection " + }, + { + "bbox": [ + 159, + 483, + 453, + 506 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 483, + 453, + 506 + ], + "type": "text", + "content": " Weather-Adaptive Diffusion " + }, + { + "bbox": [ + 159, + 483, + 453, + 506 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 483, + 453, + 506 + ], + "type": "text", + "content": " Weather Codebook" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 525, + 230, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 525, + 230, + 538 + ], + "spans": [ + { + "bbox": [ + 132, + 525, + 230, + 538 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 550, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 550, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 550, + 482, + 647 + ], + "type": "text", + "content": "Monocular 3D object detection aims to detect 3D objects using only a single camera [16, 18, 28, 46, 55, 57]. In contrast to LiDAR-based methods that rely on expensive LiDAR sensors for depth estimation [24, 44, 45, 56], and stereo-based methods that require synchronized stereo cameras, monocular 3D object detection only requires monocular images, offering the advantage of computational cost-effectiveness and requiring fewer resources. Due to this characteristic, the monocular 3D object detection is applied to a wide range of real-world applications, such as autonomous driving [12, 19, 53] and robotics [50]." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 653, + 234, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 234, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 234, + 666 + ], + "type": "text", + "content": "† Corresponding author" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 129, + 118, + 485, + 202 + ], + "blocks": [ + { + "bbox": [ + 129, + 118, + 485, + 202 + ], + "lines": [ + { + "bbox": [ + 129, + 118, + 485, + 202 + ], + "spans": [ + { + "bbox": [ + 129, + 118, + 485, + 202 + ], + "type": "image", + "image_path": "4096afacdb8773efa89658e3f125d0abdc0e7e5a72b2515d361cce17f37609ca.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 202, + 482, + 258 + ], + "lines": [ + { + "bbox": [ + 130, + 202, + 482, + 258 + ], + "spans": [ + { + "bbox": [ + 130, + 202, + 482, + 258 + ], + "type": "text", + "content": "Fig. 1: Conceptual diagram of the proposed method (foggy example). (a) In the training phase, weather codebook learns the clear knowledge to transfer it to weather-adaptive diffusion model to enhance content related to the weather conditions. (b) By doing so, even with input images under various weather conditions (e.g., foggy images), monocular 3D object detection becomes adaptable to various weather scenarios." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 270, + 482, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 270, + 482, + 378 + ], + "spans": [ + { + "bbox": [ + 130, + 270, + 482, + 378 + ], + "type": "text", + "content": "However, existing monocular 3D object detectors mainly focus on ideal autonomous driving environments (i.e., clear weather). There are challenges in applying them to real-world scenarios with adverse weather conditions, such as fog and rain. Among these, fog poses the most significant challenge compared to other weather [14, 17]. This is due to the dense and diffuse nature of fog, which strongly scatters and absorbs light, leading to difficulties in object detection [2]. Since monocular 3D object detection relies solely on visual information from a single image, unlike LiDAR, it is crucial to design detectors to achieve enhanced performance in challenging visibility scenarios." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 379, + 482, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 379, + 482, + 510 + ], + "spans": [ + { + "bbox": [ + 130, + 379, + 482, + 510 + ], + "type": "text", + "content": "In this paper, we propose MonoWAD, a novel weather-robust monocular 3D object detector to address the aforementioned issues. As mentioned earlier, due to the inherent challenges posed by foggy weather among various adverse weather conditions, we focus primarily on clear and foggy weather (results for other weather conditions such as rainy and sunset are presented in Section 4). For weather-robust object detection, clear weather requires relatively modest improvements to enhance the visual representation of features. In contrast, significant enhancements in this feature representation are required for foggy weather. To address this, we consider two key aspects: (1) how to quantify the degree of improvement needed for the input image, and (2) how to guide the representation of the input image." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 510, + 482, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 640 + ], + "type": "text", + "content": "To address the two key aspects, as shown in Fig. 1, our MonoWAD consists of a weather codebook and weather-adaptive diffusion model. First, we introduce the weather codebook to generate a weather-reference feature that contains knowledge about reference weather in a given scene. The reference weather acts as a guide, indicating the degree of weather improvement required. Since the clear weather contains a richer visual representation of objects, we adopt it as a reference weather. At this time, we devise a clear knowledge recalling (CKR) loss to guide the weather codebook to memorize information about clear weather and generate a weather-reference feature for any input (clear or foggy). As a result, our detector can understand where improvements are needed in the input features based on the weather-reference feature." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": "Second, we propose a weather-adaptive diffusion model to effectively enhance feature representations in accordance with weather conditions. Given input fea-" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "text", + "content": "Y. Oh et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 247 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 247 + ], + "type": "text", + "content": "ture (clear or foggy), the weather-adaptive diffusion model dynamically enhances the representation of the input feature based on the weather-reference feature. The weather-reference feature plays a role of attention, determining the extent to which the input features need improvement. At this time, we define the difference between clear and foggy weather (i.e., weather changes) as the fog distribution to adopt it as the noise for our diffusion model. With fog distribution, our weather-adaptive diffusion model can enhance the feature representation according to the weather conditions through multiple steps of reverse processes. To achieve this goal, we introduce a weather-adaptive enhancement (WAE) loss. As a result, our MonoWAD performs weather-robust detection by adaptively improving feature representation according to the weather conditions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 248, + 480, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 248, + 480, + 342 + ], + "spans": [ + { + "bbox": [ + 130, + 248, + 480, + 342 + ], + "type": "text", + "content": "To adaptively enhance the feature representation through the difference between weather conditions, we generate a new foggy KITTI dataset based on the KITTI dataset [12]. Comprehensive experimental results on several datasets [11, 12] show that our MonoWAD outperforms the existing state-of-the-art monocular 3D object detectors [16,28,37,40,55] under foggy weather, which is the most challenging weather condition. While our method primarily focuses on foggy, experiments conducted under various weather conditions (e.g., foggy, rainy, and sunset) have demonstrated its applicability to other weather scenarios." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 146, + 343, + 444, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 343, + 444, + 355 + ], + "spans": [ + { + "bbox": [ + 146, + 343, + 444, + 355 + ], + "type": "text", + "content": "The main contributions of our paper can be summarized as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 363, + 479, + 445 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 138, + 363, + 479, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 363, + 479, + 384 + ], + "spans": [ + { + "bbox": [ + 138, + 363, + 479, + 384 + ], + "type": "text", + "content": "- We introduce a new weather-robust monocular 3D object detector, called MonoWAD, that is robust to various weather conditions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 387, + 479, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 387, + 479, + 409 + ], + "spans": [ + { + "bbox": [ + 138, + 387, + 479, + 409 + ], + "type": "text", + "content": "- We design a weather codebook with clear knowledge recalling loss for learning about clear weather, providing reference information for enhancement." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 411, + 479, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 411, + 479, + 445 + ], + "spans": [ + { + "bbox": [ + 138, + 411, + 479, + 445 + ], + "type": "text", + "content": "- We propose weather-adaptive diffusion model with weather-adaptive enhancement loss to dynamically enhance the feature representation of the input images according to the weather conditions." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 464, + 237, + 477 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 464, + 237, + 477 + ], + "spans": [ + { + "bbox": [ + 132, + 464, + 237, + 477 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 490, + 323, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 490, + 323, + 502 + ], + "spans": [ + { + "bbox": [ + 132, + 490, + 323, + 502 + ], + "type": "text", + "content": "2.1 Monocular 3D Object Detection" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 510, + 480, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 480, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 480, + 640 + ], + "type": "text", + "content": "Monocular 3D object detection task can be categorized into two directions according to the type of data used in the training phase: (1) using only a monocular image and (2) incorporating additional data, such as depth along with a monocular image. The first category relies on the geometric relationship between 2D and 3D [3,7,25,26,28,31,34,40,46,55,57]. For example, Deep3Dbox [34] utilizes the geometric information of 2D bounding boxes to predict 3D bounding boxes. In [3], M3D-RPN was proposed to understand the 3D scene from the depth-aware convolution. MonoRCNN [46] predicted 3D bounding boxes through geometry-based distance decomposition, and MonoCon [25] learns mono context for 3D object detection. MonoDETR [55] introduces a depth-guided transformer that utilizes geometric depth cues without requiring additional data." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 642, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 642, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 642, + 480, + 665 + ], + "type": "text", + "content": "Moreover, since monocular image contains limited information for estimating 3D object cues, monocular 3D object detectors have adopted the additional" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "type": "text", + "content": "MonoWAD" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": "data for more robust detection [4,8,16,27,32,37,41,48]. For example, the depth-conditioned dynamic message propagation (DDMP) [48] was proposed to integrate prior depth information with the image context. CaDDN [41] was introduced to utilize the depth distribution of predicted categories for each pixel to project the context information onto 3D space, deriving 3D bounding boxes. MonoDTR [16] employs the transformer architecture to integrate depth features and context, thus estimating more accurate depth information." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 200, + 482, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 200, + 482, + 272 + ], + "spans": [ + { + "bbox": [ + 130, + 200, + 482, + 272 + ], + "type": "text", + "content": "Despite recent progress, the existing monocular 3D object detectors mainly rely on the benchmark data collected under clear weather conditions. However, it is essential to account for challenging adverse weather conditions, such as fog, to more accurately reflect real-world scenarios. In this paper, we aim to introduce weather-robust 3D object detection by enhancing visual features with the proposed weather-adaptive diffusion model and weather codebook." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 281, + 376, + 294 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 281, + 376, + 294 + ], + "spans": [ + { + "bbox": [ + 131, + 281, + 376, + 294 + ], + "type": "text", + "content": "2.2 Computer Vision Tasks for Foggy Weather" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 301, + 482, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 301, + 482, + 504 + ], + "spans": [ + { + "bbox": [ + 130, + 301, + 482, + 504 + ], + "type": "text", + "content": "There have been a lot of studies on improving performance in various weather conditions for real-world application of computer vision technology [2, 13, 20, 21, 33, 38, 43, 51, 54]. In particular, fog is considered one of the most critical issues due to its significant degradation of visual information compared to other weather conditions. To deal with the foggy weather conditions, the authors in [43] generated synthetic fog images (so-called Foggy Cityscapes) from clear weather images, and utilized them for training semantic segmentation and 2D object detection. Similar to Foggy Cityscapes, Martin et al. [13] naturally synthesizes fog into LIDAR to enhance the performance of LiDAR-based 3D object detectors in foggy weather. Bijelic et al. [2] use real dataset including foggy weather, for 2D detection with the multimodal fusion networks. Mai et al. [33] synthesize fog for LiDAR and stereo images and perform fusion-based 3D object detection using the SLS-Fusion network. Xin et al. [51] focused on 2D detection in foggy weather by applying domain adaptation. In this context, we focus on monocular 3D object detection in scenarios that rely solely on visual information in challenging foggy environments. To address this challenge, we propose a weather-robust diffusion model that dynamically improves features based on reference feature." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 514, + 246, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 514, + 246, + 525 + ], + "spans": [ + { + "bbox": [ + 131, + 514, + 246, + 525 + ], + "type": "text", + "content": "2.3 Diffusion Models" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "Recently, the diffusion model [15] has attracted considerable interest in computer vision due to its impressive progress in image generation [22, 29, 35, 42], as well as its potential application to other vision tasks such as segmentation [1] and image captioning [30]. Inspired by the remarkable generative ability, we design the diffusion model for robust monocular 3D object detection by considering a foggy effect (which is one of the challenging adverse weather conditions for monocular 3D object detection) as a form of noise in the model. That is, we propose a method in which visual features obscured by fog are progressively improved by training a diffusion model based on the forward/reverse diffusion process. In particular, we present an adaptive method that allows the diffusion model to control the degree of improvement by weather conditions." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "text", + "content": "Y. Oh et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 116, + 479, + 213 + ], + "blocks": [ + { + "bbox": [ + 134, + 116, + 479, + 213 + ], + "lines": [ + { + "bbox": [ + 134, + 116, + 479, + 213 + ], + "spans": [ + { + "bbox": [ + 134, + 116, + 479, + 213 + ], + "type": "image", + "image_path": "2105952e910218bb8044c128b12c704c3f65278f2fb9df10aeaedb76fc42a5ae.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 219, + 482, + 264 + ], + "lines": [ + { + "bbox": [ + 130, + 219, + 482, + 264 + ], + "spans": [ + { + "bbox": [ + 130, + 219, + 482, + 264 + ], + "type": "text", + "content": "Fig. 2: Overview of our MonoWAD in the inference phase. It mainly contains three parts: weather codebook, weather-adaptive diffusion model, and detection block. Through the weather codebook and weather-adaptive diffusion model, our method can maintain robustness against various weather conditions (i.e., clear or foggy)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 275, + 261, + 289 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 275, + 261, + 289 + ], + "spans": [ + { + "bbox": [ + 132, + 275, + 261, + 289 + ], + "type": "text", + "content": "3 Proposed Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 304, + 482, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 304, + 482, + 422 + ], + "spans": [ + { + "bbox": [ + 130, + 304, + 482, + 422 + ], + "type": "text", + "content": "Fig. 2 shows the overall framework of the proposed MonoWAD in the inference phase. A backbone network receives an input image (clear image " + }, + { + "bbox": [ + 130, + 304, + 482, + 422 + ], + "type": "inline_equation", + "content": "I_{c}" + }, + { + "bbox": [ + 130, + 304, + 482, + 422 + ], + "type": "text", + "content": " or foggy image " + }, + { + "bbox": [ + 130, + 304, + 482, + 422 + ], + "type": "inline_equation", + "content": "I_{f}" + }, + { + "bbox": [ + 130, + 304, + 482, + 422 + ], + "type": "text", + "content": ") to encode the corresponding input feature (input clear feature " + }, + { + "bbox": [ + 130, + 304, + 482, + 422 + ], + "type": "inline_equation", + "content": "x^{c}" + }, + { + "bbox": [ + 130, + 304, + 482, + 422 + ], + "type": "text", + "content": " or input foggy feature " + }, + { + "bbox": [ + 130, + 304, + 482, + 422 + ], + "type": "inline_equation", + "content": "x^{f}" + }, + { + "bbox": [ + 130, + 304, + 482, + 422 + ], + "type": "text", + "content": "). It interacts with the weather codebook " + }, + { + "bbox": [ + 130, + 304, + 482, + 422 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}" + }, + { + "bbox": [ + 130, + 304, + 482, + 422 + ], + "type": "text", + "content": " to generate weather-reference feature " + }, + { + "bbox": [ + 130, + 304, + 482, + 422 + ], + "type": "inline_equation", + "content": "x^{r}" + }, + { + "bbox": [ + 130, + 304, + 482, + 422 + ], + "type": "text", + "content": ", indicating the amount of enhancement for the given input feature. Subsequently, the weather-adaptive diffusion model attempts to enhance the input feature over " + }, + { + "bbox": [ + 130, + 304, + 482, + 422 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 304, + 482, + 422 + ], + "type": "text", + "content": " timesteps to obtain an enhanced feature " + }, + { + "bbox": [ + 130, + 304, + 482, + 422 + ], + "type": "inline_equation", + "content": "\\tilde{x}^{c}" + }, + { + "bbox": [ + 130, + 304, + 482, + 422 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 130, + 304, + 482, + 422 + ], + "type": "inline_equation", + "content": "\\tilde{x}^{f}" + }, + { + "bbox": [ + 130, + 304, + 482, + 422 + ], + "type": "text", + "content": ". Finally, monocular 3D object detection is performed through the detection block. Note that, in the training phase, our MonoWAD use the clear feature " + }, + { + "bbox": [ + 130, + 304, + 482, + 422 + ], + "type": "inline_equation", + "content": "x^{c}" + }, + { + "bbox": [ + 130, + 304, + 482, + 422 + ], + "type": "text", + "content": " to train our diffusion as well as the detection block." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 424, + 482, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 424, + 482, + 472 + ], + "spans": [ + { + "bbox": [ + 130, + 424, + 482, + 472 + ], + "type": "text", + "content": "We address two key issues: (1) how to guide weather-reference feature " + }, + { + "bbox": [ + 130, + 424, + 482, + 472 + ], + "type": "inline_equation", + "content": "x^r" + }, + { + "bbox": [ + 130, + 424, + 482, + 472 + ], + "type": "text", + "content": " to serve as a reference feature, and (2) how to guide the weather-adaptive diffusion model to effectively enhance feature representations based on weather conditions. Details are in the following subsections." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 486, + 258, + 497 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 486, + 258, + 497 + ], + "spans": [ + { + "bbox": [ + 132, + 486, + 258, + 497 + ], + "type": "text", + "content": "3.1 Weather Codebook" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 509, + 482, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 509, + 482, + 592 + ], + "spans": [ + { + "bbox": [ + 130, + 509, + 482, + 592 + ], + "type": "text", + "content": "In foggy weather conditions, the overall visual quality of the scene is generally poor, requiring a significant enhancement. Conversely, in clear weather, the amount of improvement is expected to be relatively minimal compared to foggy conditions. Therefore, inspired by [9,36], we devise a weather codebook " + }, + { + "bbox": [ + 130, + 509, + 482, + 592 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}" + }, + { + "bbox": [ + 130, + 509, + 482, + 592 + ], + "type": "text", + "content": " to provide the reference knowledge about the weather for appropriate enhancement based on the weather conditions. At this time, as the clear weather contains abundant visual representations, we use it as a reference weather knowledge." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "As shown in Fig. 3, the reference knowledge embedding procedure involves receiving paired clear-foggy features during the training phase. The weather codebook " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " consists of " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " learnable slots, denoted as " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{Z} = \\{z_k\\}_{k=1}^K (z_k \\in \\mathbb{R}^{1 \\times c})" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " represents the dimensionality of each slot. The paired clear feature " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "x^c" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " and foggy feature " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "x^f" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " pass through a convolution layer to generate " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\hat{x}^c \\in \\mathbb{R}^{h \\times w \\times c}" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\hat{x}^f \\in \\mathbb{R}^{h \\times w \\times c}" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " denotes width and " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " indicates height). Each element of the" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "type": "text", + "content": "MonoWAD" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 130, + 116, + 482, + 271 + ], + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 271 + ], + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 271 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 271 + ], + "type": "image", + "image_path": "a461609d2f407e75e3eec09b04ff65dc201e192f796c9785c68517f768250a5b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 274, + 482, + 319 + ], + "lines": [ + { + "bbox": [ + 130, + 274, + 482, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 274, + 482, + 319 + ], + "type": "text", + "content": "Fig. 3: Illustration of the proposed (a) weather-invariant guiding (WIG) loss and (b) clear knowledge embedding (CKE) loss. The clear knowledge recalling (CKR) loss, obtained from combining WIG and CKE, aims to memorize the knowledge of the clear weather and recall the same clear knowledge from the foggy weather." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 327, + 482, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 327, + 482, + 366 + ], + "spans": [ + { + "bbox": [ + 130, + 327, + 482, + 366 + ], + "type": "text", + "content": "feature denoted as " + }, + { + "bbox": [ + 130, + 327, + 482, + 366 + ], + "type": "inline_equation", + "content": "\\hat{x}_{ij}^{c}\\in \\mathbb{R}^{1\\times c}" + }, + { + "bbox": [ + 130, + 327, + 482, + 366 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 327, + 482, + 366 + ], + "type": "inline_equation", + "content": "\\hat{x}_{ij}^{f}\\in \\mathbb{R}^{1\\times c}" + }, + { + "bbox": [ + 130, + 327, + 482, + 366 + ], + "type": "text", + "content": ". Subsequently, we obtain weather-reference feature for clear weather, " + }, + { + "bbox": [ + 130, + 327, + 482, + 366 + ], + "type": "inline_equation", + "content": "x^{r(c)}\\in \\mathbb{R}^{h\\times w\\times c}" + }, + { + "bbox": [ + 130, + 327, + 482, + 366 + ], + "type": "text", + "content": ", by conducting element-wise quantization process " + }, + { + "bbox": [ + 130, + 327, + 482, + 366 + ], + "type": "inline_equation", + "content": "\\mathbf{q}(\\cdot)" + }, + { + "bbox": [ + 130, + 327, + 482, + 366 + ], + "type": "text", + "content": ", calculated as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 219, + 374, + 481, + 401 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 219, + 374, + 481, + 401 + ], + "spans": [ + { + "bbox": [ + 219, + 374, + 481, + 401 + ], + "type": "interline_equation", + "content": "x ^ {r (c)} = \\mathbf {q} (\\hat {x} ^ {c}) := \\left(\\underset {z _ {k} \\in \\mathcal {Z}} {\\arg \\min } \\| \\hat {x} _ {i j} ^ {c} - z _ {k} \\|\\right). \\tag {1}", + "image_path": "360b9acbaf54ece69814cc013999def8d7d8f4142b097bb72fef42e782f3227f.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 411, + 482, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 411, + 482, + 484 + ], + "spans": [ + { + "bbox": [ + 131, + 411, + 482, + 484 + ], + "type": "text", + "content": "Utilizing " + }, + { + "bbox": [ + 131, + 411, + 482, + 484 + ], + "type": "inline_equation", + "content": "x^{c}" + }, + { + "bbox": [ + 131, + 411, + 482, + 484 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 411, + 482, + 484 + ], + "type": "inline_equation", + "content": "x^{r(c)}" + }, + { + "bbox": [ + 131, + 411, + 482, + 484 + ], + "type": "text", + "content": ", we introduce a clear knowledge embedding (CKE) loss " + }, + { + "bbox": [ + 131, + 411, + 482, + 484 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{cke}" + }, + { + "bbox": [ + 131, + 411, + 482, + 484 + ], + "type": "text", + "content": " to guide " + }, + { + "bbox": [ + 131, + 411, + 482, + 484 + ], + "type": "inline_equation", + "content": "x^{r(c)}" + }, + { + "bbox": [ + 131, + 411, + 482, + 484 + ], + "type": "text", + "content": " to follow the representation of " + }, + { + "bbox": [ + 131, + 411, + 482, + 484 + ], + "type": "inline_equation", + "content": "x^{c}" + }, + { + "bbox": [ + 131, + 411, + 482, + 484 + ], + "type": "text", + "content": ". To this end, we perform global average pooling (GAP) for " + }, + { + "bbox": [ + 131, + 411, + 482, + 484 + ], + "type": "inline_equation", + "content": "x^{c}" + }, + { + "bbox": [ + 131, + 411, + 482, + 484 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 411, + 482, + 484 + ], + "type": "inline_equation", + "content": "x^{r(c)}" + }, + { + "bbox": [ + 131, + 411, + 482, + 484 + ], + "type": "text", + "content": " with softmax, generating " + }, + { + "bbox": [ + 131, + 411, + 482, + 484 + ], + "type": "inline_equation", + "content": "s^{c}" + }, + { + "bbox": [ + 131, + 411, + 482, + 484 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 411, + 482, + 484 + ], + "type": "inline_equation", + "content": "s^{r(c)}" + }, + { + "bbox": [ + 131, + 411, + 482, + 484 + ], + "type": "text", + "content": ", respectively. Each element in the vector indicates the probability of the significance of each channel. With " + }, + { + "bbox": [ + 131, + 411, + 482, + 484 + ], + "type": "inline_equation", + "content": "s^{c}" + }, + { + "bbox": [ + 131, + 411, + 482, + 484 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 411, + 482, + 484 + ], + "type": "inline_equation", + "content": "s^{r(c)}" + }, + { + "bbox": [ + 131, + 411, + 482, + 484 + ], + "type": "text", + "content": ", we employ KL divergence " + }, + { + "bbox": [ + 131, + 411, + 482, + 484 + ], + "type": "inline_equation", + "content": "D_{KL}(\\cdot)" + }, + { + "bbox": [ + 131, + 411, + 482, + 484 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 131, + 411, + 482, + 484 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{cke}" + }, + { + "bbox": [ + 131, + 411, + 482, + 484 + ], + "type": "text", + "content": " to compare the probability distributions, formulated as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 255, + 491, + 481, + 507 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 491, + 481, + 507 + ], + "spans": [ + { + "bbox": [ + 255, + 491, + 481, + 507 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {c k e} = D _ {K L} \\left(s ^ {c} \\| s ^ {r (c)}\\right). \\tag {2}", + "image_path": "0a93c855431d12e1a4043463fbad9efc8262acb0281af0d8a7c85aa666462d72.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 515, + 481, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 515, + 481, + 539 + ], + "spans": [ + { + "bbox": [ + 130, + 515, + 481, + 539 + ], + "type": "text", + "content": "Through " + }, + { + "bbox": [ + 130, + 515, + 481, + 539 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{cke}" + }, + { + "bbox": [ + 130, + 515, + 481, + 539 + ], + "type": "text", + "content": ", the weather codebook " + }, + { + "bbox": [ + 130, + 515, + 481, + 539 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}" + }, + { + "bbox": [ + 130, + 515, + 481, + 539 + ], + "type": "text", + "content": " can memorize the knowledge of the clear weather, allowing it to effectively reconstruct the clear weather knowledge." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 540, + 482, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 540, + 482, + 598 + ], + "spans": [ + { + "bbox": [ + 130, + 540, + 482, + 598 + ], + "type": "text", + "content": "Additionally, as the paired clear-foggy images are identical except for weather conditions, the quantization process of the foggy feature " + }, + { + "bbox": [ + 130, + 540, + 482, + 598 + ], + "type": "inline_equation", + "content": "x^{f}" + }, + { + "bbox": [ + 130, + 540, + 482, + 598 + ], + "type": "text", + "content": " with the weather codebook should generate an equivalent weather-reference feature. For obtaining weather-reference feature for foggy " + }, + { + "bbox": [ + 130, + 540, + 482, + 598 + ], + "type": "inline_equation", + "content": "x^{r(f)}" + }, + { + "bbox": [ + 130, + 540, + 482, + 598 + ], + "type": "text", + "content": ", the element-wise quantization process is also conducted for " + }, + { + "bbox": [ + 130, + 540, + 482, + 598 + ], + "type": "inline_equation", + "content": "\\hat{x}^{f}" + }, + { + "bbox": [ + 130, + 540, + 482, + 598 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 540, + 482, + 598 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}" + }, + { + "bbox": [ + 130, + 540, + 482, + 598 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 218, + 605, + 481, + 633 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 218, + 605, + 481, + 633 + ], + "spans": [ + { + "bbox": [ + 218, + 605, + 481, + 633 + ], + "type": "interline_equation", + "content": "x ^ {r (f)} = \\mathbf {q} (\\hat {x} ^ {f}) := \\left(\\underset {z _ {k} \\in \\mathcal {Z}} {\\arg \\min } \\| \\hat {x} _ {i j} ^ {f} - z _ {k} \\|\\right). \\tag {3}", + "image_path": "b4aced1e4aa45c19799416b79b6d8d72c104b5537be1a60db1e445a559c5a777.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "text", + "content": "Next, we introduce the weather-invariant guiding (WIG) loss " + }, + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{wig}" + }, + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "text", + "content": " to guide " + }, + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}" + }, + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "text", + "content": " that the weather codebook recalls the same clear knowledge of clear feature for" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "Y. Oh et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 138, + 116, + 475, + 247 + ], + "blocks": [ + { + "bbox": [ + 138, + 116, + 475, + 247 + ], + "lines": [ + { + "bbox": [ + 138, + 116, + 475, + 247 + ], + "spans": [ + { + "bbox": [ + 138, + 116, + 475, + 247 + ], + "type": "image", + "image_path": "30c2bd0b9fc9be3b440b6c3e119c54193994e8d55260f719d12000e097d5ad66.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 249, + 480, + 283 + ], + "lines": [ + { + "bbox": [ + 132, + 249, + 480, + 283 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 480, + 283 + ], + "type": "text", + "content": "Fig. 4: Training process of the weather-adaptive diffusion model, which consists of two processes: (1) Adding fog variant " + }, + { + "bbox": [ + 132, + 249, + 480, + 283 + ], + "type": "inline_equation", + "content": "\\epsilon_{n}" + }, + { + "bbox": [ + 132, + 249, + 480, + 283 + ], + "type": "text", + "content": " from input clear feature " + }, + { + "bbox": [ + 132, + 249, + 480, + 283 + ], + "type": "inline_equation", + "content": "x^{c}" + }, + { + "bbox": [ + 132, + 249, + 480, + 283 + ], + "type": "text", + "content": " (forward process) and (2) enhancing representation with weather-reference feature " + }, + { + "bbox": [ + 132, + 249, + 480, + 283 + ], + "type": "inline_equation", + "content": "x^{r}" + }, + { + "bbox": [ + 132, + 249, + 480, + 283 + ], + "type": "text", + "content": " (reverse process)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 294, + 372, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 294, + 372, + 305 + ], + "spans": [ + { + "bbox": [ + 132, + 294, + 372, + 305 + ], + "type": "text", + "content": "the foggy feature, which can be represented as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 251, + 315, + 481, + 338 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 251, + 315, + 481, + 338 + ], + "spans": [ + { + "bbox": [ + 251, + 315, + 481, + 338 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {w i g} = \\left\\| x ^ {r (c)} - x ^ {r (f)} \\right\\| _ {2} ^ {2}. \\tag {4}", + "image_path": "9e37d9cc15c2ec8dc0480daa9096a7d1bffe9c533c614bae5ad4aaa6da6ece7d.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 346, + 480, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 346, + 480, + 370 + ], + "spans": [ + { + "bbox": [ + 132, + 346, + 480, + 370 + ], + "type": "text", + "content": "Finally, the clear knowledge recalling (CKR) loss " + }, + { + "bbox": [ + 132, + 346, + 480, + 370 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{ckr}" + }, + { + "bbox": [ + 132, + 346, + 480, + 370 + ], + "type": "text", + "content": " is obtained by adding " + }, + { + "bbox": [ + 132, + 346, + 480, + 370 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{cke}" + }, + { + "bbox": [ + 132, + 346, + 480, + 370 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 346, + 480, + 370 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{wig}" + }, + { + "bbox": [ + 132, + 346, + 480, + 370 + ], + "type": "text", + "content": ", which is defined as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 261, + 380, + 480, + 392 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 261, + 380, + 480, + 392 + ], + "spans": [ + { + "bbox": [ + 261, + 380, + 480, + 392 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {c k r} = \\mathcal {L} _ {c k e} + \\mathcal {L} _ {w i g}. \\tag {5}", + "image_path": "1f81b253088cd8e2a3d143f75f6f17373662016d561a7db9b5720a6654ba76d6.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 400, + 480, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 400, + 480, + 449 + ], + "spans": [ + { + "bbox": [ + 130, + 400, + 480, + 449 + ], + "type": "text", + "content": "In the training phase, the weight parameters of embedding " + }, + { + "bbox": [ + 130, + 400, + 480, + 449 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 400, + 480, + 449 + ], + "type": "text", + "content": " slots of weather codebook " + }, + { + "bbox": [ + 130, + 400, + 480, + 449 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}" + }, + { + "bbox": [ + 130, + 400, + 480, + 449 + ], + "type": "text", + "content": " are initialized randomly and they are updated through Eq. (5). In the inference phase, all parameters are fixed to recall clear weather, generating weather-reference features for any weather conditions." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 466, + 337, + 478 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 466, + 337, + 478 + ], + "spans": [ + { + "bbox": [ + 132, + 466, + 337, + 478 + ], + "type": "text", + "content": "3.2 Weather-Adaptive Diffusion Model" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 486, + 480, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 480, + 533 + ], + "type": "text", + "content": "Through Section 3.1, we now know that the weather codebook outputs the weather-reference feature " + }, + { + "bbox": [ + 130, + 486, + 480, + 533 + ], + "type": "inline_equation", + "content": "x^{r(c)}" + }, + { + "bbox": [ + 130, + 486, + 480, + 533 + ], + "type": "text", + "content": " for the clear weather and " + }, + { + "bbox": [ + 130, + 486, + 480, + 533 + ], + "type": "inline_equation", + "content": "x^{r(f)}" + }, + { + "bbox": [ + 130, + 486, + 480, + 533 + ], + "type": "text", + "content": " for the foggy weather through Eq. (5). From now on, since our method can receive any input images (clear or foggy), we denote the weather-reference feature as " + }, + { + "bbox": [ + 130, + 486, + 480, + 533 + ], + "type": "inline_equation", + "content": "x^r" + }, + { + "bbox": [ + 130, + 486, + 480, + 533 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 534, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 534, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 534, + 481, + 665 + ], + "type": "text", + "content": "Fig. 4 shows the training process of the proposed weather-adaptive diffusion model. The key idea of the diffusion model [15,42] is to gradually enhance " + }, + { + "bbox": [ + 130, + 534, + 481, + 665 + ], + "type": "inline_equation", + "content": "x_{T}" + }, + { + "bbox": [ + 130, + 534, + 481, + 665 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 534, + 481, + 665 + ], + "type": "inline_equation", + "content": "x_{0}" + }, + { + "bbox": [ + 130, + 534, + 481, + 665 + ], + "type": "text", + "content": " with fixed Markov Chain of " + }, + { + "bbox": [ + 130, + 534, + 481, + 665 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 534, + 481, + 665 + ], + "type": "text", + "content": " timesteps. To this end, the forward and reverse processes are conducted in the training phase, and only the reverse process is used for the inference phase. Motivated by [15,42], we construct the weather-adaptive diffusion model to enhance representation related to the weather conditions. To this end, unlike traditional diffusion methods [15,42] that adopt the Gaussian noise to the image or latent space, we adopt " + }, + { + "bbox": [ + 130, + 534, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{F} = x^{f} - x^{c}" + }, + { + "bbox": [ + 130, + 534, + 481, + 665 + ], + "type": "text", + "content": ", called fog distribution, to guide our diffusion model to be aware of the foggy weather. Ideally, the information contained within " + }, + { + "bbox": [ + 130, + 534, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 130, + 534, + 481, + 665 + ], + "type": "text", + "content": " should include the information of fog, as it represents the difference between the foggy scene and the same scene with clear weather." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "type": "text", + "content": "MonoWAD" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "By doing so, our diffusion model learns the variation in weather by repeatedly adding and removing the fog. Note that, as we take clear feature " + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "inline_equation", + "content": "x^{c}" + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "inline_equation", + "content": "x_{0}" + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": " to make a reference input for our diffusion model, we newly denote " + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "inline_equation", + "content": "x_{0}" + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "inline_equation", + "content": "x_{0}^{c}" + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 480, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 480, + 188 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 480, + 188 + ], + "type": "text", + "content": "For the forward process at the " + }, + { + "bbox": [ + 130, + 152, + 480, + 188 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 152, + 480, + 188 + ], + "type": "text", + "content": "-th timestep, " + }, + { + "bbox": [ + 130, + 152, + 480, + 188 + ], + "type": "inline_equation", + "content": "q(x_{t}^{c}|x_{t - 1}^{c})" + }, + { + "bbox": [ + 130, + 152, + 480, + 188 + ], + "type": "text", + "content": " takes the previous feature " + }, + { + "bbox": [ + 130, + 152, + 480, + 188 + ], + "type": "inline_equation", + "content": "x_{t - 1}^{c}" + }, + { + "bbox": [ + 130, + 152, + 480, + 188 + ], + "type": "text", + "content": " and the noise related to the fog (i.e., " + }, + { + "bbox": [ + 130, + 152, + 480, + 188 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 130, + 152, + 480, + 188 + ], + "type": "text", + "content": ") as inputs to generate " + }, + { + "bbox": [ + 130, + 152, + 480, + 188 + ], + "type": "inline_equation", + "content": "x_{t}^{c}" + }, + { + "bbox": [ + 130, + 152, + 480, + 188 + ], + "type": "text", + "content": ". This procedure is repeated over " + }, + { + "bbox": [ + 130, + 152, + 480, + 188 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 152, + 480, + 188 + ], + "type": "text", + "content": " timesteps, which can be represented as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 224, + 196, + 481, + 212 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 224, + 196, + 481, + 212 + ], + "spans": [ + { + "bbox": [ + 224, + 196, + 481, + 212 + ], + "type": "interline_equation", + "content": "q (x _ {t} ^ {c} | x _ {t - 1} ^ {c}) = \\mathcal {F} (x _ {t} ^ {c}; \\sqrt {1 - \\beta_ {t}} x _ {t - 1} ^ {c}, \\beta_ {t} \\mathbf {I}), \\tag {6}", + "image_path": "bd001a190354d7f0ffa897a77988dac58c60f336c37b0ec43eec0be0d77931c3.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 236, + 220, + 481, + 251 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 236, + 220, + 481, + 251 + ], + "spans": [ + { + "bbox": [ + 236, + 220, + 481, + 251 + ], + "type": "interline_equation", + "content": "q \\left(x _ {1} ^ {c}, \\dots , x _ {T} ^ {c} \\mid x ^ {c}\\right) = \\prod_ {t = 1} ^ {T} q \\left(x _ {t} ^ {c} \\mid x _ {t - 1} ^ {c}\\right), \\tag {7}", + "image_path": "c91e59ccffc39d83ad3ab4faaeb2b0e88b02cd344d3cf7df6e3dc03bfafe8405.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 256, + 307, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 256, + 307, + 266 + ], + "spans": [ + { + "bbox": [ + 131, + 256, + 307, + 266 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 256, + 307, + 266 + ], + "type": "inline_equation", + "content": "\\beta_{t}" + }, + { + "bbox": [ + 131, + 256, + 307, + 266 + ], + "type": "text", + "content": " denotes the variance schedule." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 268, + 479, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 268, + 479, + 339 + ], + "spans": [ + { + "bbox": [ + 130, + 268, + 479, + 339 + ], + "type": "text", + "content": "Next, the reverse process at the " + }, + { + "bbox": [ + 130, + 268, + 479, + 339 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 268, + 479, + 339 + ], + "type": "text", + "content": "-th timestep aims to estimate fog variant " + }, + { + "bbox": [ + 130, + 268, + 479, + 339 + ], + "type": "inline_equation", + "content": "\\epsilon_{n}" + }, + { + "bbox": [ + 130, + 268, + 479, + 339 + ], + "type": "text", + "content": " using " + }, + { + "bbox": [ + 130, + 268, + 479, + 339 + ], + "type": "inline_equation", + "content": "x_{t}^{c}" + }, + { + "bbox": [ + 130, + 268, + 479, + 339 + ], + "type": "text", + "content": " to enhance the feature representation of the foggy. To this end, we adopt conditional autoencoder " + }, + { + "bbox": [ + 130, + 268, + 479, + 339 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(x_{t}^{c}, t, x^{r})" + }, + { + "bbox": [ + 130, + 268, + 479, + 339 + ], + "type": "text", + "content": " that receives " + }, + { + "bbox": [ + 130, + 268, + 479, + 339 + ], + "type": "inline_equation", + "content": "x_{t}^{c}" + }, + { + "bbox": [ + 130, + 268, + 479, + 339 + ], + "type": "text", + "content": " and reference feature " + }, + { + "bbox": [ + 130, + 268, + 479, + 339 + ], + "type": "inline_equation", + "content": "x^{r}" + }, + { + "bbox": [ + 130, + 268, + 479, + 339 + ], + "type": "text", + "content": " from the weather codebook " + }, + { + "bbox": [ + 130, + 268, + 479, + 339 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}" + }, + { + "bbox": [ + 130, + 268, + 479, + 339 + ], + "type": "text", + "content": ". Specifically, " + }, + { + "bbox": [ + 130, + 268, + 479, + 339 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 130, + 268, + 479, + 339 + ], + "type": "text", + "content": " estimates mean " + }, + { + "bbox": [ + 130, + 268, + 479, + 339 + ], + "type": "inline_equation", + "content": "\\mu_{\\theta}" + }, + { + "bbox": [ + 130, + 268, + 479, + 339 + ], + "type": "text", + "content": " and variance " + }, + { + "bbox": [ + 130, + 268, + 479, + 339 + ], + "type": "inline_equation", + "content": "\\pmb{\\Sigma}_{\\theta}" + }, + { + "bbox": [ + 130, + 268, + 479, + 339 + ], + "type": "text", + "content": " of the fog distribution at the " + }, + { + "bbox": [ + 130, + 268, + 479, + 339 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 268, + 479, + 339 + ], + "type": "text", + "content": "-th timestep, denoted as " + }, + { + "bbox": [ + 130, + 268, + 479, + 339 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{F}}(\\cdot)" + }, + { + "bbox": [ + 130, + 268, + 479, + 339 + ], + "type": "text", + "content": ". The reverse process is also repeated over " + }, + { + "bbox": [ + 130, + 268, + 479, + 339 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 268, + 479, + 339 + ], + "type": "text", + "content": " timesteps, which can be represented as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 199, + 349, + 481, + 363 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 199, + 349, + 481, + 363 + ], + "spans": [ + { + "bbox": [ + 199, + 349, + 481, + 363 + ], + "type": "interline_equation", + "content": "p _ {\\theta} \\left(x _ {t - 1} ^ {c} \\mid x _ {t} ^ {c}, x ^ {r}\\right) = \\tilde {\\mathcal {F}} \\left(x _ {t - 1} ^ {c}; \\boldsymbol {\\mu} _ {\\theta} \\left(x _ {t} ^ {c}, t, x ^ {r}\\right), \\boldsymbol {\\Sigma} _ {\\theta} \\left(x _ {t} ^ {c}, t\\right)\\right), \\tag {8}", + "image_path": "6057680afb5f25776cbac504aeb8323e59a86b211189f159c612434906bae31d.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 218, + 372, + 481, + 403 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 218, + 372, + 481, + 403 + ], + "spans": [ + { + "bbox": [ + 218, + 372, + 481, + 403 + ], + "type": "interline_equation", + "content": "p _ {\\theta} \\left(x ^ {c}, \\dots , x _ {T} ^ {c}\\right) = p \\left(x _ {T} ^ {c}\\right) \\prod_ {t = 1} ^ {T} p _ {\\theta} \\left(x _ {t - 1} ^ {c} \\mid x _ {t} ^ {c}, x ^ {r}\\right), \\tag {9}", + "image_path": "72cdbed9cec0cab6145c160b83e8d6a84d6b48ada2a2d6c6a40011363d6a2e49.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 408, + 384, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 408, + 384, + 420 + ], + "spans": [ + { + "bbox": [ + 131, + 408, + 384, + 420 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 408, + 384, + 420 + ], + "type": "inline_equation", + "content": "p_{\\theta}(x_{t - 1}^{c}|x_{t}^{c},x^{r})" + }, + { + "bbox": [ + 131, + 408, + 384, + 420 + ], + "type": "text", + "content": " includes cross-attention layer in " + }, + { + "bbox": [ + 131, + 408, + 384, + 420 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 420, + 479, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 420, + 479, + 467 + ], + "spans": [ + { + "bbox": [ + 130, + 420, + 479, + 467 + ], + "type": "text", + "content": "The cross-attention layer receives the flattened clear feature " + }, + { + "bbox": [ + 130, + 420, + 479, + 467 + ], + "type": "inline_equation", + "content": "x_{t}^{c}" + }, + { + "bbox": [ + 130, + 420, + 479, + 467 + ], + "type": "text", + "content": " and weather-reference feature " + }, + { + "bbox": [ + 130, + 420, + 479, + 467 + ], + "type": "inline_equation", + "content": "x^{r}" + }, + { + "bbox": [ + 130, + 420, + 479, + 467 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 130, + 420, + 479, + 467 + ], + "type": "inline_equation", + "content": "\\bar{x}_{t}^{c}" + }, + { + "bbox": [ + 130, + 420, + 479, + 467 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 420, + 479, + 467 + ], + "type": "inline_equation", + "content": "\\bar{x}^{r}" + }, + { + "bbox": [ + 130, + 420, + 479, + 467 + ], + "type": "text", + "content": ", respectively. The similarity between " + }, + { + "bbox": [ + 130, + 420, + 479, + 467 + ], + "type": "inline_equation", + "content": "\\bar{x}^{r}" + }, + { + "bbox": [ + 130, + 420, + 479, + 467 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 420, + 479, + 467 + ], + "type": "inline_equation", + "content": "\\bar{x}_{t}^{c}" + }, + { + "bbox": [ + 130, + 420, + 479, + 467 + ], + "type": "text", + "content": " is calculated in the cross-attention layer to transfer the enhancement to " + }, + { + "bbox": [ + 130, + 420, + 479, + 467 + ], + "type": "inline_equation", + "content": "\\bar{x}_{t}^{c}" + }, + { + "bbox": [ + 130, + 420, + 479, + 467 + ], + "type": "text", + "content": ", which can be formulated as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 216, + 474, + 481, + 502 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 216, + 474, + 481, + 502 + ], + "spans": [ + { + "bbox": [ + 216, + 474, + 481, + 502 + ], + "type": "interline_equation", + "content": "\\operatorname {A t t e n t i o n} (Q, K, V) = \\operatorname {s o f t m a x} \\left(\\frac {Q K ^ {T}}{\\sqrt {d}}\\right) V, \\tag {10}", + "image_path": "f0d79338f1089bff362c442d060480b5b81b0ddafb29e572570cd8644fa2de5a.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 509, + 479, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 509, + 479, + 533 + ], + "spans": [ + { + "bbox": [ + 131, + 509, + 479, + 533 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 509, + 479, + 533 + ], + "type": "inline_equation", + "content": "Q = W_{i}^{q} \\cdot \\bar{x}_{t}^{c}, K = W_{i}^{k} \\cdot \\bar{x}^{r}, V = W_{i}^{v} \\cdot \\bar{x}^{r}" + }, + { + "bbox": [ + 131, + 509, + 479, + 533 + ], + "type": "text", + "content": " with learnable parameters " + }, + { + "bbox": [ + 131, + 509, + 479, + 533 + ], + "type": "inline_equation", + "content": "W_{i}^{q}, W_{i}^{k}, W_{i}^{v}" + }, + { + "bbox": [ + 131, + 509, + 479, + 533 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 534, + 479, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 534, + 479, + 570 + ], + "spans": [ + { + "bbox": [ + 130, + 534, + 479, + 570 + ], + "type": "text", + "content": "To ensure that the estimated fog variant " + }, + { + "bbox": [ + 130, + 534, + 479, + 570 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 130, + 534, + 479, + 570 + ], + "type": "text", + "content": " is similar to the fog variant " + }, + { + "bbox": [ + 130, + 534, + 479, + 570 + ], + "type": "inline_equation", + "content": "\\epsilon_{n}" + }, + { + "bbox": [ + 130, + 534, + 479, + 570 + ], + "type": "text", + "content": " applied to make " + }, + { + "bbox": [ + 130, + 534, + 479, + 570 + ], + "type": "inline_equation", + "content": "x_{t}^{c}" + }, + { + "bbox": [ + 130, + 534, + 479, + 570 + ], + "type": "text", + "content": " from the input clear feature " + }, + { + "bbox": [ + 130, + 534, + 479, + 570 + ], + "type": "inline_equation", + "content": "x^{c}" + }, + { + "bbox": [ + 130, + 534, + 479, + 570 + ], + "type": "text", + "content": ", we propose a weather-adaptive enhancement loss " + }, + { + "bbox": [ + 130, + 534, + 479, + 570 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{wae}" + }, + { + "bbox": [ + 130, + 534, + 479, + 570 + ], + "type": "text", + "content": ", which is formulated as:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 218, + 578, + 481, + 598 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 218, + 578, + 481, + 598 + ], + "spans": [ + { + "bbox": [ + 218, + 578, + 481, + 598 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {w a e} = \\mathbb {E} _ {x ^ {c}, \\boldsymbol {\\epsilon} _ {n} \\sim \\mathcal {F}, t} \\left[ \\| \\boldsymbol {\\epsilon} _ {n} - \\boldsymbol {\\epsilon} _ {\\theta} \\left(x _ {t} ^ {c}, t, x ^ {r}\\right) \\| _ {2} ^ {2} \\right]. \\tag {11}", + "image_path": "8667d4e763ce7207fc9b8617caa7a69c6c54a232cb4ced5863c383cc46361bed.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "content": "With " + }, + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{wae}" + }, + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "content": " can estimate the fog variant by leveraging the fog distribution as noise for our diffusion model through multiple forward/reverse processes. Additionally, the cross-attention layer within " + }, + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "content": " dynamically enhances the feature representation based on combining knowledge of the input feature (foggy or clear) and the weather-reference feature. Since our diffusion model learns the" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "text", + "content": "Y. Oh et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": "degree of improvement needed corresponding to weather conditions, it can improve the representation of any input, whether clear or foggy in inference phase. This leads our monocular 3D object detector can handle both clear and foggy input images, resulting in weather-robust detection." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 173, + 261, + 184 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 173, + 261, + 184 + ], + "spans": [ + { + "bbox": [ + 131, + 173, + 261, + 184 + ], + "type": "text", + "content": "3.3 Total Loss Function" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 193, + 425, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 193, + 425, + 205 + ], + "spans": [ + { + "bbox": [ + 131, + 193, + 425, + 205 + ], + "type": "text", + "content": "The total loss function of our MonoWAD is represented as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 233, + 216, + 480, + 228 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 216, + 480, + 228 + ], + "spans": [ + { + "bbox": [ + 233, + 216, + 480, + 228 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {T o t a l}} = \\mathcal {L} _ {O D} + \\lambda_ {1} \\mathcal {L} _ {\\text {c k r}} + \\lambda_ {2} \\mathcal {L} _ {\\text {w a e}}, \\tag {12}", + "image_path": "3e018cff2e7c01c2757a374a387feb053c8b0b7a0a1c642173b1d0dff90a5282.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 237, + 479, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 237, + 479, + 285 + ], + "spans": [ + { + "bbox": [ + 131, + 237, + 479, + 285 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 237, + 479, + 285 + ], + "type": "inline_equation", + "content": "\\lambda_{1}" + }, + { + "bbox": [ + 131, + 237, + 479, + 285 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 237, + 479, + 285 + ], + "type": "inline_equation", + "content": "\\lambda_{2}" + }, + { + "bbox": [ + 131, + 237, + 479, + 285 + ], + "type": "text", + "content": " denote balancing hyper-parameters. In our experiments, we set " + }, + { + "bbox": [ + 131, + 237, + 479, + 285 + ], + "type": "inline_equation", + "content": "\\lambda_{1} = \\lambda_{2} = 1" + }, + { + "bbox": [ + 131, + 237, + 479, + 285 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 131, + 237, + 479, + 285 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{OD}" + }, + { + "bbox": [ + 131, + 237, + 479, + 285 + ], + "type": "text", + "content": " is the detection loss for 3D object detection [16,55]. It includes loss functions of classification, regression, and depth loss, that are similar to prior works [8,16,31]. The overall weight parameters are updated through " + }, + { + "bbox": [ + 131, + 237, + 479, + 285 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{Total}" + }, + { + "bbox": [ + 131, + 237, + 479, + 285 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 303, + 230, + 317 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 303, + 230, + 317 + ], + "spans": [ + { + "bbox": [ + 131, + 303, + 230, + 317 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 328, + 322, + 340 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 328, + 322, + 340 + ], + "spans": [ + { + "bbox": [ + 131, + 328, + 322, + 340 + ], + "type": "text", + "content": "4.1 Dataset and Evaluation Metrics" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 348, + 480, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 348, + 480, + 467 + ], + "spans": [ + { + "bbox": [ + 130, + 348, + 480, + 467 + ], + "type": "text", + "content": "Datasets. We utilize the KITTI 3D object detection dataset [12], the most widely adopted for 3D object detection. It contains 7,481 training images and 7,518 test images under clear weather conditions. Due to unavailable ground-truth annotations for test images and limited evaluation on the test server, we follow [6] by splitting 3,712 training images and 3,769 validation images. In addition, as our work requires paired images to learn about weather changes, we generate foggy images from all images in the KITTI dataset, called foggy KITTI, that emulate the foggy scene. Following the protocols of [2,43], we generate foggy images based on object distances using depth maps estimated by DORN [10]. Please refer to the supplementary materials for details." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 468, + 480, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 468, + 480, + 528 + ], + "spans": [ + { + "bbox": [ + 130, + 468, + 480, + 528 + ], + "type": "text", + "content": "In addition, as our work is focused on robust monocular 3D object detection in various weather conditions, we further adopt the Virtual KITTI dataset [5,11], which contains photo-realistic synthetic images under various weather conditions (e.g., foggy, rainy, sunset). It is associated with the original real-world KITTI dataset, which has 3D annotations for each weather." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 533, + 479, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 533, + 479, + 594 + ], + "spans": [ + { + "bbox": [ + 131, + 533, + 479, + 594 + ], + "type": "text", + "content": "Evaluation Metrics. We adopt average precision in both 3D detection " + }, + { + "bbox": [ + 131, + 533, + 479, + 594 + ], + "type": "inline_equation", + "content": "(\\mathrm{AP}_{3D})" + }, + { + "bbox": [ + 131, + 533, + 479, + 594 + ], + "type": "text", + "content": " and bird-eye view detection " + }, + { + "bbox": [ + 131, + 533, + 479, + 594 + ], + "type": "inline_equation", + "content": "(\\mathrm{AP}_{BEV})" + }, + { + "bbox": [ + 131, + 533, + 479, + 594 + ], + "type": "text", + "content": " under three difficulty levels ('Easy', 'Moderate', 'Hard') according to size, occlusion, and truncation. Following [47], we use 40 recall position metric " + }, + { + "bbox": [ + 131, + 533, + 479, + 594 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{40}" + }, + { + "bbox": [ + 131, + 533, + 479, + 594 + ], + "type": "text", + "content": " and report scores for the car category under the IoU threshold 0.7 for the KITTI dataset and 0.5 for the Virtual KITTI dataset." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 609, + 279, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 609, + 279, + 621 + ], + "spans": [ + { + "bbox": [ + 131, + 609, + 279, + 621 + ], + "type": "text", + "content": "4.2 Implementation Details" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 629, + 479, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 479, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 479, + 666 + ], + "type": "text", + "content": "We use DLA-102 [52] as our backbone and adopt the transformer architecture of [16] for the detection block. We train MonoWAD on a single RTX 4090 GPU with a batch size of 4 over 120 epochs using Adam optimizer [23] (initial learning" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "type": "text", + "content": "MonoWAD" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 161, + 479, + 243 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "type": "text", + "content": "Table 1: Detection results of car category on KITTI validation set under foggy weather and clear weather conditions. The results of the state-of-the-art methods under foggy weather are obtained through our reproduction with the official source code. Bold/underlined fonts indicate the best/second-best results." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 161, + 479, + 243 + ], + "lines": [ + { + "bbox": [ + 135, + 161, + 479, + 243 + ], + "spans": [ + { + "bbox": [ + 135, + 161, + 479, + 243 + ], + "type": "table", + "html": "
MethodFoggy (AP3D)Clear (AP3D)Average
EasyMod.HardEasyMod.HardEasyMod.Hard
GUPNet [28] (ICCV'21)2.742.192.1622.7616.4613.7212.759.337.94
DID-M3D [37] (ECCV'22)1.150.610.6422.9816.1214.0312.078.377.34
MonoGround [40] (CVPR'22)0.000.000.0625.2418.6915.5812.629.357.82
MonoDTR [16] (CVPR'22)16.8911.869.8724.5218.5715.5120.7115.2212.69
MonoDETR [55] (ICCV'23)7.405.744.5328.8420.6116.3818.1213.1810.46
MonoWAD (Ours)27.1719.5716.2129.1021.0817.7328.1420.3316.97
", + "image_path": "43d987fd65abcd4539352a30f7ddc7a32010e6856d7d49d74baa9d01a9cba868.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 159, + 277, + 451, + 361 + ], + "blocks": [ + { + "bbox": [ + 132, + 252, + 479, + 274 + ], + "lines": [ + { + "bbox": [ + 132, + 252, + 479, + 274 + ], + "spans": [ + { + "bbox": [ + 132, + 252, + 479, + 274 + ], + "type": "text", + "content": "Table 2: Detection results of car category on KITTI test set under foggy weather. Bold/underlined fonts indicate the best/second-best results." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 159, + 277, + 451, + 361 + ], + "lines": [ + { + "bbox": [ + 159, + 277, + 451, + 361 + ], + "spans": [ + { + "bbox": [ + 159, + 277, + 451, + 361 + ], + "type": "table", + "html": "
MethodAP3DAPBEV
EasyMod.HardEasyMod.Hard
GUPNet [28] (ICCV'21)3.012.421.134.903.022.91
DID-M3D [37] (ECCV'22)3.102.392.195.343.012.91
MonoGround [40] (CVPR'22)0.140.200.220.230.380.39
MonoDTR [16] (CVPR'22)11.077.415.2615.7610.157.53
MonoDETR [55] (ICCV'23)9.335.544.0613.158.066.30
MonoWAD (Ours)19.7513.3211.0427.9519.0615.61
", + "image_path": "34ef2b7aba8cdc8c396a824d68a1df5f4a92af2375954e1aaeb4d4e5c784a916.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 372, + 482, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 372, + 482, + 456 + ], + "spans": [ + { + "bbox": [ + 130, + 372, + 482, + 456 + ], + "type": "text", + "content": "rate of " + }, + { + "bbox": [ + 130, + 372, + 482, + 456 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 130, + 372, + 482, + 456 + ], + "type": "text", + "content": "). For the weather codebook, we utilize embedding slot " + }, + { + "bbox": [ + 130, + 372, + 482, + 456 + ], + "type": "inline_equation", + "content": "K = 4096" + }, + { + "bbox": [ + 130, + 372, + 482, + 456 + ], + "type": "text", + "content": " and set the dimension of each slot " + }, + { + "bbox": [ + 130, + 372, + 482, + 456 + ], + "type": "inline_equation", + "content": "D = 256" + }, + { + "bbox": [ + 130, + 372, + 482, + 456 + ], + "type": "text", + "content": ". We set 4 heads for cross-attention of the weather-adaptive diffusion model, and the timesteps for the forward and reverse process of diffusion default to 15. The number of channels for the diffusion output features is " + }, + { + "bbox": [ + 130, + 372, + 482, + 456 + ], + "type": "inline_equation", + "content": "C = 256" + }, + { + "bbox": [ + 130, + 372, + 482, + 456 + ], + "type": "text", + "content": ". As recent monocular 3D object detectors [16,28,37,40,55] have not been explored under adverse weather conditions, we implemented them with available official source code to faithfully reproduce them." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 467, + 222, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 467, + 222, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 467, + 222, + 479 + ], + "type": "text", + "content": "4.3 Comparison" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 485, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 485, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 485, + 482, + 666 + ], + "type": "text", + "content": "Results on KITTI 3D Dataset. We compared MonoWAD with state-of-the-art monocular 3D object detectors [16,28,37,40,55] that do not use additional data (e.g., depth maps or LiDAR) during inference on the KITTI and foggy KITTI validation sets. Table 1 shows the " + }, + { + "bbox": [ + 130, + 485, + 482, + 666 + ], + "type": "inline_equation", + "content": "AP_{3D}" + }, + { + "bbox": [ + 130, + 485, + 482, + 666 + ], + "type": "text", + "content": " results, and " + }, + { + "bbox": [ + 130, + 485, + 482, + 666 + ], + "type": "inline_equation", + "content": "AP_{BEV}" + }, + { + "bbox": [ + 130, + 485, + 482, + 666 + ], + "type": "text", + "content": " results are in the supplementary materials. While recent methods have shown improved performance in clear weather, their performance drops in foggy weather, limiting their applicability in real-world applications (e.g., autonomous driving and robotics). In contrast, MonoWAD showed stable 3D detection performance under both foggy and clear weather. Since our weather codebook learns knowledge about clear weather and the weather-adaptive diffusion model can enhance the feature representation of the input images under both clear and foggy weather, MonoWAD shows a more weather-robust detection performance than that of the existing methods. Also, to explore the weather robustness of MonoWAD, we conducted experiments with mixed foggy and clear weather conditions at various ratios (please see the supplementary materials)." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "text", + "content": "Y. Oh et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 140, + 476, + 222 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "type": "text", + "content": "Table 3: Detection results of car category on Virtual KITTI under foggy, rainy, sunset conditions. Bold/underlined fonts indicate the best/second-best results." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 140, + 476, + 222 + ], + "lines": [ + { + "bbox": [ + 135, + 140, + 476, + 222 + ], + "spans": [ + { + "bbox": [ + 135, + 140, + 476, + 222 + ], + "type": "table", + "html": "
MethodFoggy (AP3D)Rainy (AP3D)Sunset (AP3D)
EasyMod.HardEasyMod.HardEasyMod.Hard
GUPNet [28] (ICCV'21)1.761.571.572.341.241.212.771.641.65
DID-M3D [37] (ECCV'22)0.910.390.390.400.130.130.340.100.10
MonoGround [40] (CVPR'22)0.290.300.255.492.822.777.684.244.20
MonoDTR [16] (CVPR'22)8.795.755.7211.736.256.749.865.425.42
MonoDETR [55] (ICCV'23)4.502.992.966.613.463.427.084.174.16
MonoWAD (Ours)13.338.568.5014.128.338.2413.387.897.80
", + "image_path": "8f1f5b8491bac96f0f5e3211761154bb7fe958c2abbea8ff8288f4ecb58b66f1.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 171, + 248, + 439, + 305 + ], + "blocks": [ + { + "bbox": [ + 132, + 226, + 480, + 247 + ], + "lines": [ + { + "bbox": [ + 132, + 226, + 480, + 247 + ], + "spans": [ + { + "bbox": [ + 132, + 226, + 480, + 247 + ], + "type": "text", + "content": "Table 4: Effect of the proposed method on KITTI validation set for car category. WC denotes our weather codebook, WAD indicates our weather-adaptive diffusion model." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 171, + 248, + 439, + 305 + ], + "lines": [ + { + "bbox": [ + 171, + 248, + 439, + 305 + ], + "spans": [ + { + "bbox": [ + 171, + 248, + 439, + 305 + ], + "type": "table", + "html": "
MethodWCWADFoggy (AP3D)Clear (AP3D)
EasyMod.HardEasyMod.Hard
Baseline--13.759.618.1022.6317.1614.28
MonoWAD (Ours)-✓25.6218.6615.5626.3419.1716.15
✓✓27.1719.5716.2129.1021.0817.73
", + "image_path": "92ef456f169f950eafccc5b80a73a7c9bb1b71ae5ff26f4d5d6731618d513186.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 168, + 323, + 442, + 398 + ], + "blocks": [ + { + "bbox": [ + 132, + 311, + 479, + 322 + ], + "lines": [ + { + "bbox": [ + 132, + 311, + 479, + 322 + ], + "spans": [ + { + "bbox": [ + 132, + 311, + 479, + 322 + ], + "type": "text", + "content": "Table 5: Detection results on KITTI validation set by changing diffusion timestep " + }, + { + "bbox": [ + 132, + 311, + 479, + 322 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 132, + 311, + 479, + 322 + ], + "type": "text", + "content": " ." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 168, + 323, + 442, + 398 + ], + "lines": [ + { + "bbox": [ + 168, + 323, + 442, + 398 + ], + "spans": [ + { + "bbox": [ + 168, + 323, + 442, + 398 + ], + "type": "table", + "html": "
Timestep TFoggy (AP3D)Clear (AP3D)
EasyMod.HardEasyMod.Hard
-13.759.618.1022.6317.1614.28
523.5717.9114.9626.0319.2116.01
1025.2818.4915.4226.7919.9016.78
1527.1719.5716.2129.1021.0817.73
2024.5418.2915.1024.8518.5415.34
", + "image_path": "7ac33ed1df4e17379d7d4fd068da7ffec2ff38f541fb5f02817366a9c2141fb6.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 407, + 479, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 407, + 479, + 455 + ], + "spans": [ + { + "bbox": [ + 132, + 407, + 479, + 455 + ], + "type": "text", + "content": "We further compared 3D detection performances on the foggy KITTI test set (Table 2). Similar to Table 1, existing monocular 3D object detectors show lower performance under foggy weather. Through the results, our method shows the weather-robust monocular 3D performance under foggy and clear weather." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 461, + 480, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 461, + 480, + 545 + ], + "spans": [ + { + "bbox": [ + 132, + 461, + 480, + 545 + ], + "type": "text", + "content": "Results on Virtual KITTI Dataset. We also conducted experiments on the Virtual KITTI dataset to validate the generalization ability of our method across various weather conditions (i.e., foggy, rainy, sunset). The results are shown in Table 3. MonoWAD also shows the highest performance in foggy weather. Moreover, it outperforms the existing method in the rainy and sunset conditions. The results demonstrate that MonoWAD is robust and insensitive to various weather conditions that can be faced in real-world scenarios, not just clear weather." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 560, + 244, + 572 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 560, + 244, + 572 + ], + "spans": [ + { + "bbox": [ + 132, + 560, + 244, + 572 + ], + "type": "text", + "content": "4.4 Ablation Studies" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 578, + 479, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 479, + 625 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 479, + 625 + ], + "type": "text", + "content": "We conducted ablation studies to examine (1) the effect of each proposed component (i.e., weather codebook and weather-adaptive diffusion model) and (2) the effect of the weather-adaptive diffusion model by varying the timestep " + }, + { + "bbox": [ + 132, + 578, + 479, + 625 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 132, + 578, + 479, + 625 + ], + "type": "text", + "content": ". These experiments were performed on the KITTI and foggy KITTI 3D validation sets." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 641, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 641, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 641, + 480, + 665 + ], + "type": "text", + "content": "Effect of the Proposed Modules. The results regarding the effectiveness of the proposed weather codebook and weather-adaptive diffusion model are pre-" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "type": "text", + "content": "MonoWAD" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 149, + 476, + 224 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 6: Performance comparison on KITTI validation set under foggy and clear weather conditions. We compare our MonoWAD with MonoDTR [16], which demonstrates superior performance in foggy weather using state-of-the-art dehazing methods." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 149, + 476, + 224 + ], + "lines": [ + { + "bbox": [ + 135, + 149, + 476, + 224 + ], + "spans": [ + { + "bbox": [ + 135, + 149, + 476, + 224 + ], + "type": "table", + "html": "
MethodDehazing MethodFoggy (AP3D)Clear (AP3D)Average
EasyMod.HardEasyMod.HardEasyMod.Hard
MonoDTR(B) [16] (CVPR'22)-16.8911.869.8724.5218.5715.5120.7115.2212.69
B + RIDCP [49] (CVPR'23)Image-level17.2312.4110.4424.0217.8914.7820.6315.1512.61
B + DENet [39] (ACCV'22)Feature-level22.3517.4414.477.105.704.5314.7311.579.50
B + Yang et al. [51] (ACCV'22)Feature-level22.8715.2112.1717.9613.1010.6420.4214.1611.41
MonoWAD (Ours)-27.1719.5716.2129.1021.0817.7328.1420.3316.97
", + "image_path": "c1a042afafb71c984ae10b776ed7bb8d251369f7e429fb46a9e0001f466d4e66.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 168, + 270, + 443, + 357 + ], + "blocks": [ + { + "bbox": [ + 132, + 233, + 482, + 267 + ], + "lines": [ + { + "bbox": [ + 132, + 233, + 482, + 267 + ], + "spans": [ + { + "bbox": [ + 132, + 233, + 482, + 267 + ], + "type": "text", + "content": "Table 7: Comparison of diffusion models on KITTI validation set for car category: " + }, + { + "bbox": [ + 132, + 233, + 482, + 267 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 132, + 233, + 482, + 267 + ], + "type": "text", + "content": " is baseline detection block (transformer encoder-decoder), CDM (Conditional Diffusion Model [42]), WC (weather codebook), and WAD (weather adaptive diffusion model)." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 168, + 270, + 443, + 357 + ], + "lines": [ + { + "bbox": [ + 168, + 270, + 443, + 357 + ], + "spans": [ + { + "bbox": [ + 168, + 270, + 443, + 357 + ], + "type": "table", + "html": "
MethodFoggy (AP3D)Clear (AP3D)
EasyMod.HardEasyMod.Hard
B+DDPM [15]5.323.842.7718.3112.7110.20
B+CDM [42]2.742.102.0020.5014.5111.63
B+WC+CDM [42]17.5112.7410.4021.0515.1112.54
B+WAD25.6218.6615.5626.3419.1716.15
MonoWAD(B+WC+WAD)27.1719.5716.2129.1021.0817.73
", + "image_path": "f612a81c020046afd3345f974015b5172f1ffaa62f69e2d8a3c7124aaae1bd75.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 369, + 482, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 369, + 482, + 441 + ], + "spans": [ + { + "bbox": [ + 130, + 369, + 482, + 441 + ], + "type": "text", + "content": "sented in Table 4. Since our weather-adaptive diffusion model is designed to enhance feature representation, it alone has shown significant performance improvement. When the weather codebook is additionally considered, the weather-adaptive diffusion model can leverage the knowledge of the weather-reference feature for clear weather, leading to enhanced performance. It makes MonoWAD can be robust to various weather conditions for monocular 3D object detection." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 449, + 482, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 482, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 482, + 521 + ], + "type": "text", + "content": "Effect of Timestep " + }, + { + "bbox": [ + 130, + 449, + 482, + 521 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 449, + 482, + 521 + ], + "type": "text", + "content": ". We also conduct experiments by varying timestep " + }, + { + "bbox": [ + 130, + 449, + 482, + 521 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 449, + 482, + 521 + ], + "type": "text", + "content": " of the weather-adaptive diffusion model. Timestep is the number of steps for the forward and reverse process of the diffusion model. Table 5 indicates that the highest monocular 3D detection performance is achieved when " + }, + { + "bbox": [ + 130, + 449, + 482, + 521 + ], + "type": "inline_equation", + "content": "T = 15" + }, + { + "bbox": [ + 130, + 449, + 482, + 521 + ], + "type": "text", + "content": " under both clear and foggy weather conditions. Our MonoWAD consistently outperforms the baseline and other existing methods across all timesteps." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 533, + 218, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 533, + 218, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 533, + 218, + 544 + ], + "type": "text", + "content": "4.5 Discussions" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 552, + 485, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 552, + 485, + 673 + ], + "spans": [ + { + "bbox": [ + 130, + 552, + 485, + 673 + ], + "type": "text", + "content": "Comparison with Dehazing Methods. We investigate the weather-robustness of our method for monocular 3D object detection compared to the other monocular detectors using dehazing methods. To this end, we compared MonoWAD with MonoDTR [16], which shows the best performance in foggy weather through the application of state-of-the-art image-level and feature-level dehazing methods [39,49,51]. The results are shown in Table 6. Recent dehazing methods are primarily focused on specific weather conditions, such as foggy weather. While they have shown some improvement under foggy conditions, they exhibit reduced performances under clear weather conditions. In contrast, our MonoWAD shows robust performance across both foggy and clear weather conditions." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "text", + "content": "Y. Oh et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 137, + 116, + 478, + 230 + ], + "blocks": [ + { + "bbox": [ + 137, + 116, + 478, + 230 + ], + "lines": [ + { + "bbox": [ + 137, + 116, + 478, + 230 + ], + "spans": [ + { + "bbox": [ + 137, + 116, + 478, + 230 + ], + "type": "image", + "image_path": "1ce9cb7a8af6a016833dac592e5eb58cf51c6f754a3faf5c18414187a51fa69e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 232, + 480, + 265 + ], + "lines": [ + { + "bbox": [ + 132, + 232, + 480, + 265 + ], + "spans": [ + { + "bbox": [ + 132, + 232, + 480, + 265 + ], + "type": "text", + "content": "Fig. 5: Comparison of 3D detection examples (green: ground-truth, red: predicted 3D bounding-box) between our MonoWAD and two detectors, MonoDTR [16] and MonoDETR [55], that show the most improved performances among existing methods." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 130, + 274, + 483, + 471 + ], + "blocks": [ + { + "bbox": [ + 130, + 274, + 483, + 471 + ], + "lines": [ + { + "bbox": [ + 130, + 274, + 483, + 471 + ], + "spans": [ + { + "bbox": [ + 130, + 274, + 483, + 471 + ], + "type": "image", + "image_path": "4df94299132fc98cb15113316d4d2591056ee507d5585a7166dd4da1fed82b6a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 144, + 472, + 468, + 484 + ], + "lines": [ + { + "bbox": [ + 144, + 472, + 468, + 484 + ], + "spans": [ + { + "bbox": [ + 144, + 472, + 468, + 484 + ], + "type": "text", + "content": "Fig. 6: 3D detection results on real-world images of various weather conditions." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 493, + 482, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 493, + 482, + 566 + ], + "spans": [ + { + "bbox": [ + 130, + 493, + 482, + 566 + ], + "type": "text", + "content": "Effect of Weather-Adaptive Diffusion Model. Table 7 shows the effectiveness of MonoWAD with existing diffusion models [15,42] on the KITTI and foggy KITTI validation sets. Existing methods adopt Gaussian noise for forward and reverse processes, but they can not fully understand about weather. In contrast, our weather-adaptive diffusion understands weather variances, allowing our MonoWAD to surpass existing methods in clear and foggy weather." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 584, + 266, + 596 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 584, + 266, + 596 + ], + "spans": [ + { + "bbox": [ + 132, + 584, + 266, + 596 + ], + "type": "text", + "content": "4.6 Visualization Results" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "Results on KITTI Dataset. We visualize several 3D detection results on the KITTI 3D dataset, comparing MonoWAD with MonoDTR and MonoDETR, which exhibit the highest performances among existing methods under foggy condition (Fig. 5). Existing methods struggle to detect objects obscured by fog, indicating limitations in detecting only fully visible objects. In contrast, even" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "type": "text", + "content": "MonoWAD" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 154, + 114, + 246, + 190 + ], + "blocks": [ + { + "bbox": [ + 154, + 114, + 246, + 190 + ], + "lines": [ + { + "bbox": [ + 154, + 114, + 246, + 190 + ], + "spans": [ + { + "bbox": [ + 154, + 114, + 246, + 190 + ], + "type": "image", + "image_path": "9a7aeccd9d723fbb3756503070f2e04bc5fb3e1ddc1cc2d526c39a3329162339.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 180, + 191, + 432, + 202 + ], + "lines": [ + { + "bbox": [ + 180, + 191, + 432, + 202 + ], + "spans": [ + { + "bbox": [ + 180, + 191, + 432, + 202 + ], + "type": "text", + "content": "Fig. 7: t-SNE visualization results (Red: Clear, Blue: Foggy)." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 261, + 114, + 352, + 190 + ], + "blocks": [ + { + "bbox": [ + 261, + 114, + 352, + 190 + ], + "lines": [ + { + "bbox": [ + 261, + 114, + 352, + 190 + ], + "spans": [ + { + "bbox": [ + 261, + 114, + 352, + 190 + ], + "type": "image", + "image_path": "a32b11d56c1326e5fd5cc8acac38664ea01b1127ef3278cbfcfff7f6b61d76a0.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 369, + 114, + 460, + 190 + ], + "blocks": [ + { + "bbox": [ + 369, + 114, + 460, + 190 + ], + "lines": [ + { + "bbox": [ + 369, + 114, + 460, + 190 + ], + "spans": [ + { + "bbox": [ + 369, + 114, + 460, + 190 + ], + "type": "image", + "image_path": "2590033f084d124308f014d93a14eb5c6c55e66ce963acf9541378c89e17472c.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 211, + 480, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 211, + 480, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 211, + 480, + 236 + ], + "type": "text", + "content": "in dense fog, MonoWAD effectively detects both close and fog-obscured objects with the aid of the weather codebook and weather-adaptive diffusion model." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 241, + 479, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 241, + 479, + 314 + ], + "spans": [ + { + "bbox": [ + 130, + 241, + 479, + 314 + ], + "type": "text", + "content": "Results on Real-World Images. We further visualize 3D detection results on real-world images from the Seeing Through Fog dataset [2] under various weather conditions (i.e., foggy, rainy, snowy). In Fig. 6, MonoWAD shows robust detection under diverse weather conditions. This demonstrates that our proposed method maintains weather-robustness even in real-world scenarios by dynamically enhancing the input scenarios." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 319, + 480, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 319, + 480, + 403 + ], + "spans": [ + { + "bbox": [ + 130, + 319, + 480, + 403 + ], + "type": "text", + "content": "t-SNE Visualization. We conducted t-SNE visualization to analyze feature representations of MonoDTR, MonoDETR, and our MonoWAD on the KITTI and foggy KITTI validation set. As depicted in Fig. 7, the existing methods exhibit distinct feature representations for foggy and clear weather conditions. In contrast, MonoWAD, leveraging weather-robust feature learning from the weather codebook and weather-adaptive diffusion model, demonstrates similar feature representations for both clear and foggy weather conditions." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 414, + 218, + 425 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 414, + 218, + 425 + ], + "spans": [ + { + "bbox": [ + 131, + 414, + 218, + 425 + ], + "type": "text", + "content": "4.7 Limitations" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 430, + 480, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 430, + 480, + 525 + ], + "spans": [ + { + "bbox": [ + 130, + 430, + 480, + 525 + ], + "type": "text", + "content": "The experimental results show the weather-robustness of our method. However, due to the iterative nature of the diffusion model, our method shows " + }, + { + "bbox": [ + 130, + 430, + 480, + 525 + ], + "type": "inline_equation", + "content": "144\\mathrm{ms}/" + }, + { + "bbox": [ + 130, + 430, + 480, + 525 + ], + "type": "text", + "content": " image at timestep " + }, + { + "bbox": [ + 130, + 430, + 480, + 525 + ], + "type": "inline_equation", + "content": "T = 15" + }, + { + "bbox": [ + 130, + 430, + 480, + 525 + ], + "type": "text", + "content": " (110ms/image at " + }, + { + "bbox": [ + 130, + 430, + 480, + 525 + ], + "type": "inline_equation", + "content": "T = 5" + }, + { + "bbox": [ + 130, + 430, + 480, + 525 + ], + "type": "text", + "content": "), slower than the latest work, MonoDETR (38ms/image). Moreover, our method dynamically enhances the representation of the input feature based on the weather-reference feature and weather difference which needs paired images. Thus, exploring a method to achieve faster processing speeds while maintaining weather-robust performance without paired images could be an interesting direction for our future work." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 536, + 220, + 548 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 536, + 220, + 548 + ], + "spans": [ + { + "bbox": [ + 131, + 536, + 220, + 548 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 557, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 481, + 665 + ], + "type": "text", + "content": "We proposed MonoWAD, a novel weather-robust monocular 3D object detector to handle various weather conditions. Addressing challenges in applying existing monocular 3D object detectors to real-world scenarios with various weather, we design a weather codebook with clear knowledge recalling loss to memorize the knowledge of the clear weather and to generate a weather-reference feature from both clear and foggy features. Also, we design a weather-adaptive diffusion model with weather-adaptive enhancement loss to enhance feature representation according to the weather conditions. As a result, our MonoWAD can detect objects occluded by fog and perform well in clear weather." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "Y. Oh et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 226 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 226 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 226 + ], + "type": "text", + "content": "This work was supported by the NRF grant funded by the Korea government (MSIT) (No. RS-2023-00252391), and by IITP grant funded by the Korea government (MSIT) (No. RS-2022-00155911: Artificial Intelligence Convergence Innovation Human Resources Development (Kyung Hee University), IITP-2023-RS-2023-00266615: Convergence Security Core Talent Training Business Support Program, No. 2022-0-00124: Development of Artificial Intelligence Technology for Self-Improving Competency-Aware Learning Capabilities)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 243, + 197, + 255 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 243, + 197, + 255 + ], + "spans": [ + { + "bbox": [ + 133, + 243, + 197, + 255 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 269, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 138, + 269, + 480, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 269, + 480, + 300 + ], + "spans": [ + { + "bbox": [ + 138, + 269, + 480, + 300 + ], + "type": "text", + "content": "1. Baranchuk, D., Rubachev, I., Voynov, A., Khrulkov, V., Babenko, A.: Label-efficient semantic segmentation with diffusion models. In: Int. Conf. Learn. Represent. (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 302, + 480, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 302, + 480, + 334 + ], + "spans": [ + { + "bbox": [ + 138, + 302, + 480, + 334 + ], + "type": "text", + "content": "2. Bijelic, M., Gruber, T., Mannan, F., Kraus, F., Ritter, W., Dietmayer, K., Heide, F.: Seeing through fog without seeing fog: Deep multimodal sensor fusion in unseen adverse weather. In: IEEE Conf. Comput. Vis. Pattern Recog. (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 335, + 480, + 356 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 335, + 480, + 356 + ], + "spans": [ + { + "bbox": [ + 138, + 335, + 480, + 356 + ], + "type": "text", + "content": "3. Brazil, G., Liu, X.: M3d-rpn: Monocular 3d region proposal network for object detection. In: Int. Conf. Comput. Vis. (2019)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 357, + 480, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 357, + 480, + 378 + ], + "spans": [ + { + "bbox": [ + 138, + 357, + 480, + 378 + ], + "type": "text", + "content": "4. Brazil, G., Pons-Moll, G., Liu, X., Schiele, B.: Kinematic 3d object detection in monocular video. In: Eur. Conf. Comput. Vis. Springer (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 380, + 410, + 389 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 380, + 410, + 389 + ], + "spans": [ + { + "bbox": [ + 138, + 380, + 410, + 389 + ], + "type": "text", + "content": "5. Cabon, Y., Murray, N., Humenberger, M.: Virtual kitti 2 (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 391, + 480, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 391, + 480, + 422 + ], + "spans": [ + { + "bbox": [ + 138, + 391, + 480, + 422 + ], + "type": "text", + "content": "6. Chen, X., Kundu, K., Zhang, Z., Ma, H., Fidler, S., Urtasun, R.: Monocular 3d object detection for autonomous driving. In: IEEE Conf. Comput. Vis. Pattern Recog. (2016)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 423, + 480, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 423, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 138, + 423, + 480, + 445 + ], + "type": "text", + "content": "7. Chen, Y., Tai, L., Sun, K., Li, M.: Monopair: Monocular 3d object detection using pairwise spatial relationships. In: IEEE Conf. Comput. Vis. Pattern Recog. (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 446, + 480, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 446, + 480, + 477 + ], + "spans": [ + { + "bbox": [ + 138, + 446, + 480, + 477 + ], + "type": "text", + "content": "8. Ding, M., Huo, Y., Yi, H., Wang, Z., Shi, J., Lu, Z., Luo, P.: Learning depth-guided convolutions for monocular 3d object detection. In: IEEE Conf. Comput. Vis. Pattern Recog. Worksh. (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 479, + 480, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 479, + 480, + 499 + ], + "spans": [ + { + "bbox": [ + 138, + 479, + 480, + 499 + ], + "type": "text", + "content": "9. Esser, P., Rombach, R., Ommer, B.: Taming transformers for high-resolution image synthesis. In: IEEE Conf. Comput. Vis. Pattern Recog. (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 501, + 480, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 501, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 138, + 501, + 480, + 533 + ], + "type": "text", + "content": "10. Fu, H., Gong, M., Wang, C., Batmanghelich, K., Tao, D.: Deep ordinal regression network for monocular depth estimation. In: IEEE Conf. Comput. Vis. Pattern Recog. (2018)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 534, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 534, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 138, + 534, + 480, + 555 + ], + "type": "text", + "content": "1. Gaidon, A., Wang, Q., Cabon, Y., Vig, E.: Virtual worlds as proxy for multi-object tracking analysis. In: IEEE Conf. Comput. Vis. Pattern Recog. (2016)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 556, + 480, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 556, + 480, + 586 + ], + "spans": [ + { + "bbox": [ + 138, + 556, + 480, + 586 + ], + "type": "text", + "content": "2. Geiger, A., Lenz, P., Urtasun, R.: Are we ready for autonomous driving? the kitti vision benchmark suite. In: IEEE Conf. Comput. Vis. Pattern Recog. (2012). https://doi.org/10.1109/CVPR.2012.6248074" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 588, + 480, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 588, + 480, + 620 + ], + "spans": [ + { + "bbox": [ + 138, + 588, + 480, + 620 + ], + "type": "text", + "content": "3. Hahner, M., Sakaridis, C., Dai, D., Van Gool, L.: Fog simulation on real lidar point clouds for 3d object detection in adverse weather. In: Int. Conf. Comput. Vis. (2021)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 622, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 622, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 138, + 622, + 480, + 643 + ], + "type": "text", + "content": "4. Hamilton, B., Tefft, B., Arnold, L., Grabowski, J.: Hidden highways: Fog and traffic crashes on america's roads (november 2014). Montana 40 (2006)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "text", + "content": "5. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. In: Adv. Neural Inform. Process. Syst. vol. 33 (2020)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "type": "text", + "content": "MonoWAD" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "text", + "content": "16. Huang, K.C., Wu, T.H., Su, H.T., Hsu, W.H.: Monodtr: Monocular 3d object detection with depth-aware transformer. In: IEEE Conf. Comput. Vis. Pattern Recog. (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 480, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 480, + 183 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 480, + 183 + ], + "type": "text", + "content": "17. Juneja, A., Kumar, V., Singla, S.K.: A systematic review on foggy datasets: Applications and challenges. Archives of Computational Methods in Engineering 29(3) (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 184, + 480, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 184, + 480, + 216 + ], + "spans": [ + { + "bbox": [ + 132, + 184, + 480, + 216 + ], + "type": "text", + "content": "18. Kim, J.U., Kim, H.I., Ro, Y.M.: Stereoscopic vision recalling memory for monocular 3d object detection. IEEE Trans. Image Process. 32 (2023). https://doi.org/10.1109/TIP.2023.3274479" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 217, + 480, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 217, + 480, + 239 + ], + "spans": [ + { + "bbox": [ + 132, + 217, + 480, + 239 + ], + "type": "text", + "content": "19. Kim, J.U., Park, S., Ro, Y.M.: Robust small-scale pedestrian detection with cued recall via memory learning. In: Int. Conf. Comput. Vis. (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 240, + 480, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 240, + 480, + 273 + ], + "spans": [ + { + "bbox": [ + 132, + 240, + 480, + 273 + ], + "type": "text", + "content": "20. Kim, J.U., Park, S., Ro, Y.M.: Uncertainty-guided cross-modal learning for robust multispectral pedestrian detection. IEEE Trans. Circuit Syst. Video Technol. 32(3) (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 274, + 480, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 274, + 480, + 306 + ], + "spans": [ + { + "bbox": [ + 132, + 274, + 480, + 306 + ], + "type": "text", + "content": "21. Kim, J.U., Park, S., Ro, Y.M.: Towards versatile pedestrian detector with multisensory-matching and multispectral recalling memory. In: AAAI. vol. 36 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 307, + 480, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 307, + 480, + 340 + ], + "spans": [ + { + "bbox": [ + 132, + 307, + 480, + 340 + ], + "type": "text", + "content": "22. Kim, S.W., Brown, B., Yin, K., Kreis, K., Schwarz, K., Li, D., Rombach, R., Torralba, A., Fidler, S.: Neuralfield-ldm: Scene generation with hierarchical latent diffusion models. In: IEEE Conf. Comput. Vis. Pattern Recog. (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 341, + 480, + 362 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 341, + 480, + 362 + ], + "spans": [ + { + "bbox": [ + 132, + 341, + 480, + 362 + ], + "type": "text", + "content": "23. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 363, + 480, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 363, + 480, + 396 + ], + "spans": [ + { + "bbox": [ + 132, + 363, + 480, + 396 + ], + "type": "text", + "content": "24. Lang, A.H., Vora, S., Caesar, H., Zhou, L., Yang, J., Beijbom, O.: Pointpillars: Fast encoders for object detection from point clouds. In: IEEE Conf. Comput. Vis. Pattern Recog. (2019)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 396, + 480, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 396, + 480, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 396, + 480, + 418 + ], + "type": "text", + "content": "25. Liu, X., Xue, N., Wu, T.: Learning auxiliary monocular contexts helps monocular 3d object detection. In: AAAI. vol. 36 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 419, + 480, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 419, + 480, + 441 + ], + "spans": [ + { + "bbox": [ + 132, + 419, + 480, + 441 + ], + "type": "text", + "content": "26. Liu, Z., Wu, Z., Toth, R.: Smoke: Single-stage monocular 3d object detection via keypoint estimation. In: IEEE Conf. Comput. Vis. Pattern Recog. Worksh. (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 442, + 480, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 442, + 480, + 464 + ], + "spans": [ + { + "bbox": [ + 132, + 442, + 480, + 464 + ], + "type": "text", + "content": "27. Liu, Z., Zhou, D., Lu, F., Fang, J., Zhang, L.: Autoshape: Real-time shape-aware monocular 3d object detection. In: Int. Conf. Comput. Vis. (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 464, + 480, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 464, + 480, + 497 + ], + "spans": [ + { + "bbox": [ + 132, + 464, + 480, + 497 + ], + "type": "text", + "content": "28. Lu, Y., Ma, X., Yang, L., Zhang, T., Liu, Y., Chu, Q., Yan, J., Ouyang, W.: Geometry uncertainty projection network for monocular 3d object detection. In: Int. Conf. Comput. Vis. (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 498, + 480, + 530 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 498, + 480, + 530 + ], + "spans": [ + { + "bbox": [ + 132, + 498, + 480, + 530 + ], + "type": "text", + "content": "29. Lugmayr, A., Danelljan, M., Romero, A., Yu, F., Timofte, R., Van Gool, L.: Repaint: Inpainting using denoising diffusion probabilistic models. In: IEEE Conf. Comput. Vis. Pattern Recog. (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 531, + 480, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 531, + 480, + 563 + ], + "spans": [ + { + "bbox": [ + 132, + 531, + 480, + 563 + ], + "type": "text", + "content": "30. Luo, J., Li, Y., Pan, Y., Yao, T., Feng, J., Chao, H., Mei, T.: Semantic-conditional diffusion networks for image captioning. In: IEEE Conf. Comput. Vis. Pattern Recog. (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 564, + 480, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 564, + 480, + 586 + ], + "spans": [ + { + "bbox": [ + 132, + 564, + 480, + 586 + ], + "type": "text", + "content": "31. Luo, S., Dai, H., Shao, L., Ding, Y.: M3dssd: Monocular 3d single stage object detector. In: IEEE Conf. Comput. Vis. Pattern Recog. (2021)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 587, + 480, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 587, + 480, + 609 + ], + "spans": [ + { + "bbox": [ + 132, + 587, + 480, + 609 + ], + "type": "text", + "content": "32. Ma, X., Liu, S., Xia, Z., Zhang, H., Zeng, X., Ouyang, W.: Rethinking pseudo-lidar representation. In: Eur. Conf. Comput. Vis. Springer (2020)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 610, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 610, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 610, + 480, + 632 + ], + "type": "text", + "content": "33. Mai, N.A.M., Duthon, P., Khoudour, L., Crouzil, A., Velastin, S.A.: 3d object detection with sls-fusion network in foggy weather conditions. Sensors 21(20) (2021)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "text", + "content": "34. Mousavian, A., Anguelov, D., Flynn, J., Kosecka, J.: 3d bounding box estimation using deep learning and geometry. In: IEEE Conf. Comput. Vis. Pattern Recog. (2017)" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "type": "text", + "content": "Y. Oh et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 666 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "35. Ni, H., Shi, C., Li, K., Huang, S.X., Min, M.R.: Conditional image-to-video generation with latent flow diffusion models. In: IEEE Conf. Comput. Vis. Pattern Recog. (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 481, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 481, + 172 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 481, + 172 + ], + "type": "text", + "content": "36. van den Oord, A., Vinyals, O., Kavukcuoglu, K.: Neural discrete representation learning (2018)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 172, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 172, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 130, + 172, + 481, + 194 + ], + "type": "text", + "content": "37. Peng, L., Wu, X., Yang, Z., Liu, H., Cai, D.: Did-m3d: Decoupling instance depth for monocular 3d object detection. In: Eur. Conf. Comput. Vis. Springer (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 194, + 481, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 194, + 481, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 194, + 481, + 236 + ], + "type": "text", + "content": "38. Pfeuffer, A., Dietmayer, K.: Robust semantic segmentation in adverse weather conditions by means of sensor data fusion. In: 2019 22th International Conference on Information Fusion (FUSION) (2019). https://doi.org/10.23919/FUSION43075.2019.9011192" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 237, + 481, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 237, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 130, + 237, + 481, + 270 + ], + "type": "text", + "content": "39. Qin, Q., Chang, K., Huang, M., Li, G.: Denet: Detection-driven enhancement network for object detection under adverse weather conditions. In: Asian Conf. Comput. Vis. (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 270, + 481, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 270, + 481, + 293 + ], + "spans": [ + { + "bbox": [ + 130, + 270, + 481, + 293 + ], + "type": "text", + "content": "40. Qin, Z., Li, X.: Monoground: Detecting monocular 3d objects from the ground. In: IEEE Conf. Comput. Vis. Pattern Recog. (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 293, + 481, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 293, + 481, + 326 + ], + "spans": [ + { + "bbox": [ + 130, + 293, + 481, + 326 + ], + "type": "text", + "content": "41. Reading, C., Harakeh, A., Chae, J., Waslander, S.L.: Categorical depth distribution network for monocular 3d object detection. In: IEEE Conf. Comput. Vis. Pattern Recog. (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 326, + 481, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 326, + 481, + 358 + ], + "spans": [ + { + "bbox": [ + 130, + 326, + 481, + 358 + ], + "type": "text", + "content": "42. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: IEEE Conf. Comput. Vis. Pattern Recog. (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 358, + 481, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 358, + 481, + 379 + ], + "spans": [ + { + "bbox": [ + 130, + 358, + 481, + 379 + ], + "type": "text", + "content": "43. Sakaridis, C., Dai, D., Van Gool, L.: Semantic foggy scene understanding with synthetic data. Int. J. Comput. Vis. 126 (2018)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 380, + 481, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 380, + 481, + 413 + ], + "spans": [ + { + "bbox": [ + 130, + 380, + 481, + 413 + ], + "type": "text", + "content": "44. Shi, S., Guo, C., Jiang, L., Wang, Z., Shi, J., Wang, X., Li, H.: Pv-rcnn: Point-voxel feature set abstraction for 3d object detection. In: IEEE Conf. Comput. Vis. Pattern Recog. (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 413, + 481, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 413, + 481, + 435 + ], + "spans": [ + { + "bbox": [ + 130, + 413, + 481, + 435 + ], + "type": "text", + "content": "45. Shi, S., Wang, X., Li, H.: Pointcnn: 3d object proposal generation and detection from point cloud. In: IEEE Conf. Comput. Vis. Pattern Recog. (2019)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 435, + 481, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 435, + 481, + 468 + ], + "spans": [ + { + "bbox": [ + 130, + 435, + 481, + 468 + ], + "type": "text", + "content": "46. Shi, X., Ye, Q., Chen, X., Chen, C., Chen, Z., Kim, T.K.: Geometry-based distance decomposition for monocular 3d object detection. In: Int. Conf. Comput. Vis. (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 468, + 481, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 468, + 481, + 490 + ], + "spans": [ + { + "bbox": [ + 130, + 468, + 481, + 490 + ], + "type": "text", + "content": "47. Simonelli, A., Bulo, S.R., Porzi, L., Lopez-Antequera, M., Kontschieder, P.: Disentangling monocular 3d object detection. In: Int. Conf. Comput. Vis. (2019)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 491, + 481, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 491, + 481, + 522 + ], + "spans": [ + { + "bbox": [ + 130, + 491, + 481, + 522 + ], + "type": "text", + "content": "48. Wang, L., Du, L., Ye, X., Fu, Y., Guo, G., Xue, X., Feng, J., Zhang, L.: Depth-conditioned dynamic message propagation for monocular 3d object detection. In: IEEE Conf. Comput. Vis. Pattern Recog. (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 522, + 481, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 522, + 481, + 555 + ], + "spans": [ + { + "bbox": [ + 130, + 522, + 481, + 555 + ], + "type": "text", + "content": "49. Wu, R.Q., Duan, Z.P., Guo, C.L., Chai, Z., Li, C.: Ridcp: Revitalizing real image dehazing via high-quality codebook priors. In: IEEE Conf. Comput. Vis. Pattern Recog. (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 555, + 481, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 555, + 481, + 578 + ], + "spans": [ + { + "bbox": [ + 130, + 555, + 481, + 578 + ], + "type": "text", + "content": "50. Yang, S., Scherer, S.: Cubeslam: Monocular 3-d object slam. IEEE Transactions on Robotics 35(4) (2019). https://doi.org/10.1109/TR0.2019.2909168" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 578, + 481, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 578, + 481, + 610 + ], + "spans": [ + { + "bbox": [ + 130, + 578, + 481, + 610 + ], + "type": "text", + "content": "51. Yang, X., Mi, M.B., Yuan, Y., Wang, X., Tan, R.T.: Object detection in foggy scenes by embedding depth and reconstruction into domain adaptation. In: Asian Conf. Comput. Vis. (2022)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 130, + 610, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 610, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 130, + 610, + 481, + 632 + ], + "type": "text", + "content": "52. Yu, F., Wang, D., Shelhamer, E., Darrell, T.: Deep layer aggregation. In: IEEE Conf. Comput. Vis. Pattern Recog. (2018)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 130, + 632, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 632, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 632, + 481, + 666 + ], + "type": "text", + "content": "53. Yurtsever, E., Lambert, J., Carballo, A., Takeda, K.: A survey of autonomous driving: Common practices and emerging technologies. IEEE Access 8 (2020). https://doi.org/10.1109/ACCESS.2020.2983149" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "type": "text", + "content": "MonoWAD" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 249 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 481, + 160 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 481, + 160 + ], + "type": "text", + "content": "54. Zhang, C., Wang, H., Cai, Y., Chen, L., Li, Y., Sotelo, M.A., Li, Z.: Robust-fusionnet: Deep multimodal sensor fusion for 3-d object detection under severe weather conditions. IEEE Transactions on Instrumentation and Measurement 71 (2022). https://doi.org/10.1109/TIM.2022.3191724" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 161, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 161, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 161, + 481, + 194 + ], + "type": "text", + "content": "55. Zhang, R., Qiu, H., Wang, T., Guo, Z., Cui, Z., Qiao, Y., Li, H., Gao, P.: Monodetr: Depth-guided transformer for monocular 3d object detection. In: Int. Conf. Comput. Vis. (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 194, + 481, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 481, + 216 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 481, + 216 + ], + "type": "text", + "content": "56. Zhou, Y., Tuzel, O.: Voxelnet: End-to-end learning for point cloud based 3d object detection. In: IEEE Conf. Comput. Vis. Pattern Recog. (2018)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 216, + 481, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 481, + 249 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 481, + 249 + ], + "type": "text", + "content": "57. Zhou, Y., He, Y., Zhu, H., Wang, C., Li, H., Jiang, Q.: Monocular 3d object detection: An extrinsic parameter free approach. In: IEEE Conf. Comput. Vis. Pattern Recog. (2021)" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "type": "text", + "content": "Y. Oh et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Monocular Occupancy Prediction for Scalable Indoor Scenes/325d434b-3564-4d20-a82e-4ec4e345eb58_content_list.json b/2024/Monocular Occupancy Prediction for Scalable Indoor Scenes/325d434b-3564-4d20-a82e-4ec4e345eb58_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..61c31dafefd87ad5eb8a4dca7d4615d22b8705f7 --- /dev/null +++ b/2024/Monocular Occupancy Prediction for Scalable Indoor Scenes/325d434b-3564-4d20-a82e-4ec4e345eb58_content_list.json @@ -0,0 +1,1824 @@ +[ + { + "type": "text", + "text": "Monocular Occupancy Prediction for Scalable Indoor Scenes", + "text_level": 1, + "bbox": [ + 238, + 140, + 763, + 183 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hongxiao $\\mathrm{Yu}^{1,2}$ Yuqi Wang $^{1,2}$ , Yuntao Chen $^{3}$ , and Zhaoxiang Zhang $^{1,2,3}$", + "bbox": [ + 215, + 210, + 785, + 228 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 School of Artificial Intelligence, University of Chinese Academy of Sciences (UCAS)", + "bbox": [ + 215, + 238, + 782, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ NLPR, MAIS, Institute of Automation, Chinese Academy of Sciences (CASIA)", + "bbox": [ + 228, + 253, + 769, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 Centre for Artificial Intelligence and Robotics (HKISI_CAS) {yuhongxiao2023, wangyuqi2020, zhaoxiang.zhang}@ia.ac.cn, chenyuntao08@gmail.com", + "bbox": [ + 281, + 268, + 720, + 309 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Camera-based 3D occupancy prediction has recently garnered increasing attention in outdoor driving scenes. However, research in indoor scenes remains relatively unexplored. The core differences in indoor scenes lie in the complexity of scene scale and the variance in object size. In this paper, we propose a novel method, named ISO, for predicting indoor scene occupancy using monocular images. ISO harnesses the advantages of a pretrained depth model to achieve accurate depth predictions. Furthermore, we introduce the Dual Feature Line of Sight Projection (D-FLoSP) module within ISO, which enhances the learning of 3D voxel features. To foster further research in this domain, we introduce Occ-ScanNet, a large-scale occupancy benchmark for indoor scenes. With a dataset size 40 times larger than the NYUv2 dataset, it facilitates future scalable research in indoor scene analysis. Experimental results on both NYUv2 and Occ-ScanNet demonstrate that our method achieves state-of-the-art performance. The dataset and code are made publicly at https://github.com/hongxiaoy/ISO.git.", + "bbox": [ + 259, + 342, + 738, + 564 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: 3D occupancy prediction $\\cdot$ Camera-based scene understanding $\\cdot$ Semantic scene completion", + "bbox": [ + 259, + 575, + 740, + 604 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 628, + 375, + 643 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D scene understanding is a crucial task in computer vision, becoming increasingly important for applications such as robotic navigation [10], augmented reality [2], and autonomous driving [47]. While humans possess a natural ability to comprehend 3D environments through vision, enabling this ability in computers poses a significant challenge due to limitations like restricted fields of view, sparse sensing capabilities, and measurement noise. However, with the rapid development of deep learning and the availability of large-scale 3D driving datasets, camera-based 3D object detection [14, 15, 20, 21, 35, 37, 39, 43] and occupancy prediction [5, 22, 33, 38, 40] have developed rapidly, achieving significant improvements in performance.", + "bbox": [ + 212, + 657, + 787, + 809 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Occupancy prediction has gained popularity recently due to its effectiveness in representing both background and foreground objects within a scene using a", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "unified representation. Although significant progress has been made in outdoor driving scenarios, research in indoor scenarios remains limited. Indoor scenes differ from outdoor driving scenes in two key aspects: (1) scene-scale complexity: Indoor rooms often exhibit a more diverse range of sizes compared to outdoor environments, where driving scenarios typically focus on a fixed 3D space for perception. This diversity, ranging from spacious living rooms to narrow kitchens, poses higher precision requirements for depth prediction. (2) object complexity: Indoor scenes feature a higher density and greater variety of objects. Unlike outdoor objects such as vehicles and pedestrians, which typically have consistent sizes within their respective categories and are relatively well-separated, indoor furniture and other objects often exhibit significant variations in scale and are closely positioned to each other. This increased complexity necessitates more sophisticated 3D perception techniques to accurately capture and understand the intricate geometry and relationships among objects within indoor spaces. Furthermore, existing works [5,45] focusing on indoor scenes primarily utilize the NYUv2 dataset [29], which lacks a more scalable and generalizable benchmark for comprehensive evaluation.", + "bbox": [ + 212, + 146, + 787, + 402 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the above issues, we firstly introduce a more scalable benchmark, Occ-ScanNet, for 3D occupancy prediction in indoor scenes. This benchmark builds upon the large-scale ScanNet [8] dataset, offering a 40 times more samples compared to the NYUv2 [29] dataset, thus greatly expanding the scope and diversity of indoor scene study. To effectively address the unique complexities of indoor scenes, we propose a novel method named ISO (Indoor Scene Occupancy). ISO leverages a powerful pretrained depth model and integrates a D-FLoSP (Dual Feature Line of Sight Projection) module. This module enables precise depth estimation and facilitates learning voxel features for accurate predictions. To accommodate the varying sizes of indoor objects, we introduce a multi-scale feature fusion module. This module enhances the learning of object features across different scales. Leveraging a powerful depth model, our method is adept at handling diverse indoor scenes, offering a robust and versatile solution.", + "bbox": [ + 212, + 405, + 787, + 602 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our main contributions can be summarized as follows.", + "bbox": [ + 238, + 604, + 632, + 618 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce a new benchmark called Occ-ScanNet for monocular 3D occupancy prediction in indoor scenes. With a dataset size 40 times larger than the NYUv2 dataset, it significantly enhances the potential for scalable research in indoor scene analysis.", + "- We propose a novel approach called $ISO$ , which primarily comprises the DFLoSP (Dual Feature Line of Sight Projection) module and a multi-scale feature fusion module. Together, these components effectively address the challenges posed by variations in scene and object sizes, enabling more accurate and robust 3D occupancy prediction.", + "- Experiments on the Occ-ScanNet and NYUv2 datasets demonstrate that our approach achieves state-of-the-art performance in monocular 3D occupancy prediction. Furthermore, our method exhibits significant scalability potential." + ], + "bbox": [ + 223, + 638, + 784, + 838 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "H. Yu et al.", + "bbox": [ + 271, + 114, + 351, + 127 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2 Related works", + "text_level": 1, + "bbox": [ + 215, + 143, + 392, + 160 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Monocular 3D Semantic Scene Completion", + "text_level": 1, + "bbox": [ + 215, + 176, + 616, + 191 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Monocular 3D Semantic Scene Completion (SSC) aims to infer the complete 3D structure and corresponding semantics from a single image. This monocular setting was first introduced by Monoscene [5], advancing upon prior SSC methods [18,23,30,48,51] by relying solely on vision, without additional 3D inputs. Monoscene [5] introduces the Features Line of Sight Projection (FLoSP), inspired by optics, to obtain 3D features through ray projection. However, the shared 2D features lifted to 3D rays via FLoSP results in depth ambiguity. Consequently, subsequent efforts have placed greater emphasis on leveraging depth information. VoxFormer [19] proposed a novel query proposal network based on 2D convolutions, generating sparse queries from image depth, which showed impressive performance in driving scenes. Meanwhile, NDC-Scene [45] further devised a Depth-Adaptive Dual Decoder to concurrently upsample and merge the 2D and 3D feature maps, thereby enhancing overall performance.", + "bbox": [ + 212, + 200, + 787, + 398 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2 Multiview 3D Occupancy Prediction", + "text_level": 1, + "bbox": [ + 215, + 420, + 565, + 435 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Multiview 3D Occupancy Prediction predicts the semantic occupancy of the surrounding 3D scene given multiview images, has recently garnered significant attention in the field of autonomous driving. TPVFormer [17] pioneers exploration in driving scenes, employing sparse LiDAR labels for supervision. It introduces a tri-perspective view (TPV) representation that accompanies BEV with two additional perpendicular planes. However, due to the sparse supervision provided by LiDAR, subsequent works [33, 34, 36, 40] have focused on providing more dense occupancy benchmarks. FB-OCC [22] integrates the lifting of 2D to 3D and querying of 3D to 2D, achieving a more efficient 3D feature transformation. Meanwhile, PanoOcc [38] unifies detection and semantic occupancy tasks, enabling comprehensive panoramic scene understanding. Recent methods also explore improving model efficiency [46] and utilizing weaker forms of supervision [16, 25].", + "bbox": [ + 212, + 445, + 787, + 642 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3 3D Reconstruction from Image", + "text_level": 1, + "bbox": [ + 215, + 664, + 517, + 679 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3D reconstruction is a technique that involves recovering a 3D representation of objects [1, 11, 26], scenes [9, 41, 49], or even human bodies [12, 13, 50] from camera images. The task of 3D reconstruction can be classified into two categories: monocular reconstruction and multi-view reconstruction, depending on the number of images utilized. In the context of indoor scenes, 3D scene reconstruction aims to determine the surface geometry of the entire scene, often without incorporating semantic information. Atlas [24] proposes an end-to-end 3D reconstruction framework using TSDF regression from RGB images, bypassing traditional depth map estimation for efficient semantic segmentation. Panoptic-Reconstruction [7] unifies geometric reconstruction, 3D semantic, and", + "bbox": [ + 212, + 689, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Monocular Occupancy Prediction for Scalable Indoor Scenes", + "bbox": [ + 328, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/cb27d372c27774b713bc3f76a42488bc37ffc0f9440739cdc29e72df35393e5f.jpg", + "image_caption": [ + "Fig. 1: The core design of ISO centers around the transformation of features from 2D to 3D spaces, encompassing the Depth Branch and the D-FLoSP module. A depth branch is initially integrated, it leverages a pre-trained depth model to estimate a pixel-wise depth map which is processed by the DepthNet to generate the final depth distribution. An element-wise multiplication between the voxel depth and features followed by summation are subsequently performed to derive the initial 3D voxel feature. The 3D feature is further processed to predict the 3D scene occupancy." + ], + "image_footnote": [], + "bbox": [ + 215, + 150, + 785, + 287 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "instance segmentation tasks. It predicts complete geometric reconstruction, semantic, and instance segmentations from a single RGB image's camera frustum. SCFusion [42], a real-time scene reconstruction framework, integrates continuous depth data using neural architecture for occupancy maps. It efficiently combines semantic completion with voxel states for simultaneous scene reconstruction and semantic understanding in real-time 3D environments.", + "bbox": [ + 212, + 429, + 787, + 521 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 547, + 330, + 565 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Problem definition. We focus on the problem of monocular 3D Occupancy Prediction. Specifically, this task takes a single RGB image $\\mathbf{I}^{RGB}$ as input and output a voxel-wise occupancy along with semantic categories $\\mathbf{Y}^{X\\times Y\\times Z\\times C}$ . $X$ , $Y$ , and $Z$ represent the dimensions of the predicted 3D scene, while $C$ denotes the total number of semantic categories.", + "bbox": [ + 212, + 585, + 787, + 662 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Model overview. In this section, we introduce the overall architecture of ISO, as shown in Fig. 1. Following Monoscene [5], we utilize 2D Unet and 3D Unet architectures to handle 2D and 3D features. Our core design focuses on transforming features from 2D to 3D, incorporating the Depth Branch and D-FLoSP module. Specifically, we incorporate a depth branch to estimate the pixel-wise depth map using a pre-trained depth model. Processed by the DepthNet, the model outputs a refined depth distribution. Then, voxel depth and features are multiplied element-wise and summed to obtain the initial 3D voxel feature volume $\\mathbf{X}^{3\\mathbf{d}} \\in \\mathbb{R}^{X \\times Y \\times Z \\times C}$ . Finally, after processing the 3D voxel features, the model outputs the 3D scene occupancy.", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "H. Yu et al.", + "bbox": [ + 271, + 114, + 352, + 126 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Depth Branch", + "text_level": 1, + "bbox": [ + 215, + 146, + 380, + 161 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we introduce how to effectively estimate depth information from a single image.", + "bbox": [ + 212, + 167, + 782, + 198 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Coarse depth estimation. Learning depth from scratch can be quite challenging. However, thanks to the recent rapid advancements in depth estimation [3,4,27,44], we can leverage pre-trained depth models to initially estimate a coarse depth map.", + "bbox": [ + 212, + 217, + 782, + 276 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Compared to past models that could only estimate relative depth, we opted for Depth-Anything [44] as the pre-trained depth model, because it excels in predicting metric depth, as the Eq. (1) shows:", + "bbox": [ + 212, + 277, + 784, + 324 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {D} ^ {\\mathrm {m e t r i c}} = \\mathbf {N} _ {\\mathrm {d e p t h}} (\\mathbf {I} ^ {\\mathrm {r g b}}) \\in \\mathbb {R} ^ {1 \\times H \\times W}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 372, + 328, + 785, + 348 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{N}_{\\mathrm{depth}}$ denotes the pre-trained depth model [3, 44], and $\\mathbf{I}^{\\mathrm{rgb}}$ represents the input image. $H$ and $W$ represents the height and width of the input image.", + "bbox": [ + 214, + 354, + 782, + 386 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Depth refinement. The coarse depth estimation is subsequently refined through model learning. The depth from pre-trained depth model is not precise enough for a higher mIoU, so we design a fine-tuning strategy. Specifically, the predicted metric depth $\\mathbf{D}^{\\mathrm{metric}}$ is then concatenated with the image feature $\\mathbf{X}$ that has the same spatial scale. The augmented feature is continue processed by a following DepthNet to get a refined depth distribution $\\mathbf{D}_{s=1}^{\\mathrm{dist}}$ , as Eq. (2) shows:", + "bbox": [ + 214, + 404, + 789, + 496 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {D} _ {\\mathrm {s} = 1} ^ {\\text {d i s t}} = \\mathbf {F} _ {\\text {d e p t h}} \\left(\\operatorname {C o n c a t} \\left(\\mathbf {D} ^ {\\text {m e t r i c}}, \\mathbf {X} _ {\\mathrm {s} = 1} ^ {2 \\mathrm {d}}\\right)\\right), \\quad \\mathbf {D} _ {\\mathrm {s} = 1} ^ {\\text {d i s t}} \\in \\mathbb {R} ^ {N _ {\\text {b i n s}} \\times H \\times W}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 267, + 513, + 784, + 547 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{F}_{\\mathrm{depth}}$ is the DepthNet, $\\mathbf{X}_{s=1}^{2d}$ is the 2D image feature, and $N_{\\mathrm{bins}}$ is the number of discrete depth bins. $s$ denotes the level of feature scale, where $s = 1$ implies that the feature is at a 1:1 original image ratio.", + "bbox": [ + 212, + 554, + 782, + 599 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We down-samples the scale $\\frac{1}{1}$ distribution to get other smaller-scale 2D image feature maps' depth distribution. We represent the procedure as Eq. (3):", + "bbox": [ + 212, + 599, + 782, + 631 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {D} _ {\\mathrm {s} = k} ^ {\\text {d i s t}} = \\operatorname {D o w n S a m p l e} \\left(\\mathbf {D} _ {\\mathrm {s} = 1} ^ {\\text {d i s t}}\\right), k = 2, 4, 8. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 346, + 636, + 784, + 655 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The refined depth is supervised by the ground truth $\\mathbf{D}^{\\mathrm{GT}}$ . The ground truth has single depth value at each pixel, so we convert the single value to a one-hot vector of length $N_{\\mathrm{bins}}$ . The refined depth prediction can be optimized using the BCE loss function as Eq. (4) shows.", + "bbox": [ + 212, + 662, + 782, + 724 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\mathrm {d e p t h}} = - \\frac {1}{N _ {\\mathrm {b i n s}} \\times H \\times W} \\\\ \\sum_ {d = 1} ^ {N _ {\\mathrm {b i n s}}} \\sum_ {h = 1} ^ {H} \\sum_ {w = 1} ^ {W} \\left[ \\mathbf {D} _ {N _ {\\mathrm {b i n s}}, d, h, w} ^ {\\mathrm {G T}} \\cdot \\log \\left(\\mathbf {D} _ {N _ {\\mathrm {b i n s}}, d, h, w} ^ {\\mathrm {d i s t}}\\right) \\right. \\\\ \\left. + \\left(1 - \\mathbf {D} _ {N _ {\\mathrm {b i n s}}, d, h, w} ^ {\\mathrm {G T}}\\right) \\cdot \\log \\left(1 - \\mathbf {D} _ {N _ {\\mathrm {b i n s}}, d, h, w} ^ {\\mathrm {d i s t}}\\right) \\right] \\quad (4) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 228, + 744, + 785, + 843 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Monocular Occupancy Prediction for Scalable Indoor Scenes", + "bbox": [ + 328, + 114, + 732, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Dual Feature Line of Sight Projection (D-FLoSP)", + "text_level": 1, + "bbox": [ + 215, + 146, + 671, + 162 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Compared to the FLoSP module in MonoScene [5], which projects 3D voxels to 2D, and 3D feature vectors are the corresponding pixel feature vectors. It only considers the projection of 3D rays and overlooks depth information, our proposed D-FLoSP module can more effectively integrate depth information.", + "bbox": [ + 212, + 180, + 782, + 241 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Specifically, after getting depth distribution in the camera frame, we implement the FLoSP module to depth distribution. Each 3D voxel centroid position $x^{c}$ in the world frame can be projected to the camera frame and the pixel frame using camera pose and intrinsic matrix. We assign the projected depth values to discrete depth bins index by Eqs. (5) and (6), according to [32],", + "bbox": [ + 212, + 243, + 784, + 318 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\delta = \\frac {2 \\left(d _ {\\max } - d _ {\\min }\\right)}{N _ {b i n s} \\left(1 + N _ {b i n s}\\right)}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 382, + 333, + 784, + 364 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nl = - 0. 5 + 0. 5 \\sqrt {1 + \\frac {8 \\left(d - d _ {\\operatorname* {m i n}}\\right)}{\\delta}}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 385, + 369, + 784, + 401 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "and the depth distribution probability it lies in serves as the depth probability of that voxel. The process is illustrated by Eq. (7):", + "bbox": [ + 212, + 415, + 782, + 445 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {D} _ {\\mathrm {s} = k} ^ {\\mathrm {3 d}} = \\Phi_ {\\rho \\left(x ^ {c}\\right)} ^ {\\mathrm {3 d}} \\left(\\mathbf {D} _ {\\mathrm {s} = k} ^ {\\mathrm {d i s t}}\\right) \\in \\mathbb {R} ^ {X \\times Y \\times Z \\times C}, k = 1, 2, 4, 8, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 323, + 459, + 784, + 479 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\Phi_{a}^{3\\mathrm{d}}(b)$ is the 3D sampling of b at coordinates a, and $\\rho (\\cdot)$ is the perspective projection. Thus, the depth distribution is lifted from 2D to 3D. The bin-based representation with quantization is used for mapping continuous depth value to a discrete form, each bin will be assigned a different depth distribution score to make voxel aware of depth. The original depth precision is a continuous value, and hard to integrate it with the discrete form of voxel.", + "bbox": [ + 212, + 494, + 784, + 584 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Following [5], we also employ the FLoSP module to generate a 3D voxel feature from a 2D feature map. Similar to what we did with the voxel centroids, the projected voxel can sample corresponding $1:k$ scaled feature maps $F_{1:k}$ from the 2D UNet decoder. This process is illustrated by Equation Eq. (8):", + "bbox": [ + 212, + 585, + 784, + 648 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {X} _ {\\mathrm {s} = k} ^ {3 \\mathrm {d}} = \\Phi_ {\\rho \\left(x ^ {c}\\right)} ^ {2 \\mathrm {d}} \\left(\\mathbf {X} _ {\\mathrm {s} = k} ^ {2 \\mathrm {d}}\\right) \\in \\mathbb {R} ^ {X \\times Y \\times Z \\times C}, k = 1, 2, 4, 8, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 323, + 662, + 784, + 681 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\Phi_{a}^{2\\mathrm{d}}(b)$ is the 2D sampling of b at coordinates a, and $\\rho (\\cdot)$ is the perspective projection.", + "bbox": [ + 212, + 696, + 782, + 727 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The final 3D feature map $\\mathbf{X}^{3\\mathrm{d}}$ are summed in Eq. (9):", + "bbox": [ + 238, + 729, + 630, + 744 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {X} ^ {\\mathrm {3 d}} = \\sum_ {\\mathrm {s} \\in \\{1, 2, 4, 8 \\}} \\mathbf {X} _ {\\mathrm {s} = k} ^ {\\mathrm {3 d}} \\odot \\mathbf {D} _ {\\mathrm {s} = k} ^ {\\mathrm {3 d}}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 392, + 760, + 784, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\odot$ is the element-wise multiplication. The output map $\\mathbf{X}^{3\\mathrm{d}}$ is then served as the input of 3D UNet.", + "bbox": [ + 212, + 809, + 782, + 839 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "H. Yu et al.", + "bbox": [ + 271, + 114, + 351, + 127 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4e4b51667dfd71e9661d016ed42f74358aa31bb99bf8b59fd60904e8f7f647cb.jpg", + "image_caption": [ + "(a) Depth ranges in NYUv2 and our Occ-ScanNet Benchmark" + ], + "image_footnote": [], + "bbox": [ + 271, + 162, + 429, + 287 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/225cd6aab82aef724a36995945316cee74bdd5373a4c636b37ae5faa67dde27d.jpg", + "image_caption": [ + "(b) Dataset attributes of NYUv2 and our OccScanNet Benchmark", + "Fig. 2: Comparison of NYUv2 and Occ-ScanNet Benchmark. In (a), the depth ranges of NYUv2 and Occ-ScanNet are distinguished by dark and light green, respectively, with the horizontal axis indicating the minimum depth and the vertical axis showing the maximum depth of scenes. (b) quantitatively demonstrates that Occ-ScanNet possesses a significantly larger data scale compared to the NYUv2 dataset." + ], + "image_footnote": [], + "bbox": [ + 558, + 162, + 717, + 286 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4 Occ-ScanNet Benchmark", + "text_level": 1, + "bbox": [ + 215, + 426, + 493, + 441 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Compared to the previous widely used indoor scene benchmark NYUv2 [29], which contains only 795 / 654 for train / test samples, our benchmark boasts 45,755 / 19,764 samples. Our benchmark significantly surpasses NYUv2 [29] in both data quantity and richness of scene depth, as illustrated in Fig. 2. This dataset is available at https://huggingface.co/datasets/hongxiaoy/OccScanNet.", + "bbox": [ + 212, + 455, + 787, + 546 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1 Overview", + "text_level": 1, + "bbox": [ + 215, + 565, + 339, + 579 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Occ-ScanNet benchmark features a train/Validation split of 45,755 / 19,764 samples. As shown in Fig. 3, Occ-ScanNet exhibits rich diversity in scenes and viewpoints. This diverse dataset not only challenges the task of predicting scene occupancy but also fosters future research endeavors towards developing larger-scale and more versatile occupancy models.", + "bbox": [ + 212, + 587, + 785, + 662 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2 Occupancy Label Generation", + "text_level": 1, + "bbox": [ + 215, + 681, + 503, + 698 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We adhere to the data formulation used in NYUv2 [29] dataset. The pipeline of occupancy generation process is showed in Fig. 4. For generating 3D voxel labels, we initially followed the generation process in CompleteScanNet and then employed manual inspection.", + "bbox": [ + 212, + 704, + 785, + 763 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "More specifically, from the official ScanNet [8] scenes, we extracted various data components, including color images, depth images, camera intrinsic matrices, and camera poses. A total of 100 frames were sampled from each scene and were randomly divided into training and validation sets with a $7/3$ split ratio per scene.", + "bbox": [ + 212, + 763, + 785, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Monocular Occupancy Prediction for Scalable Indoor Scenes", + "bbox": [ + 328, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c16bba8f8ca001189e36dac9dacbdb0c0ff618762516ff7dee305a17b86e6b84.jpg", + "image_caption": [ + "Fig. 3: Samples Visualization in Occ-ScanNet Benchmark. The original RGB image is shown in column 1,3 and 5, the corresponding scene voxel labels is shown in column 2, 4 and 6. The first two rows are different views from different scenes and the last two rows each is three different views from the same scene." + ], + "image_footnote": [], + "bbox": [ + 225, + 152, + 776, + 428 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Manual inspection. The manual inspection involves verifying camera poses, camera intrinsic, and voxel positions. We filter out erroneous samples to prevent convergence issues in the model training process.", + "bbox": [ + 212, + 535, + 787, + 583 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Voxel label assignment. For each frame, only a specific area in front of the camera is defined for analysis. Therefore, the selection of the voxel origin is paramount, as it determines the subsequent coordinates of other voxel. Subsequently, each voxel is assigned a label based on its nearest voxel in the CompleteScanNet dataset. Additionally, we excluded frames with a ratio of unknown or empty labels exceeding $95\\%$ or frames where the number of semantic label classes was less than two. Consequently, we obtained the generated 3D voxel labels for each frame.", + "bbox": [ + 212, + 604, + 787, + 726 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5 Experiments", + "text_level": 1, + "bbox": [ + 215, + 750, + 377, + 767 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.1 Setup", + "text_level": 1, + "bbox": [ + 215, + 782, + 310, + 799 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Datasets. The NYUv2 [29] dataset provides scenes represented in $240 \\times 144 \\times 240$ voxels grids, labeled with 13 classes, including 1 for free space, 1 for unknown,", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "H. Yu et al.", + "bbox": [ + 271, + 114, + 351, + 126 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/600bf412f151f2798f467dfc92b731ba91e8e72b94db45aad736c5d837165504.jpg", + "image_caption": [ + "Fig. 4: Pipeline of Occ-ScanNet dataset label generation. Color images, depth images, camera intrinsic and poses are extracted from ScanNet scenes. For each scene, 100 frames were sampled and randomly split into training and validation sets with a $7/3$ ratio. Frames with invalid camera poses or exceeding scene boundaries were filtered out. Only the area in front of the camera was analyzed, necessitating careful selection of the voxel origin. Voxel were labeled based on their nearest voxel in the CompleteScanNet dataset. Frames with $>95\\%$ unknown/empty labels or $<2$ semantic classes were excluded, resulting in generated 3D voxel labels for each frame." + ], + "image_footnote": [], + "bbox": [ + 222, + 150, + 777, + 200 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "and 11 for specific semantics (ceiling, floor, wall, window, chair, bed, sofa, table, tvs, furniture, objects). The dataset consists of 795 / 654 scenes in the train / test splits. The model is trained and evaluated on down-sampled $60 \\times 36 \\times 60$ voxels.", + "bbox": [ + 212, + 369, + 787, + 429 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Occ-ScanNet dataset provides scenes represented in $60 \\times 60 \\times 36$ voxel grids, labeled with 12 classes including 1 for free space, and 11 for specific semantics (ceiling, floor, wall, window, chair, bed, sofa, table, tvs, furniture, objects). The dataset comprises 45,755 / 19,764 frames in the train / val splits. The model is trained and evaluated on the original scale.", + "bbox": [ + 212, + 433, + 787, + 508 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Occ-ScanNet-mini dataset has the same class number and voxel scene size as Occ-ScanNet, except that this mini dataset consists of 4,639 / 2,007 frames in the train / val splits. The model is also trained and evaluated on the original scale.", + "bbox": [ + 212, + 512, + 787, + 571 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Implementation Details. We employ a pre-trained EfficientNet-B7 [31] as the encoder in our 2D UNet architecture. In the depth branch, we utilize a pre-trained DepthAnything model [44], which remains frozen during training. Additionally, we integrate a depth loss specific to the depth branch, complementing the other losses outlined in Monoscene [5]. Our model is trained on two datasets: NYUv2 [29] and Occ-ScanNet. For NYUv2, ISO undergoes 30 epochs of training using AdamW optimizers. Initially, the learning rate is set to 5e-6 for the DepthNet and 1e-4 for other components, with a learning rate reduction by 0.1 at epoch 20. Training on NYUv2 takes approximately 7 hours using 2 NVIDIA L20 GPUs (2 items per GPU). On the Occ-ScanNet-mini dataset, ISO is trained for 60 epochs under similar learning rate settings, decreasing the rate by 0.1 at epoch 40. This training process takes around 12 hours using 8 L20 GPUs. For Occ-ScanNet, ISO is trained for 10 epochs with the same learning rate schedule as used for Occ-ScanNet-mini, requiring approximately a day using 8 L20 GPUs.", + "bbox": [ + 212, + 613, + 787, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Monocular Occupancy Prediction for Scalable Indoor Scenes", + "bbox": [ + 328, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5.2 Main Results", + "text_level": 1, + "bbox": [ + 215, + 146, + 372, + 159 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/2a104e60b4f82f6c7f71fdac10f084e75fa5921047543bae1749de859d180f6f.jpg", + "table_caption": [ + "Table 1: Performance on the Occ-ScanNet dataset" + ], + "table_footnote": [], + "table_body": "
MethodInputIoUceilingfloorwallwindowchairbedsofatabletvsfurnitureobjects
MonoScene* [5]xrgb41.6015.1744.7122.4112.5526.1127.0335.9128.326.5732.1619.84
ISO(Ours)xrgb42.1619.8841.8822.3716.9829.0942.4342.0029.6010.6236.3624.61
", + "bbox": [ + 228, + 227, + 772, + 338 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Occ-ScanNet performance. We first evaluate our model's performance on the large-scale Occ-ScanNet dataset. As shown in Tab. 1, the results indicate that our method significantly outperforms MonoScene [5]. The * denotes results obtained using their code trained on our dataset. In Fig. 5, we also conducted qualitative visualization comparisons.", + "bbox": [ + 212, + 396, + 787, + 472 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "NYUv2 performance. The NYUv2 [29] dataset serves as a widely used benchmark for indoor scene evaluation. The results in Tab. 2 demonstrate that our method achieves state-of-the-art performance. Our method also demonstrates substantial performance improvement on the NYUv2 dataset, in addition to the gains observed on Occ-ScanNet. As illustrated in Fig. 6, our ISO demonstrates superior room layout prediction compared to [5], particularly evident in rows 1 and 2. This improvement can be attributed to incorporating depth knowledge into the model. Additionally, our method exhibits the capability of learning small objects, such as pictures on the wall in row 3.", + "bbox": [ + 212, + 500, + 789, + 635 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/f7253140a0003ccaf15ae943cff655b57cc59fb75729c5c34e91a9ab0459850d.jpg", + "table_caption": [ + "Table 2: Performance on the NYUv2 dataset" + ], + "table_footnote": [], + "table_body": "
MethodInputIoUceilingfloorwallwindowchairbedsofatabletvsfurnitureobjects
■■■■■■■■■■■
LMSCNetrgb [28]xocc33.934.4988.414.630.253.9432.0315.446.570.0214.514.3915.88
AICNetrgb [18]xrgb, xdepth30.037.5882.979.150.056.9335.8722.9211.110.7115.906.4518.15
3DSketchrgb [6]xrgb, xTSDF38.648.5390.459.945.6710.6442.2929.2113.889.3823.838.1922.91
MonoScene [5]xrgb42.518.8993.5012.0612.5713.7248.1936.1115.1315.2227.9612.9426.94
NDC-Scene [45]xrgb44.1712.0293.5113.1113.7715.8349.5739.8717.1724.5731.0014.9629.03
ISO(Ours)xrgb47.1114.2193.4715.8915.1418.3550.0140.8218.2525.9034.0817.6731.25
", + "bbox": [ + 215, + 689, + 792, + 809 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "H. Yu et al.", + "bbox": [ + 271, + 114, + 351, + 126 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/04e2c5714d704f272ab8c860434ace6dba790a5c5dbad8506865878756310863.jpg", + "image_caption": [ + "Image" + ], + "image_footnote": [], + "bbox": [ + 243, + 174, + 370, + 250 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/b56c565fcb1a40a72e382e16616d20db61fc9f40eeac725137d84d8009862ef3.jpg", + "image_caption": [ + "MonoScene" + ], + "image_footnote": [], + "bbox": [ + 377, + 176, + 491, + 247 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/d7fc5ba70886aa3bf20a8c1e7c897294e9dc3e9e11d5f7a942a351a5a6d7a1e0.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 506, + 175, + 622, + 247 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/cd182c385abf9d22677067034b20841ffaa1c7229ac8fee7df544e42ff29b5ee.jpg", + "image_caption": [ + "Ground Truth" + ], + "image_footnote": [], + "bbox": [ + 633, + 176, + 750, + 247 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/fdc44390d075e3fc6576247a1bd8f6e1d0312141487072c44920d171df5c8cc1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 243, + 270, + 370, + 345 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/1ebc98688a3d3fd3a3173e2685869d24e0e043b6de44b4a6d2f0e53ac916e198.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 379, + 272, + 490, + 345 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/3ac2892e3db2e908440b580e701c4dc413377e7e64b972d05bf3372b2341814f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 273, + 622, + 343 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/d1400f44a4c56cceb48fbcb19e51e9f955459cff2e343a5887fad4d17eb142a0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 630, + 273, + 751, + 342 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/5cc79ad1a3418afb90628108bbd37d04b82c455cad379098b54a1c683844ec18.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 243, + 366, + 370, + 441 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/15e5ef3661021822b1013b521dc01a98e033212ebacb07213fc7d9182cb23f7c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 375, + 371, + 493, + 435 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/6db8146507af64f3ad0e265ce1b04282f92a486f53ede1b967eb929a0a72a9d0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 503, + 371, + 625, + 436 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/94828630daa0a3789f33db9c02a70978ce6f2b4eab3a1c1da1a90f57e8435cee.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 632, + 375, + 751, + 435 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/10396fe5d4ea378ddc1107ded1ed79d07052a88a88841e2e2835d29180f65811.jpg", + "image_caption": [ + "Fig. 5: Qualitative Analysis on the Occ-ScanNet Dataset. The input image is displayed on the left, while the predicted scene is shown in the middle two columns, and ground truth on the right column." + ], + "image_footnote": [], + "bbox": [ + 243, + 460, + 370, + 536 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/ac93ecc616302a419ebe3b5a45700ca5a7f49fc81925b8ef6119eae7527bce78.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 377, + 467, + 496, + 532 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/f0898202e01e8619bc49b2aaa306d734b49a00b17d953826e5404469616c0a16.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 460, + 625, + 539 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/30279eba2499b123db6f91a3fa0b0c4272933fffbf5a95082748f4553709b126.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 632, + 460, + 751, + 537 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "5.3 Ablation Study", + "text_level": 1, + "bbox": [ + 214, + 643, + 390, + 659 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We conduct ablation studies on the NYUv2 [29] and Occ-ScanNet-mini datasets to validate the effectiveness of our model design.", + "bbox": [ + 212, + 667, + 787, + 698 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Depth fusion. Inspired by BEVDepth [20], we initially adopt the integration of BEV depth information, denoted as $^{+}$ bev-depth'. However, we observed that while BEV depth works well for 3D detection in outdoor scenes, directly applying it to indoor scenes is less effective. We speculate that this discrepancy arises because height information is more crucial in indoor scenes, where indoor heights often correspond to different 3D structures and semantics. Therefore, we adopt voxel depth fusion, denoted as $^{+}$ voxel-depth'. Specifically, we project each 3D voxel centroid to a frustum grid to sample the depth distribution probability", + "bbox": [ + 212, + 718, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Monocular Occupancy Prediction for Scalable Indoor Scenes", + "bbox": [ + 328, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/e1e790fceece663066ce4a35a679e75bee7a43e5396d4b9cfad477e8820e4e36.jpg", + "image_caption": [ + "Fig. 6: Qualitative Analysis on the NYUv2 Dataset: The input is displayed on the left, with the camera viewing frustum shown in every image. ISO adeptly captures the scene layout and delineates various semantic instances more accurately. Notably, it excels in reconstructing the corners of walls (row 2) and accurately representing small objects like pictures on the wall (row 3)." + ], + "image_footnote": [], + "bbox": [ + 238, + 148, + 764, + 378 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "of the voxel. We then weight the 2D feature at that pixel coordinate as the voxel feature. The results in Tab. 3 further validate that the voxel depth fusion approach is more effective.", + "bbox": [ + 212, + 496, + 784, + 541 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "The BEV method usually performs better in detecting ceilings and floors but performs poorly in detecting furniture such as windows, walls, televisions, beds, sofas, and tables. Ceilings and floors usually have relatively flat and regular geometric shapes, which makes their projections clearer and more consistent from a BEV perspective. In contrast, objects such as windows, walls, and televisions typically have more complex geometric shapes. In the BEV perspective, the projections of these objects may become blurry or overlapping, making it difficult to accurately distinguish and detect.", + "bbox": [ + 212, + 542, + 787, + 662 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/a0a0d6e1b1e3bccd7e8972fb9e8506736b3a0886d6898eb0af319b27a83b2b27.jpg", + "table_caption": [ + "Table 3: Ablation study of depth fusion on the NYUv2 and Occ-ScanNet-mini dataset" + ], + "table_footnote": [], + "table_body": "
MethodInputNYUv2Occ-ScanNet-mini
IoUmIoUIoUmIoU
baselinexrgb42.2727.1350.9438.95
+ bev-depthxrgb42.6727.1451.5838.48
+ voxel-depthxrgb47.1131.2551.0339.08
", + "bbox": [ + 307, + 719, + 697, + 808 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "H. Yu et al.", + "bbox": [ + 271, + 114, + 351, + 126 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/9bbba87470286f0793b8ea3a417066382d7328d93c2cb7ab62af7c0492c2a150.jpg", + "table_caption": [ + "Table 4: Ablation of depth information on the NYUv2 dataset" + ], + "table_footnote": [], + "table_body": "
depth-methodlearnedmulti-scaleIoUmIoU
GT✓53.9834.47
ZoeDepth✓45.1829.15
ZoeDepth✓✓45.2429.40
DepthAnything44.7729.28
DepthAnything✓45.4829.57
DepthAnything✓✓46.9431.02
DepthAnything✓47.1131.25
", + "bbox": [ + 339, + 170, + 660, + 282 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Different depth pre-trained models. We conducted ablation experiments on the depth method using the Occ-ScanNet-mini dataset. Initially, we observed the upper bound of model performance when ground truth (GT) depth information was provided. Subsequently, we initialized our depth module using ZoeDepth [3] and Depth-Anything [44]. Experimental results indicate that utilizing Depth-Anything [44] yields better performance. Moreover, our approach allows for further enhancement by fine-tuning a pre-trained depth module, resulting in improved performance.", + "bbox": [ + 215, + 315, + 785, + 436 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Multi-scale depth. In Tab. 4, we investigate the impact of using multi-scale depth distribution on the NYUv2 [29] dataset. Specifically, the multi-scale depth distribution is used to weight the multi-scale 3D voxel feature. Meanwhile, we tried a method that only uses the single-scale depth distribution map to generate the 3D voxel depth matrix. The used depth distribution map has a scale of $\\frac{1}{8}$ original image scale. Only this single-scale 3D voxel matrix is used to weigh the summed-up 3D voxel feature out of the FLoSP module. Without multi-scale depth, all 3D features are simply added up and a single depth distribution is used for weighting. This method ignores the different importance that features at different scales may have, as all features at all scales are treated equally and all depth information is weighted by a unified depth distribution. With multiscale depth, the 3D features of each scale are first weighted with the depth distribution of their corresponding scale, and then these weighted features are added up. This method takes into account the differences in the importance of features at different scales, as each scale's feature is weighted based on its corresponding depth distribution.", + "bbox": [ + 215, + 474, + 785, + 715 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "In real-world scenarios, various objects and structures (such as ceilings, walls, furniture, etc.) often exhibit different scales. Hence, accounting for scale variations is crucial for accurately modeling these objects. Multi-scale depth can effectively capture these scale changes by assigning different weights to features at each scale. Depth information is indispensable for 3D scene completion. Leveraging the depth distribution information at each scale allows us to weight the corresponding 3D features, enabling the model to better understand and represent the objects in the scene and their relationships.", + "bbox": [ + 215, + 719, + 785, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Monocular Occupancy Prediction for Scalable Indoor Scenes", + "bbox": [ + 328, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5.4 Discussion", + "text_level": 1, + "bbox": [ + 215, + 145, + 349, + 160 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Data scaling up. In Tab. 5, we delve into the impact of data volume on our method's performance by scaling up the number of scene samples. We trained the model on Occ-ScanNet using $10\\%$ and $100\\%$ scene samples respectively, and tested the model performance. The results indicate that larger dataset exhibits more significant gains.", + "bbox": [ + 212, + 171, + 787, + 247 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/9081432db3b01e93e94c944b68deffaf5eb915ce6e250fce8593cf994bb1c78f.jpg", + "table_caption": [ + "Table 5: Scaling up comparison on the Occ-ScanNet" + ], + "table_footnote": [], + "table_body": "
DataDepth BranchOurs
IoUmIoU
10%✓21.799.57
100%✓42.1628.71
", + "bbox": [ + 377, + 301, + 619, + 375 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Limitations and future works. Although our model can estimate 3D occupancy effectively with the assistance of depth information, semantic learning still faces challenges due to class imbalance issues. Furthermore, our proposed Occ-ScanNet only considers 11 common semantic classes, which may not fully capture the diversity of categories present in real-world scenarios. Despite the significant increase in data volume compared to previous datasets, Occ-ScanNet remains limited in the number of scenes it covers. In future work, we will focus more on semantic exploration and conduct tests in more generalized scenarios.", + "bbox": [ + 212, + 424, + 787, + 546 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 569, + 359, + 585 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we introduce ISO, a novel method for monocular 3D occupancy prediction. ISO utilizes a D-FLoSP (Dual Feature Line of Sight Projection) and a multi-scale feature fusion strategy to address variations in scene and object sizes, specifically tailored for indoor scenes. Additionally, we present a new benchmark, Occ-ScanNet, aimed at fostering scalable studies of indoor scenes. We hope that this dataset will attract more attention to research on indoor occupancy prediction. While achieving promising results in 3D structure prediction, accurately identifying semantics remains a significant challenge in the future.", + "bbox": [ + 212, + 601, + 787, + 724 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 215, + 746, + 393, + 763 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "This work was supported in part by the National Key R&D Program of China (No. 2022ZD0160102), the National Natural Science Foundation of China (No. U21B2042, No. 62320106010), and in part by the 2035 Innovation Program of CAS, and the InnoHK program.", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "H. Yu et al.", + "bbox": [ + 271, + 114, + 351, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Arshad, M.S., Beksi, W.J.: List: Learning implicitly from spatial transformers for single-view 3d reconstruction. In: ICCV (2023)", + "2. Azuma, R.T.: A survey of augmented reality. Presence: teleoperators & virtual environments (1997)", + "3. Bhat, S.F., Birkl, R., Wofk, D., Wonka, P., Müller, M.: Zoedepth: Zero-shot transfer by combining relative and metric depth. arXiv preprint arXiv:2302.12288 (2023)", + "4. Birkl, R., Wofk, D., Müller, M.: Midas v3. 1-a model zoo for robust monocular relative depth estimation. arXiv preprint arXiv:2307.14460 (2023)", + "5. Cao, A.Q., de Charette, R.: Monoscene: Monocular 3d semantic scene completion. In: CVPR (2022)", + "6. Chen, X., Lin, K.Y., Qian, C., Zeng, G., Li, H.: 3d sketch-aware semantic scene completion via semi-supervised structure prior. In: CVPR (2020)", + "7. Dahnert, M., Hou, J., Nießner, M., Dai, A.: Panoptic 3d scene reconstruction from a single rgb image. NeurIPS (2021)", + "8. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: Scannet: Richly-annotated 3d reconstructions of indoor scenes. In: CVPR (2017)", + "9. Denninger, M., Triebel, R.: 3d scene reconstruction from a single viewpoint. In: ECCV (2020)", + "10. DeSouza, G.N., Kak, A.C.: Vision for mobile robot navigation: A survey. TPAMI (2002)", + "1. Fan, H., Su, H., Guibas, L.J.: A point set generation network for 3d object reconstruction from a single image. In: CVPR (2017)", + "2. Fieraru, M., Zanfir, M., Oneata, E., Popa, A.I., Olaru, V., Sminchisescu, C.: Reconstructing three-dimensional models of interacting humans. arXiv preprint arXiv:2308.01854 (2023)", + "3. Goel, S., Pavlakos, G., Rajasegaran, J., Kanazawa, A., Malik, J.: Humans in 4d: Reconstructing and tracking humans with transformers. arXiv preprint arXiv:2305.20091 (2023)", + "4. Guan, H., Song, C., Zhang, Z.: Gramo: geometric resampling augmentation for monocular 3d object detection. Frontiers of Computer Science (2024)", + "5. Huang, J., Huang, G., Zhu, Z., Ye, Y., Du, D.: Bevdet: High-performance multicamera 3d object detection in bird-eye-view. arXiv preprint arXiv:2112.11790 (2021)", + "6. Huang, Y., Zheng, W., Zhang, B., Zhou, J., Lu, J.: Selfocc: Self-supervised vision-based 3d occupancy prediction. In: CVPR (2024)", + "7. Huang, Y., Zheng, W., Zhang, Y., Zhou, J., Lu, J.: Tri-perspective view for vision-based 3d semantic occupancy prediction. In: CVPR (2023)", + "8. Li, J., Han, K., Wang, P., Liu, Y., Yuan, X.: Anisotropic convolutional networks for 3d semantic scene completion. In: CVPR (2020)", + "9. Li, Y., Yu, Z., Choy, C., Xiao, C., Alvarez, J.M., Fidler, S., Feng, C., Anandkumar, A.: Voxformer: Sparse voxel transformer for camera-based 3d semantic scene completion. In: CVPR (2023)", + "20. Li, Y., Ge, Z., Yu, G., Yang, J., Wang, Z., Shi, Y., Sun, J., Li, Z.: Bevdepth: Acquisition of reliable depth for multi-view 3d object detection. In: AAAI (2023)", + "21. Li, Z., Wang, W., Li, H., Xie, E., Sima, C., Lu, T., Qiao, Y., Dai, J.: Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In: ECCV (2022)" + ], + "bbox": [ + 225, + 178, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Monocular Occupancy Prediction for Scalable Indoor Scenes", + "bbox": [ + 328, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "22. Li, Z., Yu, Z., Austin, D., Fang, M., Lan, S., Kautz, J., Alvarez, J.M.: Fb-occ: 3d occupancy prediction based on forward-backward view transformation. arXiv preprint arXiv:2307.01492 (2023)", + "23. Liu, S., Hu, Y., Zeng, Y., Tang, Q., Jin, B., Han, Y., Li, X.: See and think: Disentangling semantic scene completion. NeurIPS (2018)", + "24. Murez, Z., Van As, T., Bartolozzi, J., Sinha, A., Badrinarayanan, V., Rabinovich, A.: Atlas: End-to-end 3d scene reconstruction from posed images. In: ECCV (2020)", + "25. Pan, M., Liu, J., Zhang, R., Huang, P., Li, X., Liu, L., Zhang, S.: Renderocc: Visioncentric 3d occupancy prediction with 2d rendering supervision. arXiv preprint arXiv:2309.09502 (2023)", + "26. Park, J.J., Florence, P., Straub, J., Newcombe, R., Lovegrove, S.: Deepsdf: Learning continuous signed distance functions for shape representation. In: CVPR (2019)", + "27. Ranftl, R., Lasinger, K., Hafner, D., Schindler, K., Koltun, V.: Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. TPAMI (2020)", + "28. Roldao, L., de Charette, R., Verroust-Blondet, A.: Lmscnet: Lightweight multiscale 3d semantic completion. In: 3DV (2020)", + "29. Silberman, N., Hoiem, D., Kohli, P., Fergus, R.: Indoor segmentation and support inference from rgbd images. In: ECCV (2012)", + "30. Song, S., Yu, F., Zeng, A., Chang, A.X., Savva, M., Funkhouser, T.: Semantic scene completion from a single depth image. In: CVPR (2017)", + "31. Tan, M., Le, Q.: Efficientnet: Rethinking model scaling for convolutional neural networks. In: ICML (2019)", + "32. Tang, Y., Dorn, S., Savani, C.: Center3d: Center-based monocular 3d object detection with joint depth understanding. In: DAGM GCPR (2020)", + "33. Tian, X., Jiang, T., Yun, L., Mao, Y., Yang, H., Wang, Y., Wang, Y., Zhao, H.: Occ3d: A large-scale 3d occupancy prediction benchmark for autonomous driving. NeurIPS (2023)", + "34. Tong, W., Sima, C., Wang, T., Chen, L., Wu, S., Deng, H., Gu, Y., Lu, L., Luo, P., Lin, D., et al.: Scene as occupancy. In: ICCV (2023)", + "35. Wang, S., Liu, Y., Wang, T., Li, Y., Zhang, X.: Exploring object-centric temporal modeling for efficient multi-view 3d object detection. In: ICCV (2023)", + "36. Wang, X., Zhu, Z., Xu, W., Zhang, Y., Wei, Y., Chi, X., Ye, Y., Du, D., Lu, J., Wang, X.: Openoccupancy: A large scale benchmark for surrounding semantic occupancy perception. In: ICCV (2023)", + "37. Wang, Y., Guizilini, V.C., Zhang, T., Wang, Y., Zhao, H., Solomon, J.: Detr3d: 3d object detection from multi-view images via 3d-to-2d queries. In: CoRL (2022)", + "38. Wang, Y., Chen, Y., Liao, X., Fan, L., Zhang, Z.: Panoocc: Unified occupancy representation for camera-based 3d panoptic segmentation. In: CVPR (2024)", + "39. Wang, Y., Chen, Y., Zhang, Z.: Frustumformer: Adaptive instance-aware resampling for multi-view 3d detection. In: CVPR (2023)", + "40. Wei, Y., Zhao, L., Zheng, W., Zhu, Z., Zhou, J., Lu, J.: Surroundocc: Multi-camera 3d occupancy prediction for autonomous driving. In: ICCV (2023)", + "41. Wu, Q., Wang, K., Li, K., Zheng, J., Cai, J.: Objectsdf++: Improved object-compositional neural implicit surfaces. In: ICCV (2023)", + "42. Wu, S.C., Tateno, K., Navab, N., Tombari, F.: Scfusion: Real-time incremental scene reconstruction with semantic completion. In: 3DV (2020)", + "43. Yang, C., Chen, Y., Tian, H., Tao, C., Zhu, X., Zhang, Z., Huang, G., Li, H., Qiao, Y., Lu, L., et al.: Bevformer v2: Adapting modern image backbones to bird's-eye-view recognition via perspective supervision. In: CVPR (2023)" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "H. Yu et al.", + "bbox": [ + 271, + 114, + 351, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "44. Yang, L., Kang, B., Huang, Z., Xu, X., Feng, J., Zhao, H.: Depth anything: Unleashing the power of large-scale unlabeled data. In: CVPR (2024)", + "45. Yao, J., Li, C., Sun, K., Cai, Y., Li, H., Ouyang, W., Li, H.: Ndc-scene: Boost monocular 3d semantic scene completion in normalized device coordinates space. In: ICCV (2023)", + "46. Yu, Z., Shu, C., Deng, J., Lu, K., Liu, Z., Yu, J., Yang, D., Li, H., Chen, Y.: Flashocc: Fast and memory-efficient occupancy prediction via channel-to-height plugin. arXiv preprint arXiv:2311.12058 (2023)", + "47. Yurtsever, E., Lambert, J., Carballo, A., Takeda, K.: A survey of autonomous driving: Common practices and emerging technologies. IEEE access (2020)", + "48. Zhang, J., Zhao, H., Yao, A., Chen, Y., Zhang, L., Liao, H.: Efficient semantic scene completion network with spatial group convolution. In: ECCV (2018)", + "49. Zhang, X., Bi, S., Sunkavalli, K., Su, H., Xu, Z.: Nerfusion: Fusing radiance fields for large-scale scene reconstruction. In: CVPR (2022)", + "50. Zheng, Z., Yu, T., Wei, Y., Dai, Q., Liu, Y.: Deephuman: 3d human reconstruction from a single image. In: ICCV (2019)", + "51. Zhong, M., Zeng, G.: Semantic point completion network for 3d semantic scene completion. In: ECAI (2020)" + ], + "bbox": [ + 215, + 146, + 784, + 396 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Monocular Occupancy Prediction for Scalable Indoor Scenes", + "bbox": [ + 328, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/Monocular Occupancy Prediction for Scalable Indoor Scenes/325d434b-3564-4d20-a82e-4ec4e345eb58_model.json b/2024/Monocular Occupancy Prediction for Scalable Indoor Scenes/325d434b-3564-4d20-a82e-4ec4e345eb58_model.json new file mode 100644 index 0000000000000000000000000000000000000000..69d0207d39fb2e33fac018e20fbf6ce9dc50745c --- /dev/null +++ b/2024/Monocular Occupancy Prediction for Scalable Indoor Scenes/325d434b-3564-4d20-a82e-4ec4e345eb58_model.json @@ -0,0 +1,2456 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.24, + 0.141, + 0.764, + 0.184 + ], + "angle": 0, + "content": "Monocular Occupancy Prediction for Scalable Indoor Scenes" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.212, + 0.786, + 0.229 + ], + "angle": 0, + "content": "Hongxiao \\(\\mathrm{Yu}^{1,2}\\) Yuqi Wang\\(^{1,2}\\), Yuntao Chen\\(^{3}\\), and Zhaoxiang Zhang\\(^{1,2,3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.239, + 0.784, + 0.253 + ], + "angle": 0, + "content": "1 School of Artificial Intelligence, University of Chinese Academy of Sciences (UCAS)" + }, + { + "type": "text", + "bbox": [ + 0.23, + 0.254, + 0.771, + 0.269 + ], + "angle": 0, + "content": "\\(^{2}\\) NLPR, MAIS, Institute of Automation, Chinese Academy of Sciences (CASIA)" + }, + { + "type": "text", + "bbox": [ + 0.282, + 0.269, + 0.721, + 0.31 + ], + "angle": 0, + "content": "3 Centre for Artificial Intelligence and Robotics (HKISI_CAS) {yuhongxiao2023, wangyuqi2020, zhaoxiang.zhang}@ia.ac.cn, chenyuntao08@gmail.com" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.343, + 0.74, + 0.565 + ], + "angle": 0, + "content": "Abstract. Camera-based 3D occupancy prediction has recently garnered increasing attention in outdoor driving scenes. However, research in indoor scenes remains relatively unexplored. The core differences in indoor scenes lie in the complexity of scene scale and the variance in object size. In this paper, we propose a novel method, named ISO, for predicting indoor scene occupancy using monocular images. ISO harnesses the advantages of a pretrained depth model to achieve accurate depth predictions. Furthermore, we introduce the Dual Feature Line of Sight Projection (D-FLoSP) module within ISO, which enhances the learning of 3D voxel features. To foster further research in this domain, we introduce Occ-ScanNet, a large-scale occupancy benchmark for indoor scenes. With a dataset size 40 times larger than the NYUv2 dataset, it facilitates future scalable research in indoor scene analysis. Experimental results on both NYUv2 and Occ-ScanNet demonstrate that our method achieves state-of-the-art performance. The dataset and code are made publicly at https://github.com/hongxiaoy/ISO.git." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.577, + 0.741, + 0.606 + ], + "angle": 0, + "content": "Keywords: 3D occupancy prediction \\(\\cdot\\) Camera-based scene understanding \\(\\cdot\\) Semantic scene completion" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.629, + 0.376, + 0.645 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.81 + ], + "angle": 0, + "content": "3D scene understanding is a crucial task in computer vision, becoming increasingly important for applications such as robotic navigation [10], augmented reality [2], and autonomous driving [47]. While humans possess a natural ability to comprehend 3D environments through vision, enabling this ability in computers poses a significant challenge due to limitations like restricted fields of view, sparse sensing capabilities, and measurement noise. However, with the rapid development of deep learning and the availability of large-scale 3D driving datasets, camera-based 3D object detection [14, 15, 20, 21, 35, 37, 39, 43] and occupancy prediction [5, 22, 33, 38, 40] have developed rapidly, achieving significant improvements in performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Occupancy prediction has gained popularity recently due to its effectiveness in representing both background and foreground objects within a scene using a" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.128 + ], + "angle": 0, + "content": "H. Yu et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.403 + ], + "angle": 0, + "content": "unified representation. Although significant progress has been made in outdoor driving scenarios, research in indoor scenarios remains limited. Indoor scenes differ from outdoor driving scenes in two key aspects: (1) scene-scale complexity: Indoor rooms often exhibit a more diverse range of sizes compared to outdoor environments, where driving scenarios typically focus on a fixed 3D space for perception. This diversity, ranging from spacious living rooms to narrow kitchens, poses higher precision requirements for depth prediction. (2) object complexity: Indoor scenes feature a higher density and greater variety of objects. Unlike outdoor objects such as vehicles and pedestrians, which typically have consistent sizes within their respective categories and are relatively well-separated, indoor furniture and other objects often exhibit significant variations in scale and are closely positioned to each other. This increased complexity necessitates more sophisticated 3D perception techniques to accurately capture and understand the intricate geometry and relationships among objects within indoor spaces. Furthermore, existing works [5,45] focusing on indoor scenes primarily utilize the NYUv2 dataset [29], which lacks a more scalable and generalizable benchmark for comprehensive evaluation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.406, + 0.788, + 0.603 + ], + "angle": 0, + "content": "To address the above issues, we firstly introduce a more scalable benchmark, Occ-ScanNet, for 3D occupancy prediction in indoor scenes. This benchmark builds upon the large-scale ScanNet [8] dataset, offering a 40 times more samples compared to the NYUv2 [29] dataset, thus greatly expanding the scope and diversity of indoor scene study. To effectively address the unique complexities of indoor scenes, we propose a novel method named ISO (Indoor Scene Occupancy). ISO leverages a powerful pretrained depth model and integrates a D-FLoSP (Dual Feature Line of Sight Projection) module. This module enables precise depth estimation and facilitates learning voxel features for accurate predictions. To accommodate the varying sizes of indoor objects, we introduce a multi-scale feature fusion module. This module enhances the learning of object features across different scales. Leveraging a powerful depth model, our method is adept at handling diverse indoor scenes, offering a robust and versatile solution." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.605, + 0.633, + 0.619 + ], + "angle": 0, + "content": "Our main contributions can be summarized as follows." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.639, + 0.785, + 0.699 + ], + "angle": 0, + "content": "- We introduce a new benchmark called Occ-ScanNet for monocular 3D occupancy prediction in indoor scenes. With a dataset size 40 times larger than the NYUv2 dataset, it significantly enhances the potential for scalable research in indoor scene analysis." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.702, + 0.785, + 0.776 + ], + "angle": 0, + "content": "- We propose a novel approach called \\( ISO \\), which primarily comprises the DFLoSP (Dual Feature Line of Sight Projection) module and a multi-scale feature fusion module. Together, these components effectively address the challenges posed by variations in scene and object sizes, enabling more accurate and robust 3D occupancy prediction." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.78, + 0.785, + 0.839 + ], + "angle": 0, + "content": "- Experiments on the Occ-ScanNet and NYUv2 datasets demonstrate that our approach achieves state-of-the-art performance in monocular 3D occupancy prediction. Furthermore, our method exhibits significant scalability potential." + }, + { + "type": "list", + "bbox": [ + 0.225, + 0.639, + 0.785, + 0.839 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.33, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Monocular Occupancy Prediction for Scalable Indoor Scenes" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.145, + 0.393, + 0.161 + ], + "angle": 0, + "content": "2 Related works" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.177, + 0.617, + 0.192 + ], + "angle": 0, + "content": "2.1 Monocular 3D Semantic Scene Completion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.202, + 0.788, + 0.399 + ], + "angle": 0, + "content": "Monocular 3D Semantic Scene Completion (SSC) aims to infer the complete 3D structure and corresponding semantics from a single image. This monocular setting was first introduced by Monoscene [5], advancing upon prior SSC methods [18,23,30,48,51] by relying solely on vision, without additional 3D inputs. Monoscene [5] introduces the Features Line of Sight Projection (FLoSP), inspired by optics, to obtain 3D features through ray projection. However, the shared 2D features lifted to 3D rays via FLoSP results in depth ambiguity. Consequently, subsequent efforts have placed greater emphasis on leveraging depth information. VoxFormer [19] proposed a novel query proposal network based on 2D convolutions, generating sparse queries from image depth, which showed impressive performance in driving scenes. Meanwhile, NDC-Scene [45] further devised a Depth-Adaptive Dual Decoder to concurrently upsample and merge the 2D and 3D feature maps, thereby enhancing overall performance." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.421, + 0.566, + 0.436 + ], + "angle": 0, + "content": "2.2 Multiview 3D Occupancy Prediction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.446, + 0.788, + 0.643 + ], + "angle": 0, + "content": "Multiview 3D Occupancy Prediction predicts the semantic occupancy of the surrounding 3D scene given multiview images, has recently garnered significant attention in the field of autonomous driving. TPVFormer [17] pioneers exploration in driving scenes, employing sparse LiDAR labels for supervision. It introduces a tri-perspective view (TPV) representation that accompanies BEV with two additional perpendicular planes. However, due to the sparse supervision provided by LiDAR, subsequent works [33, 34, 36, 40] have focused on providing more dense occupancy benchmarks. FB-OCC [22] integrates the lifting of 2D to 3D and querying of 3D to 2D, achieving a more efficient 3D feature transformation. Meanwhile, PanoOcc [38] unifies detection and semantic occupancy tasks, enabling comprehensive panoramic scene understanding. Recent methods also explore improving model efficiency [46] and utilizing weaker forms of supervision [16, 25]." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.665, + 0.519, + 0.68 + ], + "angle": 0, + "content": "2.3 3D Reconstruction from Image" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.69, + 0.788, + 0.841 + ], + "angle": 0, + "content": "3D reconstruction is a technique that involves recovering a 3D representation of objects [1, 11, 26], scenes [9, 41, 49], or even human bodies [12, 13, 50] from camera images. The task of 3D reconstruction can be classified into two categories: monocular reconstruction and multi-view reconstruction, depending on the number of images utilized. In the context of indoor scenes, 3D scene reconstruction aims to determine the surface geometry of the entire scene, often without incorporating semantic information. Atlas [24] proposes an end-to-end 3D reconstruction framework using TSDF regression from RGB images, bypassing traditional depth map estimation for efficient semantic segmentation. Panoptic-Reconstruction [7] unifies geometric reconstruction, 3D semantic, and" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.353, + 0.127 + ], + "angle": 0, + "content": "H. Yu et al." + }, + { + "type": "image", + "bbox": [ + 0.216, + 0.151, + 0.787, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.3, + 0.788, + 0.399 + ], + "angle": 0, + "content": "Fig. 1: The core design of ISO centers around the transformation of features from 2D to 3D spaces, encompassing the Depth Branch and the D-FLoSP module. A depth branch is initially integrated, it leverages a pre-trained depth model to estimate a pixel-wise depth map which is processed by the DepthNet to generate the final depth distribution. An element-wise multiplication between the voxel depth and features followed by summation are subsequently performed to derive the initial 3D voxel feature. The 3D feature is further processed to predict the 3D scene occupancy." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.43, + 0.788, + 0.522 + ], + "angle": 0, + "content": "instance segmentation tasks. It predicts complete geometric reconstruction, semantic, and instance segmentations from a single RGB image's camera frustum. SCFusion [42], a real-time scene reconstruction framework, integrates continuous depth data using neural architecture for occupancy maps. It efficiently combines semantic completion with voxel states for simultaneous scene reconstruction and semantic understanding in real-time 3D environments." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.549, + 0.331, + 0.566 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.586, + 0.788, + 0.663 + ], + "angle": 0, + "content": "Problem definition. We focus on the problem of monocular 3D Occupancy Prediction. Specifically, this task takes a single RGB image \\(\\mathbf{I}^{RGB}\\) as input and output a voxel-wise occupancy along with semantic categories \\(\\mathbf{Y}^{X\\times Y\\times Z\\times C}\\). \\(X\\), \\(Y\\), and \\(Z\\) represent the dimensions of the predicted 3D scene, while \\(C\\) denotes the total number of semantic categories." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Model overview. In this section, we introduce the overall architecture of ISO, as shown in Fig. 1. Following Monoscene [5], we utilize 2D Unet and 3D Unet architectures to handle 2D and 3D features. Our core design focuses on transforming features from 2D to 3D, incorporating the Depth Branch and D-FLoSP module. Specifically, we incorporate a depth branch to estimate the pixel-wise depth map using a pre-trained depth model. Processed by the DepthNet, the model outputs a refined depth distribution. Then, voxel depth and features are multiplied element-wise and summed to obtain the initial 3D voxel feature volume \\(\\mathbf{X}^{3\\mathbf{d}} \\in \\mathbb{R}^{X \\times Y \\times Z \\times C}\\). Finally, after processing the 3D voxel features, the model outputs the 3D scene occupancy." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.329, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Monocular Occupancy Prediction for Scalable Indoor Scenes" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.381, + 0.162 + ], + "angle": 0, + "content": "3.1 Depth Branch" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.169, + 0.784, + 0.199 + ], + "angle": 0, + "content": "In this section, we introduce how to effectively estimate depth information from a single image." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.218, + 0.784, + 0.277 + ], + "angle": 0, + "content": "Coarse depth estimation. Learning depth from scratch can be quite challenging. However, thanks to the recent rapid advancements in depth estimation [3,4,27,44], we can leverage pre-trained depth models to initially estimate a coarse depth map." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.279, + 0.785, + 0.325 + ], + "angle": 0, + "content": "Compared to past models that could only estimate relative depth, we opted for Depth-Anything [44] as the pre-trained depth model, because it excels in predicting metric depth, as the Eq. (1) shows:" + }, + { + "type": "equation", + "bbox": [ + 0.374, + 0.329, + 0.786, + 0.349 + ], + "angle": 0, + "content": "\\[\n\\mathbf {D} ^ {\\mathrm {m e t r i c}} = \\mathbf {N} _ {\\mathrm {d e p t h}} (\\mathbf {I} ^ {\\mathrm {r g b}}) \\in \\mathbb {R} ^ {1 \\times H \\times W}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.356, + 0.784, + 0.387 + ], + "angle": 0, + "content": "where \\(\\mathbf{N}_{\\mathrm{depth}}\\) denotes the pre-trained depth model [3, 44], and \\(\\mathbf{I}^{\\mathrm{rgb}}\\) represents the input image. \\(H\\) and \\(W\\) represents the height and width of the input image." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.405, + 0.79, + 0.497 + ], + "angle": 0, + "content": "Depth refinement. The coarse depth estimation is subsequently refined through model learning. The depth from pre-trained depth model is not precise enough for a higher mIoU, so we design a fine-tuning strategy. Specifically, the predicted metric depth \\(\\mathbf{D}^{\\mathrm{metric}}\\) is then concatenated with the image feature \\(\\mathbf{X}\\) that has the same spatial scale. The augmented feature is continue processed by a following DepthNet to get a refined depth distribution \\(\\mathbf{D}_{s=1}^{\\mathrm{dist}}\\), as Eq. (2) shows:" + }, + { + "type": "equation", + "bbox": [ + 0.269, + 0.515, + 0.785, + 0.548 + ], + "angle": 0, + "content": "\\[\n\\mathbf {D} _ {\\mathrm {s} = 1} ^ {\\text {d i s t}} = \\mathbf {F} _ {\\text {d e p t h}} \\left(\\operatorname {C o n c a t} \\left(\\mathbf {D} ^ {\\text {m e t r i c}}, \\mathbf {X} _ {\\mathrm {s} = 1} ^ {2 \\mathrm {d}}\\right)\\right), \\quad \\mathbf {D} _ {\\mathrm {s} = 1} ^ {\\text {d i s t}} \\in \\mathbb {R} ^ {N _ {\\text {b i n s}} \\times H \\times W}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.555, + 0.784, + 0.601 + ], + "angle": 0, + "content": "where \\(\\mathbf{F}_{\\mathrm{depth}}\\) is the DepthNet, \\(\\mathbf{X}_{s=1}^{2d}\\) is the 2D image feature, and \\(N_{\\mathrm{bins}}\\) is the number of discrete depth bins. \\(s\\) denotes the level of feature scale, where \\(s = 1\\) implies that the feature is at a 1:1 original image ratio." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.601, + 0.784, + 0.632 + ], + "angle": 0, + "content": "We down-samples the scale \\( \\frac{1}{1} \\) distribution to get other smaller-scale 2D image feature maps' depth distribution. We represent the procedure as Eq. (3):" + }, + { + "type": "equation", + "bbox": [ + 0.348, + 0.637, + 0.785, + 0.656 + ], + "angle": 0, + "content": "\\[\n\\mathbf {D} _ {\\mathrm {s} = k} ^ {\\text {d i s t}} = \\operatorname {D o w n S a m p l e} \\left(\\mathbf {D} _ {\\mathrm {s} = 1} ^ {\\text {d i s t}}\\right), k = 2, 4, 8. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.663, + 0.784, + 0.725 + ], + "angle": 0, + "content": "The refined depth is supervised by the ground truth \\(\\mathbf{D}^{\\mathrm{GT}}\\). The ground truth has single depth value at each pixel, so we convert the single value to a one-hot vector of length \\(N_{\\mathrm{bins}}\\). The refined depth prediction can be optimized using the BCE loss function as Eq. (4) shows." + }, + { + "type": "equation", + "bbox": [ + 0.23, + 0.746, + 0.786, + 0.844 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\mathrm {d e p t h}} = - \\frac {1}{N _ {\\mathrm {b i n s}} \\times H \\times W} \\\\ \\sum_ {d = 1} ^ {N _ {\\mathrm {b i n s}}} \\sum_ {h = 1} ^ {H} \\sum_ {w = 1} ^ {W} \\left[ \\mathbf {D} _ {N _ {\\mathrm {b i n s}}, d, h, w} ^ {\\mathrm {G T}} \\cdot \\log \\left(\\mathbf {D} _ {N _ {\\mathrm {b i n s}}, d, h, w} ^ {\\mathrm {d i s t}}\\right) \\right. \\\\ \\left. + \\left(1 - \\mathbf {D} _ {N _ {\\mathrm {b i n s}}, d, h, w} ^ {\\mathrm {G T}}\\right) \\cdot \\log \\left(1 - \\mathbf {D} _ {N _ {\\mathrm {b i n s}}, d, h, w} ^ {\\mathrm {d i s t}}\\right) \\right] \\quad (4) \\\\ \\end{array}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.128 + ], + "angle": 0, + "content": "H. Yu et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.672, + 0.164 + ], + "angle": 0, + "content": "3.2 Dual Feature Line of Sight Projection (D-FLoSP)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.181, + 0.784, + 0.242 + ], + "angle": 0, + "content": "Compared to the FLoSP module in MonoScene [5], which projects 3D voxels to 2D, and 3D feature vectors are the corresponding pixel feature vectors. It only considers the projection of 3D rays and overlooks depth information, our proposed D-FLoSP module can more effectively integrate depth information." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.244, + 0.785, + 0.319 + ], + "angle": 0, + "content": "Specifically, after getting depth distribution in the camera frame, we implement the FLoSP module to depth distribution. Each 3D voxel centroid position \\( x^{c} \\) in the world frame can be projected to the camera frame and the pixel frame using camera pose and intrinsic matrix. We assign the projected depth values to discrete depth bins index by Eqs. (5) and (6), according to [32]," + }, + { + "type": "equation", + "bbox": [ + 0.383, + 0.334, + 0.785, + 0.366 + ], + "angle": 0, + "content": "\\[\n\\delta = \\frac {2 \\left(d _ {\\max } - d _ {\\min }\\right)}{N _ {b i n s} \\left(1 + N _ {b i n s}\\right)}, \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.386, + 0.37, + 0.785, + 0.402 + ], + "angle": 0, + "content": "\\[\nl = - 0. 5 + 0. 5 \\sqrt {1 + \\frac {8 \\left(d - d _ {\\operatorname* {m i n}}\\right)}{\\delta}}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.416, + 0.784, + 0.446 + ], + "angle": 0, + "content": "and the depth distribution probability it lies in serves as the depth probability of that voxel. The process is illustrated by Eq. (7):" + }, + { + "type": "equation", + "bbox": [ + 0.324, + 0.46, + 0.785, + 0.48 + ], + "angle": 0, + "content": "\\[\n\\mathbf {D} _ {\\mathrm {s} = k} ^ {\\mathrm {3 d}} = \\Phi_ {\\rho \\left(x ^ {c}\\right)} ^ {\\mathrm {3 d}} \\left(\\mathbf {D} _ {\\mathrm {s} = k} ^ {\\mathrm {d i s t}}\\right) \\in \\mathbb {R} ^ {X \\times Y \\times Z \\times C}, k = 1, 2, 4, 8, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.495, + 0.785, + 0.585 + ], + "angle": 0, + "content": "where \\(\\Phi_{a}^{3\\mathrm{d}}(b)\\) is the 3D sampling of b at coordinates a, and \\(\\rho (\\cdot)\\) is the perspective projection. Thus, the depth distribution is lifted from 2D to 3D. The bin-based representation with quantization is used for mapping continuous depth value to a discrete form, each bin will be assigned a different depth distribution score to make voxel aware of depth. The original depth precision is a continuous value, and hard to integrate it with the discrete form of voxel." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.587, + 0.785, + 0.649 + ], + "angle": 0, + "content": "Following [5], we also employ the FLoSP module to generate a 3D voxel feature from a 2D feature map. Similar to what we did with the voxel centroids, the projected voxel can sample corresponding \\(1:k\\) scaled feature maps \\(F_{1:k}\\) from the 2D UNet decoder. This process is illustrated by Equation Eq. (8):" + }, + { + "type": "equation", + "bbox": [ + 0.324, + 0.663, + 0.785, + 0.682 + ], + "angle": 0, + "content": "\\[\n\\mathbf {X} _ {\\mathrm {s} = k} ^ {3 \\mathrm {d}} = \\Phi_ {\\rho \\left(x ^ {c}\\right)} ^ {2 \\mathrm {d}} \\left(\\mathbf {X} _ {\\mathrm {s} = k} ^ {2 \\mathrm {d}}\\right) \\in \\mathbb {R} ^ {X \\times Y \\times Z \\times C}, k = 1, 2, 4, 8, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.697, + 0.784, + 0.728 + ], + "angle": 0, + "content": "where \\(\\Phi_{a}^{2\\mathrm{d}}(b)\\) is the 2D sampling of b at coordinates a, and \\(\\rho (\\cdot)\\) is the perspective projection." + }, + { + "type": "text", + "bbox": [ + 0.239, + 0.73, + 0.632, + 0.745 + ], + "angle": 0, + "content": "The final 3D feature map \\(\\mathbf{X}^{3\\mathrm{d}}\\) are summed in Eq. (9):" + }, + { + "type": "equation", + "bbox": [ + 0.393, + 0.761, + 0.785, + 0.793 + ], + "angle": 0, + "content": "\\[\n\\mathbf {X} ^ {\\mathrm {3 d}} = \\sum_ {\\mathrm {s} \\in \\{1, 2, 4, 8 \\}} \\mathbf {X} _ {\\mathrm {s} = k} ^ {\\mathrm {3 d}} \\odot \\mathbf {D} _ {\\mathrm {s} = k} ^ {\\mathrm {3 d}}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.784, + 0.84 + ], + "angle": 0, + "content": "where \\(\\odot\\) is the element-wise multiplication. The output map \\(\\mathbf{X}^{3\\mathrm{d}}\\) is then served as the input of 3D UNet." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.33, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Monocular Occupancy Prediction for Scalable Indoor Scenes" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.272, + 0.163, + 0.431, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.299, + 0.498, + 0.321 + ], + "angle": 0, + "content": "(a) Depth ranges in NYUv2 and our Occ-ScanNet Benchmark" + }, + { + "type": "image", + "bbox": [ + 0.56, + 0.163, + 0.718, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.506, + 0.299, + 0.786, + 0.321 + ], + "angle": 0, + "content": "(b) Dataset attributes of NYUv2 and our OccScanNet Benchmark" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.332, + 0.788, + 0.404 + ], + "angle": 0, + "content": "Fig. 2: Comparison of NYUv2 and Occ-ScanNet Benchmark. In (a), the depth ranges of NYUv2 and Occ-ScanNet are distinguished by dark and light green, respectively, with the horizontal axis indicating the minimum depth and the vertical axis showing the maximum depth of scenes. (b) quantitatively demonstrates that Occ-ScanNet possesses a significantly larger data scale compared to the NYUv2 dataset." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.428, + 0.495, + 0.443 + ], + "angle": 0, + "content": "4 Occ-ScanNet Benchmark" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.456, + 0.788, + 0.547 + ], + "angle": 0, + "content": "Compared to the previous widely used indoor scene benchmark NYUv2 [29], which contains only 795 / 654 for train / test samples, our benchmark boasts 45,755 / 19,764 samples. Our benchmark significantly surpasses NYUv2 [29] in both data quantity and richness of scene depth, as illustrated in Fig. 2. This dataset is available at https://huggingface.co/datasets/hongxiaoy/OccScanNet." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.566, + 0.341, + 0.58 + ], + "angle": 0, + "content": "4.1 Overview" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.588, + 0.787, + 0.664 + ], + "angle": 0, + "content": "Occ-ScanNet benchmark features a train/Validation split of 45,755 / 19,764 samples. As shown in Fig. 3, Occ-ScanNet exhibits rich diversity in scenes and viewpoints. This diverse dataset not only challenges the task of predicting scene occupancy but also fosters future research endeavors towards developing larger-scale and more versatile occupancy models." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.683, + 0.504, + 0.699 + ], + "angle": 0, + "content": "4.2 Occupancy Label Generation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.787, + 0.764 + ], + "angle": 0, + "content": "We adhere to the data formulation used in NYUv2 [29] dataset. The pipeline of occupancy generation process is showed in Fig. 4. For generating 3D voxel labels, we initially followed the generation process in CompleteScanNet and then employed manual inspection." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.787, + 0.841 + ], + "angle": 0, + "content": "More specifically, from the official ScanNet [8] scenes, we extracted various data components, including color images, depth images, camera intrinsic matrices, and camera poses. A total of 100 frames were sampled from each scene and were randomly divided into training and validation sets with a \\(7/3\\) split ratio per scene." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.127 + ], + "angle": 0, + "content": "H. Yu et al." + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.154, + 0.777, + 0.429 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.449, + 0.788, + 0.507 + ], + "angle": 0, + "content": "Fig. 3: Samples Visualization in Occ-ScanNet Benchmark. The original RGB image is shown in column 1,3 and 5, the corresponding scene voxel labels is shown in column 2, 4 and 6. The first two rows are different views from different scenes and the last two rows each is three different views from the same scene." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.536, + 0.788, + 0.584 + ], + "angle": 0, + "content": "Manual inspection. The manual inspection involves verifying camera poses, camera intrinsic, and voxel positions. We filter out erroneous samples to prevent convergence issues in the model training process." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.605, + 0.788, + 0.727 + ], + "angle": 0, + "content": "Voxel label assignment. For each frame, only a specific area in front of the camera is defined for analysis. Therefore, the selection of the voxel origin is paramount, as it determines the subsequent coordinates of other voxel. Subsequently, each voxel is assigned a label based on its nearest voxel in the CompleteScanNet dataset. Additionally, we excluded frames with a ratio of unknown or empty labels exceeding \\(95\\%\\) or frames where the number of semantic label classes was less than two. Consequently, we obtained the generated 3D voxel labels for each frame." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.751, + 0.378, + 0.768 + ], + "angle": 0, + "content": "5 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.784, + 0.312, + 0.8 + ], + "angle": 0, + "content": "5.1 Setup" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Datasets. The NYUv2 [29] dataset provides scenes represented in \\( 240 \\times 144 \\times 240 \\) voxels grids, labeled with 13 classes, including 1 for free space, 1 for unknown," + } + ], + [ + { + "type": "header", + "bbox": [ + 0.33, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Monocular Occupancy Prediction for Scalable Indoor Scenes" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.151, + 0.778, + 0.202 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.222, + 0.788, + 0.333 + ], + "angle": 0, + "content": "Fig. 4: Pipeline of Occ-ScanNet dataset label generation. Color images, depth images, camera intrinsic and poses are extracted from ScanNet scenes. For each scene, 100 frames were sampled and randomly split into training and validation sets with a \\(7/3\\) ratio. Frames with invalid camera poses or exceeding scene boundaries were filtered out. Only the area in front of the camera was analyzed, necessitating careful selection of the voxel origin. Voxel were labeled based on their nearest voxel in the CompleteScanNet dataset. Frames with \\(>95\\%\\) unknown/empty labels or \\(<2\\) semantic classes were excluded, resulting in generated 3D voxel labels for each frame." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.37, + 0.788, + 0.43 + ], + "angle": 0, + "content": "and 11 for specific semantics (ceiling, floor, wall, window, chair, bed, sofa, table, tvs, furniture, objects). The dataset consists of 795 / 654 scenes in the train / test splits. The model is trained and evaluated on down-sampled \\(60 \\times 36 \\times 60\\) voxels." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.434, + 0.788, + 0.51 + ], + "angle": 0, + "content": "Occ-ScanNet dataset provides scenes represented in \\(60 \\times 60 \\times 36\\) voxel grids, labeled with 12 classes including 1 for free space, and 11 for specific semantics (ceiling, floor, wall, window, chair, bed, sofa, table, tvs, furniture, objects). The dataset comprises 45,755 / 19,764 frames in the train / val splits. The model is trained and evaluated on the original scale." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.513, + 0.788, + 0.573 + ], + "angle": 0, + "content": "Occ-ScanNet-mini dataset has the same class number and voxel scene size as Occ-ScanNet, except that this mini dataset consists of 4,639 / 2,007 frames in the train / val splits. The model is also trained and evaluated on the original scale." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.789, + 0.84 + ], + "angle": 0, + "content": "Implementation Details. We employ a pre-trained EfficientNet-B7 [31] as the encoder in our 2D UNet architecture. In the depth branch, we utilize a pre-trained DepthAnything model [44], which remains frozen during training. Additionally, we integrate a depth loss specific to the depth branch, complementing the other losses outlined in Monoscene [5]. Our model is trained on two datasets: NYUv2 [29] and Occ-ScanNet. For NYUv2, ISO undergoes 30 epochs of training using AdamW optimizers. Initially, the learning rate is set to 5e-6 for the DepthNet and 1e-4 for other components, with a learning rate reduction by 0.1 at epoch 20. Training on NYUv2 takes approximately 7 hours using 2 NVIDIA L20 GPUs (2 items per GPU). On the Occ-ScanNet-mini dataset, ISO is trained for 60 epochs under similar learning rate settings, decreasing the rate by 0.1 at epoch 40. This training process takes around 12 hours using 8 L20 GPUs. For Occ-ScanNet, ISO is trained for 10 epochs with the same learning rate schedule as used for Occ-ScanNet-mini, requiring approximately a day using 8 L20 GPUs." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.127 + ], + "angle": 0, + "content": "H. Yu et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.373, + 0.16 + ], + "angle": 0, + "content": "5.2 Main Results" + }, + { + "type": "table_caption", + "bbox": [ + 0.33, + 0.202, + 0.674, + 0.216 + ], + "angle": 0, + "content": "Table 1: Performance on the Occ-ScanNet dataset" + }, + { + "type": "table", + "bbox": [ + 0.229, + 0.228, + 0.773, + 0.339 + ], + "angle": 0, + "content": "
MethodInputIoUceilingfloorwallwindowchairbedsofatabletvsfurnitureobjects
MonoScene* [5]xrgb41.6015.1744.7122.4112.5526.1127.0335.9128.326.5732.1619.84
ISO(Ours)xrgb42.1619.8841.8822.3716.9829.0942.4342.0029.6010.6236.3624.61
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.397, + 0.789, + 0.473 + ], + "angle": 0, + "content": "Occ-ScanNet performance. We first evaluate our model's performance on the large-scale Occ-ScanNet dataset. As shown in Tab. 1, the results indicate that our method significantly outperforms MonoScene [5]. The * denotes results obtained using their code trained on our dataset. In Fig. 5, we also conducted qualitative visualization comparisons." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.5, + 0.79, + 0.636 + ], + "angle": 0, + "content": "NYUv2 performance. The NYUv2 [29] dataset serves as a widely used benchmark for indoor scene evaluation. The results in Tab. 2 demonstrate that our method achieves state-of-the-art performance. Our method also demonstrates substantial performance improvement on the NYUv2 dataset, in addition to the gains observed on Occ-ScanNet. As illustrated in Fig. 6, our ISO demonstrates superior room layout prediction compared to [5], particularly evident in rows 1 and 2. This improvement can be attributed to incorporating depth knowledge into the model. Additionally, our method exhibits the capability of learning small objects, such as pictures on the wall in row 3." + }, + { + "type": "table_caption", + "bbox": [ + 0.348, + 0.664, + 0.656, + 0.677 + ], + "angle": 0, + "content": "Table 2: Performance on the NYUv2 dataset" + }, + { + "type": "table", + "bbox": [ + 0.216, + 0.69, + 0.794, + 0.81 + ], + "angle": 0, + "content": "
MethodInputIoUceilingfloorwallwindowchairbedsofatabletvsfurnitureobjects
■■■■■■■■■■■
LMSCNetrgb [28]xocc33.934.4988.414.630.253.9432.0315.446.570.0214.514.3915.88
AICNetrgb [18]xrgb, xdepth30.037.5882.979.150.056.9335.8722.9211.110.7115.906.4518.15
3DSketchrgb [6]xrgb, xTSDF38.648.5390.459.945.6710.6442.2929.2113.889.3823.838.1922.91
MonoScene [5]xrgb42.518.8993.5012.0612.5713.7248.1936.1115.1315.2227.9612.9426.94
NDC-Scene [45]xrgb44.1712.0293.5113.1113.7715.8349.5739.8717.1724.5731.0014.9629.03
ISO(Ours)xrgb47.1114.2193.4715.8915.1418.3550.0140.8218.2525.9034.0817.6731.25
" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.33, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Monocular Occupancy Prediction for Scalable Indoor Scenes" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image_caption", + "bbox": [ + 0.279, + 0.154, + 0.327, + 0.168 + ], + "angle": 0, + "content": "Image" + }, + { + "type": "image", + "bbox": [ + 0.245, + 0.175, + 0.372, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.39, + 0.154, + 0.475, + 0.167 + ], + "angle": 0, + "content": "MonoScene" + }, + { + "type": "image", + "bbox": [ + 0.379, + 0.177, + 0.492, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.538, + 0.154, + 0.574, + 0.166 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.176, + 0.623, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.638, + 0.153, + 0.736, + 0.167 + ], + "angle": 0, + "content": "Ground Truth" + }, + { + "type": "image", + "bbox": [ + 0.635, + 0.177, + 0.75, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.245, + 0.271, + 0.372, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.38, + 0.273, + 0.491, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.274, + 0.624, + 0.344 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.632, + 0.275, + 0.752, + 0.343 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.245, + 0.367, + 0.372, + 0.442 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.376, + 0.372, + 0.495, + 0.436 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.372, + 0.626, + 0.437 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.633, + 0.376, + 0.752, + 0.436 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.245, + 0.462, + 0.372, + 0.537 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.378, + 0.468, + 0.497, + 0.533 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.462, + 0.626, + 0.54 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.633, + 0.462, + 0.752, + 0.538 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.574, + 0.788, + 0.618 + ], + "angle": 0, + "content": "Fig. 5: Qualitative Analysis on the Occ-ScanNet Dataset. The input image is displayed on the left, while the predicted scene is shown in the middle two columns, and ground truth on the right column." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.645, + 0.391, + 0.66 + ], + "angle": 0, + "content": "5.3 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.668, + 0.788, + 0.699 + ], + "angle": 0, + "content": "We conduct ablation studies on the NYUv2 [29] and Occ-ScanNet-mini datasets to validate the effectiveness of our model design." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.719, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Depth fusion. Inspired by BEVDepth [20], we initially adopt the integration of BEV depth information, denoted as \\( ^{+} \\)bev-depth'. However, we observed that while BEV depth works well for 3D detection in outdoor scenes, directly applying it to indoor scenes is less effective. We speculate that this discrepancy arises because height information is more crucial in indoor scenes, where indoor heights often correspond to different 3D structures and semantics. Therefore, we adopt voxel depth fusion, denoted as \\( ^{+} \\)voxel-depth'. Specifically, we project each 3D voxel centroid to a frustum grid to sample the depth distribution probability" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.127 + ], + "angle": 0, + "content": "H. Yu et al." + }, + { + "type": "image", + "bbox": [ + 0.239, + 0.149, + 0.765, + 0.379 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.395, + 0.788, + 0.466 + ], + "angle": 0, + "content": "Fig. 6: Qualitative Analysis on the NYUv2 Dataset: The input is displayed on the left, with the camera viewing frustum shown in every image. ISO adeptly captures the scene layout and delineates various semantic instances more accurately. Notably, it excels in reconstructing the corners of walls (row 2) and accurately representing small objects like pictures on the wall (row 3)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.497, + 0.785, + 0.542 + ], + "angle": 0, + "content": "of the voxel. We then weight the 2D feature at that pixel coordinate as the voxel feature. The results in Tab. 3 further validate that the voxel depth fusion approach is more effective." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.544, + 0.788, + 0.664 + ], + "angle": 0, + "content": "The BEV method usually performs better in detecting ceilings and floors but performs poorly in detecting furniture such as windows, walls, televisions, beds, sofas, and tables. Ceilings and floors usually have relatively flat and regular geometric shapes, which makes their projections clearer and more consistent from a BEV perspective. In contrast, objects such as windows, walls, and televisions typically have more complex geometric shapes. In the BEV perspective, the projections of these objects may become blurry or overlapping, making it difficult to accurately distinguish and detect." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.694, + 0.787, + 0.709 + ], + "angle": 0, + "content": "Table 3: Ablation study of depth fusion on the NYUv2 and Occ-ScanNet-mini dataset" + }, + { + "type": "table", + "bbox": [ + 0.308, + 0.72, + 0.698, + 0.809 + ], + "angle": 0, + "content": "
MethodInputNYUv2Occ-ScanNet-mini
IoUmIoUIoUmIoU
baselinexrgb42.2727.1350.9438.95
+ bev-depthxrgb42.6727.1451.5838.48
+ voxel-depthxrgb47.1131.2551.0339.08
" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.33, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Monocular Occupancy Prediction for Scalable Indoor Scenes" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.29, + 0.145, + 0.712, + 0.159 + ], + "angle": 0, + "content": "Table 4: Ablation of depth information on the NYUv2 dataset" + }, + { + "type": "table", + "bbox": [ + 0.341, + 0.171, + 0.661, + 0.283 + ], + "angle": 0, + "content": "
depth-methodlearnedmulti-scaleIoUmIoU
GT✓53.9834.47
ZoeDepth✓45.1829.15
ZoeDepth✓✓45.2429.40
DepthAnything44.7729.28
DepthAnything✓45.4829.57
DepthAnything✓✓46.9431.02
DepthAnything✓47.1131.25
" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.316, + 0.787, + 0.438 + ], + "angle": 0, + "content": "Different depth pre-trained models. We conducted ablation experiments on the depth method using the Occ-ScanNet-mini dataset. Initially, we observed the upper bound of model performance when ground truth (GT) depth information was provided. Subsequently, we initialized our depth module using ZoeDepth [3] and Depth-Anything [44]. Experimental results indicate that utilizing Depth-Anything [44] yields better performance. Moreover, our approach allows for further enhancement by fine-tuning a pre-trained depth module, resulting in improved performance." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.475, + 0.787, + 0.717 + ], + "angle": 0, + "content": "Multi-scale depth. In Tab. 4, we investigate the impact of using multi-scale depth distribution on the NYUv2 [29] dataset. Specifically, the multi-scale depth distribution is used to weight the multi-scale 3D voxel feature. Meanwhile, we tried a method that only uses the single-scale depth distribution map to generate the 3D voxel depth matrix. The used depth distribution map has a scale of \\(\\frac{1}{8}\\) original image scale. Only this single-scale 3D voxel matrix is used to weigh the summed-up 3D voxel feature out of the FLoSP module. Without multi-scale depth, all 3D features are simply added up and a single depth distribution is used for weighting. This method ignores the different importance that features at different scales may have, as all features at all scales are treated equally and all depth information is weighted by a unified depth distribution. With multiscale depth, the 3D features of each scale are first weighted with the depth distribution of their corresponding scale, and then these weighted features are added up. This method takes into account the differences in the importance of features at different scales, as each scale's feature is weighted based on its corresponding depth distribution." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.72, + 0.787, + 0.84 + ], + "angle": 0, + "content": "In real-world scenarios, various objects and structures (such as ceilings, walls, furniture, etc.) often exhibit different scales. Hence, accounting for scale variations is crucial for accurately modeling these objects. Multi-scale depth can effectively capture these scale changes by assigning different weights to features at each scale. Depth information is indispensable for 3D scene completion. Leveraging the depth distribution information at each scale allows us to weight the corresponding 3D features, enabling the model to better understand and represent the objects in the scene and their relationships." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.127 + ], + "angle": 0, + "content": "H. Yu et al." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.146, + 0.35, + 0.161 + ], + "angle": 0, + "content": "5.4 Discussion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.172, + 0.788, + 0.248 + ], + "angle": 0, + "content": "Data scaling up. In Tab. 5, we delve into the impact of data volume on our method's performance by scaling up the number of scene samples. We trained the model on Occ-ScanNet using \\(10\\%\\) and \\(100\\%\\) scene samples respectively, and tested the model performance. The results indicate that larger dataset exhibits more significant gains." + }, + { + "type": "table_caption", + "bbox": [ + 0.323, + 0.274, + 0.681, + 0.29 + ], + "angle": 0, + "content": "Table 5: Scaling up comparison on the Occ-ScanNet" + }, + { + "type": "table", + "bbox": [ + 0.379, + 0.302, + 0.62, + 0.375 + ], + "angle": 0, + "content": "
DataDepth BranchOurs
IoUmIoU
10%✓21.799.57
100%✓42.1628.71
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.425, + 0.788, + 0.547 + ], + "angle": 0, + "content": "Limitations and future works. Although our model can estimate 3D occupancy effectively with the assistance of depth information, semantic learning still faces challenges due to class imbalance issues. Furthermore, our proposed Occ-ScanNet only considers 11 common semantic classes, which may not fully capture the diversity of categories present in real-world scenarios. Despite the significant increase in data volume compared to previous datasets, Occ-ScanNet remains limited in the number of scenes it covers. In future work, we will focus more on semantic exploration and conduct tests in more generalized scenarios." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.57, + 0.36, + 0.587 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.602, + 0.788, + 0.725 + ], + "angle": 0, + "content": "In this paper, we introduce ISO, a novel method for monocular 3D occupancy prediction. ISO utilizes a D-FLoSP (Dual Feature Line of Sight Projection) and a multi-scale feature fusion strategy to address variations in scene and object sizes, specifically tailored for indoor scenes. Additionally, we present a new benchmark, Occ-ScanNet, aimed at fostering scalable studies of indoor scenes. We hope that this dataset will attract more attention to research on indoor occupancy prediction. While achieving promising results in 3D structure prediction, accurately identifying semantics remains a significant challenge in the future." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.747, + 0.395, + 0.765 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "This work was supported in part by the National Key R&D Program of China (No. 2022ZD0160102), the National Natural Science Foundation of China (No. U21B2042, No. 62320106010), and in part by the 2035 Innovation Program of CAS, and the InnoHK program." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.329, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Monocular Occupancy Prediction for Scalable Indoor Scenes" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.179, + 0.785, + 0.207 + ], + "angle": 0, + "content": "1. Arshad, M.S., Beksi, W.J.: List: Learning implicitly from spatial transformers for single-view 3d reconstruction. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.209, + 0.785, + 0.235 + ], + "angle": 0, + "content": "2. Azuma, R.T.: A survey of augmented reality. Presence: teleoperators & virtual environments (1997)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.236, + 0.785, + 0.263 + ], + "angle": 0, + "content": "3. Bhat, S.F., Birkl, R., Wofk, D., Wonka, P., Müller, M.: Zoedepth: Zero-shot transfer by combining relative and metric depth. arXiv preprint arXiv:2302.12288 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.265, + 0.785, + 0.292 + ], + "angle": 0, + "content": "4. Birkl, R., Wofk, D., Müller, M.: Midas v3. 1-a model zoo for robust monocular relative depth estimation. arXiv preprint arXiv:2307.14460 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.294, + 0.785, + 0.32 + ], + "angle": 0, + "content": "5. Cao, A.Q., de Charette, R.: Monoscene: Monocular 3d semantic scene completion. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.321, + 0.785, + 0.348 + ], + "angle": 0, + "content": "6. Chen, X., Lin, K.Y., Qian, C., Zeng, G., Li, H.: 3d sketch-aware semantic scene completion via semi-supervised structure prior. In: CVPR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.349, + 0.785, + 0.376 + ], + "angle": 0, + "content": "7. Dahnert, M., Hou, J., Nießner, M., Dai, A.: Panoptic 3d scene reconstruction from a single rgb image. NeurIPS (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.378, + 0.785, + 0.404 + ], + "angle": 0, + "content": "8. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: Scannet: Richly-annotated 3d reconstructions of indoor scenes. In: CVPR (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.406, + 0.785, + 0.432 + ], + "angle": 0, + "content": "9. Denninger, M., Triebel, R.: 3d scene reconstruction from a single viewpoint. In: ECCV (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.434, + 0.785, + 0.46 + ], + "angle": 0, + "content": "10. DeSouza, G.N., Kak, A.C.: Vision for mobile robot navigation: A survey. TPAMI (2002)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.462, + 0.785, + 0.489 + ], + "angle": 0, + "content": "1. Fan, H., Su, H., Guibas, L.J.: A point set generation network for 3d object reconstruction from a single image. In: CVPR (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.49, + 0.785, + 0.531 + ], + "angle": 0, + "content": "2. Fieraru, M., Zanfir, M., Oneata, E., Popa, A.I., Olaru, V., Sminchisescu, C.: Reconstructing three-dimensional models of interacting humans. arXiv preprint arXiv:2308.01854 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.532, + 0.785, + 0.573 + ], + "angle": 0, + "content": "3. Goel, S., Pavlakos, G., Rajasegaran, J., Kanazawa, A., Malik, J.: Humans in 4d: Reconstructing and tracking humans with transformers. arXiv preprint arXiv:2305.20091 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.575, + 0.785, + 0.601 + ], + "angle": 0, + "content": "4. Guan, H., Song, C., Zhang, Z.: Gramo: geometric resampling augmentation for monocular 3d object detection. Frontiers of Computer Science (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.602, + 0.785, + 0.643 + ], + "angle": 0, + "content": "5. Huang, J., Huang, G., Zhu, Z., Ye, Y., Du, D.: Bevdet: High-performance multicamera 3d object detection in bird-eye-view. arXiv preprint arXiv:2112.11790 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.645, + 0.785, + 0.672 + ], + "angle": 0, + "content": "6. Huang, Y., Zheng, W., Zhang, B., Zhou, J., Lu, J.: Selfocc: Self-supervised vision-based 3d occupancy prediction. In: CVPR (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.673, + 0.785, + 0.699 + ], + "angle": 0, + "content": "7. Huang, Y., Zheng, W., Zhang, Y., Zhou, J., Lu, J.: Tri-perspective view for vision-based 3d semantic occupancy prediction. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.701, + 0.785, + 0.728 + ], + "angle": 0, + "content": "8. Li, J., Han, K., Wang, P., Liu, Y., Yuan, X.: Anisotropic convolutional networks for 3d semantic scene completion. In: CVPR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.729, + 0.785, + 0.77 + ], + "angle": 0, + "content": "9. Li, Y., Yu, Z., Choy, C., Xiao, C., Alvarez, J.M., Fidler, S., Feng, C., Anandkumar, A.: Voxformer: Sparse voxel transformer for camera-based 3d semantic scene completion. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.771, + 0.785, + 0.798 + ], + "angle": 0, + "content": "20. Li, Y., Ge, Z., Yu, G., Yang, J., Wang, Z., Shi, Y., Sun, J., Li, Z.: Bevdepth: Acquisition of reliable depth for multi-view 3d object detection. In: AAAI (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "21. Li, Z., Wang, W., Li, H., Xie, E., Sima, C., Lu, T., Qiao, Y., Dai, J.: Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In: ECCV (2022)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.179, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.127 + ], + "angle": 0, + "content": "H. Yu et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "22. Li, Z., Yu, Z., Austin, D., Fang, M., Lan, S., Kautz, J., Alvarez, J.M.: Fb-occ: 3d occupancy prediction based on forward-backward view transformation. arXiv preprint arXiv:2307.01492 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.785, + 0.217 + ], + "angle": 0, + "content": "23. Liu, S., Hu, Y., Zeng, Y., Tang, Q., Jin, B., Han, Y., Li, X.: See and think: Disentangling semantic scene completion. NeurIPS (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.218, + 0.785, + 0.244 + ], + "angle": 0, + "content": "24. Murez, Z., Van As, T., Bartolozzi, J., Sinha, A., Badrinarayanan, V., Rabinovich, A.: Atlas: End-to-end 3d scene reconstruction from posed images. In: ECCV (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.245, + 0.785, + 0.285 + ], + "angle": 0, + "content": "25. Pan, M., Liu, J., Zhang, R., Huang, P., Li, X., Liu, L., Zhang, S.: Renderocc: Visioncentric 3d occupancy prediction with 2d rendering supervision. arXiv preprint arXiv:2309.09502 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.785, + 0.314 + ], + "angle": 0, + "content": "26. Park, J.J., Florence, P., Straub, J., Newcombe, R., Lovegrove, S.: Deepsdf: Learning continuous signed distance functions for shape representation. In: CVPR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.785, + 0.355 + ], + "angle": 0, + "content": "27. Ranftl, R., Lasinger, K., Hafner, D., Schindler, K., Koltun, V.: Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. TPAMI (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.357, + 0.785, + 0.383 + ], + "angle": 0, + "content": "28. Roldao, L., de Charette, R., Verroust-Blondet, A.: Lmscnet: Lightweight multiscale 3d semantic completion. In: 3DV (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.785, + 0.411 + ], + "angle": 0, + "content": "29. Silberman, N., Hoiem, D., Kohli, P., Fergus, R.: Indoor segmentation and support inference from rgbd images. In: ECCV (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.412, + 0.785, + 0.438 + ], + "angle": 0, + "content": "30. Song, S., Yu, F., Zeng, A., Chang, A.X., Savva, M., Funkhouser, T.: Semantic scene completion from a single depth image. In: CVPR (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.44, + 0.785, + 0.466 + ], + "angle": 0, + "content": "31. Tan, M., Le, Q.: Efficientnet: Rethinking model scaling for convolutional neural networks. In: ICML (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.467, + 0.785, + 0.493 + ], + "angle": 0, + "content": "32. Tang, Y., Dorn, S., Savani, C.: Center3d: Center-based monocular 3d object detection with joint depth understanding. In: DAGM GCPR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.495, + 0.785, + 0.535 + ], + "angle": 0, + "content": "33. Tian, X., Jiang, T., Yun, L., Mao, Y., Yang, H., Wang, Y., Wang, Y., Zhao, H.: Occ3d: A large-scale 3d occupancy prediction benchmark for autonomous driving. NeurIPS (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.536, + 0.785, + 0.563 + ], + "angle": 0, + "content": "34. Tong, W., Sima, C., Wang, T., Chen, L., Wu, S., Deng, H., Gu, Y., Lu, L., Luo, P., Lin, D., et al.: Scene as occupancy. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.565, + 0.785, + 0.591 + ], + "angle": 0, + "content": "35. Wang, S., Liu, Y., Wang, T., Li, Y., Zhang, X.: Exploring object-centric temporal modeling for efficient multi-view 3d object detection. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.592, + 0.785, + 0.632 + ], + "angle": 0, + "content": "36. Wang, X., Zhu, Z., Xu, W., Zhang, Y., Wei, Y., Chi, X., Ye, Y., Du, D., Lu, J., Wang, X.: Openoccupancy: A large scale benchmark for surrounding semantic occupancy perception. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.633, + 0.785, + 0.66 + ], + "angle": 0, + "content": "37. Wang, Y., Guizilini, V.C., Zhang, T., Wang, Y., Zhao, H., Solomon, J.: Detr3d: 3d object detection from multi-view images via 3d-to-2d queries. In: CoRL (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.661, + 0.785, + 0.688 + ], + "angle": 0, + "content": "38. Wang, Y., Chen, Y., Liao, X., Fan, L., Zhang, Z.: Panoocc: Unified occupancy representation for camera-based 3d panoptic segmentation. In: CVPR (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.689, + 0.785, + 0.715 + ], + "angle": 0, + "content": "39. Wang, Y., Chen, Y., Zhang, Z.: Frustumformer: Adaptive instance-aware resampling for multi-view 3d detection. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.716, + 0.785, + 0.743 + ], + "angle": 0, + "content": "40. Wei, Y., Zhao, L., Zheng, W., Zhu, Z., Zhou, J., Lu, J.: Surroundocc: Multi-camera 3d occupancy prediction for autonomous driving. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.744, + 0.785, + 0.771 + ], + "angle": 0, + "content": "41. Wu, Q., Wang, K., Li, K., Zheng, J., Cai, J.: Objectsdf++: Improved object-compositional neural implicit surfaces. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.772, + 0.785, + 0.798 + ], + "angle": 0, + "content": "42. Wu, S.C., Tateno, K., Navab, N., Tombari, F.: Scfusion: Real-time incremental scene reconstruction with semantic completion. In: 3DV (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "43. Yang, C., Chen, Y., Tian, H., Tao, C., Zhu, X., Zhang, Z., Huang, G., Li, H., Qiao, Y., Lu, L., et al.: Bevformer v2: Adapting modern image backbones to bird's-eye-view recognition via perspective supervision. In: CVPR (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.33, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Monocular Occupancy Prediction for Scalable Indoor Scenes" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "44. Yang, L., Kang, B., Huang, Z., Xu, X., Feng, J., Zhao, H.: Depth anything: Unleashing the power of large-scale unlabeled data. In: CVPR (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.175, + 0.785, + 0.217 + ], + "angle": 0, + "content": "45. Yao, J., Li, C., Sun, K., Cai, Y., Li, H., Ouyang, W., Li, H.: Ndc-scene: Boost monocular 3d semantic scene completion in normalized device coordinates space. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.218, + 0.785, + 0.259 + ], + "angle": 0, + "content": "46. Yu, Z., Shu, C., Deng, J., Lu, K., Liu, Z., Yu, J., Yang, D., Li, H., Chen, Y.: Flashocc: Fast and memory-efficient occupancy prediction via channel-to-height plugin. arXiv preprint arXiv:2311.12058 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.259, + 0.785, + 0.286 + ], + "angle": 0, + "content": "47. Yurtsever, E., Lambert, J., Carballo, A., Takeda, K.: A survey of autonomous driving: Common practices and emerging technologies. IEEE access (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.286, + 0.785, + 0.314 + ], + "angle": 0, + "content": "48. Zhang, J., Zhao, H., Yao, A., Chen, Y., Zhang, L., Liao, H.: Efficient semantic scene completion network with spatial group convolution. In: ECCV (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.314, + 0.785, + 0.342 + ], + "angle": 0, + "content": "49. Zhang, X., Bi, S., Sunkavalli, K., Su, H., Xu, Z.: Nerfusion: Fusing radiance fields for large-scale scene reconstruction. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.342, + 0.785, + 0.37 + ], + "angle": 0, + "content": "50. Zheng, Z., Yu, T., Wei, Y., Dai, Q., Liu, Y.: Deephuman: 3d human reconstruction from a single image. In: ICCV (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.37, + 0.785, + 0.397 + ], + "angle": 0, + "content": "51. Zhong, M., Zeng, G.: Semantic point completion network for 3d semantic scene completion. In: ECAI (2020)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.397 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Monocular Occupancy Prediction for Scalable Indoor Scenes/325d434b-3564-4d20-a82e-4ec4e345eb58_origin.pdf b/2024/Monocular Occupancy Prediction for Scalable Indoor Scenes/325d434b-3564-4d20-a82e-4ec4e345eb58_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..53fc8433898ec4cfbc8b7f8472f2a114dbd9f003 --- /dev/null +++ b/2024/Monocular Occupancy Prediction for Scalable Indoor Scenes/325d434b-3564-4d20-a82e-4ec4e345eb58_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d58f7e7baa7239ee9b1c2a6feedb53420ef603a1d417acbf68f7451da60da1aa +size 3336218 diff --git a/2024/Monocular Occupancy Prediction for Scalable Indoor Scenes/full.md b/2024/Monocular Occupancy Prediction for Scalable Indoor Scenes/full.md new file mode 100644 index 0000000000000000000000000000000000000000..c2cefa262a213819f117b8b0c1ca9f0600148e61 --- /dev/null +++ b/2024/Monocular Occupancy Prediction for Scalable Indoor Scenes/full.md @@ -0,0 +1,323 @@ +# Monocular Occupancy Prediction for Scalable Indoor Scenes + +Hongxiao $\mathrm{Yu}^{1,2}$ Yuqi Wang $^{1,2}$ , Yuntao Chen $^{3}$ , and Zhaoxiang Zhang $^{1,2,3}$ + +1 School of Artificial Intelligence, University of Chinese Academy of Sciences (UCAS) + +$^{2}$ NLPR, MAIS, Institute of Automation, Chinese Academy of Sciences (CASIA) + +3 Centre for Artificial Intelligence and Robotics (HKISI_CAS) {yuhongxiao2023, wangyuqi2020, zhaoxiang.zhang}@ia.ac.cn, chenyuntao08@gmail.com + +Abstract. Camera-based 3D occupancy prediction has recently garnered increasing attention in outdoor driving scenes. However, research in indoor scenes remains relatively unexplored. The core differences in indoor scenes lie in the complexity of scene scale and the variance in object size. In this paper, we propose a novel method, named ISO, for predicting indoor scene occupancy using monocular images. ISO harnesses the advantages of a pretrained depth model to achieve accurate depth predictions. Furthermore, we introduce the Dual Feature Line of Sight Projection (D-FLoSP) module within ISO, which enhances the learning of 3D voxel features. To foster further research in this domain, we introduce Occ-ScanNet, a large-scale occupancy benchmark for indoor scenes. With a dataset size 40 times larger than the NYUv2 dataset, it facilitates future scalable research in indoor scene analysis. Experimental results on both NYUv2 and Occ-ScanNet demonstrate that our method achieves state-of-the-art performance. The dataset and code are made publicly at https://github.com/hongxiaoy/ISO.git. + +Keywords: 3D occupancy prediction $\cdot$ Camera-based scene understanding $\cdot$ Semantic scene completion + +# 1 Introduction + +3D scene understanding is a crucial task in computer vision, becoming increasingly important for applications such as robotic navigation [10], augmented reality [2], and autonomous driving [47]. While humans possess a natural ability to comprehend 3D environments through vision, enabling this ability in computers poses a significant challenge due to limitations like restricted fields of view, sparse sensing capabilities, and measurement noise. However, with the rapid development of deep learning and the availability of large-scale 3D driving datasets, camera-based 3D object detection [14, 15, 20, 21, 35, 37, 39, 43] and occupancy prediction [5, 22, 33, 38, 40] have developed rapidly, achieving significant improvements in performance. + +Occupancy prediction has gained popularity recently due to its effectiveness in representing both background and foreground objects within a scene using a + +unified representation. Although significant progress has been made in outdoor driving scenarios, research in indoor scenarios remains limited. Indoor scenes differ from outdoor driving scenes in two key aspects: (1) scene-scale complexity: Indoor rooms often exhibit a more diverse range of sizes compared to outdoor environments, where driving scenarios typically focus on a fixed 3D space for perception. This diversity, ranging from spacious living rooms to narrow kitchens, poses higher precision requirements for depth prediction. (2) object complexity: Indoor scenes feature a higher density and greater variety of objects. Unlike outdoor objects such as vehicles and pedestrians, which typically have consistent sizes within their respective categories and are relatively well-separated, indoor furniture and other objects often exhibit significant variations in scale and are closely positioned to each other. This increased complexity necessitates more sophisticated 3D perception techniques to accurately capture and understand the intricate geometry and relationships among objects within indoor spaces. Furthermore, existing works [5,45] focusing on indoor scenes primarily utilize the NYUv2 dataset [29], which lacks a more scalable and generalizable benchmark for comprehensive evaluation. + +To address the above issues, we firstly introduce a more scalable benchmark, Occ-ScanNet, for 3D occupancy prediction in indoor scenes. This benchmark builds upon the large-scale ScanNet [8] dataset, offering a 40 times more samples compared to the NYUv2 [29] dataset, thus greatly expanding the scope and diversity of indoor scene study. To effectively address the unique complexities of indoor scenes, we propose a novel method named ISO (Indoor Scene Occupancy). ISO leverages a powerful pretrained depth model and integrates a D-FLoSP (Dual Feature Line of Sight Projection) module. This module enables precise depth estimation and facilitates learning voxel features for accurate predictions. To accommodate the varying sizes of indoor objects, we introduce a multi-scale feature fusion module. This module enhances the learning of object features across different scales. Leveraging a powerful depth model, our method is adept at handling diverse indoor scenes, offering a robust and versatile solution. + +Our main contributions can be summarized as follows. + +- We introduce a new benchmark called Occ-ScanNet for monocular 3D occupancy prediction in indoor scenes. With a dataset size 40 times larger than the NYUv2 dataset, it significantly enhances the potential for scalable research in indoor scene analysis. +- We propose a novel approach called $ISO$ , which primarily comprises the DFLoSP (Dual Feature Line of Sight Projection) module and a multi-scale feature fusion module. Together, these components effectively address the challenges posed by variations in scene and object sizes, enabling more accurate and robust 3D occupancy prediction. +- Experiments on the Occ-ScanNet and NYUv2 datasets demonstrate that our approach achieves state-of-the-art performance in monocular 3D occupancy prediction. Furthermore, our method exhibits significant scalability potential. + +# 2 Related works + +# 2.1 Monocular 3D Semantic Scene Completion + +Monocular 3D Semantic Scene Completion (SSC) aims to infer the complete 3D structure and corresponding semantics from a single image. This monocular setting was first introduced by Monoscene [5], advancing upon prior SSC methods [18,23,30,48,51] by relying solely on vision, without additional 3D inputs. Monoscene [5] introduces the Features Line of Sight Projection (FLoSP), inspired by optics, to obtain 3D features through ray projection. However, the shared 2D features lifted to 3D rays via FLoSP results in depth ambiguity. Consequently, subsequent efforts have placed greater emphasis on leveraging depth information. VoxFormer [19] proposed a novel query proposal network based on 2D convolutions, generating sparse queries from image depth, which showed impressive performance in driving scenes. Meanwhile, NDC-Scene [45] further devised a Depth-Adaptive Dual Decoder to concurrently upsample and merge the 2D and 3D feature maps, thereby enhancing overall performance. + +# 2.2 Multiview 3D Occupancy Prediction + +Multiview 3D Occupancy Prediction predicts the semantic occupancy of the surrounding 3D scene given multiview images, has recently garnered significant attention in the field of autonomous driving. TPVFormer [17] pioneers exploration in driving scenes, employing sparse LiDAR labels for supervision. It introduces a tri-perspective view (TPV) representation that accompanies BEV with two additional perpendicular planes. However, due to the sparse supervision provided by LiDAR, subsequent works [33, 34, 36, 40] have focused on providing more dense occupancy benchmarks. FB-OCC [22] integrates the lifting of 2D to 3D and querying of 3D to 2D, achieving a more efficient 3D feature transformation. Meanwhile, PanoOcc [38] unifies detection and semantic occupancy tasks, enabling comprehensive panoramic scene understanding. Recent methods also explore improving model efficiency [46] and utilizing weaker forms of supervision [16, 25]. + +# 2.3 3D Reconstruction from Image + +3D reconstruction is a technique that involves recovering a 3D representation of objects [1, 11, 26], scenes [9, 41, 49], or even human bodies [12, 13, 50] from camera images. The task of 3D reconstruction can be classified into two categories: monocular reconstruction and multi-view reconstruction, depending on the number of images utilized. In the context of indoor scenes, 3D scene reconstruction aims to determine the surface geometry of the entire scene, often without incorporating semantic information. Atlas [24] proposes an end-to-end 3D reconstruction framework using TSDF regression from RGB images, bypassing traditional depth map estimation for efficient semantic segmentation. Panoptic-Reconstruction [7] unifies geometric reconstruction, 3D semantic, and + +![](images/cb27d372c27774b713bc3f76a42488bc37ffc0f9440739cdc29e72df35393e5f.jpg) +Fig. 1: The core design of ISO centers around the transformation of features from 2D to 3D spaces, encompassing the Depth Branch and the D-FLoSP module. A depth branch is initially integrated, it leverages a pre-trained depth model to estimate a pixel-wise depth map which is processed by the DepthNet to generate the final depth distribution. An element-wise multiplication between the voxel depth and features followed by summation are subsequently performed to derive the initial 3D voxel feature. The 3D feature is further processed to predict the 3D scene occupancy. + +instance segmentation tasks. It predicts complete geometric reconstruction, semantic, and instance segmentations from a single RGB image's camera frustum. SCFusion [42], a real-time scene reconstruction framework, integrates continuous depth data using neural architecture for occupancy maps. It efficiently combines semantic completion with voxel states for simultaneous scene reconstruction and semantic understanding in real-time 3D environments. + +# 3 Method + +Problem definition. We focus on the problem of monocular 3D Occupancy Prediction. Specifically, this task takes a single RGB image $\mathbf{I}^{RGB}$ as input and output a voxel-wise occupancy along with semantic categories $\mathbf{Y}^{X\times Y\times Z\times C}$ . $X$ , $Y$ , and $Z$ represent the dimensions of the predicted 3D scene, while $C$ denotes the total number of semantic categories. + +Model overview. In this section, we introduce the overall architecture of ISO, as shown in Fig. 1. Following Monoscene [5], we utilize 2D Unet and 3D Unet architectures to handle 2D and 3D features. Our core design focuses on transforming features from 2D to 3D, incorporating the Depth Branch and D-FLoSP module. Specifically, we incorporate a depth branch to estimate the pixel-wise depth map using a pre-trained depth model. Processed by the DepthNet, the model outputs a refined depth distribution. Then, voxel depth and features are multiplied element-wise and summed to obtain the initial 3D voxel feature volume $\mathbf{X}^{3\mathbf{d}} \in \mathbb{R}^{X \times Y \times Z \times C}$ . Finally, after processing the 3D voxel features, the model outputs the 3D scene occupancy. + +# 3.1 Depth Branch + +In this section, we introduce how to effectively estimate depth information from a single image. + +Coarse depth estimation. Learning depth from scratch can be quite challenging. However, thanks to the recent rapid advancements in depth estimation [3,4,27,44], we can leverage pre-trained depth models to initially estimate a coarse depth map. + +Compared to past models that could only estimate relative depth, we opted for Depth-Anything [44] as the pre-trained depth model, because it excels in predicting metric depth, as the Eq. (1) shows: + +$$ +\mathbf {D} ^ {\mathrm {m e t r i c}} = \mathbf {N} _ {\mathrm {d e p t h}} (\mathbf {I} ^ {\mathrm {r g b}}) \in \mathbb {R} ^ {1 \times H \times W}, \tag {1} +$$ + +where $\mathbf{N}_{\mathrm{depth}}$ denotes the pre-trained depth model [3, 44], and $\mathbf{I}^{\mathrm{rgb}}$ represents the input image. $H$ and $W$ represents the height and width of the input image. + +Depth refinement. The coarse depth estimation is subsequently refined through model learning. The depth from pre-trained depth model is not precise enough for a higher mIoU, so we design a fine-tuning strategy. Specifically, the predicted metric depth $\mathbf{D}^{\mathrm{metric}}$ is then concatenated with the image feature $\mathbf{X}$ that has the same spatial scale. The augmented feature is continue processed by a following DepthNet to get a refined depth distribution $\mathbf{D}_{s=1}^{\mathrm{dist}}$ , as Eq. (2) shows: + +$$ +\mathbf {D} _ {\mathrm {s} = 1} ^ {\text {d i s t}} = \mathbf {F} _ {\text {d e p t h}} \left(\operatorname {C o n c a t} \left(\mathbf {D} ^ {\text {m e t r i c}}, \mathbf {X} _ {\mathrm {s} = 1} ^ {2 \mathrm {d}}\right)\right), \quad \mathbf {D} _ {\mathrm {s} = 1} ^ {\text {d i s t}} \in \mathbb {R} ^ {N _ {\text {b i n s}} \times H \times W}, \tag {2} +$$ + +where $\mathbf{F}_{\mathrm{depth}}$ is the DepthNet, $\mathbf{X}_{s=1}^{2d}$ is the 2D image feature, and $N_{\mathrm{bins}}$ is the number of discrete depth bins. $s$ denotes the level of feature scale, where $s = 1$ implies that the feature is at a 1:1 original image ratio. + +We down-samples the scale $\frac{1}{1}$ distribution to get other smaller-scale 2D image feature maps' depth distribution. We represent the procedure as Eq. (3): + +$$ +\mathbf {D} _ {\mathrm {s} = k} ^ {\text {d i s t}} = \operatorname {D o w n S a m p l e} \left(\mathbf {D} _ {\mathrm {s} = 1} ^ {\text {d i s t}}\right), k = 2, 4, 8. \tag {3} +$$ + +The refined depth is supervised by the ground truth $\mathbf{D}^{\mathrm{GT}}$ . The ground truth has single depth value at each pixel, so we convert the single value to a one-hot vector of length $N_{\mathrm{bins}}$ . The refined depth prediction can be optimized using the BCE loss function as Eq. (4) shows. + +$$ +\begin{array}{l} \mathcal {L} _ {\mathrm {d e p t h}} = - \frac {1}{N _ {\mathrm {b i n s}} \times H \times W} \\ \sum_ {d = 1} ^ {N _ {\mathrm {b i n s}}} \sum_ {h = 1} ^ {H} \sum_ {w = 1} ^ {W} \left[ \mathbf {D} _ {N _ {\mathrm {b i n s}}, d, h, w} ^ {\mathrm {G T}} \cdot \log \left(\mathbf {D} _ {N _ {\mathrm {b i n s}}, d, h, w} ^ {\mathrm {d i s t}}\right) \right. \\ \left. + \left(1 - \mathbf {D} _ {N _ {\mathrm {b i n s}}, d, h, w} ^ {\mathrm {G T}}\right) \cdot \log \left(1 - \mathbf {D} _ {N _ {\mathrm {b i n s}}, d, h, w} ^ {\mathrm {d i s t}}\right) \right] \quad (4) \\ \end{array} +$$ + +# 3.2 Dual Feature Line of Sight Projection (D-FLoSP) + +Compared to the FLoSP module in MonoScene [5], which projects 3D voxels to 2D, and 3D feature vectors are the corresponding pixel feature vectors. It only considers the projection of 3D rays and overlooks depth information, our proposed D-FLoSP module can more effectively integrate depth information. + +Specifically, after getting depth distribution in the camera frame, we implement the FLoSP module to depth distribution. Each 3D voxel centroid position $x^{c}$ in the world frame can be projected to the camera frame and the pixel frame using camera pose and intrinsic matrix. We assign the projected depth values to discrete depth bins index by Eqs. (5) and (6), according to [32], + +$$ +\delta = \frac {2 \left(d _ {\max } - d _ {\min }\right)}{N _ {b i n s} \left(1 + N _ {b i n s}\right)}, \tag {5} +$$ + +$$ +l = - 0. 5 + 0. 5 \sqrt {1 + \frac {8 \left(d - d _ {\operatorname* {m i n}}\right)}{\delta}}, \tag {6} +$$ + +and the depth distribution probability it lies in serves as the depth probability of that voxel. The process is illustrated by Eq. (7): + +$$ +\mathbf {D} _ {\mathrm {s} = k} ^ {\mathrm {3 d}} = \Phi_ {\rho \left(x ^ {c}\right)} ^ {\mathrm {3 d}} \left(\mathbf {D} _ {\mathrm {s} = k} ^ {\mathrm {d i s t}}\right) \in \mathbb {R} ^ {X \times Y \times Z \times C}, k = 1, 2, 4, 8, \tag {7} +$$ + +where $\Phi_{a}^{3\mathrm{d}}(b)$ is the 3D sampling of b at coordinates a, and $\rho (\cdot)$ is the perspective projection. Thus, the depth distribution is lifted from 2D to 3D. The bin-based representation with quantization is used for mapping continuous depth value to a discrete form, each bin will be assigned a different depth distribution score to make voxel aware of depth. The original depth precision is a continuous value, and hard to integrate it with the discrete form of voxel. + +Following [5], we also employ the FLoSP module to generate a 3D voxel feature from a 2D feature map. Similar to what we did with the voxel centroids, the projected voxel can sample corresponding $1:k$ scaled feature maps $F_{1:k}$ from the 2D UNet decoder. This process is illustrated by Equation Eq. (8): + +$$ +\mathbf {X} _ {\mathrm {s} = k} ^ {3 \mathrm {d}} = \Phi_ {\rho \left(x ^ {c}\right)} ^ {2 \mathrm {d}} \left(\mathbf {X} _ {\mathrm {s} = k} ^ {2 \mathrm {d}}\right) \in \mathbb {R} ^ {X \times Y \times Z \times C}, k = 1, 2, 4, 8, \tag {8} +$$ + +where $\Phi_{a}^{2\mathrm{d}}(b)$ is the 2D sampling of b at coordinates a, and $\rho (\cdot)$ is the perspective projection. + +The final 3D feature map $\mathbf{X}^{3\mathrm{d}}$ are summed in Eq. (9): + +$$ +\mathbf {X} ^ {\mathrm {3 d}} = \sum_ {\mathrm {s} \in \{1, 2, 4, 8 \}} \mathbf {X} _ {\mathrm {s} = k} ^ {\mathrm {3 d}} \odot \mathbf {D} _ {\mathrm {s} = k} ^ {\mathrm {3 d}}, \tag {9} +$$ + +where $\odot$ is the element-wise multiplication. The output map $\mathbf{X}^{3\mathrm{d}}$ is then served as the input of 3D UNet. + +![](images/4e4b51667dfd71e9661d016ed42f74358aa31bb99bf8b59fd60904e8f7f647cb.jpg) +(a) Depth ranges in NYUv2 and our Occ-ScanNet Benchmark + +![](images/225cd6aab82aef724a36995945316cee74bdd5373a4c636b37ae5faa67dde27d.jpg) +(b) Dataset attributes of NYUv2 and our OccScanNet Benchmark +Fig. 2: Comparison of NYUv2 and Occ-ScanNet Benchmark. In (a), the depth ranges of NYUv2 and Occ-ScanNet are distinguished by dark and light green, respectively, with the horizontal axis indicating the minimum depth and the vertical axis showing the maximum depth of scenes. (b) quantitatively demonstrates that Occ-ScanNet possesses a significantly larger data scale compared to the NYUv2 dataset. + +# 4 Occ-ScanNet Benchmark + +Compared to the previous widely used indoor scene benchmark NYUv2 [29], which contains only 795 / 654 for train / test samples, our benchmark boasts 45,755 / 19,764 samples. Our benchmark significantly surpasses NYUv2 [29] in both data quantity and richness of scene depth, as illustrated in Fig. 2. This dataset is available at https://huggingface.co/datasets/hongxiaoy/OccScanNet. + +# 4.1 Overview + +Occ-ScanNet benchmark features a train/Validation split of 45,755 / 19,764 samples. As shown in Fig. 3, Occ-ScanNet exhibits rich diversity in scenes and viewpoints. This diverse dataset not only challenges the task of predicting scene occupancy but also fosters future research endeavors towards developing larger-scale and more versatile occupancy models. + +# 4.2 Occupancy Label Generation + +We adhere to the data formulation used in NYUv2 [29] dataset. The pipeline of occupancy generation process is showed in Fig. 4. For generating 3D voxel labels, we initially followed the generation process in CompleteScanNet and then employed manual inspection. + +More specifically, from the official ScanNet [8] scenes, we extracted various data components, including color images, depth images, camera intrinsic matrices, and camera poses. A total of 100 frames were sampled from each scene and were randomly divided into training and validation sets with a $7/3$ split ratio per scene. + +![](images/c16bba8f8ca001189e36dac9dacbdb0c0ff618762516ff7dee305a17b86e6b84.jpg) +Fig. 3: Samples Visualization in Occ-ScanNet Benchmark. The original RGB image is shown in column 1,3 and 5, the corresponding scene voxel labels is shown in column 2, 4 and 6. The first two rows are different views from different scenes and the last two rows each is three different views from the same scene. + +Manual inspection. The manual inspection involves verifying camera poses, camera intrinsic, and voxel positions. We filter out erroneous samples to prevent convergence issues in the model training process. + +Voxel label assignment. For each frame, only a specific area in front of the camera is defined for analysis. Therefore, the selection of the voxel origin is paramount, as it determines the subsequent coordinates of other voxel. Subsequently, each voxel is assigned a label based on its nearest voxel in the CompleteScanNet dataset. Additionally, we excluded frames with a ratio of unknown or empty labels exceeding $95\%$ or frames where the number of semantic label classes was less than two. Consequently, we obtained the generated 3D voxel labels for each frame. + +# 5 Experiments + +# 5.1 Setup + +Datasets. The NYUv2 [29] dataset provides scenes represented in $240 \times 144 \times 240$ voxels grids, labeled with 13 classes, including 1 for free space, 1 for unknown, + +![](images/600bf412f151f2798f467dfc92b731ba91e8e72b94db45aad736c5d837165504.jpg) +Fig. 4: Pipeline of Occ-ScanNet dataset label generation. Color images, depth images, camera intrinsic and poses are extracted from ScanNet scenes. For each scene, 100 frames were sampled and randomly split into training and validation sets with a $7/3$ ratio. Frames with invalid camera poses or exceeding scene boundaries were filtered out. Only the area in front of the camera was analyzed, necessitating careful selection of the voxel origin. Voxel were labeled based on their nearest voxel in the CompleteScanNet dataset. Frames with $>95\%$ unknown/empty labels or $<2$ semantic classes were excluded, resulting in generated 3D voxel labels for each frame. + +and 11 for specific semantics (ceiling, floor, wall, window, chair, bed, sofa, table, tvs, furniture, objects). The dataset consists of 795 / 654 scenes in the train / test splits. The model is trained and evaluated on down-sampled $60 \times 36 \times 60$ voxels. + +Occ-ScanNet dataset provides scenes represented in $60 \times 60 \times 36$ voxel grids, labeled with 12 classes including 1 for free space, and 11 for specific semantics (ceiling, floor, wall, window, chair, bed, sofa, table, tvs, furniture, objects). The dataset comprises 45,755 / 19,764 frames in the train / val splits. The model is trained and evaluated on the original scale. + +Occ-ScanNet-mini dataset has the same class number and voxel scene size as Occ-ScanNet, except that this mini dataset consists of 4,639 / 2,007 frames in the train / val splits. The model is also trained and evaluated on the original scale. + +Implementation Details. We employ a pre-trained EfficientNet-B7 [31] as the encoder in our 2D UNet architecture. In the depth branch, we utilize a pre-trained DepthAnything model [44], which remains frozen during training. Additionally, we integrate a depth loss specific to the depth branch, complementing the other losses outlined in Monoscene [5]. Our model is trained on two datasets: NYUv2 [29] and Occ-ScanNet. For NYUv2, ISO undergoes 30 epochs of training using AdamW optimizers. Initially, the learning rate is set to 5e-6 for the DepthNet and 1e-4 for other components, with a learning rate reduction by 0.1 at epoch 20. Training on NYUv2 takes approximately 7 hours using 2 NVIDIA L20 GPUs (2 items per GPU). On the Occ-ScanNet-mini dataset, ISO is trained for 60 epochs under similar learning rate settings, decreasing the rate by 0.1 at epoch 40. This training process takes around 12 hours using 8 L20 GPUs. For Occ-ScanNet, ISO is trained for 10 epochs with the same learning rate schedule as used for Occ-ScanNet-mini, requiring approximately a day using 8 L20 GPUs. + +# 5.2 Main Results + +Table 1: Performance on the Occ-ScanNet dataset + +
MethodInputIoUceilingfloorwallwindowchairbedsofatabletvsfurnitureobjects
MonoScene* [5]xrgb41.6015.1744.7122.4112.5526.1127.0335.9128.326.5732.1619.84
ISO(Ours)xrgb42.1619.8841.8822.3716.9829.0942.4342.0029.6010.6236.3624.61
+ +Occ-ScanNet performance. We first evaluate our model's performance on the large-scale Occ-ScanNet dataset. As shown in Tab. 1, the results indicate that our method significantly outperforms MonoScene [5]. The * denotes results obtained using their code trained on our dataset. In Fig. 5, we also conducted qualitative visualization comparisons. + +NYUv2 performance. The NYUv2 [29] dataset serves as a widely used benchmark for indoor scene evaluation. The results in Tab. 2 demonstrate that our method achieves state-of-the-art performance. Our method also demonstrates substantial performance improvement on the NYUv2 dataset, in addition to the gains observed on Occ-ScanNet. As illustrated in Fig. 6, our ISO demonstrates superior room layout prediction compared to [5], particularly evident in rows 1 and 2. This improvement can be attributed to incorporating depth knowledge into the model. Additionally, our method exhibits the capability of learning small objects, such as pictures on the wall in row 3. + +Table 2: Performance on the NYUv2 dataset + +
MethodInputIoUceilingfloorwallwindowchairbedsofatabletvsfurnitureobjects
■■■■■■■■■■■
LMSCNetrgb [28]xocc33.934.4988.414.630.253.9432.0315.446.570.0214.514.3915.88
AICNetrgb [18]xrgb, xdepth30.037.5882.979.150.056.9335.8722.9211.110.7115.906.4518.15
3DSketchrgb [6]xrgb, xTSDF38.648.5390.459.945.6710.6442.2929.2113.889.3823.838.1922.91
MonoScene [5]xrgb42.518.8993.5012.0612.5713.7248.1936.1115.1315.2227.9612.9426.94
NDC-Scene [45]xrgb44.1712.0293.5113.1113.7715.8349.5739.8717.1724.5731.0014.9629.03
ISO(Ours)xrgb47.1114.2193.4715.8915.1418.3550.0140.8218.2525.9034.0817.6731.25
+ +![](images/04e2c5714d704f272ab8c860434ace6dba790a5c5dbad8506865878756310863.jpg) +Image + +![](images/b56c565fcb1a40a72e382e16616d20db61fc9f40eeac725137d84d8009862ef3.jpg) +MonoScene + +![](images/d7fc5ba70886aa3bf20a8c1e7c897294e9dc3e9e11d5f7a942a351a5a6d7a1e0.jpg) +Ours + +![](images/cd182c385abf9d22677067034b20841ffaa1c7229ac8fee7df544e42ff29b5ee.jpg) +Ground Truth + +![](images/fdc44390d075e3fc6576247a1bd8f6e1d0312141487072c44920d171df5c8cc1.jpg) + +![](images/1ebc98688a3d3fd3a3173e2685869d24e0e043b6de44b4a6d2f0e53ac916e198.jpg) + +![](images/3ac2892e3db2e908440b580e701c4dc413377e7e64b972d05bf3372b2341814f.jpg) + +![](images/d1400f44a4c56cceb48fbcb19e51e9f955459cff2e343a5887fad4d17eb142a0.jpg) + +![](images/5cc79ad1a3418afb90628108bbd37d04b82c455cad379098b54a1c683844ec18.jpg) + +![](images/15e5ef3661021822b1013b521dc01a98e033212ebacb07213fc7d9182cb23f7c.jpg) + +![](images/6db8146507af64f3ad0e265ce1b04282f92a486f53ede1b967eb929a0a72a9d0.jpg) + +![](images/94828630daa0a3789f33db9c02a70978ce6f2b4eab3a1c1da1a90f57e8435cee.jpg) + +![](images/10396fe5d4ea378ddc1107ded1ed79d07052a88a88841e2e2835d29180f65811.jpg) +Fig. 5: Qualitative Analysis on the Occ-ScanNet Dataset. The input image is displayed on the left, while the predicted scene is shown in the middle two columns, and ground truth on the right column. + +![](images/ac93ecc616302a419ebe3b5a45700ca5a7f49fc81925b8ef6119eae7527bce78.jpg) + +![](images/f0898202e01e8619bc49b2aaa306d734b49a00b17d953826e5404469616c0a16.jpg) + +![](images/30279eba2499b123db6f91a3fa0b0c4272933fffbf5a95082748f4553709b126.jpg) + +# 5.3 Ablation Study + +We conduct ablation studies on the NYUv2 [29] and Occ-ScanNet-mini datasets to validate the effectiveness of our model design. + +Depth fusion. Inspired by BEVDepth [20], we initially adopt the integration of BEV depth information, denoted as $^{+}$ bev-depth'. However, we observed that while BEV depth works well for 3D detection in outdoor scenes, directly applying it to indoor scenes is less effective. We speculate that this discrepancy arises because height information is more crucial in indoor scenes, where indoor heights often correspond to different 3D structures and semantics. Therefore, we adopt voxel depth fusion, denoted as $^{+}$ voxel-depth'. Specifically, we project each 3D voxel centroid to a frustum grid to sample the depth distribution probability + +![](images/e1e790fceece663066ce4a35a679e75bee7a43e5396d4b9cfad477e8820e4e36.jpg) +Fig. 6: Qualitative Analysis on the NYUv2 Dataset: The input is displayed on the left, with the camera viewing frustum shown in every image. ISO adeptly captures the scene layout and delineates various semantic instances more accurately. Notably, it excels in reconstructing the corners of walls (row 2) and accurately representing small objects like pictures on the wall (row 3). + +of the voxel. We then weight the 2D feature at that pixel coordinate as the voxel feature. The results in Tab. 3 further validate that the voxel depth fusion approach is more effective. + +The BEV method usually performs better in detecting ceilings and floors but performs poorly in detecting furniture such as windows, walls, televisions, beds, sofas, and tables. Ceilings and floors usually have relatively flat and regular geometric shapes, which makes their projections clearer and more consistent from a BEV perspective. In contrast, objects such as windows, walls, and televisions typically have more complex geometric shapes. In the BEV perspective, the projections of these objects may become blurry or overlapping, making it difficult to accurately distinguish and detect. + +Table 3: Ablation study of depth fusion on the NYUv2 and Occ-ScanNet-mini dataset + +
MethodInputNYUv2Occ-ScanNet-mini
IoUmIoUIoUmIoU
baselinexrgb42.2727.1350.9438.95
+ bev-depthxrgb42.6727.1451.5838.48
+ voxel-depthxrgb47.1131.2551.0339.08
+ +Table 4: Ablation of depth information on the NYUv2 dataset + +
depth-methodlearnedmulti-scaleIoUmIoU
GT✓53.9834.47
ZoeDepth✓45.1829.15
ZoeDepth✓✓45.2429.40
DepthAnything44.7729.28
DepthAnything✓45.4829.57
DepthAnything✓✓46.9431.02
DepthAnything✓47.1131.25
+ +Different depth pre-trained models. We conducted ablation experiments on the depth method using the Occ-ScanNet-mini dataset. Initially, we observed the upper bound of model performance when ground truth (GT) depth information was provided. Subsequently, we initialized our depth module using ZoeDepth [3] and Depth-Anything [44]. Experimental results indicate that utilizing Depth-Anything [44] yields better performance. Moreover, our approach allows for further enhancement by fine-tuning a pre-trained depth module, resulting in improved performance. + +Multi-scale depth. In Tab. 4, we investigate the impact of using multi-scale depth distribution on the NYUv2 [29] dataset. Specifically, the multi-scale depth distribution is used to weight the multi-scale 3D voxel feature. Meanwhile, we tried a method that only uses the single-scale depth distribution map to generate the 3D voxel depth matrix. The used depth distribution map has a scale of $\frac{1}{8}$ original image scale. Only this single-scale 3D voxel matrix is used to weigh the summed-up 3D voxel feature out of the FLoSP module. Without multi-scale depth, all 3D features are simply added up and a single depth distribution is used for weighting. This method ignores the different importance that features at different scales may have, as all features at all scales are treated equally and all depth information is weighted by a unified depth distribution. With multiscale depth, the 3D features of each scale are first weighted with the depth distribution of their corresponding scale, and then these weighted features are added up. This method takes into account the differences in the importance of features at different scales, as each scale's feature is weighted based on its corresponding depth distribution. + +In real-world scenarios, various objects and structures (such as ceilings, walls, furniture, etc.) often exhibit different scales. Hence, accounting for scale variations is crucial for accurately modeling these objects. Multi-scale depth can effectively capture these scale changes by assigning different weights to features at each scale. Depth information is indispensable for 3D scene completion. Leveraging the depth distribution information at each scale allows us to weight the corresponding 3D features, enabling the model to better understand and represent the objects in the scene and their relationships. + +# 5.4 Discussion + +Data scaling up. In Tab. 5, we delve into the impact of data volume on our method's performance by scaling up the number of scene samples. We trained the model on Occ-ScanNet using $10\%$ and $100\%$ scene samples respectively, and tested the model performance. The results indicate that larger dataset exhibits more significant gains. + +Table 5: Scaling up comparison on the Occ-ScanNet + +
DataDepth BranchOurs
IoUmIoU
10%✓21.799.57
100%✓42.1628.71
+ +Limitations and future works. Although our model can estimate 3D occupancy effectively with the assistance of depth information, semantic learning still faces challenges due to class imbalance issues. Furthermore, our proposed Occ-ScanNet only considers 11 common semantic classes, which may not fully capture the diversity of categories present in real-world scenarios. Despite the significant increase in data volume compared to previous datasets, Occ-ScanNet remains limited in the number of scenes it covers. In future work, we will focus more on semantic exploration and conduct tests in more generalized scenarios. + +# 6 Conclusion + +In this paper, we introduce ISO, a novel method for monocular 3D occupancy prediction. ISO utilizes a D-FLoSP (Dual Feature Line of Sight Projection) and a multi-scale feature fusion strategy to address variations in scene and object sizes, specifically tailored for indoor scenes. Additionally, we present a new benchmark, Occ-ScanNet, aimed at fostering scalable studies of indoor scenes. We hope that this dataset will attract more attention to research on indoor occupancy prediction. While achieving promising results in 3D structure prediction, accurately identifying semantics remains a significant challenge in the future. + +# Acknowledgement + +This work was supported in part by the National Key R&D Program of China (No. 2022ZD0160102), the National Natural Science Foundation of China (No. U21B2042, No. 62320106010), and in part by the 2035 Innovation Program of CAS, and the InnoHK program. + +# References + +1. Arshad, M.S., Beksi, W.J.: List: Learning implicitly from spatial transformers for single-view 3d reconstruction. In: ICCV (2023) +2. Azuma, R.T.: A survey of augmented reality. Presence: teleoperators & virtual environments (1997) +3. Bhat, S.F., Birkl, R., Wofk, D., Wonka, P., Müller, M.: Zoedepth: Zero-shot transfer by combining relative and metric depth. arXiv preprint arXiv:2302.12288 (2023) +4. Birkl, R., Wofk, D., Müller, M.: Midas v3. 1-a model zoo for robust monocular relative depth estimation. arXiv preprint arXiv:2307.14460 (2023) +5. Cao, A.Q., de Charette, R.: Monoscene: Monocular 3d semantic scene completion. In: CVPR (2022) +6. Chen, X., Lin, K.Y., Qian, C., Zeng, G., Li, H.: 3d sketch-aware semantic scene completion via semi-supervised structure prior. In: CVPR (2020) +7. Dahnert, M., Hou, J., Nießner, M., Dai, A.: Panoptic 3d scene reconstruction from a single rgb image. NeurIPS (2021) +8. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: Scannet: Richly-annotated 3d reconstructions of indoor scenes. In: CVPR (2017) +9. Denninger, M., Triebel, R.: 3d scene reconstruction from a single viewpoint. In: ECCV (2020) +10. DeSouza, G.N., Kak, A.C.: Vision for mobile robot navigation: A survey. TPAMI (2002) +1. Fan, H., Su, H., Guibas, L.J.: A point set generation network for 3d object reconstruction from a single image. In: CVPR (2017) +2. Fieraru, M., Zanfir, M., Oneata, E., Popa, A.I., Olaru, V., Sminchisescu, C.: Reconstructing three-dimensional models of interacting humans. arXiv preprint arXiv:2308.01854 (2023) +3. Goel, S., Pavlakos, G., Rajasegaran, J., Kanazawa, A., Malik, J.: Humans in 4d: Reconstructing and tracking humans with transformers. arXiv preprint arXiv:2305.20091 (2023) +4. Guan, H., Song, C., Zhang, Z.: Gramo: geometric resampling augmentation for monocular 3d object detection. Frontiers of Computer Science (2024) +5. Huang, J., Huang, G., Zhu, Z., Ye, Y., Du, D.: Bevdet: High-performance multicamera 3d object detection in bird-eye-view. arXiv preprint arXiv:2112.11790 (2021) +6. Huang, Y., Zheng, W., Zhang, B., Zhou, J., Lu, J.: Selfocc: Self-supervised vision-based 3d occupancy prediction. In: CVPR (2024) +7. Huang, Y., Zheng, W., Zhang, Y., Zhou, J., Lu, J.: Tri-perspective view for vision-based 3d semantic occupancy prediction. In: CVPR (2023) +8. Li, J., Han, K., Wang, P., Liu, Y., Yuan, X.: Anisotropic convolutional networks for 3d semantic scene completion. In: CVPR (2020) +9. Li, Y., Yu, Z., Choy, C., Xiao, C., Alvarez, J.M., Fidler, S., Feng, C., Anandkumar, A.: Voxformer: Sparse voxel transformer for camera-based 3d semantic scene completion. In: CVPR (2023) +20. Li, Y., Ge, Z., Yu, G., Yang, J., Wang, Z., Shi, Y., Sun, J., Li, Z.: Bevdepth: Acquisition of reliable depth for multi-view 3d object detection. In: AAAI (2023) +21. Li, Z., Wang, W., Li, H., Xie, E., Sima, C., Lu, T., Qiao, Y., Dai, J.: Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In: ECCV (2022) + +22. Li, Z., Yu, Z., Austin, D., Fang, M., Lan, S., Kautz, J., Alvarez, J.M.: Fb-occ: 3d occupancy prediction based on forward-backward view transformation. arXiv preprint arXiv:2307.01492 (2023) +23. Liu, S., Hu, Y., Zeng, Y., Tang, Q., Jin, B., Han, Y., Li, X.: See and think: Disentangling semantic scene completion. NeurIPS (2018) +24. Murez, Z., Van As, T., Bartolozzi, J., Sinha, A., Badrinarayanan, V., Rabinovich, A.: Atlas: End-to-end 3d scene reconstruction from posed images. In: ECCV (2020) +25. Pan, M., Liu, J., Zhang, R., Huang, P., Li, X., Liu, L., Zhang, S.: Renderocc: Visioncentric 3d occupancy prediction with 2d rendering supervision. arXiv preprint arXiv:2309.09502 (2023) +26. Park, J.J., Florence, P., Straub, J., Newcombe, R., Lovegrove, S.: Deepsdf: Learning continuous signed distance functions for shape representation. In: CVPR (2019) +27. Ranftl, R., Lasinger, K., Hafner, D., Schindler, K., Koltun, V.: Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. TPAMI (2020) +28. Roldao, L., de Charette, R., Verroust-Blondet, A.: Lmscnet: Lightweight multiscale 3d semantic completion. In: 3DV (2020) +29. Silberman, N., Hoiem, D., Kohli, P., Fergus, R.: Indoor segmentation and support inference from rgbd images. In: ECCV (2012) +30. Song, S., Yu, F., Zeng, A., Chang, A.X., Savva, M., Funkhouser, T.: Semantic scene completion from a single depth image. In: CVPR (2017) +31. Tan, M., Le, Q.: Efficientnet: Rethinking model scaling for convolutional neural networks. In: ICML (2019) +32. Tang, Y., Dorn, S., Savani, C.: Center3d: Center-based monocular 3d object detection with joint depth understanding. In: DAGM GCPR (2020) +33. Tian, X., Jiang, T., Yun, L., Mao, Y., Yang, H., Wang, Y., Wang, Y., Zhao, H.: Occ3d: A large-scale 3d occupancy prediction benchmark for autonomous driving. NeurIPS (2023) +34. Tong, W., Sima, C., Wang, T., Chen, L., Wu, S., Deng, H., Gu, Y., Lu, L., Luo, P., Lin, D., et al.: Scene as occupancy. In: ICCV (2023) +35. Wang, S., Liu, Y., Wang, T., Li, Y., Zhang, X.: Exploring object-centric temporal modeling for efficient multi-view 3d object detection. In: ICCV (2023) +36. Wang, X., Zhu, Z., Xu, W., Zhang, Y., Wei, Y., Chi, X., Ye, Y., Du, D., Lu, J., Wang, X.: Openoccupancy: A large scale benchmark for surrounding semantic occupancy perception. In: ICCV (2023) +37. Wang, Y., Guizilini, V.C., Zhang, T., Wang, Y., Zhao, H., Solomon, J.: Detr3d: 3d object detection from multi-view images via 3d-to-2d queries. In: CoRL (2022) +38. Wang, Y., Chen, Y., Liao, X., Fan, L., Zhang, Z.: Panoocc: Unified occupancy representation for camera-based 3d panoptic segmentation. In: CVPR (2024) +39. Wang, Y., Chen, Y., Zhang, Z.: Frustumformer: Adaptive instance-aware resampling for multi-view 3d detection. In: CVPR (2023) +40. Wei, Y., Zhao, L., Zheng, W., Zhu, Z., Zhou, J., Lu, J.: Surroundocc: Multi-camera 3d occupancy prediction for autonomous driving. In: ICCV (2023) +41. Wu, Q., Wang, K., Li, K., Zheng, J., Cai, J.: Objectsdf++: Improved object-compositional neural implicit surfaces. In: ICCV (2023) +42. Wu, S.C., Tateno, K., Navab, N., Tombari, F.: Scfusion: Real-time incremental scene reconstruction with semantic completion. In: 3DV (2020) +43. Yang, C., Chen, Y., Tian, H., Tao, C., Zhu, X., Zhang, Z., Huang, G., Li, H., Qiao, Y., Lu, L., et al.: Bevformer v2: Adapting modern image backbones to bird's-eye-view recognition via perspective supervision. In: CVPR (2023) + +44. Yang, L., Kang, B., Huang, Z., Xu, X., Feng, J., Zhao, H.: Depth anything: Unleashing the power of large-scale unlabeled data. In: CVPR (2024) +45. Yao, J., Li, C., Sun, K., Cai, Y., Li, H., Ouyang, W., Li, H.: Ndc-scene: Boost monocular 3d semantic scene completion in normalized device coordinates space. In: ICCV (2023) +46. Yu, Z., Shu, C., Deng, J., Lu, K., Liu, Z., Yu, J., Yang, D., Li, H., Chen, Y.: Flashocc: Fast and memory-efficient occupancy prediction via channel-to-height plugin. arXiv preprint arXiv:2311.12058 (2023) +47. Yurtsever, E., Lambert, J., Carballo, A., Takeda, K.: A survey of autonomous driving: Common practices and emerging technologies. IEEE access (2020) +48. Zhang, J., Zhao, H., Yao, A., Chen, Y., Zhang, L., Liao, H.: Efficient semantic scene completion network with spatial group convolution. In: ECCV (2018) +49. Zhang, X., Bi, S., Sunkavalli, K., Su, H., Xu, Z.: Nerfusion: Fusing radiance fields for large-scale scene reconstruction. In: CVPR (2022) +50. Zheng, Z., Yu, T., Wei, Y., Dai, Q., Liu, Y.: Deephuman: 3d human reconstruction from a single image. In: ICCV (2019) +51. Zhong, M., Zeng, G.: Semantic point completion network for 3d semantic scene completion. In: ECAI (2020) \ No newline at end of file diff --git a/2024/Monocular Occupancy Prediction for Scalable Indoor Scenes/images.zip b/2024/Monocular Occupancy Prediction for Scalable Indoor Scenes/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..b9ab416ec05671c12d588ecec228b8167c76503b --- /dev/null +++ b/2024/Monocular Occupancy Prediction for Scalable Indoor Scenes/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ce7b980313a267490a07d464220bbb8b2d92f000251265a535f97ffcf1e3cd32 +size 541269 diff --git a/2024/Monocular Occupancy Prediction for Scalable Indoor Scenes/layout.json b/2024/Monocular Occupancy Prediction for Scalable Indoor Scenes/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..9befb81c383ccd21b4e99171578b3bde0ca442d2 --- /dev/null +++ b/2024/Monocular Occupancy Prediction for Scalable Indoor Scenes/layout.json @@ -0,0 +1,8754 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 146, + 111, + 467, + 145 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 111, + 467, + 145 + ], + "spans": [ + { + "bbox": [ + 146, + 111, + 467, + 145 + ], + "type": "text", + "content": "Monocular Occupancy Prediction for Scalable Indoor Scenes" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 132, + 167, + 481, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 167, + 481, + 181 + ], + "spans": [ + { + "bbox": [ + 132, + 167, + 481, + 181 + ], + "type": "text", + "content": "Hongxiao " + }, + { + "bbox": [ + 132, + 167, + 481, + 181 + ], + "type": "inline_equation", + "content": "\\mathrm{Yu}^{1,2}" + }, + { + "bbox": [ + 132, + 167, + 481, + 181 + ], + "type": "text", + "content": " Yuqi Wang" + }, + { + "bbox": [ + 132, + 167, + 481, + 181 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 132, + 167, + 481, + 181 + ], + "type": "text", + "content": ", Yuntao Chen" + }, + { + "bbox": [ + 132, + 167, + 481, + 181 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 132, + 167, + 481, + 181 + ], + "type": "text", + "content": ", and Zhaoxiang Zhang" + }, + { + "bbox": [ + 132, + 167, + 481, + 181 + ], + "type": "inline_equation", + "content": "^{1,2,3}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 132, + 189, + 479, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 189, + 479, + 200 + ], + "spans": [ + { + "bbox": [ + 132, + 189, + 479, + 200 + ], + "type": "text", + "content": "1 School of Artificial Intelligence, University of Chinese Academy of Sciences (UCAS)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 140, + 201, + 471, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 201, + 471, + 213 + ], + "spans": [ + { + "bbox": [ + 140, + 201, + 471, + 213 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 140, + 201, + 471, + 213 + ], + "type": "text", + "content": " NLPR, MAIS, Institute of Automation, Chinese Academy of Sciences (CASIA)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 172, + 213, + 441, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 172, + 213, + 441, + 245 + ], + "spans": [ + { + "bbox": [ + 172, + 213, + 441, + 245 + ], + "type": "text", + "content": "3 Centre for Artificial Intelligence and Robotics (HKISI_CAS) {yuhongxiao2023, wangyuqi2020, zhaoxiang.zhang}@ia.ac.cn, chenyuntao08@gmail.com" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 159, + 271, + 452, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 271, + 452, + 447 + ], + "spans": [ + { + "bbox": [ + 159, + 271, + 452, + 447 + ], + "type": "text", + "content": "Abstract. Camera-based 3D occupancy prediction has recently garnered increasing attention in outdoor driving scenes. However, research in indoor scenes remains relatively unexplored. The core differences in indoor scenes lie in the complexity of scene scale and the variance in object size. In this paper, we propose a novel method, named ISO, for predicting indoor scene occupancy using monocular images. ISO harnesses the advantages of a pretrained depth model to achieve accurate depth predictions. Furthermore, we introduce the Dual Feature Line of Sight Projection (D-FLoSP) module within ISO, which enhances the learning of 3D voxel features. To foster further research in this domain, we introduce Occ-ScanNet, a large-scale occupancy benchmark for indoor scenes. With a dataset size 40 times larger than the NYUv2 dataset, it facilitates future scalable research in indoor scene analysis. Experimental results on both NYUv2 and Occ-ScanNet demonstrate that our method achieves state-of-the-art performance. The dataset and code are made publicly at https://github.com/hongxiaoy/ISO.git." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 159, + 456, + 453, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 456, + 453, + 479 + ], + "spans": [ + { + "bbox": [ + 159, + 456, + 453, + 479 + ], + "type": "text", + "content": "Keywords: 3D occupancy prediction " + }, + { + "bbox": [ + 159, + 456, + 453, + 479 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 456, + 453, + 479 + ], + "type": "text", + "content": " Camera-based scene understanding " + }, + { + "bbox": [ + 159, + 456, + 453, + 479 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 456, + 453, + 479 + ], + "type": "text", + "content": " Semantic scene completion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 498, + 230, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 498, + 230, + 510 + ], + "spans": [ + { + "bbox": [ + 132, + 498, + 230, + 510 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 521, + 482, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 641 + ], + "type": "text", + "content": "3D scene understanding is a crucial task in computer vision, becoming increasingly important for applications such as robotic navigation [10], augmented reality [2], and autonomous driving [47]. While humans possess a natural ability to comprehend 3D environments through vision, enabling this ability in computers poses a significant challenge due to limitations like restricted fields of view, sparse sensing capabilities, and measurement noise. However, with the rapid development of deep learning and the availability of large-scale 3D driving datasets, camera-based 3D object detection [14, 15, 20, 21, 35, 37, 39, 43] and occupancy prediction [5, 22, 33, 38, 40] have developed rapidly, achieving significant improvements in performance." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "Occupancy prediction has gained popularity recently due to its effectiveness in representing both background and foreground objects within a scene using a" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 319 + ], + "type": "text", + "content": "unified representation. Although significant progress has been made in outdoor driving scenarios, research in indoor scenarios remains limited. Indoor scenes differ from outdoor driving scenes in two key aspects: (1) scene-scale complexity: Indoor rooms often exhibit a more diverse range of sizes compared to outdoor environments, where driving scenarios typically focus on a fixed 3D space for perception. This diversity, ranging from spacious living rooms to narrow kitchens, poses higher precision requirements for depth prediction. (2) object complexity: Indoor scenes feature a higher density and greater variety of objects. Unlike outdoor objects such as vehicles and pedestrians, which typically have consistent sizes within their respective categories and are relatively well-separated, indoor furniture and other objects often exhibit significant variations in scale and are closely positioned to each other. This increased complexity necessitates more sophisticated 3D perception techniques to accurately capture and understand the intricate geometry and relationships among objects within indoor spaces. Furthermore, existing works [5,45] focusing on indoor scenes primarily utilize the NYUv2 dataset [29], which lacks a more scalable and generalizable benchmark for comprehensive evaluation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 321, + 482, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 321, + 482, + 477 + ], + "spans": [ + { + "bbox": [ + 130, + 321, + 482, + 477 + ], + "type": "text", + "content": "To address the above issues, we firstly introduce a more scalable benchmark, Occ-ScanNet, for 3D occupancy prediction in indoor scenes. This benchmark builds upon the large-scale ScanNet [8] dataset, offering a 40 times more samples compared to the NYUv2 [29] dataset, thus greatly expanding the scope and diversity of indoor scene study. To effectively address the unique complexities of indoor scenes, we propose a novel method named ISO (Indoor Scene Occupancy). ISO leverages a powerful pretrained depth model and integrates a D-FLoSP (Dual Feature Line of Sight Projection) module. This module enables precise depth estimation and facilitates learning voxel features for accurate predictions. To accommodate the varying sizes of indoor objects, we introduce a multi-scale feature fusion module. This module enhances the learning of object features across different scales. Leveraging a powerful depth model, our method is adept at handling diverse indoor scenes, offering a robust and versatile solution." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 146, + 479, + 387, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 479, + 387, + 490 + ], + "spans": [ + { + "bbox": [ + 146, + 479, + 387, + 490 + ], + "type": "text", + "content": "Our main contributions can be summarized as follows." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 137, + 506, + 480, + 664 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 138, + 506, + 480, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 506, + 480, + 553 + ], + "spans": [ + { + "bbox": [ + 138, + 506, + 480, + 553 + ], + "type": "text", + "content": "- We introduce a new benchmark called Occ-ScanNet for monocular 3D occupancy prediction in indoor scenes. With a dataset size 40 times larger than the NYUv2 dataset, it significantly enhances the potential for scalable research in indoor scene analysis." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 555, + 480, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 555, + 480, + 614 + ], + "spans": [ + { + "bbox": [ + 138, + 555, + 480, + 614 + ], + "type": "text", + "content": "- We propose a novel approach called " + }, + { + "bbox": [ + 138, + 555, + 480, + 614 + ], + "type": "inline_equation", + "content": "ISO" + }, + { + "bbox": [ + 138, + 555, + 480, + 614 + ], + "type": "text", + "content": ", which primarily comprises the DFLoSP (Dual Feature Line of Sight Projection) module and a multi-scale feature fusion module. Together, these components effectively address the challenges posed by variations in scene and object sizes, enabling more accurate and robust 3D occupancy prediction." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 137, + 617, + 480, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 617, + 480, + 664 + ], + "spans": [ + { + "bbox": [ + 137, + 617, + 480, + 664 + ], + "type": "text", + "content": "- Experiments on the Occ-ScanNet and NYUv2 datasets demonstrate that our approach achieves state-of-the-art performance in monocular 3D occupancy prediction. Furthermore, our method exhibits significant scalability potential." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "text", + "content": "H. Yu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 240, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 240, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 240, + 127 + ], + "type": "text", + "content": "2 Related works" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 377, + 152 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 377, + 152 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 377, + 152 + ], + "type": "text", + "content": "2.1 Monocular 3D Semantic Scene Completion" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 159, + 482, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 159, + 482, + 316 + ], + "spans": [ + { + "bbox": [ + 130, + 159, + 482, + 316 + ], + "type": "text", + "content": "Monocular 3D Semantic Scene Completion (SSC) aims to infer the complete 3D structure and corresponding semantics from a single image. This monocular setting was first introduced by Monoscene [5], advancing upon prior SSC methods [18,23,30,48,51] by relying solely on vision, without additional 3D inputs. Monoscene [5] introduces the Features Line of Sight Projection (FLoSP), inspired by optics, to obtain 3D features through ray projection. However, the shared 2D features lifted to 3D rays via FLoSP results in depth ambiguity. Consequently, subsequent efforts have placed greater emphasis on leveraging depth information. VoxFormer [19] proposed a novel query proposal network based on 2D convolutions, generating sparse queries from image depth, which showed impressive performance in driving scenes. Meanwhile, NDC-Scene [45] further devised a Depth-Adaptive Dual Decoder to concurrently upsample and merge the 2D and 3D feature maps, thereby enhancing overall performance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 333, + 346, + 345 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 333, + 346, + 345 + ], + "spans": [ + { + "bbox": [ + 132, + 333, + 346, + 345 + ], + "type": "text", + "content": "2.2 Multiview 3D Occupancy Prediction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 353, + 482, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 353, + 482, + 509 + ], + "spans": [ + { + "bbox": [ + 130, + 353, + 482, + 509 + ], + "type": "text", + "content": "Multiview 3D Occupancy Prediction predicts the semantic occupancy of the surrounding 3D scene given multiview images, has recently garnered significant attention in the field of autonomous driving. TPVFormer [17] pioneers exploration in driving scenes, employing sparse LiDAR labels for supervision. It introduces a tri-perspective view (TPV) representation that accompanies BEV with two additional perpendicular planes. However, due to the sparse supervision provided by LiDAR, subsequent works [33, 34, 36, 40] have focused on providing more dense occupancy benchmarks. FB-OCC [22] integrates the lifting of 2D to 3D and querying of 3D to 2D, achieving a more efficient 3D feature transformation. Meanwhile, PanoOcc [38] unifies detection and semantic occupancy tasks, enabling comprehensive panoramic scene understanding. Recent methods also explore improving model efficiency [46] and utilizing weaker forms of supervision [16, 25]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 526, + 317, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 526, + 317, + 538 + ], + "spans": [ + { + "bbox": [ + 132, + 526, + 317, + 538 + ], + "type": "text", + "content": "2.3 3D Reconstruction from Image" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "content": "3D reconstruction is a technique that involves recovering a 3D representation of objects [1, 11, 26], scenes [9, 41, 49], or even human bodies [12, 13, 50] from camera images. The task of 3D reconstruction can be classified into two categories: monocular reconstruction and multi-view reconstruction, depending on the number of images utilized. In the context of indoor scenes, 3D scene reconstruction aims to determine the surface geometry of the entire scene, often without incorporating semantic information. Atlas [24] proposes an end-to-end 3D reconstruction framework using TSDF regression from RGB images, bypassing traditional depth map estimation for efficient semantic segmentation. Panoptic-Reconstruction [7] unifies geometric reconstruction, 3D semantic, and" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 201, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 201, + 91, + 447, + 102 + ], + "type": "text", + "content": "Monocular Occupancy Prediction for Scalable Indoor Scenes" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 119, + 481, + 228 + ], + "blocks": [ + { + "bbox": [ + 132, + 119, + 481, + 228 + ], + "lines": [ + { + "bbox": [ + 132, + 119, + 481, + 228 + ], + "spans": [ + { + "bbox": [ + 132, + 119, + 481, + 228 + ], + "type": "image", + "image_path": "cb27d372c27774b713bc3f76a42488bc37ffc0f9440739cdc29e72df35393e5f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 237, + 482, + 316 + ], + "lines": [ + { + "bbox": [ + 130, + 237, + 482, + 316 + ], + "spans": [ + { + "bbox": [ + 130, + 237, + 482, + 316 + ], + "type": "text", + "content": "Fig. 1: The core design of ISO centers around the transformation of features from 2D to 3D spaces, encompassing the Depth Branch and the D-FLoSP module. A depth branch is initially integrated, it leverages a pre-trained depth model to estimate a pixel-wise depth map which is processed by the DepthNet to generate the final depth distribution. An element-wise multiplication between the voxel depth and features followed by summation are subsequently performed to derive the initial 3D voxel feature. The 3D feature is further processed to predict the 3D scene occupancy." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 340, + 482, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 340, + 482, + 413 + ], + "spans": [ + { + "bbox": [ + 130, + 340, + 482, + 413 + ], + "type": "text", + "content": "instance segmentation tasks. It predicts complete geometric reconstruction, semantic, and instance segmentations from a single RGB image's camera frustum. SCFusion [42], a real-time scene reconstruction framework, integrates continuous depth data using neural architecture for occupancy maps. It efficiently combines semantic completion with voxel states for simultaneous scene reconstruction and semantic understanding in real-time 3D environments." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 434, + 202, + 448 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 434, + 202, + 448 + ], + "spans": [ + { + "bbox": [ + 132, + 434, + 202, + 448 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 464, + 482, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 464, + 482, + 525 + ], + "spans": [ + { + "bbox": [ + 130, + 464, + 482, + 525 + ], + "type": "text", + "content": "Problem definition. We focus on the problem of monocular 3D Occupancy Prediction. Specifically, this task takes a single RGB image " + }, + { + "bbox": [ + 130, + 464, + 482, + 525 + ], + "type": "inline_equation", + "content": "\\mathbf{I}^{RGB}" + }, + { + "bbox": [ + 130, + 464, + 482, + 525 + ], + "type": "text", + "content": " as input and output a voxel-wise occupancy along with semantic categories " + }, + { + "bbox": [ + 130, + 464, + 482, + 525 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}^{X\\times Y\\times Z\\times C}" + }, + { + "bbox": [ + 130, + 464, + 482, + 525 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 130, + 464, + 482, + 525 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 130, + 464, + 482, + 525 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 464, + 482, + 525 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 130, + 464, + 482, + 525 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 464, + 482, + 525 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 130, + 464, + 482, + 525 + ], + "type": "text", + "content": " represent the dimensions of the predicted 3D scene, while " + }, + { + "bbox": [ + 130, + 464, + 482, + 525 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 464, + 482, + 525 + ], + "type": "text", + "content": " denotes the total number of semantic categories." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "Model overview. In this section, we introduce the overall architecture of ISO, as shown in Fig. 1. Following Monoscene [5], we utilize 2D Unet and 3D Unet architectures to handle 2D and 3D features. Our core design focuses on transforming features from 2D to 3D, incorporating the Depth Branch and D-FLoSP module. Specifically, we incorporate a depth branch to estimate the pixel-wise depth map using a pre-trained depth model. Processed by the DepthNet, the model outputs a refined depth distribution. Then, voxel depth and features are multiplied element-wise and summed to obtain the initial 3D voxel feature volume " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^{3\\mathbf{d}} \\in \\mathbb{R}^{X \\times Y \\times Z \\times C}" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": ". Finally, after processing the 3D voxel features, the model outputs the 3D scene occupancy." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "type": "text", + "content": "H. Yu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 233, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 233, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 233, + 128 + ], + "type": "text", + "content": "3.1 Depth Branch" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 133, + 479, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 133, + 479, + 157 + ], + "spans": [ + { + "bbox": [ + 130, + 133, + 479, + 157 + ], + "type": "text", + "content": "In this section, we introduce how to effectively estimate depth information from a single image." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 172, + 479, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 172, + 479, + 219 + ], + "spans": [ + { + "bbox": [ + 130, + 172, + 479, + 219 + ], + "type": "text", + "content": "Coarse depth estimation. Learning depth from scratch can be quite challenging. However, thanks to the recent rapid advancements in depth estimation [3,4,27,44], we can leverage pre-trained depth models to initially estimate a coarse depth map." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 220, + 480, + 257 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 220, + 480, + 257 + ], + "spans": [ + { + "bbox": [ + 130, + 220, + 480, + 257 + ], + "type": "text", + "content": "Compared to past models that could only estimate relative depth, we opted for Depth-Anything [44] as the pre-trained depth model, because it excels in predicting metric depth, as the Eq. (1) shows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 228, + 260, + 481, + 276 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 260, + 481, + 276 + ], + "spans": [ + { + "bbox": [ + 228, + 260, + 481, + 276 + ], + "type": "interline_equation", + "content": "\\mathbf {D} ^ {\\mathrm {m e t r i c}} = \\mathbf {N} _ {\\mathrm {d e p t h}} (\\mathbf {I} ^ {\\mathrm {r g b}}) \\in \\mathbb {R} ^ {1 \\times H \\times W}, \\tag {1}", + "image_path": "96b051d5209c77016a7c158dd7a823b9b6403706388485264b2a436243f3d23c.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 281, + 479, + 306 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 281, + 479, + 306 + ], + "spans": [ + { + "bbox": [ + 131, + 281, + 479, + 306 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 281, + 479, + 306 + ], + "type": "inline_equation", + "content": "\\mathbf{N}_{\\mathrm{depth}}" + }, + { + "bbox": [ + 131, + 281, + 479, + 306 + ], + "type": "text", + "content": " denotes the pre-trained depth model [3, 44], and " + }, + { + "bbox": [ + 131, + 281, + 479, + 306 + ], + "type": "inline_equation", + "content": "\\mathbf{I}^{\\mathrm{rgb}}" + }, + { + "bbox": [ + 131, + 281, + 479, + 306 + ], + "type": "text", + "content": " represents the input image. " + }, + { + "bbox": [ + 131, + 281, + 479, + 306 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 131, + 281, + 479, + 306 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 281, + 479, + 306 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 131, + 281, + 479, + 306 + ], + "type": "text", + "content": " represents the height and width of the input image." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 320, + 483, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 320, + 483, + 393 + ], + "spans": [ + { + "bbox": [ + 131, + 320, + 483, + 393 + ], + "type": "text", + "content": "Depth refinement. The coarse depth estimation is subsequently refined through model learning. The depth from pre-trained depth model is not precise enough for a higher mIoU, so we design a fine-tuning strategy. Specifically, the predicted metric depth " + }, + { + "bbox": [ + 131, + 320, + 483, + 393 + ], + "type": "inline_equation", + "content": "\\mathbf{D}^{\\mathrm{metric}}" + }, + { + "bbox": [ + 131, + 320, + 483, + 393 + ], + "type": "text", + "content": " is then concatenated with the image feature " + }, + { + "bbox": [ + 131, + 320, + 483, + 393 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 131, + 320, + 483, + 393 + ], + "type": "text", + "content": " that has the same spatial scale. The augmented feature is continue processed by a following DepthNet to get a refined depth distribution " + }, + { + "bbox": [ + 131, + 320, + 483, + 393 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_{s=1}^{\\mathrm{dist}}" + }, + { + "bbox": [ + 131, + 320, + 483, + 393 + ], + "type": "text", + "content": ", as Eq. (2) shows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 164, + 407, + 480, + 434 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 164, + 407, + 480, + 434 + ], + "spans": [ + { + "bbox": [ + 164, + 407, + 480, + 434 + ], + "type": "interline_equation", + "content": "\\mathbf {D} _ {\\mathrm {s} = 1} ^ {\\text {d i s t}} = \\mathbf {F} _ {\\text {d e p t h}} \\left(\\operatorname {C o n c a t} \\left(\\mathbf {D} ^ {\\text {m e t r i c}}, \\mathbf {X} _ {\\mathrm {s} = 1} ^ {2 \\mathrm {d}}\\right)\\right), \\quad \\mathbf {D} _ {\\mathrm {s} = 1} ^ {\\text {d i s t}} \\in \\mathbb {R} ^ {N _ {\\text {b i n s}} \\times H \\times W}, \\tag {2}", + "image_path": "97c600b4f70b781188829183dfd4c2f285fcfdd20cb398a3f316ccc4e28323df.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 439, + 479, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 439, + 479, + 475 + ], + "spans": [ + { + "bbox": [ + 130, + 439, + 479, + 475 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 439, + 479, + 475 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathrm{depth}}" + }, + { + "bbox": [ + 130, + 439, + 479, + 475 + ], + "type": "text", + "content": " is the DepthNet, " + }, + { + "bbox": [ + 130, + 439, + 479, + 475 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{s=1}^{2d}" + }, + { + "bbox": [ + 130, + 439, + 479, + 475 + ], + "type": "text", + "content": " is the 2D image feature, and " + }, + { + "bbox": [ + 130, + 439, + 479, + 475 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{bins}}" + }, + { + "bbox": [ + 130, + 439, + 479, + 475 + ], + "type": "text", + "content": " is the number of discrete depth bins. " + }, + { + "bbox": [ + 130, + 439, + 479, + 475 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 130, + 439, + 479, + 475 + ], + "type": "text", + "content": " denotes the level of feature scale, where " + }, + { + "bbox": [ + 130, + 439, + 479, + 475 + ], + "type": "inline_equation", + "content": "s = 1" + }, + { + "bbox": [ + 130, + 439, + 479, + 475 + ], + "type": "text", + "content": " implies that the feature is at a 1:1 original image ratio." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 475, + 479, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 475, + 479, + 500 + ], + "spans": [ + { + "bbox": [ + 130, + 475, + 479, + 500 + ], + "type": "text", + "content": "We down-samples the scale " + }, + { + "bbox": [ + 130, + 475, + 479, + 500 + ], + "type": "inline_equation", + "content": "\\frac{1}{1}" + }, + { + "bbox": [ + 130, + 475, + 479, + 500 + ], + "type": "text", + "content": " distribution to get other smaller-scale 2D image feature maps' depth distribution. We represent the procedure as Eq. (3):" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 212, + 504, + 480, + 519 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 212, + 504, + 480, + 519 + ], + "spans": [ + { + "bbox": [ + 212, + 504, + 480, + 519 + ], + "type": "interline_equation", + "content": "\\mathbf {D} _ {\\mathrm {s} = k} ^ {\\text {d i s t}} = \\operatorname {D o w n S a m p l e} \\left(\\mathbf {D} _ {\\mathrm {s} = 1} ^ {\\text {d i s t}}\\right), k = 2, 4, 8. \\tag {3}", + "image_path": "8f44763b020105e6e824294e805ce7407b09afd9cb8693e4a6e1d95b88f7b219.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 525, + 479, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 525, + 479, + 574 + ], + "spans": [ + { + "bbox": [ + 130, + 525, + 479, + 574 + ], + "type": "text", + "content": "The refined depth is supervised by the ground truth " + }, + { + "bbox": [ + 130, + 525, + 479, + 574 + ], + "type": "inline_equation", + "content": "\\mathbf{D}^{\\mathrm{GT}}" + }, + { + "bbox": [ + 130, + 525, + 479, + 574 + ], + "type": "text", + "content": ". The ground truth has single depth value at each pixel, so we convert the single value to a one-hot vector of length " + }, + { + "bbox": [ + 130, + 525, + 479, + 574 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{bins}}" + }, + { + "bbox": [ + 130, + 525, + 479, + 574 + ], + "type": "text", + "content": ". The refined depth prediction can be optimized using the BCE loss function as Eq. (4) shows." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 140, + 590, + 481, + 668 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 590, + 481, + 668 + ], + "spans": [ + { + "bbox": [ + 140, + 590, + 481, + 668 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\mathrm {d e p t h}} = - \\frac {1}{N _ {\\mathrm {b i n s}} \\times H \\times W} \\\\ \\sum_ {d = 1} ^ {N _ {\\mathrm {b i n s}}} \\sum_ {h = 1} ^ {H} \\sum_ {w = 1} ^ {W} \\left[ \\mathbf {D} _ {N _ {\\mathrm {b i n s}}, d, h, w} ^ {\\mathrm {G T}} \\cdot \\log \\left(\\mathbf {D} _ {N _ {\\mathrm {b i n s}}, d, h, w} ^ {\\mathrm {d i s t}}\\right) \\right. \\\\ \\left. + \\left(1 - \\mathbf {D} _ {N _ {\\mathrm {b i n s}}, d, h, w} ^ {\\mathrm {G T}}\\right) \\cdot \\log \\left(1 - \\mathbf {D} _ {N _ {\\mathrm {b i n s}}, d, h, w} ^ {\\mathrm {d i s t}}\\right) \\right] \\quad (4) \\\\ \\end{array}", + "image_path": "935379fb822ae2cbe7da4df001393b2bf066f9952f8f139d4f2e89b60e8a1561.jpg" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 201, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 201, + 91, + 448, + 102 + ], + "type": "text", + "content": "Monocular Occupancy Prediction for Scalable Indoor Scenes" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 411, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 411, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 411, + 129 + ], + "type": "text", + "content": "3.2 Dual Feature Line of Sight Projection (D-FLoSP)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 143, + 479, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 143, + 479, + 191 + ], + "spans": [ + { + "bbox": [ + 130, + 143, + 479, + 191 + ], + "type": "text", + "content": "Compared to the FLoSP module in MonoScene [5], which projects 3D voxels to 2D, and 3D feature vectors are the corresponding pixel feature vectors. It only considers the projection of 3D rays and overlooks depth information, our proposed D-FLoSP module can more effectively integrate depth information." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 193, + 480, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 193, + 480, + 252 + ], + "spans": [ + { + "bbox": [ + 130, + 193, + 480, + 252 + ], + "type": "text", + "content": "Specifically, after getting depth distribution in the camera frame, we implement the FLoSP module to depth distribution. Each 3D voxel centroid position " + }, + { + "bbox": [ + 130, + 193, + 480, + 252 + ], + "type": "inline_equation", + "content": "x^{c}" + }, + { + "bbox": [ + 130, + 193, + 480, + 252 + ], + "type": "text", + "content": " in the world frame can be projected to the camera frame and the pixel frame using camera pose and intrinsic matrix. We assign the projected depth values to discrete depth bins index by Eqs. (5) and (6), according to [32]," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 234, + 264, + 480, + 289 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 234, + 264, + 480, + 289 + ], + "spans": [ + { + "bbox": [ + 234, + 264, + 480, + 289 + ], + "type": "interline_equation", + "content": "\\delta = \\frac {2 \\left(d _ {\\max } - d _ {\\min }\\right)}{N _ {b i n s} \\left(1 + N _ {b i n s}\\right)}, \\tag {5}", + "image_path": "e17b6ac3b490fa7c45104b178eda43c7223a0d2d11e74412ee4eddd6ada2ca35.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 236, + 293, + 480, + 318 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 236, + 293, + 480, + 318 + ], + "spans": [ + { + "bbox": [ + 236, + 293, + 480, + 318 + ], + "type": "interline_equation", + "content": "l = - 0. 5 + 0. 5 \\sqrt {1 + \\frac {8 \\left(d - d _ {\\operatorname* {m i n}}\\right)}{\\delta}}, \\tag {6}", + "image_path": "0973a2c5610c2c958a8ec068501aac5e2926e1504064ddc508c1e763d832ae80.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 329, + 479, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 329, + 479, + 353 + ], + "spans": [ + { + "bbox": [ + 130, + 329, + 479, + 353 + ], + "type": "text", + "content": "and the depth distribution probability it lies in serves as the depth probability of that voxel. The process is illustrated by Eq. (7):" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 198, + 364, + 480, + 380 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 364, + 480, + 380 + ], + "spans": [ + { + "bbox": [ + 198, + 364, + 480, + 380 + ], + "type": "interline_equation", + "content": "\\mathbf {D} _ {\\mathrm {s} = k} ^ {\\mathrm {3 d}} = \\Phi_ {\\rho \\left(x ^ {c}\\right)} ^ {\\mathrm {3 d}} \\left(\\mathbf {D} _ {\\mathrm {s} = k} ^ {\\mathrm {d i s t}}\\right) \\in \\mathbb {R} ^ {X \\times Y \\times Z \\times C}, k = 1, 2, 4, 8, \\tag {7}", + "image_path": "eea41230e300bea15bd2a56a612bc0bc83aacb2a102a65b27115eac7bed54c6a.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 392, + 480, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 392, + 480, + 463 + ], + "spans": [ + { + "bbox": [ + 130, + 392, + 480, + 463 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 392, + 480, + 463 + ], + "type": "inline_equation", + "content": "\\Phi_{a}^{3\\mathrm{d}}(b)" + }, + { + "bbox": [ + 130, + 392, + 480, + 463 + ], + "type": "text", + "content": " is the 3D sampling of b at coordinates a, and " + }, + { + "bbox": [ + 130, + 392, + 480, + 463 + ], + "type": "inline_equation", + "content": "\\rho (\\cdot)" + }, + { + "bbox": [ + 130, + 392, + 480, + 463 + ], + "type": "text", + "content": " is the perspective projection. Thus, the depth distribution is lifted from 2D to 3D. The bin-based representation with quantization is used for mapping continuous depth value to a discrete form, each bin will be assigned a different depth distribution score to make voxel aware of depth. The original depth precision is a continuous value, and hard to integrate it with the discrete form of voxel." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 464, + 480, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 464, + 480, + 514 + ], + "spans": [ + { + "bbox": [ + 130, + 464, + 480, + 514 + ], + "type": "text", + "content": "Following [5], we also employ the FLoSP module to generate a 3D voxel feature from a 2D feature map. Similar to what we did with the voxel centroids, the projected voxel can sample corresponding " + }, + { + "bbox": [ + 130, + 464, + 480, + 514 + ], + "type": "inline_equation", + "content": "1:k" + }, + { + "bbox": [ + 130, + 464, + 480, + 514 + ], + "type": "text", + "content": " scaled feature maps " + }, + { + "bbox": [ + 130, + 464, + 480, + 514 + ], + "type": "inline_equation", + "content": "F_{1:k}" + }, + { + "bbox": [ + 130, + 464, + 480, + 514 + ], + "type": "text", + "content": " from the 2D UNet decoder. This process is illustrated by Equation Eq. (8):" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 198, + 525, + 480, + 540 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 525, + 480, + 540 + ], + "spans": [ + { + "bbox": [ + 198, + 525, + 480, + 540 + ], + "type": "interline_equation", + "content": "\\mathbf {X} _ {\\mathrm {s} = k} ^ {3 \\mathrm {d}} = \\Phi_ {\\rho \\left(x ^ {c}\\right)} ^ {2 \\mathrm {d}} \\left(\\mathbf {X} _ {\\mathrm {s} = k} ^ {2 \\mathrm {d}}\\right) \\in \\mathbb {R} ^ {X \\times Y \\times Z \\times C}, k = 1, 2, 4, 8, \\tag {8}", + "image_path": "4406a6378bcbe150311d815f707264122f9cfc9ae1b30999916e502cea530c00.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 552, + 479, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 552, + 479, + 576 + ], + "spans": [ + { + "bbox": [ + 130, + 552, + 479, + 576 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 552, + 479, + 576 + ], + "type": "inline_equation", + "content": "\\Phi_{a}^{2\\mathrm{d}}(b)" + }, + { + "bbox": [ + 130, + 552, + 479, + 576 + ], + "type": "text", + "content": " is the 2D sampling of b at coordinates a, and " + }, + { + "bbox": [ + 130, + 552, + 479, + 576 + ], + "type": "inline_equation", + "content": "\\rho (\\cdot)" + }, + { + "bbox": [ + 130, + 552, + 479, + 576 + ], + "type": "text", + "content": " is the perspective projection." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 146, + 578, + 386, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 578, + 386, + 590 + ], + "spans": [ + { + "bbox": [ + 146, + 578, + 386, + 590 + ], + "type": "text", + "content": "The final 3D feature map " + }, + { + "bbox": [ + 146, + 578, + 386, + 590 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^{3\\mathrm{d}}" + }, + { + "bbox": [ + 146, + 578, + 386, + 590 + ], + "type": "text", + "content": " are summed in Eq. (9):" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 240, + 602, + 480, + 628 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 240, + 602, + 480, + 628 + ], + "spans": [ + { + "bbox": [ + 240, + 602, + 480, + 628 + ], + "type": "interline_equation", + "content": "\\mathbf {X} ^ {\\mathrm {3 d}} = \\sum_ {\\mathrm {s} \\in \\{1, 2, 4, 8 \\}} \\mathbf {X} _ {\\mathrm {s} = k} ^ {\\mathrm {3 d}} \\odot \\mathbf {D} _ {\\mathrm {s} = k} ^ {\\mathrm {3 d}}, \\tag {9}", + "image_path": "00e8e576ef7673d0832feb717deb38d142fac7203c4c7c1c29052be917e67b12.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "content": " is the element-wise multiplication. The output map " + }, + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^{3\\mathrm{d}}" + }, + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "content": " is then served as the input of 3D UNet." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "text", + "content": "H. Yu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 166, + 129, + 263, + 228 + ], + "blocks": [ + { + "bbox": [ + 166, + 129, + 263, + 228 + ], + "lines": [ + { + "bbox": [ + 166, + 129, + 263, + 228 + ], + "spans": [ + { + "bbox": [ + 166, + 129, + 263, + 228 + ], + "type": "image", + "image_path": "4e4b51667dfd71e9661d016ed42f74358aa31bb99bf8b59fd60904e8f7f647cb.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 236, + 304, + 254 + ], + "lines": [ + { + "bbox": [ + 132, + 236, + 304, + 254 + ], + "spans": [ + { + "bbox": [ + 132, + 236, + 304, + 254 + ], + "type": "text", + "content": "(a) Depth ranges in NYUv2 and our Occ-ScanNet Benchmark" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 342, + 129, + 439, + 227 + ], + "blocks": [ + { + "bbox": [ + 342, + 129, + 439, + 227 + ], + "lines": [ + { + "bbox": [ + 342, + 129, + 439, + 227 + ], + "spans": [ + { + "bbox": [ + 342, + 129, + 439, + 227 + ], + "type": "image", + "image_path": "225cd6aab82aef724a36995945316cee74bdd5373a4c636b37ae5faa67dde27d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 309, + 236, + 481, + 254 + ], + "lines": [ + { + "bbox": [ + 309, + 236, + 481, + 254 + ], + "spans": [ + { + "bbox": [ + 309, + 236, + 481, + 254 + ], + "type": "text", + "content": "(b) Dataset attributes of NYUv2 and our OccScanNet Benchmark" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 262, + 482, + 319 + ], + "lines": [ + { + "bbox": [ + 130, + 262, + 482, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 262, + 482, + 319 + ], + "type": "text", + "content": "Fig. 2: Comparison of NYUv2 and Occ-ScanNet Benchmark. In (a), the depth ranges of NYUv2 and Occ-ScanNet are distinguished by dark and light green, respectively, with the horizontal axis indicating the minimum depth and the vertical axis showing the maximum depth of scenes. (b) quantitatively demonstrates that Occ-ScanNet possesses a significantly larger data scale compared to the NYUv2 dataset." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 338, + 302, + 350 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 338, + 302, + 350 + ], + "spans": [ + { + "bbox": [ + 132, + 338, + 302, + 350 + ], + "type": "text", + "content": "4 Occ-ScanNet Benchmark" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 361, + 482, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 361, + 482, + 433 + ], + "spans": [ + { + "bbox": [ + 130, + 361, + 482, + 433 + ], + "type": "text", + "content": "Compared to the previous widely used indoor scene benchmark NYUv2 [29], which contains only 795 / 654 for train / test samples, our benchmark boasts 45,755 / 19,764 samples. Our benchmark significantly surpasses NYUv2 [29] in both data quantity and richness of scene depth, as illustrated in Fig. 2. This dataset is available at https://huggingface.co/datasets/hongxiaoy/OccScanNet." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 448, + 208, + 459 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 448, + 208, + 459 + ], + "spans": [ + { + "bbox": [ + 132, + 448, + 208, + 459 + ], + "type": "text", + "content": "4.1 Overview" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 465, + 481, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 465, + 481, + 525 + ], + "spans": [ + { + "bbox": [ + 130, + 465, + 481, + 525 + ], + "type": "text", + "content": "Occ-ScanNet benchmark features a train/Validation split of 45,755 / 19,764 samples. As shown in Fig. 3, Occ-ScanNet exhibits rich diversity in scenes and viewpoints. This diverse dataset not only challenges the task of predicting scene occupancy but also fosters future research endeavors towards developing larger-scale and more versatile occupancy models." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 540, + 308, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 540, + 308, + 553 + ], + "spans": [ + { + "bbox": [ + 132, + 540, + 308, + 553 + ], + "type": "text", + "content": "4.2 Occupancy Label Generation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 558, + 481, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 481, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 481, + 605 + ], + "type": "text", + "content": "We adhere to the data formulation used in NYUv2 [29] dataset. The pipeline of occupancy generation process is showed in Fig. 4. For generating 3D voxel labels, we initially followed the generation process in CompleteScanNet and then employed manual inspection." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "text", + "content": "More specifically, from the official ScanNet [8] scenes, we extracted various data components, including color images, depth images, camera intrinsic matrices, and camera poses. A total of 100 frames were sampled from each scene and were randomly divided into training and validation sets with a " + }, + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "inline_equation", + "content": "7/3" + }, + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "text", + "content": " split ratio per scene." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 201, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 201, + 91, + 447, + 102 + ], + "type": "text", + "content": "Monocular Occupancy Prediction for Scalable Indoor Scenes" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 138, + 121, + 475, + 339 + ], + "blocks": [ + { + "bbox": [ + 138, + 121, + 475, + 339 + ], + "lines": [ + { + "bbox": [ + 138, + 121, + 475, + 339 + ], + "spans": [ + { + "bbox": [ + 138, + 121, + 475, + 339 + ], + "type": "image", + "image_path": "c16bba8f8ca001189e36dac9dacbdb0c0ff618762516ff7dee305a17b86e6b84.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 355, + 482, + 401 + ], + "lines": [ + { + "bbox": [ + 130, + 355, + 482, + 401 + ], + "spans": [ + { + "bbox": [ + 130, + 355, + 482, + 401 + ], + "type": "text", + "content": "Fig. 3: Samples Visualization in Occ-ScanNet Benchmark. The original RGB image is shown in column 1,3 and 5, the corresponding scene voxel labels is shown in column 2, 4 and 6. The first two rows are different views from different scenes and the last two rows each is three different views from the same scene." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 424, + 482, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 424, + 482, + 462 + ], + "spans": [ + { + "bbox": [ + 130, + 424, + 482, + 462 + ], + "type": "text", + "content": "Manual inspection. The manual inspection involves verifying camera poses, camera intrinsic, and voxel positions. We filter out erroneous samples to prevent convergence issues in the model training process." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 479, + 482, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 479, + 482, + 575 + ], + "spans": [ + { + "bbox": [ + 130, + 479, + 482, + 575 + ], + "type": "text", + "content": "Voxel label assignment. For each frame, only a specific area in front of the camera is defined for analysis. Therefore, the selection of the voxel origin is paramount, as it determines the subsequent coordinates of other voxel. Subsequently, each voxel is assigned a label based on its nearest voxel in the CompleteScanNet dataset. Additionally, we excluded frames with a ratio of unknown or empty labels exceeding " + }, + { + "bbox": [ + 130, + 479, + 482, + 575 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 130, + 479, + 482, + 575 + ], + "type": "text", + "content": " or frames where the number of semantic label classes was less than two. Consequently, we obtained the generated 3D voxel labels for each frame." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 594, + 231, + 608 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 594, + 231, + 608 + ], + "spans": [ + { + "bbox": [ + 132, + 594, + 231, + 608 + ], + "type": "text", + "content": "5 Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 620, + 190, + 633 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 620, + 190, + 633 + ], + "spans": [ + { + "bbox": [ + 132, + 620, + 190, + 633 + ], + "type": "text", + "content": "5.1 Setup" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "Datasets. The NYUv2 [29] dataset provides scenes represented in " + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "240 \\times 144 \\times 240" + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": " voxels grids, labeled with 13 classes, including 1 for free space, 1 for unknown," + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "type": "text", + "content": "H. Yu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 119, + 476, + 159 + ], + "blocks": [ + { + "bbox": [ + 136, + 119, + 476, + 159 + ], + "lines": [ + { + "bbox": [ + 136, + 119, + 476, + 159 + ], + "spans": [ + { + "bbox": [ + 136, + 119, + 476, + 159 + ], + "type": "image", + "image_path": "600bf412f151f2798f467dfc92b731ba91e8e72b94db45aad736c5d837165504.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 175, + 482, + 263 + ], + "lines": [ + { + "bbox": [ + 130, + 175, + 482, + 263 + ], + "spans": [ + { + "bbox": [ + 130, + 175, + 482, + 263 + ], + "type": "text", + "content": "Fig. 4: Pipeline of Occ-ScanNet dataset label generation. Color images, depth images, camera intrinsic and poses are extracted from ScanNet scenes. For each scene, 100 frames were sampled and randomly split into training and validation sets with a " + }, + { + "bbox": [ + 130, + 175, + 482, + 263 + ], + "type": "inline_equation", + "content": "7/3" + }, + { + "bbox": [ + 130, + 175, + 482, + 263 + ], + "type": "text", + "content": " ratio. Frames with invalid camera poses or exceeding scene boundaries were filtered out. Only the area in front of the camera was analyzed, necessitating careful selection of the voxel origin. Voxel were labeled based on their nearest voxel in the CompleteScanNet dataset. Frames with " + }, + { + "bbox": [ + 130, + 175, + 482, + 263 + ], + "type": "inline_equation", + "content": ">95\\%" + }, + { + "bbox": [ + 130, + 175, + 482, + 263 + ], + "type": "text", + "content": " unknown/empty labels or " + }, + { + "bbox": [ + 130, + 175, + 482, + 263 + ], + "type": "inline_equation", + "content": "<2" + }, + { + "bbox": [ + 130, + 175, + 482, + 263 + ], + "type": "text", + "content": " semantic classes were excluded, resulting in generated 3D voxel labels for each frame." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 293, + 482, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 293, + 482, + 340 + ], + "spans": [ + { + "bbox": [ + 130, + 293, + 482, + 340 + ], + "type": "text", + "content": "and 11 for specific semantics (ceiling, floor, wall, window, chair, bed, sofa, table, tvs, furniture, objects). The dataset consists of 795 / 654 scenes in the train / test splits. The model is trained and evaluated on down-sampled " + }, + { + "bbox": [ + 130, + 293, + 482, + 340 + ], + "type": "inline_equation", + "content": "60 \\times 36 \\times 60" + }, + { + "bbox": [ + 130, + 293, + 482, + 340 + ], + "type": "text", + "content": " voxels." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 343, + 482, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 343, + 482, + 403 + ], + "spans": [ + { + "bbox": [ + 130, + 343, + 482, + 403 + ], + "type": "text", + "content": "Occ-ScanNet dataset provides scenes represented in " + }, + { + "bbox": [ + 130, + 343, + 482, + 403 + ], + "type": "inline_equation", + "content": "60 \\times 60 \\times 36" + }, + { + "bbox": [ + 130, + 343, + 482, + 403 + ], + "type": "text", + "content": " voxel grids, labeled with 12 classes including 1 for free space, and 11 for specific semantics (ceiling, floor, wall, window, chair, bed, sofa, table, tvs, furniture, objects). The dataset comprises 45,755 / 19,764 frames in the train / val splits. The model is trained and evaluated on the original scale." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 406, + 482, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 406, + 482, + 453 + ], + "spans": [ + { + "bbox": [ + 130, + 406, + 482, + 453 + ], + "type": "text", + "content": "Occ-ScanNet-mini dataset has the same class number and voxel scene size as Occ-ScanNet, except that this mini dataset consists of 4,639 / 2,007 frames in the train / val splits. The model is also trained and evaluated on the original scale." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 486, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 482, + 665 + ], + "type": "text", + "content": "Implementation Details. We employ a pre-trained EfficientNet-B7 [31] as the encoder in our 2D UNet architecture. In the depth branch, we utilize a pre-trained DepthAnything model [44], which remains frozen during training. Additionally, we integrate a depth loss specific to the depth branch, complementing the other losses outlined in Monoscene [5]. Our model is trained on two datasets: NYUv2 [29] and Occ-ScanNet. For NYUv2, ISO undergoes 30 epochs of training using AdamW optimizers. Initially, the learning rate is set to 5e-6 for the DepthNet and 1e-4 for other components, with a learning rate reduction by 0.1 at epoch 20. Training on NYUv2 takes approximately 7 hours using 2 NVIDIA L20 GPUs (2 items per GPU). On the Occ-ScanNet-mini dataset, ISO is trained for 60 epochs under similar learning rate settings, decreasing the rate by 0.1 at epoch 40. This training process takes around 12 hours using 8 L20 GPUs. For Occ-ScanNet, ISO is trained for 10 epochs with the same learning rate schedule as used for Occ-ScanNet-mini, requiring approximately a day using 8 L20 GPUs." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 201, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 201, + 91, + 447, + 102 + ], + "type": "text", + "content": "Monocular Occupancy Prediction for Scalable Indoor Scenes" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 228, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 228, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 228, + 126 + ], + "type": "text", + "content": "5.2 Main Results" + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 140, + 180, + 473, + 268 + ], + "blocks": [ + { + "bbox": [ + 201, + 159, + 412, + 171 + ], + "lines": [ + { + "bbox": [ + 201, + 159, + 412, + 171 + ], + "spans": [ + { + "bbox": [ + 201, + 159, + 412, + 171 + ], + "type": "text", + "content": "Table 1: Performance on the Occ-ScanNet dataset" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 140, + 180, + 473, + 268 + ], + "lines": [ + { + "bbox": [ + 140, + 180, + 473, + 268 + ], + "spans": [ + { + "bbox": [ + 140, + 180, + 473, + 268 + ], + "type": "table", + "html": "
MethodInputIoUceilingfloorwallwindowchairbedsofatabletvsfurnitureobjects
MonoScene* [5]xrgb41.6015.1744.7122.4112.5526.1127.0335.9128.326.5732.1619.84
ISO(Ours)xrgb42.1619.8841.8822.3716.9829.0942.4342.0029.6010.6236.3624.61
", + "image_path": "2a104e60b4f82f6c7f71fdac10f084e75fa5921047543bae1749de859d180f6f.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 314, + 482, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 314, + 482, + 374 + ], + "spans": [ + { + "bbox": [ + 130, + 314, + 482, + 374 + ], + "type": "text", + "content": "Occ-ScanNet performance. We first evaluate our model's performance on the large-scale Occ-ScanNet dataset. As shown in Tab. 1, the results indicate that our method significantly outperforms MonoScene [5]. The * denotes results obtained using their code trained on our dataset. In Fig. 5, we also conducted qualitative visualization comparisons." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 396, + 483, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 396, + 483, + 503 + ], + "spans": [ + { + "bbox": [ + 130, + 396, + 483, + 503 + ], + "type": "text", + "content": "NYUv2 performance. The NYUv2 [29] dataset serves as a widely used benchmark for indoor scene evaluation. The results in Tab. 2 demonstrate that our method achieves state-of-the-art performance. Our method also demonstrates substantial performance improvement on the NYUv2 dataset, in addition to the gains observed on Occ-ScanNet. As illustrated in Fig. 6, our ISO demonstrates superior room layout prediction compared to [5], particularly evident in rows 1 and 2. This improvement can be attributed to incorporating depth knowledge into the model. Additionally, our method exhibits the capability of learning small objects, such as pictures on the wall in row 3." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 132, + 546, + 485, + 641 + ], + "blocks": [ + { + "bbox": [ + 212, + 525, + 401, + 536 + ], + "lines": [ + { + "bbox": [ + 212, + 525, + 401, + 536 + ], + "spans": [ + { + "bbox": [ + 212, + 525, + 401, + 536 + ], + "type": "text", + "content": "Table 2: Performance on the NYUv2 dataset" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 546, + 485, + 641 + ], + "lines": [ + { + "bbox": [ + 132, + 546, + 485, + 641 + ], + "spans": [ + { + "bbox": [ + 132, + 546, + 485, + 641 + ], + "type": "table", + "html": "
MethodInputIoUceilingfloorwallwindowchairbedsofatabletvsfurnitureobjects
■■■■■■■■■■■
LMSCNetrgb [28]xocc33.934.4988.414.630.253.9432.0315.446.570.0214.514.3915.88
AICNetrgb [18]xrgb, xdepth30.037.5882.979.150.056.9335.8722.9211.110.7115.906.4518.15
3DSketchrgb [6]xrgb, xTSDF38.648.5390.459.945.6710.6442.2929.2113.889.3823.838.1922.91
MonoScene [5]xrgb42.518.8993.5012.0612.5713.7248.1936.1115.1315.2227.9612.9426.94
NDC-Scene [45]xrgb44.1712.0293.5113.1113.7715.8349.5739.8717.1724.5731.0014.9629.03
ISO(Ours)xrgb47.1114.2193.4715.8915.1418.3550.0140.8218.2525.9034.0817.6731.25
", + "image_path": "f7253140a0003ccaf15ae943cff655b57cc59fb75729c5c34e91a9ab0459850d.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "type": "text", + "content": "H. Yu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 149, + 138, + 227, + 198 + ], + "blocks": [ + { + "bbox": [ + 170, + 121, + 200, + 133 + ], + "lines": [ + { + "bbox": [ + 170, + 121, + 200, + 133 + ], + "spans": [ + { + "bbox": [ + 170, + 121, + 200, + 133 + ], + "type": "text", + "content": "Image" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 149, + 138, + 227, + 198 + ], + "lines": [ + { + "bbox": [ + 149, + 138, + 227, + 198 + ], + "spans": [ + { + "bbox": [ + 149, + 138, + 227, + 198 + ], + "type": "image", + "image_path": "04e2c5714d704f272ab8c860434ace6dba790a5c5dbad8506865878756310863.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 231, + 140, + 301, + 196 + ], + "blocks": [ + { + "bbox": [ + 238, + 121, + 290, + 132 + ], + "lines": [ + { + "bbox": [ + 238, + 121, + 290, + 132 + ], + "spans": [ + { + "bbox": [ + 238, + 121, + 290, + 132 + ], + "type": "text", + "content": "MonoScene" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 231, + 140, + 301, + 196 + ], + "lines": [ + { + "bbox": [ + 231, + 140, + 301, + 196 + ], + "spans": [ + { + "bbox": [ + 231, + 140, + 301, + 196 + ], + "type": "image", + "image_path": "b56c565fcb1a40a72e382e16616d20db61fc9f40eeac725137d84d8009862ef3.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 310, + 139, + 381, + 196 + ], + "blocks": [ + { + "bbox": [ + 329, + 121, + 351, + 131 + ], + "lines": [ + { + "bbox": [ + 329, + 121, + 351, + 131 + ], + "spans": [ + { + "bbox": [ + 329, + 121, + 351, + 131 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 310, + 139, + 381, + 196 + ], + "lines": [ + { + "bbox": [ + 310, + 139, + 381, + 196 + ], + "spans": [ + { + "bbox": [ + 310, + 139, + 381, + 196 + ], + "type": "image", + "image_path": "d7fc5ba70886aa3bf20a8c1e7c897294e9dc3e9e11d5f7a942a351a5a6d7a1e0.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 388, + 140, + 459, + 196 + ], + "blocks": [ + { + "bbox": [ + 390, + 121, + 450, + 132 + ], + "lines": [ + { + "bbox": [ + 390, + 121, + 450, + 132 + ], + "spans": [ + { + "bbox": [ + 390, + 121, + 450, + 132 + ], + "type": "text", + "content": "Ground Truth" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 388, + 140, + 459, + 196 + ], + "lines": [ + { + "bbox": [ + 388, + 140, + 459, + 196 + ], + "spans": [ + { + "bbox": [ + 388, + 140, + 459, + 196 + ], + "type": "image", + "image_path": "cd182c385abf9d22677067034b20841ffaa1c7229ac8fee7df544e42ff29b5ee.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 149, + 214, + 227, + 274 + ], + "blocks": [ + { + "bbox": [ + 149, + 214, + 227, + 274 + ], + "lines": [ + { + "bbox": [ + 149, + 214, + 227, + 274 + ], + "spans": [ + { + "bbox": [ + 149, + 214, + 227, + 274 + ], + "type": "image", + "image_path": "fdc44390d075e3fc6576247a1bd8f6e1d0312141487072c44920d171df5c8cc1.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 232, + 216, + 300, + 274 + ], + "blocks": [ + { + "bbox": [ + 232, + 216, + 300, + 274 + ], + "lines": [ + { + "bbox": [ + 232, + 216, + 300, + 274 + ], + "spans": [ + { + "bbox": [ + 232, + 216, + 300, + 274 + ], + "type": "image", + "image_path": "1ebc98688a3d3fd3a3173e2685869d24e0e043b6de44b4a6d2f0e53ac916e198.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 309, + 217, + 381, + 272 + ], + "blocks": [ + { + "bbox": [ + 309, + 217, + 381, + 272 + ], + "lines": [ + { + "bbox": [ + 309, + 217, + 381, + 272 + ], + "spans": [ + { + "bbox": [ + 309, + 217, + 381, + 272 + ], + "type": "image", + "image_path": "3ac2892e3db2e908440b580e701c4dc413377e7e64b972d05bf3372b2341814f.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 386, + 217, + 460, + 271 + ], + "blocks": [ + { + "bbox": [ + 386, + 217, + 460, + 271 + ], + "lines": [ + { + "bbox": [ + 386, + 217, + 460, + 271 + ], + "spans": [ + { + "bbox": [ + 386, + 217, + 460, + 271 + ], + "type": "image", + "image_path": "d1400f44a4c56cceb48fbcb19e51e9f955459cff2e343a5887fad4d17eb142a0.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 149, + 290, + 227, + 350 + ], + "blocks": [ + { + "bbox": [ + 149, + 290, + 227, + 350 + ], + "lines": [ + { + "bbox": [ + 149, + 290, + 227, + 350 + ], + "spans": [ + { + "bbox": [ + 149, + 290, + 227, + 350 + ], + "type": "image", + "image_path": "5cc79ad1a3418afb90628108bbd37d04b82c455cad379098b54a1c683844ec18.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 230, + 294, + 302, + 345 + ], + "blocks": [ + { + "bbox": [ + 230, + 294, + 302, + 345 + ], + "lines": [ + { + "bbox": [ + 230, + 294, + 302, + 345 + ], + "spans": [ + { + "bbox": [ + 230, + 294, + 302, + 345 + ], + "type": "image", + "image_path": "15e5ef3661021822b1013b521dc01a98e033212ebacb07213fc7d9182cb23f7c.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 308, + 294, + 383, + 346 + ], + "blocks": [ + { + "bbox": [ + 308, + 294, + 383, + 346 + ], + "lines": [ + { + "bbox": [ + 308, + 294, + 383, + 346 + ], + "spans": [ + { + "bbox": [ + 308, + 294, + 383, + 346 + ], + "type": "image", + "image_path": "6db8146507af64f3ad0e265ce1b04282f92a486f53ede1b967eb929a0a72a9d0.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 387, + 297, + 460, + 345 + ], + "blocks": [ + { + "bbox": [ + 387, + 297, + 460, + 345 + ], + "lines": [ + { + "bbox": [ + 387, + 297, + 460, + 345 + ], + "spans": [ + { + "bbox": [ + 387, + 297, + 460, + 345 + ], + "type": "image", + "image_path": "94828630daa0a3789f33db9c02a70978ce6f2b4eab3a1c1da1a90f57e8435cee.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 149, + 365, + 227, + 425 + ], + "blocks": [ + { + "bbox": [ + 149, + 365, + 227, + 425 + ], + "lines": [ + { + "bbox": [ + 149, + 365, + 227, + 425 + ], + "spans": [ + { + "bbox": [ + 149, + 365, + 227, + 425 + ], + "type": "image", + "image_path": "10396fe5d4ea378ddc1107ded1ed79d07052a88a88841e2e2835d29180f65811.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 454, + 482, + 489 + ], + "lines": [ + { + "bbox": [ + 130, + 454, + 482, + 489 + ], + "spans": [ + { + "bbox": [ + 130, + 454, + 482, + 489 + ], + "type": "text", + "content": "Fig. 5: Qualitative Analysis on the Occ-ScanNet Dataset. The input image is displayed on the left, while the predicted scene is shown in the middle two columns, and ground truth on the right column." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 231, + 370, + 304, + 422 + ], + "blocks": [ + { + "bbox": [ + 231, + 370, + 304, + 422 + ], + "lines": [ + { + "bbox": [ + 231, + 370, + 304, + 422 + ], + "spans": [ + { + "bbox": [ + 231, + 370, + 304, + 422 + ], + "type": "image", + "image_path": "ac93ecc616302a419ebe3b5a45700ca5a7f49fc81925b8ef6119eae7527bce78.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 309, + 365, + 383, + 427 + ], + "blocks": [ + { + "bbox": [ + 309, + 365, + 383, + 427 + ], + "lines": [ + { + "bbox": [ + 309, + 365, + 383, + 427 + ], + "spans": [ + { + "bbox": [ + 309, + 365, + 383, + 427 + ], + "type": "image", + "image_path": "f0898202e01e8619bc49b2aaa306d734b49a00b17d953826e5404469616c0a16.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 387, + 365, + 460, + 426 + ], + "blocks": [ + { + "bbox": [ + 387, + 365, + 460, + 426 + ], + "lines": [ + { + "bbox": [ + 387, + 365, + 460, + 426 + ], + "spans": [ + { + "bbox": [ + 387, + 365, + 460, + 426 + ], + "type": "image", + "image_path": "30279eba2499b123db6f91a3fa0b0c4272933fffbf5a95082748f4553709b126.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "bbox": [ + 131, + 510, + 239, + 522 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 510, + 239, + 522 + ], + "spans": [ + { + "bbox": [ + 131, + 510, + 239, + 522 + ], + "type": "text", + "content": "5.3 Ablation Study" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 130, + 529, + 482, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 529, + 482, + 553 + ], + "spans": [ + { + "bbox": [ + 130, + 529, + 482, + 553 + ], + "type": "text", + "content": "We conduct ablation studies on the NYUv2 [29] and Occ-ScanNet-mini datasets to validate the effectiveness of our model design." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "content": "Depth fusion. Inspired by BEVDepth [20], we initially adopt the integration of BEV depth information, denoted as " + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "inline_equation", + "content": "^{+}" + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "content": "bev-depth'. However, we observed that while BEV depth works well for 3D detection in outdoor scenes, directly applying it to indoor scenes is less effective. We speculate that this discrepancy arises because height information is more crucial in indoor scenes, where indoor heights often correspond to different 3D structures and semantics. Therefore, we adopt voxel depth fusion, denoted as " + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "inline_equation", + "content": "^{+}" + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "content": "voxel-depth'. Specifically, we project each 3D voxel centroid to a frustum grid to sample the depth distribution probability" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 201, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 201, + 91, + 447, + 102 + ], + "type": "text", + "content": "Monocular Occupancy Prediction for Scalable Indoor Scenes" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 146, + 118, + 468, + 300 + ], + "blocks": [ + { + "bbox": [ + 146, + 118, + 468, + 300 + ], + "lines": [ + { + "bbox": [ + 146, + 118, + 468, + 300 + ], + "spans": [ + { + "bbox": [ + 146, + 118, + 468, + 300 + ], + "type": "image", + "image_path": "e1e790fceece663066ce4a35a679e75bee7a43e5396d4b9cfad477e8820e4e36.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 312, + 482, + 369 + ], + "lines": [ + { + "bbox": [ + 130, + 312, + 482, + 369 + ], + "spans": [ + { + "bbox": [ + 130, + 312, + 482, + 369 + ], + "type": "text", + "content": "Fig. 6: Qualitative Analysis on the NYUv2 Dataset: The input is displayed on the left, with the camera viewing frustum shown in every image. ISO adeptly captures the scene layout and delineates various semantic instances more accurately. Notably, it excels in reconstructing the corners of walls (row 2) and accurately representing small objects like pictures on the wall (row 3)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 393, + 480, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 393, + 480, + 429 + ], + "spans": [ + { + "bbox": [ + 130, + 393, + 480, + 429 + ], + "type": "text", + "content": "of the voxel. We then weight the 2D feature at that pixel coordinate as the voxel feature. The results in Tab. 3 further validate that the voxel depth fusion approach is more effective." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 430, + 482, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 430, + 482, + 525 + ], + "spans": [ + { + "bbox": [ + 130, + 430, + 482, + 525 + ], + "type": "text", + "content": "The BEV method usually performs better in detecting ceilings and floors but performs poorly in detecting furniture such as windows, walls, televisions, beds, sofas, and tables. Ceilings and floors usually have relatively flat and regular geometric shapes, which makes their projections clearer and more consistent from a BEV perspective. In contrast, objects such as windows, walls, and televisions typically have more complex geometric shapes. In the BEV perspective, the projections of these objects may become blurry or overlapping, making it difficult to accurately distinguish and detect." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 188, + 570, + 427, + 640 + ], + "blocks": [ + { + "bbox": [ + 132, + 549, + 481, + 561 + ], + "lines": [ + { + "bbox": [ + 132, + 549, + 481, + 561 + ], + "spans": [ + { + "bbox": [ + 132, + 549, + 481, + 561 + ], + "type": "text", + "content": "Table 3: Ablation study of depth fusion on the NYUv2 and Occ-ScanNet-mini dataset" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 188, + 570, + 427, + 640 + ], + "lines": [ + { + "bbox": [ + 188, + 570, + 427, + 640 + ], + "spans": [ + { + "bbox": [ + 188, + 570, + 427, + 640 + ], + "type": "table", + "html": "
MethodInputNYUv2Occ-ScanNet-mini
IoUmIoUIoUmIoU
baselinexrgb42.2727.1350.9438.95
+ bev-depthxrgb42.6727.1451.5838.48
+ voxel-depthxrgb47.1131.2551.0339.08
", + "image_path": "a0a0d6e1b1e3bccd7e8972fb9e8506736b3a0886d6898eb0af319b27a83b2b27.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "type": "text", + "content": "H. Yu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 208, + 135, + 404, + 224 + ], + "blocks": [ + { + "bbox": [ + 177, + 114, + 435, + 125 + ], + "lines": [ + { + "bbox": [ + 177, + 114, + 435, + 125 + ], + "spans": [ + { + "bbox": [ + 177, + 114, + 435, + 125 + ], + "type": "text", + "content": "Table 4: Ablation of depth information on the NYUv2 dataset" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 208, + 135, + 404, + 224 + ], + "lines": [ + { + "bbox": [ + 208, + 135, + 404, + 224 + ], + "spans": [ + { + "bbox": [ + 208, + 135, + 404, + 224 + ], + "type": "table", + "html": "
depth-methodlearnedmulti-scaleIoUmIoU
GT✓53.9834.47
ZoeDepth✓45.1829.15
ZoeDepth✓✓45.2429.40
DepthAnything44.7729.28
DepthAnything✓45.4829.57
DepthAnything✓✓46.9431.02
DepthAnything✓47.1131.25
", + "image_path": "9bbba87470286f0793b8ea3a417066382d7328d93c2cb7ab62af7c0492c2a150.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 250, + 481, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 250, + 481, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 250, + 481, + 346 + ], + "type": "text", + "content": "Different depth pre-trained models. We conducted ablation experiments on the depth method using the Occ-ScanNet-mini dataset. Initially, we observed the upper bound of model performance when ground truth (GT) depth information was provided. Subsequently, we initialized our depth module using ZoeDepth [3] and Depth-Anything [44]. Experimental results indicate that utilizing Depth-Anything [44] yields better performance. Moreover, our approach allows for further enhancement by fine-tuning a pre-trained depth module, resulting in improved performance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 376, + 481, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 376, + 481, + 567 + ], + "spans": [ + { + "bbox": [ + 132, + 376, + 481, + 567 + ], + "type": "text", + "content": "Multi-scale depth. In Tab. 4, we investigate the impact of using multi-scale depth distribution on the NYUv2 [29] dataset. Specifically, the multi-scale depth distribution is used to weight the multi-scale 3D voxel feature. Meanwhile, we tried a method that only uses the single-scale depth distribution map to generate the 3D voxel depth matrix. The used depth distribution map has a scale of " + }, + { + "bbox": [ + 132, + 376, + 481, + 567 + ], + "type": "inline_equation", + "content": "\\frac{1}{8}" + }, + { + "bbox": [ + 132, + 376, + 481, + 567 + ], + "type": "text", + "content": " original image scale. Only this single-scale 3D voxel matrix is used to weigh the summed-up 3D voxel feature out of the FLoSP module. Without multi-scale depth, all 3D features are simply added up and a single depth distribution is used for weighting. This method ignores the different importance that features at different scales may have, as all features at all scales are treated equally and all depth information is weighted by a unified depth distribution. With multiscale depth, the 3D features of each scale are first weighted with the depth distribution of their corresponding scale, and then these weighted features are added up. This method takes into account the differences in the importance of features at different scales, as each scale's feature is weighted based on its corresponding depth distribution." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 570, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 570, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 570, + 481, + 665 + ], + "type": "text", + "content": "In real-world scenarios, various objects and structures (such as ceilings, walls, furniture, etc.) often exhibit different scales. Hence, accounting for scale variations is crucial for accurately modeling these objects. Multi-scale depth can effectively capture these scale changes by assigning different weights to features at each scale. Depth information is indispensable for 3D scene completion. Leveraging the depth distribution information at each scale allows us to weight the corresponding 3D features, enabling the model to better understand and represent the objects in the scene and their relationships." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 201, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 201, + 91, + 447, + 102 + ], + "type": "text", + "content": "Monocular Occupancy Prediction for Scalable Indoor Scenes" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 115, + 214, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 115, + 214, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 115, + 214, + 127 + ], + "type": "text", + "content": "5.4 Discussion" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 136, + 482, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 136, + 482, + 196 + ], + "spans": [ + { + "bbox": [ + 130, + 136, + 482, + 196 + ], + "type": "text", + "content": "Data scaling up. In Tab. 5, we delve into the impact of data volume on our method's performance by scaling up the number of scene samples. We trained the model on Occ-ScanNet using " + }, + { + "bbox": [ + 130, + 136, + 482, + 196 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 130, + 136, + 482, + 196 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 136, + 482, + 196 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 130, + 136, + 482, + 196 + ], + "type": "text", + "content": " scene samples respectively, and tested the model performance. The results indicate that larger dataset exhibits more significant gains." + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 231, + 239, + 379, + 297 + ], + "blocks": [ + { + "bbox": [ + 197, + 217, + 416, + 229 + ], + "lines": [ + { + "bbox": [ + 197, + 217, + 416, + 229 + ], + "spans": [ + { + "bbox": [ + 197, + 217, + 416, + 229 + ], + "type": "text", + "content": "Table 5: Scaling up comparison on the Occ-ScanNet" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 231, + 239, + 379, + 297 + ], + "lines": [ + { + "bbox": [ + 231, + 239, + 379, + 297 + ], + "spans": [ + { + "bbox": [ + 231, + 239, + 379, + 297 + ], + "type": "table", + "html": "
DataDepth BranchOurs
IoUmIoU
10%✓21.799.57
100%✓42.1628.71
", + "image_path": "9081432db3b01e93e94c944b68deffaf5eb915ce6e250fce8593cf994bb1c78f.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 336, + 482, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 336, + 482, + 433 + ], + "spans": [ + { + "bbox": [ + 130, + 336, + 482, + 433 + ], + "type": "text", + "content": "Limitations and future works. Although our model can estimate 3D occupancy effectively with the assistance of depth information, semantic learning still faces challenges due to class imbalance issues. Furthermore, our proposed Occ-ScanNet only considers 11 common semantic classes, which may not fully capture the diversity of categories present in real-world scenarios. Despite the significant increase in data volume compared to previous datasets, Occ-ScanNet remains limited in the number of scenes it covers. In future work, we will focus more on semantic exploration and conduct tests in more generalized scenarios." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 451, + 220, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 451, + 220, + 464 + ], + "spans": [ + { + "bbox": [ + 132, + 451, + 220, + 464 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 476, + 482, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 476, + 482, + 574 + ], + "spans": [ + { + "bbox": [ + 130, + 476, + 482, + 574 + ], + "type": "text", + "content": "In this paper, we introduce ISO, a novel method for monocular 3D occupancy prediction. ISO utilizes a D-FLoSP (Dual Feature Line of Sight Projection) and a multi-scale feature fusion strategy to address variations in scene and object sizes, specifically tailored for indoor scenes. Additionally, we present a new benchmark, Occ-ScanNet, aimed at fostering scalable studies of indoor scenes. We hope that this dataset will attract more attention to research on indoor occupancy prediction. While achieving promising results in 3D structure prediction, accurately identifying semantics remains a significant challenge in the future." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 591, + 241, + 605 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 591, + 241, + 605 + ], + "spans": [ + { + "bbox": [ + 132, + 591, + 241, + 605 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "This work was supported in part by the National Key R&D Program of China (No. 2022ZD0160102), the National Natural Science Foundation of China (No. U21B2042, No. 62320106010), and in part by the 2035 Innovation Program of CAS, and the InnoHK program." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "type": "text", + "content": "H. Yu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 141, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 138, + 141, + 480, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 141, + 480, + 163 + ], + "spans": [ + { + "bbox": [ + 138, + 141, + 480, + 163 + ], + "type": "text", + "content": "1. Arshad, M.S., Beksi, W.J.: List: Learning implicitly from spatial transformers for single-view 3d reconstruction. In: ICCV (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 165, + 480, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 165, + 480, + 186 + ], + "spans": [ + { + "bbox": [ + 138, + 165, + 480, + 186 + ], + "type": "text", + "content": "2. Azuma, R.T.: A survey of augmented reality. Presence: teleoperators & virtual environments (1997)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 186, + 480, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 186, + 480, + 208 + ], + "spans": [ + { + "bbox": [ + 138, + 186, + 480, + 208 + ], + "type": "text", + "content": "3. Bhat, S.F., Birkl, R., Wofk, D., Wonka, P., Müller, M.: Zoedepth: Zero-shot transfer by combining relative and metric depth. arXiv preprint arXiv:2302.12288 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 209, + 480, + 231 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 209, + 480, + 231 + ], + "spans": [ + { + "bbox": [ + 138, + 209, + 480, + 231 + ], + "type": "text", + "content": "4. Birkl, R., Wofk, D., Müller, M.: Midas v3. 1-a model zoo for robust monocular relative depth estimation. arXiv preprint arXiv:2307.14460 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 232, + 480, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 232, + 480, + 253 + ], + "spans": [ + { + "bbox": [ + 138, + 232, + 480, + 253 + ], + "type": "text", + "content": "5. Cao, A.Q., de Charette, R.: Monoscene: Monocular 3d semantic scene completion. In: CVPR (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 254, + 480, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 254, + 480, + 275 + ], + "spans": [ + { + "bbox": [ + 138, + 254, + 480, + 275 + ], + "type": "text", + "content": "6. Chen, X., Lin, K.Y., Qian, C., Zeng, G., Li, H.: 3d sketch-aware semantic scene completion via semi-supervised structure prior. In: CVPR (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 276, + 480, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 276, + 480, + 297 + ], + "spans": [ + { + "bbox": [ + 138, + 276, + 480, + 297 + ], + "type": "text", + "content": "7. Dahnert, M., Hou, J., Nießner, M., Dai, A.: Panoptic 3d scene reconstruction from a single rgb image. NeurIPS (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 299, + 480, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 299, + 480, + 319 + ], + "spans": [ + { + "bbox": [ + 138, + 299, + 480, + 319 + ], + "type": "text", + "content": "8. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: Scannet: Richly-annotated 3d reconstructions of indoor scenes. In: CVPR (2017)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 321, + 480, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 321, + 480, + 342 + ], + "spans": [ + { + "bbox": [ + 138, + 321, + 480, + 342 + ], + "type": "text", + "content": "9. Denninger, M., Triebel, R.: 3d scene reconstruction from a single viewpoint. In: ECCV (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 343, + 480, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 343, + 480, + 364 + ], + "spans": [ + { + "bbox": [ + 138, + 343, + 480, + 364 + ], + "type": "text", + "content": "10. DeSouza, G.N., Kak, A.C.: Vision for mobile robot navigation: A survey. TPAMI (2002)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 365, + 480, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 365, + 480, + 387 + ], + "spans": [ + { + "bbox": [ + 138, + 365, + 480, + 387 + ], + "type": "text", + "content": "1. Fan, H., Su, H., Guibas, L.J.: A point set generation network for 3d object reconstruction from a single image. In: CVPR (2017)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 388, + 480, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 388, + 480, + 420 + ], + "spans": [ + { + "bbox": [ + 138, + 388, + 480, + 420 + ], + "type": "text", + "content": "2. Fieraru, M., Zanfir, M., Oneata, E., Popa, A.I., Olaru, V., Sminchisescu, C.: Reconstructing three-dimensional models of interacting humans. arXiv preprint arXiv:2308.01854 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 421, + 480, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 421, + 480, + 453 + ], + "spans": [ + { + "bbox": [ + 138, + 421, + 480, + 453 + ], + "type": "text", + "content": "3. Goel, S., Pavlakos, G., Rajasegaran, J., Kanazawa, A., Malik, J.: Humans in 4d: Reconstructing and tracking humans with transformers. arXiv preprint arXiv:2305.20091 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 455, + 480, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 455, + 480, + 475 + ], + "spans": [ + { + "bbox": [ + 138, + 455, + 480, + 475 + ], + "type": "text", + "content": "4. Guan, H., Song, C., Zhang, Z.: Gramo: geometric resampling augmentation for monocular 3d object detection. Frontiers of Computer Science (2024)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 476, + 480, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 476, + 480, + 509 + ], + "spans": [ + { + "bbox": [ + 138, + 476, + 480, + 509 + ], + "type": "text", + "content": "5. Huang, J., Huang, G., Zhu, Z., Ye, Y., Du, D.: Bevdet: High-performance multicamera 3d object detection in bird-eye-view. arXiv preprint arXiv:2112.11790 (2021)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 510, + 480, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 510, + 480, + 532 + ], + "spans": [ + { + "bbox": [ + 138, + 510, + 480, + 532 + ], + "type": "text", + "content": "6. Huang, Y., Zheng, W., Zhang, B., Zhou, J., Lu, J.: Selfocc: Self-supervised vision-based 3d occupancy prediction. In: CVPR (2024)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 533, + 480, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 533, + 480, + 553 + ], + "spans": [ + { + "bbox": [ + 138, + 533, + 480, + 553 + ], + "type": "text", + "content": "7. Huang, Y., Zheng, W., Zhang, Y., Zhou, J., Lu, J.: Tri-perspective view for vision-based 3d semantic occupancy prediction. In: CVPR (2023)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 555, + 480, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 555, + 480, + 576 + ], + "spans": [ + { + "bbox": [ + 138, + 555, + 480, + 576 + ], + "type": "text", + "content": "8. Li, J., Han, K., Wang, P., Liu, Y., Yuan, X.: Anisotropic convolutional networks for 3d semantic scene completion. In: CVPR (2020)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 138, + 577, + 480, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 577, + 480, + 609 + ], + "spans": [ + { + "bbox": [ + 138, + 577, + 480, + 609 + ], + "type": "text", + "content": "9. Li, Y., Yu, Z., Choy, C., Xiao, C., Alvarez, J.M., Fidler, S., Feng, C., Anandkumar, A.: Voxformer: Sparse voxel transformer for camera-based 3d semantic scene completion. In: CVPR (2023)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 138, + 610, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 610, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 610, + 480, + 632 + ], + "type": "text", + "content": "20. Li, Y., Ge, Z., Yu, G., Yang, J., Wang, Z., Shi, Y., Sun, J., Li, Z.: Bevdepth: Acquisition of reliable depth for multi-view 3d object detection. In: AAAI (2023)" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "type": "text", + "content": "21. Li, Z., Wang, W., Li, H., Xie, E., Sima, C., Lu, T., Qiao, Y., Dai, J.: Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In: ECCV (2022)" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 201, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 201, + 91, + 447, + 102 + ], + "type": "text", + "content": "Monocular Occupancy Prediction for Scalable Indoor Scenes" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "text", + "content": "22. Li, Z., Yu, Z., Austin, D., Fang, M., Lan, S., Kautz, J., Alvarez, J.M.: Fb-occ: 3d occupancy prediction based on forward-backward view transformation. arXiv preprint arXiv:2307.01492 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 480, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 480, + 171 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 480, + 171 + ], + "type": "text", + "content": "23. Liu, S., Hu, Y., Zeng, Y., Tang, Q., Jin, B., Han, Y., Li, X.: See and think: Disentangling semantic scene completion. NeurIPS (2018)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 480, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 480, + 193 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 480, + 193 + ], + "type": "text", + "content": "24. Murez, Z., Van As, T., Bartolozzi, J., Sinha, A., Badrinarayanan, V., Rabinovich, A.: Atlas: End-to-end 3d scene reconstruction from posed images. In: ECCV (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 194, + 480, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 480, + 225 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 480, + 225 + ], + "type": "text", + "content": "25. Pan, M., Liu, J., Zhang, R., Huang, P., Li, X., Liu, L., Zhang, S.: Renderocc: Visioncentric 3d occupancy prediction with 2d rendering supervision. arXiv preprint arXiv:2309.09502 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 227, + 480, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 480, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 480, + 248 + ], + "type": "text", + "content": "26. Park, J.J., Florence, P., Straub, J., Newcombe, R., Lovegrove, S.: Deepsdf: Learning continuous signed distance functions for shape representation. In: CVPR (2019)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 249, + 480, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 480, + 281 + ], + "type": "text", + "content": "27. Ranftl, R., Lasinger, K., Hafner, D., Schindler, K., Koltun, V.: Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. TPAMI (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 282, + 480, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 282, + 480, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 282, + 480, + 303 + ], + "type": "text", + "content": "28. Roldao, L., de Charette, R., Verroust-Blondet, A.: Lmscnet: Lightweight multiscale 3d semantic completion. In: 3DV (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 304, + 480, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 480, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 480, + 325 + ], + "type": "text", + "content": "29. Silberman, N., Hoiem, D., Kohli, P., Fergus, R.: Indoor segmentation and support inference from rgbd images. In: ECCV (2012)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 326, + 480, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 480, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 480, + 346 + ], + "type": "text", + "content": "30. Song, S., Yu, F., Zeng, A., Chang, A.X., Savva, M., Funkhouser, T.: Semantic scene completion from a single depth image. In: CVPR (2017)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 348, + 480, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 348, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 348, + 480, + 369 + ], + "type": "text", + "content": "31. Tan, M., Le, Q.: Efficientnet: Rethinking model scaling for convolutional neural networks. In: ICML (2019)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 369, + 480, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 480, + 390 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 480, + 390 + ], + "type": "text", + "content": "32. Tang, Y., Dorn, S., Savani, C.: Center3d: Center-based monocular 3d object detection with joint depth understanding. In: DAGM GCPR (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 392, + 480, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 480, + 423 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 480, + 423 + ], + "type": "text", + "content": "33. Tian, X., Jiang, T., Yun, L., Mao, Y., Yang, H., Wang, Y., Wang, Y., Zhao, H.: Occ3d: A large-scale 3d occupancy prediction benchmark for autonomous driving. NeurIPS (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 424, + 480, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 424, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 424, + 480, + 445 + ], + "type": "text", + "content": "34. Tong, W., Sima, C., Wang, T., Chen, L., Wu, S., Deng, H., Gu, Y., Lu, L., Luo, P., Lin, D., et al.: Scene as occupancy. In: ICCV (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 447, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 447, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 447, + 480, + 468 + ], + "type": "text", + "content": "35. Wang, S., Liu, Y., Wang, T., Li, Y., Zhang, X.: Exploring object-centric temporal modeling for efficient multi-view 3d object detection. In: ICCV (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 468, + 480, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 480, + 500 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 480, + 500 + ], + "type": "text", + "content": "36. Wang, X., Zhu, Z., Xu, W., Zhang, Y., Wei, Y., Chi, X., Ye, Y., Du, D., Lu, J., Wang, X.: Openoccupancy: A large scale benchmark for surrounding semantic occupancy perception. In: ICCV (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 501, + 480, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 480, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 480, + 522 + ], + "type": "text", + "content": "37. Wang, Y., Guizilini, V.C., Zhang, T., Wang, Y., Zhao, H., Solomon, J.: Detr3d: 3d object detection from multi-view images via 3d-to-2d queries. In: CoRL (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 523, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 523, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 523, + 480, + 544 + ], + "type": "text", + "content": "38. Wang, Y., Chen, Y., Liao, X., Fan, L., Zhang, Z.: Panoocc: Unified occupancy representation for camera-based 3d panoptic segmentation. In: CVPR (2024)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 545, + 480, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 545, + 480, + 566 + ], + "spans": [ + { + "bbox": [ + 132, + 545, + 480, + 566 + ], + "type": "text", + "content": "39. Wang, Y., Chen, Y., Zhang, Z.: Frustumformer: Adaptive instance-aware resampling for multi-view 3d detection. In: CVPR (2023)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 567, + 480, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 480, + 588 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 480, + 588 + ], + "type": "text", + "content": "40. Wei, Y., Zhao, L., Zheng, W., Zhu, Z., Zhou, J., Lu, J.: Surroundocc: Multi-camera 3d occupancy prediction for autonomous driving. In: ICCV (2023)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 589, + 480, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 589, + 480, + 610 + ], + "spans": [ + { + "bbox": [ + 132, + 589, + 480, + 610 + ], + "type": "text", + "content": "41. Wu, Q., Wang, K., Li, K., Zheng, J., Cai, J.: Objectsdf++: Improved object-compositional neural implicit surfaces. In: ICCV (2023)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 611, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 480, + 632 + ], + "type": "text", + "content": "42. Wu, S.C., Tateno, K., Navab, N., Tombari, F.: Scfusion: Real-time incremental scene reconstruction with semantic completion. In: 3DV (2020)" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "text", + "content": "43. Yang, C., Chen, Y., Tian, H., Tao, C., Zhu, X., Zhang, Z., Huang, G., Li, H., Qiao, Y., Lu, L., et al.: Bevformer v2: Adapting modern image backbones to bird's-eye-view recognition via perspective supervision. In: CVPR (2023)" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "type": "text", + "content": "H. Yu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 314 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "text", + "content": "44. Yang, L., Kang, B., Huang, Z., Xu, X., Feng, J., Zhao, H.: Depth anything: Unleashing the power of large-scale unlabeled data. In: CVPR (2024)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 138, + 480, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 138, + 480, + 171 + ], + "spans": [ + { + "bbox": [ + 132, + 138, + 480, + 171 + ], + "type": "text", + "content": "45. Yao, J., Li, C., Sun, K., Cai, Y., Li, H., Ouyang, W., Li, H.: Ndc-scene: Boost monocular 3d semantic scene completion in normalized device coordinates space. In: ICCV (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 480, + 205 + ], + "type": "text", + "content": "46. Yu, Z., Shu, C., Deng, J., Lu, K., Liu, Z., Yu, J., Yang, D., Li, H., Chen, Y.: Flashocc: Fast and memory-efficient occupancy prediction via channel-to-height plugin. arXiv preprint arXiv:2311.12058 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 205, + 480, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 480, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 480, + 226 + ], + "type": "text", + "content": "47. Yurtsever, E., Lambert, J., Carballo, A., Takeda, K.: A survey of autonomous driving: Common practices and emerging technologies. IEEE access (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 226, + 480, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 226, + 480, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 226, + 480, + 248 + ], + "type": "text", + "content": "48. Zhang, J., Zhao, H., Yao, A., Chen, Y., Zhang, L., Liao, H.: Efficient semantic scene completion network with spatial group convolution. In: ECCV (2018)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 248, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 248, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 248, + 480, + 270 + ], + "type": "text", + "content": "49. Zhang, X., Bi, S., Sunkavalli, K., Su, H., Xu, Z.: Nerfusion: Fusing radiance fields for large-scale scene reconstruction. In: CVPR (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 270, + 480, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 270, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 270, + 480, + 293 + ], + "type": "text", + "content": "50. Zheng, Z., Yu, T., Wei, Y., Dai, Q., Liu, Y.: Deephuman: 3d human reconstruction from a single image. In: ICCV (2019)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 293, + 480, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 480, + 314 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 480, + 314 + ], + "type": "text", + "content": "51. Zhong, M., Zeng, G.: Semantic point completion network for 3d semantic scene completion. In: ECAI (2020)" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 201, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 201, + 91, + 447, + 102 + ], + "type": "text", + "content": "Monocular Occupancy Prediction for Scalable Indoor Scenes" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Motion Aware Event Representation-driven Image Deblurring/a9105a02-b537-456a-9778-80b745c423ab_content_list.json b/2024/Motion Aware Event Representation-driven Image Deblurring/a9105a02-b537-456a-9778-80b745c423ab_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..5ee1847634084604f51dbec65e7c72d80d086bb2 --- /dev/null +++ b/2024/Motion Aware Event Representation-driven Image Deblurring/a9105a02-b537-456a-9778-80b745c423ab_content_list.json @@ -0,0 +1,2161 @@ +[ + { + "type": "text", + "text": "Motion Aware Event Representation-driven Image Deblurring", + "text_level": 1, + "bbox": [ + 254, + 140, + 750, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhijing Sun, Xueyang Fu, Longzhuo Huang, Aiping Liu, and Zheng-Jun Zha*", + "bbox": [ + 223, + 212, + 777, + 228 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "School of Information Science and Technology and MoE Key Laboratory of Brain-inspired Intelligent Perception and Cognition, University of Science and Technology of China, Hefei, 230026, China {sunzhijing, hlz_0}@mail.ustc.edu.cn {xyfu, aipingl, zhazj}@ustc.edu.cn", + "bbox": [ + 250, + 239, + 753, + 310 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Traditional image deblurring struggles with high-quality reconstruction due to limited motion data from single blurred images. Excitingly, the high-temporal resolution of event cameras records motion more precisely in a different modality, transforming image deblurring. However, many event camera-based methods, which only care about the final value of the polarity accumulation, ignore the influence of the absolute intensity change where events generate so fall short in perceiving motion patterns and effectively aiding image reconstruction. To overcome this, in this work, we propose a new event preprocessing technique that accumulates the deviation from the initial moment each time the event is updated. This process can distinguish the order of events to improve the perception of object motion patterns. To complement our proposed event representation, we create a recurrent module designed to meticulously extract motion features across local and global time scales. To further facilitate the event feature and image feature integration, which assists in image reconstruction, we develop a bi-directional feature alignment and fusion module. This module works to lessen inter-modal inconsistencies. Our approach has been thoroughly tested through rigorous experiments carried out on several datasets with different distributions. These trials have delivered promising results, with our method achieving top-tier performance in both quantitative and qualitative assessments. Code is available at https://github.com/ZhijingS/DA_event_deblur.", + "bbox": [ + 261, + 348, + 738, + 654 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Image deblurring $\\cdot$ Event camera $\\cdot$ Event representation", + "bbox": [ + 261, + 666, + 723, + 681 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 217, + 708, + 374, + 723 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Frame-based cameras often produce blurred images due to factors like camera shake or object motion, given their configured exposure time. Such blur, a common form of image degradation, occurs across various scenes. While photographers typically aim for clear and focused photos, clear inputs are equally critical for numerous computer vision tasks. These tasks include super-resolution, target", + "bbox": [ + 212, + 739, + 787, + 816 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author", + "bbox": [ + 232, + 825, + 388, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5bfd42e747bf5dbc6e4d8cc6284f047ed99a601db36a9a27c56293f41bf9dedb.jpg", + "image_caption": [ + "(a) GT (PSNR |SSIM)" + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 312, + 215 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/17adc58bc8c7075408d48df65f7dadc8389c5148ab985309ad083c9396351c0f.jpg", + "image_caption": [ + "(b) Blurry (28.62dB |0.819)" + ], + "image_footnote": [], + "bbox": [ + 313, + 143, + 406, + 215 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/9296ea44d1b5bdf6a1adb5c182b71bd016571a194d8379b57fef5dee10f83353.jpg", + "image_caption": [ + "(c) HINet [4] (29.14dB |0.837)", + "Fig. 1: A deblurring example on HS-ERGB [37] dataset via recent state-of-the-art methods. Compared to other methods, our approach yields a more distinct result owing to its enhanced motion perception capability." + ], + "image_footnote": [], + "bbox": [ + 408, + 143, + 500, + 215 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/13044559a5ade1896c43bb1813ecb568a3239ac67268cd28c59ff07c03370df3.jpg", + "image_caption": [ + "(d) NAFNet [3] (30.41dB |0.855)" + ], + "image_footnote": [], + "bbox": [ + 501, + 143, + 596, + 215 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/473f75e245093e693cdddd4d784fb33482225b41569de5edb3f7f85fa30dcde8.jpg", + "image_caption": [ + "(e) EFNet [33] (30.76dB |0.856)" + ], + "image_footnote": [], + "bbox": [ + 598, + 143, + 691, + 215 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/0ad319576978f72a3e13af2d0e95f04fc6e4cb45953156a234d058cded602fea.jpg", + "image_caption": [ + "(f) Ours (31.01dB [0.858]" + ], + "image_footnote": [], + "bbox": [ + 694, + 143, + 785, + 215 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "detection, and semantic segmentation, to name a few. Hence, image deblurring serves as a fundamental task in the realm of low-level computer vision, reinforcing the image quality and supporting the proper function of various algorithms.", + "bbox": [ + 212, + 337, + 784, + 383 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Traditional image-based techniques often attempt to derive a clear image from a single blurred one. They do this by leveraging natural image priors or making assumptions about blur operations [6, 8, 13, 15, 34, 44]. However, in capturing a single blurred image, much motion information is lost. This lack of motion information during exposure creates undesirable artifacts in the deblurred results, especially in complex motion scenarios, as illustrated in Figures 1c and 1d. To counter this challenge, recent methodologies use more temporal information, such as differently exposed frames [2, 46, 48] and neighboring frames [16, 28, 32]. However, these approaches have their limitations. For instance, the use of differently exposed frames is constrained by the low temporal resolution of frame-based cameras, improving the reconstruction effect primarily in slow-motion scenes. The method that uses adjacent frames to estimate the current frame's motion information can lead to accumulative errors. In summary, traditional deblurring techniques encounter hurdles due to the inability of frame-based cameras to capture motion information during exposure.", + "bbox": [ + 212, + 385, + 787, + 611 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As opposed to frame-based cameras, event cameras, inspired by biological systems, offer a promising solution to this issue. These cameras' microsecond-level temporal resolution $(\\mu \\mathrm{s})$ and asynchronous output architecture allow the recording of fast motions. This capability enables event streams to store pixel intensity changes and provide accurate motion information to deblur the corresponding frame. By using event frames as additional input, some methods attempt to guide the deblurring process through these event cameras [9,25,39,43,49]. One method [45] subdivides the event stream into ternary 2D data and generates a mask to localize high-blur regions, thereby enhancing the network's ability to distinguish different blur levels. Another method [33] symmetrically accumulates event polarities and fuses event and image features using an attention mechanism, enabling the integration of motion information into image reconstruction. However, these methods primarily employ the event stream in a manner akin to polarity accumulation. In essence, they consider only the final accumulation result of relative intensity change between subsequent moments, neglecting the ab", + "bbox": [ + 212, + 613, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Z. Sun et al.", + "bbox": [ + 271, + 114, + 357, + 127 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "solute intensity change caused by historical events. Consequently, this approach fails to differentiate between various motion patterns, injecting unnecessary ambiguity into the deblurring process, and leading to edge artifacts as displayed in Figure 1e.", + "bbox": [ + 212, + 146, + 782, + 205 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given the limited perception of motion during exposure offered by polarity accumulation, we posit that the deviation from the initial moment provides more precise motion information for image deblurring. Each time the event generates, the absolute light intensity changes. The deviation from the initial moment enables us to consider the impact of changes in absolute light intensity, even when the absolute light intensity at the initial moment is unknown. Thus, the deviation, to a degree, implies the order of arrival of the events, offering different information for varying motion patterns. As a result, we propose the Deviation Accumulation (DA) method to enhance motion perception. Specifically, every time a new event arrives, our method accumulates the deviation relative to the initial moment, thus encompassing more comprehensive motion information. Furthermore, we find that longer time scales for event accumulation can capture global motion patterns, while shorter scales precisely depict local fast-motion patterns. To better perceive motion at both global and local time scales, we have designed a Recurrent Motion Extraction (RME) module that works with corresponding event representations. Moreover, in line with multimodal learning [18-20, 29, 36, 42], we present an effective Feature Alignment and Fusion (FAF) module. This module serves to address the issue that simple fusion methods struggle with the disparate nature of event and frame modalities. As a result, these two modalities align, ensuring better information interaction. We integrate these modules into a single image deblurring network. This network extracts the motion features within the exposure time from the event stream and combines them with the texture information supplied by the image to complete the motion blur removal, as illustrated in Figure 1f.", + "bbox": [ + 212, + 207, + 787, + 568 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The key contributions of this paper are as follows:", + "bbox": [ + 238, + 569, + 601, + 583 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce an event preprocessing method designed specifically for the deblurring problem. The deviation accumulation is capable of responding distinctively to various motion patterns, thereby adding accurate motion information for image deblurring tasks.", + "- We develop a network equipped with recurrent motion extraction modules, designed to better extract motion features on local and global time scales. This design enables the network to handle multiple composite motion scenarios effectively.", + "- We propose a feature alignment and fusion module to mitigate the impacts of inter-modal inconsistencies and ensure that motion information more effectively guides the image deblurring process." + ], + "bbox": [ + 227, + 592, + 785, + 756 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our experimental results indicate that our method achieves state-of-the-art performance on synthetic, semi-synthetic, and real-world datasets, including a 0.61dB improvement in PSNR on the GoPro dataset compared to previous studies [33]. Moreover, our method produces more distinct subjective results, as demonstrated in Figure 1.", + "bbox": [ + 212, + 763, + 787, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Motion Aware Event Representation-driven Image Deblurring", + "bbox": [ + 321, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 143, + 388, + 160 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.1 Event-based Motion Deblurring", + "text_level": 1, + "bbox": [ + 215, + 176, + 526, + 193 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The utilization of event information to assist in deblurring tasks has become increasingly common in recent research [9, 12, 25, 31, 39, 43, 49]. The high temporal resolution and low latency of event cameras not only capture edge positions but also embed moving temporal information within the event stream. This information is crucial to reconstruct the latent sharp image. Many methods have been attempted to explore the relationship between event information, blurred images, and sharp images. One such method [25] includes the Event-based Double Integral (EDI) model, which leverages the principles of event generation to formulate an optimization problem that establishes a correlation between a blurry image and a latent sharp image. In recent studies, the event stream is typically converted into event frames [30, 38], which simplifies the processing of event data within traditional CNN frameworks. [9] combines the motion feature extracted from event data with the image feature, using them as the input for the decoder. [50] employs a Multi-Layer Perceptron (MLP) to predict the fused feature value from event and image features. [33] integrates them through a cross-modal attention method, which generates a query matrix from the image features, but derives key and value matrices from the event features.", + "bbox": [ + 212, + 202, + 787, + 458 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Confronting the prevalent issue of integrating image and event features is the problem of inter-modal inconsistencies, which traditional fusion techniques cannot easily mitigate. To address this, our approach seeks to align these two modalities, optimizing the use of event data. Through this alignment, our goal is to minimize the inconsistencies spawned when merging differing modalities, thus amplifying the efficacy of the deblurring process.", + "bbox": [ + 212, + 459, + 789, + 550 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Event Representation", + "text_level": 1, + "bbox": [ + 215, + 571, + 441, + 588 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Event data, characterized by its spatial sparsity and strong temporal continuity due to its asynchronous generation, necessitates a custom-designed event representation to deliver the required information. Methods [11, 26, 41] that process events individually maximize the use of high temporal resolution, as suggested by [26], which proposed an adaptive spiking neuron model to manage the fluctuating event input. The effectiveness of these algorithms, however, can be significantly hampered by the noise within the event due to independent processing. A Time Surface [1, 14, 22, 51] is created using the timestamp of the last event at each pixel, which prominently reveals the rich temporal information contained within the event. Furthermore, the event frame [5, 7, 21, 23, 45] transforms the original 4D event data into a 2D frame by accumulating events on the number or the polarity pixel-wise, facilitating extraction with conventional image processing techniques. In [45], the authors utilize ternary data for their event frames to allocate high blur region of the image.", + "bbox": [ + 212, + 598, + 787, + 809 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "However, existing event frame approaches result in abundant light intensity change information being lost in the squeezing process, making it impossible to", + "bbox": [ + 215, + 809, + 787, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Z. Sun et al.", + "bbox": [ + 271, + 114, + 357, + 127 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "distinguish the order of event arrivals and lacking the ability to detect multiple motion patterns. Some techniques attempt to recover some of this lost data by dividing the exposure time into smaller segments, though this essentially equates to sampling the original data. Consequently, we design a motion-aware event representation specifically tailored for motion deblurring.", + "bbox": [ + 212, + 146, + 787, + 223 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 250, + 330, + 265 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Overall Framework", + "text_level": 1, + "bbox": [ + 215, + 284, + 419, + 299 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The overall framework of our proposed method is depicted in Figure 2. We adopt a classical encoder-decoder architecture for our approach. Initially, the blurred image is inputted into the image branch, where we use the baseline conv block mentioned in [3] as our fundamental block, to extract relevant features. Simultaneously, the corresponding event frames generated by our deviation accumulation are fed into the Recurrent Motion Extraction (RME) module, which facilitates the extraction of features covering motion in local and global time scales. Following each block of the two branches, the Feature Alignment and Fusion (FAF) module is employed to eliminate the inconsistencies between the two modalities. By performing down-sampling operations, we are able to obtain both shallow and semantic information. In the decoder, skip connections are utilized to incorporate the extracted features for image reconstruction. The architecture of the conv block and the decoder is shown in the supplementary materials. The entire training process is conducted in an end-to-end manner, with the PSNR Loss function [4] employed to optimize the network's parameters.", + "bbox": [ + 212, + 313, + 787, + 541 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Deviation Accumulation Event Representation", + "text_level": 1, + "bbox": [ + 215, + 566, + 645, + 580 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Event cameras are bio-inspired sensors which asynchronously output signals. They trigger an event while the logarithm of the intensity change exceeds the preset threshold $c$ at any pixel, which is formulated as", + "bbox": [ + 212, + 595, + 787, + 641 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\np = \\left\\{ \\begin{array}{l l} + 1 & \\text {i f} \\log \\left(I _ {x y} (t)\\right) - \\log \\left(I _ {x y} (t - \\Delta t)\\right) \\geq c, \\\\ - 1 & \\text {i f} \\log \\left(I _ {x y} (t)\\right) - \\log \\left(I _ {x y} (t - \\Delta t)\\right) \\leq - c, \\end{array} \\right. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 321, + 654, + 785, + 694 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $p$ is the event's polarity, $I_{xy}(t)$ and $I_{xy}(t - \\Delta t)$ represent the intensity at the pixel coordinate $(x,y)$ at time $t$ and $t - \\Delta t$ , respectively. An event is a quaternion value $(x,y,t,p)$ where $x,y$ denotes the pixel coordinate, $t$ represents the time of change, the polarity $p \\in \\{+1, -1\\}$ indicates the increase and decrease. For a given blurry image $B$ , it can be formulated as the average of latent intensity images $I(t)$ during the exposure time [25]:", + "bbox": [ + 212, + 705, + 787, + 797 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nB = \\frac {1}{T} \\int_ {0} ^ {T} I (t) d t, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 437, + 809, + 785, + 844 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Motion Aware Event Representation-driven Image Deblurring", + "bbox": [ + 321, + 114, + 732, + 130 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/727fbf1276279060979dc2c666d494d80bcb46f3f42f97cec3fb0bcc253b654c.jpg", + "image_caption": [ + "Fig. 2: Overview of the architecture of our proposed method for image deblurring. The blurred image and related event frames within the exposure time are fed into the corresponding network branches respectively. The core module of the event branch: (a) Recurrent Motion Extraction (RME) module processes event frames from longer time scales to shorter, merging old features into new features through weighted sum, which results in perceiving various motion patterns under different time ranges. The image branch extracts the features with conv block and fuses them with event features using (b) Feature Alignment and Fusion (FAF) module. FAF module eliminates the inter-modal inconsistencies in a bi-directional way ensuring better information interaction." + ], + "image_footnote": [], + "bbox": [ + 232, + 143, + 772, + 324 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $T$ is the exposure time. For simplicity of understanding, we now only consider the pixel at $(x,y)$ then combining Equations (1) and (2), we have", + "bbox": [ + 214, + 486, + 787, + 518 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} B _ {x y} = \\frac {1}{T} \\int_ {0} ^ {T} \\exp (\\log (I _ {x y} (t))) d t (3) \\\\ = \\frac {1}{T} \\int_ {0} ^ {T} e x p (\\log (I _ {x y} (0)) + c \\int_ {0} ^ {t} p (s) d s) d t. (4) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 331, + 526, + 785, + 599 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The internal part of the exponential operation is the logarithmic value of the light intensity of the pixel at time $t$ , while the external part of the exponential operation is the integration and averaging of each moment within the exposure time. Based on the event data, we can discretize Equation (4):", + "bbox": [ + 214, + 607, + 787, + 669 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} B _ {x y} = \\frac {1}{N} \\sum_ {t ^ {\\prime} \\in \\mathcal {T}} \\exp \\left(\\log \\left(I _ {x y} (0)\\right) + c \\sum_ {t = 0} ^ {t ^ {\\prime}} p _ {t}\\right) (5) \\\\ = \\frac {I _ {x y} (0)}{N} \\sum_ {t ^ {\\prime} \\in \\mathcal {T}} e x p \\left(c \\sum_ {t = 0} ^ {t ^ {\\prime}} p _ {t}\\right), (6) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 351, + 680, + 785, + 768 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathcal{T} = \\{t_1, t_2, \\dots, t_N\\}$ indicates the set of time events occurs, $N$ is the number of events, and $p_t$ means the polarity of the event at time $t$ .", + "bbox": [ + 212, + 779, + 787, + 809 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The aforementioned derivation prompts us to consider not only the polarity of the event but also the impact of the alteration in the absolute value after", + "bbox": [ + 212, + 810, + 787, + 839 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Z. Sun et al.", + "bbox": [ + 271, + 114, + 357, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Algorithm 1 Deviation Accumulation (DA)", + "text_level": 1, + "bbox": [ + 217, + 146, + 540, + 162 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Input: taking pixel at $(x,y)$ for an example, the events stream $(x,y,t_i,p_i)$ in which $i\\in [1,N]$ , deviation from the initial moment $D$ , accumulation of deviation $R$ , Output: the value of DA event representation $V$ .", + "bbox": [ + 215, + 166, + 785, + 208 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1: $D = 0,R = 0$", + "2: Counting the number of events $N$ ;", + "3: for $(i = 1$ to $N)$ do", + "4: $D = D + p_{i}$", + "5: $R = R + D$", + "6: end for", + "7: $V = R / N$" + ], + "bbox": [ + 222, + 209, + 472, + 305 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "each accumulation when considering the problem of deblurring. However, since event data cannot provide $I_{xy}(0)$ , the absolute light intensity is not accessible. But we can indirectly capture this information through the deviation from the initial moment. Therefore, we design a new event representation method for the deblurring problem presented in Algorithm 1. First, we accumulate all the events in the exposure time, so we get a 2D matrix named frequency matrix representing the number of the event occurs at each pixel position. We deal with the events considering their polarities subsequently, recording the deviation at each pixel by an additional tensor. Upon the arrival of the subsequent event, we update the deviation according to polarity and then add the deviation to the corresponding position. In the end, by dividing the cumulative matrix by the frequency matrix, we obtain the average deviation change over the exposure period as the new event representation named Deviation Accumulation (DA).", + "bbox": [ + 212, + 340, + 785, + 537 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Traditional polarity accumulation methods directly accumulating event polarities lead to identical outputs whenever the number of positive and negative events within a given time frame $([0, T])$ remains unchanged. Nevertheless, by considering the deviation brought by historical events, our method can generate distinct values for events arriving in different orders, thereby improving the ability to perceive motion patterns and providing more precise motion information for the deblurring procedure. For example, as illustrated in Figure 3, (a) shows event signals generated by a motion pattern which represents that the light intensity at that pixel has risen twice and then fallen twice while (d) shows another pattern. Facing these two patterns, (b) and (e) depict the changing process of the outputs of traditional method and our method after each event is accumulated. After processing all four events, traditional polarity accumulation gets the same output of 0 for both patterns, but our DA representation yields 1 and 0.5 respectively which contains richer motion information. More examples of motion patterns our method can distinguish are shown in supplementary materials.", + "bbox": [ + 212, + 537, + 785, + 763 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Overall, the phenomenon that polarity accumulation methods represent some different motion patterns as the same value introduces ambiguity into the deblurring process. Our method has a better perception of the motion patterns and offers a more distinct motion edge than conventional approaches from the visual effect as shown in Figure 3 (c) and (f). Inspired by [33], directly segmenting", + "bbox": [ + 212, + 765, + 785, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Motion Aware Event Representation-driven Image Deblurring", + "bbox": [ + 321, + 114, + 732, + 130 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f27df0ad0618d5db90f55bedf00ddfb7b27b5a54a5d9f563739f621712895c21.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 263, + 148, + 411, + 255 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/42187f67118f8b940e6d4101d46ceb7e90f1c0c1d5e25fc0efcf2e6b161825e5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 416, + 145, + 594, + 253 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2594bfdcceba4383dab92ecf16d109c444ee6f9000430d2f4182e0ca07f17411.jpg", + "image_caption": [ + "Fig. 3: (a) and (d) are event signals of two different motion patterns. (b) shows the Polarity Accumulation (PA) can't distinguish (a) and (d) as it gets the same output of 0 after processing all events while (e) shows our Deviation Accumulation (DA) can. DA perceives more motion information than PA as shown in (c) and (f)." + ], + "image_footnote": [], + "bbox": [ + 263, + 257, + 411, + 359 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1c9c90660ce6ee749ed03a3071e895c9dbd79e1abc3052138dd0644cf4112f44.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 418, + 255, + 593, + 359 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/16ebf709d3c72c1877985cdb2c25867915be52e3966f58db11e80140a702ae40.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 617, + 147, + 736, + 255 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a6efd72a2e849957ba7e7139a202dbda424f44cc76c8a7f95b59c21f5a3f3d94.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 617, + 256, + 736, + 359 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the entire exposure time averagely is not suitable for motion deblurring, which focuses on capturing rapid motion but falls short in adequately modeling the continuous movement throughout the exposure. In this paper, we symmetrically accumulate events along the $T / 2$ axis using the DA method. Taking $[0,T / 2]$ as an example, we accumulate events on different time scales. The longest accumulating range is $T / 2$ , followed by $T / 3$ , and the shortest is $T / 6$ . A detailed schematic is provided in the supplementary materials.", + "bbox": [ + 212, + 454, + 787, + 561 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.3 Recurrent Motion Extraction Module", + "text_level": 1, + "bbox": [ + 214, + 579, + 573, + 593 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The goal of the Recurrent Motion Extraction (RME) module is to cater to both global slow motion and local fast motion in different time scales. Based on our observations of the motion process and event representations, the focus of the event representations varies in different time scales. In shorter time intervals, the global slow motion is not significant in the event representation due to the smaller number of events triggered. On the contrary slow motion can be clearly recorded in longer intervals, but at the same time, the violent moving part will produce edge spreading artifact.", + "bbox": [ + 212, + 599, + 787, + 720 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To enable the network to accurately model the various in-homogeneous motions, we devised an RME module that incorporates a recurrent mechanism. As shown in Figure 2(a), the event features are fed into the processing module from the longest time scales to the shortest and we use $e1$ , $e2$ , and $e3$ to denote them. Formally, the RME module can be presented as", + "bbox": [ + 212, + 720, + 787, + 797 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\ne _ {i} ^ {\\prime} = \\left\\{ \\begin{array}{l l} f _ {e x} \\left(e _ {i}\\right) & \\text {i f} i = 1, \\\\ \\left(1 - \\alpha\\right) \\times f _ {e x} \\left(e _ {i}\\right) + \\alpha \\times f _ {r e} \\left(e _ {i - 1} ^ {\\prime}\\right) & \\text {i f} i = 2, 3, \\end{array} \\right. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 313, + 803, + 785, + 844 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Z. Sun et al.", + "bbox": [ + 271, + 114, + 357, + 127 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $f_{ex}(\\cdot)$ and $f_{re}(\\cdot)$ are two baseline conv blocks, $f_{ex}(\\cdot)$ is used to extract information from the current representation, $f_{re}(\\cdot)$ is used to retain useful information from previous features, $\\alpha$ is the learnable weight as a trade-off between two features. By sharing the parameters of $f_{ex}(\\cdot)$ and $f_{re}(\\cdot)$ for different time scales, we force the network to be able to perceive motions at varying time scales. Each input feature is symmetric on the time axis so that the priori information about the motion direction is simultaneously fed into the network.", + "bbox": [ + 212, + 146, + 787, + 252 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "3.4 Feature Alignment and Fusion Module", + "text_level": 1, + "bbox": [ + 214, + 273, + 581, + 289 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Although the information captured by the event camera exhibits a high correlation with the frame camera during the equivalent exposure period, there is still an inevitable gap originating from the modal dissimilarities between them. This phenomenon is frequently mentioned in multi-modal learning [18,42], but in event-guided image restoration tasks people usually [9,33,45] fuse features from two modalities directly using convolutions or attention mechanisms. Therefore, we propose an effective Feature Alignment and Fusion (FAF) module to facilitate enhanced feature fusion in a more coherent environment. Concretely, as shown in Figure 2(b), this module consists of a feature alignment module and a multi-head attention module. We use the operations of multiplication and addition successively to merge the raw information of one modality into the changing process of the other modality, which fills the gap of inter-modal inconsistency. The aligned features are fed into the attention module. In each fusion process, attention plays a bi-directional role, we not only generate the Q matrix through the frame features and the K and V matrix through the event features but also use the event features to perform look-ups on the frame features. The weights of the attention module are shared so that we can fuse the motion information under a more consistent space. Therefore, FAF module can be expressed as", + "bbox": [ + 212, + 299, + 787, + 571 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nx ^ {\\prime} = \\operatorname {C o n v} _ {2} (x) + (\\operatorname {C o n v} _ {1} (x) \\times e + e), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 227, + 582, + 784, + 599 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\ne ^ {\\prime} = \\operatorname {C o n v} _ {2} (e) + (\\operatorname {C o n v} _ {1} (e) \\times x + x), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 228, + 604, + 784, + 619 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nF = \\operatorname {C o n v} _ {3} (\\operatorname {C o n c a t} ([ \\text {A t t e n t i o n} (Q _ {x ^ {\\prime}}, K _ {e ^ {\\prime}}, V _ {e ^ {\\prime}}), \\text {A t t e n t i o n} (Q _ {e ^ {\\prime}}, K _ {x ^ {\\prime}}, V _ {x ^ {\\prime}}) ])), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 228, + 625, + 784, + 640 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $x'$ and $e'$ are aligned image features $x$ and event feature $e$ respectively, Attention(\\cdot) denotes the multi-head attention operation, Concat(\\cdot) denotes the concatenation operation along the channel dimension.", + "bbox": [ + 214, + 652, + 787, + 700 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiments and Analysis", + "text_level": 1, + "bbox": [ + 214, + 722, + 504, + 739 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Experimental Settings", + "text_level": 1, + "bbox": [ + 214, + 753, + 447, + 768 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Datasets. We use three different datasets containing synthetic, semi-synthetic, and real-world frames and events to evaluate the proposed method.", + "bbox": [ + 212, + 779, + 782, + 808 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "GoPro: We train the network on the GoPro dataset which is the benchmark dataset for the image motion deblurring [24]. The blurry image is averaged from", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Motion Aware Event Representation-driven Image Deblurring", + "bbox": [ + 320, + 114, + 732, + 130 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/c910256561819228f874ee586a3a0c12b3b6df3b608f13ac97ba3cd6c9d9c00b.jpg", + "table_caption": [ + "Table 1: The quantitative results on GoPro, HS-ERGB, and REBlur test datasets. The best values are highlighted in bold. HINet+ and UFPNet+ are event-enhanced versions of HINet and UFPNet by concatenating our DA to their input." + ], + "table_footnote": [], + "table_body": "
MethodInputGoProHS-ERGBREBlurParams(M)
PSNRSSIMPSNRSSIMPSNRSSIM
LEVS [10]F20.840.54722.130.555--18.2
EDI [25]F+E29.060.94323.930.70436.520.9640.5
SRN [35]F30.260.934--35.100.96110.2
EVDI [49]F+E30.400.90625.130.707--0.4
HINet [4]F32.710.95927.320.80735.580.96588.7
Restormer [47]F32.920.96127.550.80835.500.95926.1
MSDI-Net [17]F33.280.96427.460.80936.140.968241.3
NAFNet [3]F33.710.96727.640.81136.150.96967.8
UFPNet [6]F34.060.96827.640.80936.110.96880.3
HINet+ [4]F+E34.630.96827.660.80837.920.97688.7
UFPNet+ [6]F+E35.220.97227.680.80937.970.97680.3
EFNet [33]F+E35.460.97226.680.80038.120.9758.5
OursF+E36.070.97627.930.81238.470.97813.9
", + "bbox": [ + 228, + 199, + 771, + 422 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "adjacent sharp frames, by doing so the dataset contains 3214 pairs of blurry and sharp images. According to the standard division, we adopt 2103 pairs for training and 1111 pairs for testing. For the event data, we use the GoPro raw event dataset provided by [33] which synthesizes events with a random threshold $c$ following the Gaussian distribution by ESIM simulator [27]. Following the DA method, the raw event data is preprocessed into 6 event frames for each image.", + "bbox": [ + 212, + 462, + 784, + 551 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "HS-ERGB: The HS-ERGB dataset [37] consists of sharp videos and real-world events, we use [50] released normal blur version which synthesizes blurry frames by averaging 49 interpolating images. Each blurry frame in the dataset has a known exposure time, allowing us to generate the DA representation for each frame within its respective exposure duration.", + "bbox": [ + 212, + 551, + 785, + 628 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "REBlur: The REBlur dataset collects a number of sequences of real-world event data, corresponding with blurry images and reliable ground-truth sharp images [33]. It contains 12 kinds of linear and nonlinear motions. Following the original setting, we use 486 pairs of blurry-sharp images with associated events for training and 983 pairs for testing.", + "bbox": [ + 212, + 628, + 785, + 703 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Implementation details. We train the end-to-end network without pre-training on a single Tesla V100 GPU. We extract patches of size $256 \\times 256$ for training images and corresponding event frames, and the batch size is set to 2 by default. The AdamW optimizer is employed ( $\\beta_{1} = 0.9$ and $\\beta_{2} = 0.99$ ) with the initial learning rate 0.0001 with the cosine annealing schedule where the $T_{max}$ is 400K iteration. For data augmentation, we perform horizontal and vertical flips, the random noise and the hot pixels are added into event frames in order to simulate the real-world situation. For HS-ERGB and REBlur datasets, We fine-tune the model trained on GoPro with the training set of them. The fine-tuning process", + "bbox": [ + 212, + 704, + 787, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Z. Sun et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/20a836d4e920f402f1279f0912bc458bdbdbfc258f2ce5277dd485ce5c876056.jpg", + "image_caption": [ + "(a) Blur image in GoPro testset" + ], + "image_footnote": [], + "bbox": [ + 259, + 146, + 431, + 277 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "(a) Blur image in GoPro testset", + "bbox": [ + 261, + 279, + 429, + 290 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/090bccd521fd0515526e1fdddda2a1d27d3b2e0539b66ab88db218c349e85371.jpg", + "image_caption": [ + "(b) HINet" + ], + "image_footnote": [], + "bbox": [ + 434, + 146, + 509, + 204 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/de240c96888191bad7fd23d6f5d1fd42c49aab6b15b4ad351ec42ba2462276a4.jpg", + "image_caption": [ + "(c) UFPNet" + ], + "image_footnote": [], + "bbox": [ + 511, + 146, + 586, + 204 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/855a3626dcaa19b3a6f81c7a38abba6894d8210cbb0351136d26db20c5ded06a.jpg", + "image_caption": [ + "(d) EFNet" + ], + "image_footnote": [], + "bbox": [ + 589, + 146, + 663, + 204 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/9d59b2a9ca86a9e8467a6a4624297134ae54fe4303e5a9a48ebb08b3760aa938.jpg", + "image_caption": [ + "(e) NAFNet" + ], + "image_footnote": [], + "bbox": [ + 666, + 146, + 741, + 204 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/33624c427a9def027a5e47ac4beec07fe892844f016eb3e4c139b6b2dff08fd6.jpg", + "image_caption": [ + "(f) HINet+" + ], + "image_footnote": [], + "bbox": [ + 434, + 220, + 509, + 277 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/b844597b0b2628d9b494fbb49079e9538da6c357ce669b0189408d786989910e.jpg", + "image_caption": [ + "(g) UFPNet+" + ], + "image_footnote": [], + "bbox": [ + 511, + 220, + 586, + 277 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/525fd323a7e502680d99c7d455846726cc7a4848aa71c023a1ac82cea2cdfb8e.jpg", + "image_caption": [ + "(h) Ours" + ], + "image_footnote": [], + "bbox": [ + 589, + 220, + 663, + 277 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/1352bdc01951e52d0cda0812c9706632ddf092d3b3b441a507c212605ee64399.jpg", + "image_caption": [ + "(i) GT" + ], + "image_footnote": [], + "bbox": [ + 668, + 220, + 743, + 277 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/0843f6c080552a51c92468fb751bc0e4305eec693b7bb36b6c1b2a4b492e83be.jpg", + "image_caption": [ + "Fig. 4: Qualitative comparisons on the GoPro dataset. The notation is the same as in Table 1. Zoom in for better view.", + "(a) Blur image in REBlur testset", + ": Qualitative comparisons on the REBlur dataset. The notation is the same as Table 1. Zoom in for better view." + ], + "image_footnote": [], + "bbox": [ + 261, + 348, + 431, + 479 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/e3739c1a5a39ccec854698002b1ed5041ff11d891789b8f9a04cefc4b762dc8d.jpg", + "image_caption": [ + "(b) HINet" + ], + "image_footnote": [], + "bbox": [ + 434, + 348, + 509, + 406 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/1cee75ff1bd3522ee52b1985f367cea90277b0673d9543de97be08f469dad2dc.jpg", + "image_caption": [ + "(c) UFPNet" + ], + "image_footnote": [], + "bbox": [ + 514, + 348, + 586, + 406 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/49d03350932205f4c7456d7b8004e01f93a025b15a3769694659b63a66aebda9.jpg", + "image_caption": [ + "(d) EFNet" + ], + "image_footnote": [], + "bbox": [ + 589, + 348, + 665, + 406 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/b60a52f299b5b271275e91e328b201333ba9fc08023bb258ff4899f4394d7338.jpg", + "image_caption": [ + "(e) NAFNet" + ], + "image_footnote": [], + "bbox": [ + 668, + 348, + 741, + 406 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "consisted of 4000 iterations while keeping the other configurations consistent with the previous experiments.", + "bbox": [ + 214, + 563, + 784, + 594 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.2 Comparison with State-of-the-Art Methods", + "text_level": 1, + "bbox": [ + 214, + 617, + 619, + 633 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Table 1 shows the comparison results of our method and other SOTA methods on GoPro [24], HS-ERGB [50] and REBlur [33]. We compare our network with not only event-based methods, but also with some SOTA image-only deblurring networks enhancing with event input. The results illustrate that: (1) Our method outperforms all other methods in synthetic, semi-synthetic and real-world datasets, which indicates that with the proper utilization of events the designed approach is able to handle blur with different distributions. Specifically, the average gains of our method over the second-best one are 0.61dB, 0.25dB and 0.35dB in terms of PSNR on GoPro, HS-ERGB and REBlur datasets, respectively. (2) Through the enhancement of our proposed event representation, the HINet+ and the UFPNet+ increase 1.92dB and 1.16dB in PSNR on the benchmark GoPro dataset compared with their image-only version, respectively. Taking HINet as an example, with the same operation of [33], HINet with DA", + "bbox": [ + 212, + 643, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Motion Aware Event Representation-driven Image Deblurring", + "bbox": [ + 321, + 114, + 732, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/734fd1b1a209722f46554a3beff3f560d60b9d4c52ce867556462ddd29643728.jpg", + "image_caption": [ + "(a) GT" + ], + "image_footnote": [], + "bbox": [ + 267, + 143, + 343, + 202 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/23665851adff685957098865bc95f18a700e064978f64467d82b3f335ad0e2c5.jpg", + "image_caption": [ + "(b) Blurry" + ], + "image_footnote": [], + "bbox": [ + 344, + 143, + 421, + 202 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/e436f01102b7a8802bc8f0f4fda16ca5e7ee47f048512df78a52a0737fd0df72.jpg", + "image_caption": [ + "(c) UFPNet" + ], + "image_footnote": [], + "bbox": [ + 424, + 143, + 500, + 202 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/116209ed0274009ad194fa824ab167736e9ae06fc2d09cdee0ae569b63bfb553.jpg", + "image_caption": [ + "(d) NAFNet" + ], + "image_footnote": [], + "bbox": [ + 501, + 143, + 578, + 202 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/75b923889f3b670525027bcd31c23cc3db0dd31392f76380b22b20f16a83888f.jpg", + "image_caption": [ + "(e) EFNet" + ], + "image_footnote": [], + "bbox": [ + 581, + 143, + 656, + 202 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/1e7c2eabaea83f858b8621cc979ffad7fca20e9e2d345effe15da8aa95aa0aa0.jpg", + "image_caption": [ + "(f) Ours" + ], + "image_footnote": [], + "bbox": [ + 658, + 143, + 735, + 202 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/41e0651326b76cccb5d5710c1442995461b95c4ddbe1fcacc18b94d134bc4605.jpg", + "image_caption": [ + "Fig. 6: Qualitative comparisons on the HS-ERGB dataset." + ], + "image_footnote": [], + "bbox": [ + 267, + 256, + 382, + 345 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/a720aa58896079b0e0b39e7458067a7fbec6479614a71146e05ad3f99b7d9700.jpg", + "image_caption": [ + "(a) Blurry" + ], + "image_footnote": [], + "bbox": [ + 267, + 348, + 380, + 434 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/68e51ec54743ea9c8e64884178afd1a439dff5366a8d45a80122461dcc428e57.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 383, + 257, + 500, + 345 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/aca0016fe3ce7794d334f49c234da809ebe1c6ab4418653c1f8df20c1c5fd89a.jpg", + "image_caption": [ + "(b) SCER [33]" + ], + "image_footnote": [], + "bbox": [ + 385, + 348, + 500, + 433 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/c83f0d2d8eb537da6533e4b9f86be810fee0e31af3b33665379c2553ef6c7061.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 257, + 617, + 345 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/e006b54c00ea90658069b0fd4689d9fe01929f4c59460730601fbe9ea5ab731a.jpg", + "image_caption": [ + "(c) DA (Ours)" + ], + "image_footnote": [], + "bbox": [ + 503, + 348, + 617, + 434 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/2355dc71134c668061af47e87b50aa1f1aa24daa40fa40843e9198ed91743256.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 620, + 257, + 733, + 345 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/f1a387eda2a305341be0a7284808fbc25adaa1a60547ed3ebee3eb97710f9f44.jpg", + "image_caption": [ + "(d) GT", + "Fig. 7: Visual comparisons on GoPro dataset with different event representation. Compared to SCER representation, our DA representation reduces edge artifacts." + ], + "image_footnote": [], + "bbox": [ + 620, + 348, + 733, + 434 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "event representation exceeds 0.94dB over the result in [33], which presents the fact that DA extracts more conducive information for motion deblurring.", + "bbox": [ + 212, + 517, + 784, + 546 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We also show the visual comparison results to verify the effectiveness of our method in Figure 4, Figure 5 and Figure 6 with representative samples from each dataset. As illustrated in the figures, our model reconstructs better details and has minor spatial distortions compared with other methods. This indicates that our event representation perceives more motion information and the modules effectively embed it to reconstruct a high-quality image with enhanced details.", + "bbox": [ + 212, + 547, + 784, + 638 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.3 Ablation Studies", + "text_level": 1, + "bbox": [ + 214, + 662, + 400, + 676 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We conduct plenty ablation experiments on GoPro dataset to verify the effectiveness of our event representation and proposed method.", + "bbox": [ + 212, + 688, + 784, + 718 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Effectiveness of Event Representation. To demonstrate the adaptability of our event representation for motion deblurring tasks, as shown in Table 2, we replace our DA representation with several other types of representations and trained the network accordingly. The results reveal certain shortcomings of the alternative representations. Firstly, the stack representation simply accumulates all events on their polarities which leaves out the temporal information. Secondly, the SBT approach [40], which subdivides the entire exposure time equally, proves to be inadequate for modeling the varying degrees of blur. Thirdly, the SCER", + "bbox": [ + 212, + 719, + 785, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Z. Sun et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/5d434a7baeecb04e0bb393f3d63371deeecd9117414eb116d1158d1296fd0d7d.jpg", + "table_caption": [ + "Table 2: Ablation study of each module effects." + ], + "table_footnote": [], + "table_body": "
Event RepresentationRMEFusionPSNRSSIM
Stack33.240.953
SBT [40]✓FAF(Ours)35.630.973
SCER [33]35.810.975
Concat.34.020.967
Add34.740.970
Attention35.750.974
ACA(Ours)✓Only Align Event+Attention35.780.975
Only Align Image+Attention35.740.974
35.420.973
ACA(Ours)✗FAF(Ours)36.070.976
✓
", + "bbox": [ + 241, + 171, + 756, + 339 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/81f7908c63a90b433e90594dc13fa912cb37f5504f5f993a8d50ce93a37c0d8f.jpg", + "image_caption": [ + "(a) w/o RME" + ], + "image_footnote": [], + "bbox": [ + 248, + 353, + 415, + 426 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/11be934ee6cb38b10dfb35c3fc2d91777a02b74c7898ac526a9795643b474ccb.jpg", + "image_caption": [ + "(b) w/o RME" + ], + "image_footnote": [], + "bbox": [ + 416, + 353, + 584, + 426 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/1f18276d5b5d8ae61f7cddeb2945d238a52e8902dde8319fdaa11eaef552ba1d.jpg", + "image_caption": [ + "(c) w/o RME" + ], + "image_footnote": [], + "bbox": [ + 586, + 353, + 754, + 426 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/7815bdf963277e3713ad059dbd509e05af04828114043170576ce6d39efadbb0.jpg", + "image_caption": [ + "(d) w/RME", + "Fig. 8: Visual comparison results of w/ and w/o RME module. All of these scenarios involve movement in different directions and speeds, RME module can successfully capture motion features in these complex motion scenarios. Zoom in for better view." + ], + "image_footnote": [], + "bbox": [ + 250, + 440, + 415, + 513 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/7a552d1a9227850a52be4560f38c65a17ba9c1bf2257581d64e68c82f2f56e03.jpg", + "image_caption": [ + "(e) w/RME" + ], + "image_footnote": [], + "bbox": [ + 416, + 440, + 584, + 513 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/a03e2927b922c9edc7b8ca0699305e75b5d34138e48b931449aba8642488c107.jpg", + "image_caption": [ + "(f) w/RME" + ], + "image_footnote": [], + "bbox": [ + 586, + 440, + 753, + 513 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "method [33], which directly accumulates event polarities, is found to be incapable of capturing the influence of absolute intensity changes. This limitation is illustrated in Figure 3. In contrast, our DA representation not only possesses a solid physical foundation for event generation but also perceives more motion patterns. As a result, it achieves a 0.26dB improvement in performance compared to the best alternative representation. We also provide the visual comparison on the image reconstruction results in Figure 7. Our DA representation reduces motion-generated edge artifacts, resulting in a better deblurring effect.", + "bbox": [ + 212, + 611, + 787, + 733 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Effectiveness of Recurrent Motion Extraction Module. To assess the RME module's contribution, we replace it with a baseline conv block and adjust the parameters for fair comparison. In this approach, we concatenate event frames of different time scales and feed them into the conv block. The findings, as presented in the last two rows of Table 2, clearly indicate that the contribution of $0.65\\mathrm{dB}$ increment in PSNR stems from the network structure design, rather than an increase in parameter quantity. This shows the importance of incorpo", + "bbox": [ + 212, + 734, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Motion Aware Event Representation-driven Image Deblurring", + "bbox": [ + 321, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "rating the module for effectively leveraging motion information in multi-timescales and improving performance in motion deblurring tasks. We also present the visual comparisons when removing the RME module in Figure 8, the RME module adeptly captures both fast local and slow global motion characteristics.", + "bbox": [ + 212, + 146, + 787, + 205 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Effectiveness of Feature Alignment and Fusion Module. To illustrate the effectiveness of our FAF module, we compare it with several common fusion methods, including concatenation, addition, and cross-modality attention. As shown in Table 2, it is arduous to utilize the motion information contained in event feature through the direct concatenation or addition. While attention mechanisms can capture relative parts in the two features, the inherent inconsistencies within the two modalities still lead to a 0.32dB performance gap when compared to our FAF module. We also use a t-SNE visualization in Figure 9 to demonstrate the effectiveness of our feature alignment, where two originally in-congruent modal features are mixed up together after the FAF module.", + "bbox": [ + 212, + 207, + 787, + 358 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Furthermore, we also verify the importance of bi-directional alignment by conducting experiments solely on aligning event or image features. Since both events and frames provide crucial information for image restoration, such as edge motion information and textual information, respectively, it is unreasonable to let one modality dominate the alignment process. As shown in Table 2, the bi-directional alignment outperforms the highest score achieved by one-way correction by 0.29dB. This demonstrates the necessity of fully utilizing the mutual information between the two modalities in order to achieve better image reconstruction results.", + "bbox": [ + 212, + 359, + 488, + 630 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/5fb4c0b4afc0a3f1a56e00ab48428f673c8a69bd8eea8b47eb91ccc2480c19a4.jpg", + "image_caption": [ + "Fig. 9: t-SNE visualization of original and aligned features. The distributions of the aligned two features overlap more compared to the original features. This indicates the effectiveness of FAF module to reduce the inter-modal inconsistency." + ], + "image_footnote": [], + "bbox": [ + 516, + 375, + 769, + 545 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 655, + 359, + 671 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we propose an event-based image deblurring method. Firstly, we rethink the limit of traditional event preprocessing method and propose a new deviation accumulation (DA) technique. With deviation, the new approach can distinguish the order of events and perceive more motion patterns. Moreover, we design an image deblurring network equipped with recurrent motion extraction (RME) module and feature alignment and fusion (FAF) module to fit our proposed event representation. The RME module can better perceive motions in different time scales, and the FAF module promotes better integration of the two modalities. Finally, extensive experimental results over several datasets demonstrate the effectiveness of the proposed algorithm.", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Z. Sun et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements. This work was supported by National Natural Science Foundation of China (NSFC) under Grants 62225207 and 62276243.", + "bbox": [ + 215, + 146, + 784, + 176 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 202, + 321, + 218 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Benosman, R., Clercq, C., Lagorce, X., Ieng, S.H., Bartolozzi, C.: Event-based visual flow. IEEE transactions on neural networks and learning systems 25(2), 407-417 (2013)", + "2. Chang, M., Feng, H., Xu, Z., Li, Q.: Low-light image restoration with short-and long-exposure raw pairs. IEEE Transactions on Multimedia 24, 702-714 (2021)", + "3. Chen, L., Chu, X., Zhang, X., Sun, J.: Simple baselines for image restoration. In: Proceedings of the European conference on computer vision (ECCV). pp. 17-33. Springer (2022)", + "4. Chen, L., Lu, X., Zhang, J., Chu, X., Chen, C.: Hinet: Half instance normalization network for image restoration. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 182-192 (2021)", + "5. Cook, M., Gugelmann, L., Jug, F., Krautz, C., Steger, A.: Interacting maps for fast visual interpretation. In: The 2011 International Joint Conference on Neural Networks. pp. 770-776. IEEE (2011)", + "6. Fang, Z., Wu, F., Dong, W., Li, X., Wu, J., Shi, G.: Self-supervised non-uniform kernel estimation with flow-based motion prior for blind image deblurring. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18105-18114 (2023)", + "7. Gehrig, D., Rebecq, H., Gallego, G., Scaramuzza, D.: Eklt: Asynchronous photometric feature tracking using events and frames. International Journal of Computer Vision 128(3), 601-618 (2020)", + "8. Ji, S.W., Lee, J., Kim, S.W., Hong, J.P., Baek, S.J., Jung, S.W., Ko, S.J.: Xydeblur: divide and conquer for single image deblurring. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17421-17430 (2022)", + "9. Jiang, Z., Zhang, Y., Zou, D., Ren, J., Lv, J., Liu, Y.: Learning event-based motion deblurring. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3320-3329 (2020)", + "10. Jin, M., Meishvili, G., Favaro, P.: Learning to extract a video sequence from a single motion-blurred image. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 6334-6342 (2018)", + "1. Kim, H., Handa, A., Benosman, R., Ieng, S.H., Davison, A.J.: Simultaneous mosaicing and tracking with an event camera. J. Solid State Circ 43, 566-576 (2008)", + "2. Kim, T., Lee, J., Wang, L., Yoon, K.J.: Event-guided deblurring of unknown exposure time videos. In: Proceedings of the European conference on computer vision (ECCV). pp. 519-538. Springer (2022)", + "3. Krishnan, D., Tay, T., Fergus, R.: Blind deconvolution using a normalized sparsity measure. In: CVPR 2011. pp. 233-240. IEEE (2011)", + "4. Lagorce, X., Orchard, G., Galluppi, F., Shi, B.E., Benosman, R.B.: Hots: a hierarchy of event-based time-surfaces for pattern recognition. IEEE transactions on pattern analysis and machine intelligence 39(7), 1346-1359 (2016)", + "5. Levin, A., Weiss, Y., Durand, F., Freeman, W.T.: Understanding and evaluating blind deconvolution algorithms. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 1964-1971. IEEE (2009)" + ], + "bbox": [ + 225, + 237, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Motion Aware Event Representation-driven Image Deblurring", + "bbox": [ + 321, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "16. Li, D., Shi, X., Zhang, Y., Cheung, K.C., See, S., Wang, X., Qin, H., Li, H.: A simple baseline for video restoration with grouped spatial-temporal shift. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9822-9832 (2023)", + "17. Li, D., Zhang, Y., Cheung, K.C., Wang, X., Qin, H., Li, H.: Learning degradation representations for image deblurring. In: Proceedings of the European conference on computer vision (ECCV). pp. 736-753. Springer (2022)", + "18. Li, Y., Yu, A.W., Meng, T., Caine, B., Ngiam, J., Peng, D., Shen, J., Lu, Y., Zhou, D., Le, Q.V., et al.: Deepfusion: Lidar-camera deep fusion for multi-modal 3d object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17182-17191 (2022)", + "19. Liang, M., Yang, B., Wang, S., Urtasun, R.: Deep continuous fusion for multi-sensor 3d object detection. In: Proceedings of the European conference on computer vision (ECCV). pp. 641-656 (2018)", + "20. Liang, X., Qian, Y., Guo, Q., Cheng, H., Liang, J.: Af: An association-based fusion method for multi-modal classification. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(12), 9236-9254 (2021)", + "21. Liu, M., Delbruck, T.: Adaptive time-slice block-matching optical flow algorithm for dynamic vision sensors. BMVC (2018)", + "22. Manderscheid, J., Sironi, A., Bourdis, N., Migliore, D., Lepetit, V.: Speed invariant time surface for learning to detect corner points with event-based cameras. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10245-10254 (2019)", + "23. Maqueda, A.I., Loquercio, A., Gallego, G., García, N., Scaramuzza, D.: Event-based vision meets deep learning on steering prediction for self-driving cars. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5419-5427 (2018)", + "24. Nah, S., Hyun Kim, T., Mu Lee, K.: Deep multi-scale convolutional neural network for dynamic scene deblurring. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3883-3891 (2017)", + "25. Pan, L., Scheerlinck, C., Yu, X., Hartley, R., Liu, M., Dai, Y.: Bringing a blurry frame alive at high frame-rate with an event camera. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6820-6829 (2019)", + "26. Paredes-Vallés, F., Scheper, K.Y., De Croon, G.C.: Unsupervised learning of a hierarchical spiking neural network for optical flow estimation: From events to global motion perception. IEEE transactions on pattern analysis and machine intelligence 42(8), 2051-2064 (2019)", + "27. Rebecq, H., Gehrig, D., Scaramuzza, D.: Esim: an open event camera simulator. In: Conference on robot learning. pp. 969-982. PMLR (2018)", + "28. Ren, D., Shang, W., Yang, Y., Zuo, W.: Aggregating long-term sharp features via hybrid transformers for video deblurring. arXiv preprint arXiv:2309.07054 (2023)", + "29. Ren, S., Du, Y., Lv, J., Han, G., He, S.: Learning from the master: Distilling cross-modal advanced knowledge for lip reading. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13325-13333 (2021)", + "30. Shang, W., Ren, D., Zou, D., Ren, J.S., Luo, P., Zuo, W.: Bringing events into video deblurring with non-consecutively blurry frames. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4531-4540 (2021)", + "31. Song, C., Bajaj, C., Huang, Q.: Deblursr: Event-based motion deblurring under the spiking representation. arXiv preprint arXiv:2303.08977 (2023)" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Z. Sun et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "32. Su, S., Delbracio, M., Wang, J., Sapiro, G., Heidrich, W., Wang, O.: Deep video deblurring for hand-held cameras. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1279-1288 (2017)", + "33. Sun, L., Sakaridis, C., Liang, J., Jiang, Q., Yang, K., Sun, P., Ye, Y., Wang, K., Gool, L.V.: Event-based fusion for motion deblurring with cross-modal attention. In: Proceedings of the European conference on computer vision (ECCV). pp. 412-428. Springer (2022)", + "34. Sun, L., Cho, S., Wang, J., Hays, J.: Edge-based blur kernel estimation using patch priors. In: IEEE international conference on computational photography (ICCP). pp. 1-8. IEEE (2013)", + "35. Tao, X., Gao, H., Shen, X., Wang, J., Jia, J.: Scale-recurrent network for deep image deblurring. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 8174-8182 (2018)", + "36. Tian, Y., Krishnan, D., Isola, P.: Contrastive multiview coding. In: Proceedings of the European conference on computer vision (ECCV). pp. 776-794. Springer (2020)", + "37. Tulyakov, S., Gehrig, D., Georgoulis, S., Erbach, J., Gehrig, M., Li, Y., Scaramuzza, D.: Time lens: Event-based video frame interpolation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 16155-16164 (2021)", + "38. Vitoria, P., Georgoulis, S., Tulyakov, S., Bochicchio, A., Erbach, J., Li, Y.: Event-based image deblurring with dynamic motion awareness. In: Proceedings of the European conference on computer vision (ECCV). pp. 95-112. Springer (2022)", + "39. Wang, B., He, J., Yu, L., Xia, G.S., Yang, W.: Event enhanced high-quality image recovery. In: Proceedings of the European conference on computer vision (ECCV). pp. 155-171. Springer (2020)", + "40. Wang, L., Ho, Y.S., Yoon, K.J., et al.: Event-based high dynamic range image and very high frame rate video generation using conditional generative adversarial networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10081-10090 (2019)", + "41. Weikersdorfer, D., Conradt, J.: Event-based particle filtering for robot self-localization. In: 2012 IEEE International Conference on Robotics and Biomimetics (ROBIO). pp. 866-870. IEEE (2012)", + "42. Xia, W., Li, X., Deng, A., Xiong, H., Dou, D., Hu, D.: Robust cross-modal knowledge distillation for unconstrained videos. arXiv preprint arXiv:2304.07775 (2023)", + "43. Xu, F., Yu, L., Wang, B., Yang, W., Xia, G.S., Jia, X., Qiao, Z., Liu, J.: Motion deblurring with real events. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2583-2592 (2021)", + "44. Xu, L., Jia, J.: Two-phase kernel estimation for robust motion deblurring. In: Proceedings of the European conference on computer vision (ECCV). pp. 157-170. Springer (2010)", + "45. Yang, D., Yamac, M.: Motion aware double attention network for dynamic scene deblurring. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1113-1123 (2022)", + "46. Yuan, L., Sun, J., Quan, L., Shum, H.Y.: Image deblurring with blurred/noisy image pairs. In: ACM SIGGRAPH 2007 papers, pp. 1-es (2007)", + "47. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H.: Restormer: Efficient transformer for high-resolution image restoration. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5728-5739 (2022)" + ], + "bbox": [ + 212, + 146, + 787, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Motion Aware Event Representation-driven Image Deblurring", + "bbox": [ + 321, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "48. Zhang, S., Zhen, A., Stevenson, R.L.: Deep motion blur removal using noisy/blurry image pairs. Journal of Electronic Imaging 30(3), 033022-033022 (2021)", + "49. Zhang, X., Yu, L.: Unifying motion deblurring and frame interpolation with events. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17765-17774 (2022)", + "50. Zhang, X., Yu, L., Yang, W., Liu, J., Xia, G.S.: Generalizing event-based motion deblurring in real-world scenarios. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10734-10744 (2023)", + "51. Zhou, Y., Gallego, G., Rebecq, H., Kneip, L., Li, H., Scaramuzza, D.: Semi-dense 3d reconstruction with a stereo event camera. In: Proceedings of the European conference on computer vision (ECCV). pp. 235-251 (2018)" + ], + "bbox": [ + 215, + 146, + 785, + 300 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Z. Sun et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/Motion Aware Event Representation-driven Image Deblurring/a9105a02-b537-456a-9778-80b745c423ab_model.json b/2024/Motion Aware Event Representation-driven Image Deblurring/a9105a02-b537-456a-9778-80b745c423ab_model.json new file mode 100644 index 0000000000000000000000000000000000000000..f0163a233f1c86f2c6dc1416cced453528046589 --- /dev/null +++ b/2024/Motion Aware Event Representation-driven Image Deblurring/a9105a02-b537-456a-9778-80b745c423ab_model.json @@ -0,0 +1,3096 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.255, + 0.141, + 0.75, + 0.187 + ], + "angle": 0, + "content": "Motion Aware Event Representation-driven Image Deblurring" + }, + { + "type": "text", + "bbox": [ + 0.224, + 0.213, + 0.779, + 0.229 + ], + "angle": 0, + "content": "Zhijing Sun, Xueyang Fu, Longzhuo Huang, Aiping Liu, and Zheng-Jun Zha*" + }, + { + "type": "text", + "bbox": [ + 0.25, + 0.24, + 0.754, + 0.311 + ], + "angle": 0, + "content": "School of Information Science and Technology and MoE Key Laboratory of Brain-inspired Intelligent Perception and Cognition, University of Science and Technology of China, Hefei, 230026, China {sunzhijing, hlz_0}@mail.ustc.edu.cn {xyfu, aipingl, zhazj}@ustc.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.349, + 0.74, + 0.655 + ], + "angle": 0, + "content": "Abstract. Traditional image deblurring struggles with high-quality reconstruction due to limited motion data from single blurred images. Excitingly, the high-temporal resolution of event cameras records motion more precisely in a different modality, transforming image deblurring. However, many event camera-based methods, which only care about the final value of the polarity accumulation, ignore the influence of the absolute intensity change where events generate so fall short in perceiving motion patterns and effectively aiding image reconstruction. To overcome this, in this work, we propose a new event preprocessing technique that accumulates the deviation from the initial moment each time the event is updated. This process can distinguish the order of events to improve the perception of object motion patterns. To complement our proposed event representation, we create a recurrent module designed to meticulously extract motion features across local and global time scales. To further facilitate the event feature and image feature integration, which assists in image reconstruction, we develop a bi-directional feature alignment and fusion module. This module works to lessen inter-modal inconsistencies. Our approach has been thoroughly tested through rigorous experiments carried out on several datasets with different distributions. These trials have delivered promising results, with our method achieving top-tier performance in both quantitative and qualitative assessments. Code is available at https://github.com/ZhijingS/DA_event_deblur." + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.667, + 0.724, + 0.682 + ], + "angle": 0, + "content": "Keywords: Image deblurring \\(\\cdot\\) Event camera \\(\\cdot\\) Event representation" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.709, + 0.375, + 0.724 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.741, + 0.788, + 0.817 + ], + "angle": 0, + "content": "Frame-based cameras often produce blurred images due to factors like camera shake or object motion, given their configured exposure time. Such blur, a common form of image degradation, occurs across various scenes. While photographers typically aim for clear and focused photos, clear inputs are equally critical for numerous computer vision tasks. These tasks include super-resolution, target" + }, + { + "type": "page_footnote", + "bbox": [ + 0.233, + 0.826, + 0.39, + 0.84 + ], + "angle": 0, + "content": "*Corresponding author" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.128 + ], + "angle": 0, + "content": "Z. Sun et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.313, + 0.217 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.242, + 0.218, + 0.287, + 0.251 + ], + "angle": 0, + "content": "(a) GT (PSNR |SSIM)" + }, + { + "type": "image", + "bbox": [ + 0.315, + 0.144, + 0.407, + 0.217 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.329, + 0.218, + 0.391, + 0.251 + ], + "angle": 0, + "content": "(b) Blurry (28.62dB |0.819)" + }, + { + "type": "image", + "bbox": [ + 0.41, + 0.144, + 0.501, + 0.217 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.416, + 0.218, + 0.498, + 0.251 + ], + "angle": 0, + "content": "(c) HINet [4] (29.14dB |0.837)" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.144, + 0.597, + 0.217 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.504, + 0.218, + 0.597, + 0.251 + ], + "angle": 0, + "content": "(d) NAFNet [3] (30.41dB |0.855)" + }, + { + "type": "image", + "bbox": [ + 0.599, + 0.144, + 0.692, + 0.217 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.601, + 0.218, + 0.688, + 0.251 + ], + "angle": 0, + "content": "(e) EFNet [33] (30.76dB |0.856)" + }, + { + "type": "image", + "bbox": [ + 0.695, + 0.144, + 0.787, + 0.217 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.713, + 0.218, + 0.766, + 0.25 + ], + "angle": 0, + "content": "(f) Ours (31.01dB [0.858]" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.261, + 0.788, + 0.304 + ], + "angle": 0, + "content": "Fig. 1: A deblurring example on HS-ERGB [37] dataset via recent state-of-the-art methods. Compared to other methods, our approach yields a more distinct result owing to its enhanced motion perception capability." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.338, + 0.785, + 0.384 + ], + "angle": 0, + "content": "detection, and semantic segmentation, to name a few. Hence, image deblurring serves as a fundamental task in the realm of low-level computer vision, reinforcing the image quality and supporting the proper function of various algorithms." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.386, + 0.789, + 0.612 + ], + "angle": 0, + "content": "Traditional image-based techniques often attempt to derive a clear image from a single blurred one. They do this by leveraging natural image priors or making assumptions about blur operations [6, 8, 13, 15, 34, 44]. However, in capturing a single blurred image, much motion information is lost. This lack of motion information during exposure creates undesirable artifacts in the deblurred results, especially in complex motion scenarios, as illustrated in Figures 1c and 1d. To counter this challenge, recent methodologies use more temporal information, such as differently exposed frames [2, 46, 48] and neighboring frames [16, 28, 32]. However, these approaches have their limitations. For instance, the use of differently exposed frames is constrained by the low temporal resolution of frame-based cameras, improving the reconstruction effect primarily in slow-motion scenes. The method that uses adjacent frames to estimate the current frame's motion information can lead to accumulative errors. In summary, traditional deblurring techniques encounter hurdles due to the inability of frame-based cameras to capture motion information during exposure." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.789, + 0.841 + ], + "angle": 0, + "content": "As opposed to frame-based cameras, event cameras, inspired by biological systems, offer a promising solution to this issue. These cameras' microsecond-level temporal resolution \\((\\mu \\mathrm{s})\\) and asynchronous output architecture allow the recording of fast motions. This capability enables event streams to store pixel intensity changes and provide accurate motion information to deblur the corresponding frame. By using event frames as additional input, some methods attempt to guide the deblurring process through these event cameras [9,25,39,43,49]. One method [45] subdivides the event stream into ternary 2D data and generates a mask to localize high-blur regions, thereby enhancing the network's ability to distinguish different blur levels. Another method [33] symmetrically accumulates event polarities and fuses event and image features using an attention mechanism, enabling the integration of motion information into image reconstruction. However, these methods primarily employ the event stream in a manner akin to polarity accumulation. In essence, they consider only the final accumulation result of relative intensity change between subsequent moments, neglecting the ab" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.322, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Motion Aware Event Representation-driven Image Deblurring" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.207 + ], + "angle": 0, + "content": "solute intensity change caused by historical events. Consequently, this approach fails to differentiate between various motion patterns, injecting unnecessary ambiguity into the deblurring process, and leading to edge artifacts as displayed in Figure 1e." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.208, + 0.789, + 0.569 + ], + "angle": 0, + "content": "Given the limited perception of motion during exposure offered by polarity accumulation, we posit that the deviation from the initial moment provides more precise motion information for image deblurring. Each time the event generates, the absolute light intensity changes. The deviation from the initial moment enables us to consider the impact of changes in absolute light intensity, even when the absolute light intensity at the initial moment is unknown. Thus, the deviation, to a degree, implies the order of arrival of the events, offering different information for varying motion patterns. As a result, we propose the Deviation Accumulation (DA) method to enhance motion perception. Specifically, every time a new event arrives, our method accumulates the deviation relative to the initial moment, thus encompassing more comprehensive motion information. Furthermore, we find that longer time scales for event accumulation can capture global motion patterns, while shorter scales precisely depict local fast-motion patterns. To better perceive motion at both global and local time scales, we have designed a Recurrent Motion Extraction (RME) module that works with corresponding event representations. Moreover, in line with multimodal learning [18-20, 29, 36, 42], we present an effective Feature Alignment and Fusion (FAF) module. This module serves to address the issue that simple fusion methods struggle with the disparate nature of event and frame modalities. As a result, these two modalities align, ensuring better information interaction. We integrate these modules into a single image deblurring network. This network extracts the motion features within the exposure time from the event stream and combines them with the texture information supplied by the image to complete the motion blur removal, as illustrated in Figure 1f." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.57, + 0.602, + 0.584 + ], + "angle": 0, + "content": "The key contributions of this paper are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.593, + 0.787, + 0.651 + ], + "angle": 0, + "content": "- We introduce an event preprocessing method designed specifically for the deblurring problem. The deviation accumulation is capable of responding distinctively to various motion patterns, thereby adding accurate motion information for image deblurring tasks." + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.652, + 0.787, + 0.711 + ], + "angle": 0, + "content": "- We develop a network equipped with recurrent motion extraction modules, designed to better extract motion features on local and global time scales. This design enables the network to handle multiple composite motion scenarios effectively." + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.712, + 0.787, + 0.757 + ], + "angle": 0, + "content": "- We propose a feature alignment and fusion module to mitigate the impacts of inter-modal inconsistencies and ensure that motion information more effectively guides the image deblurring process." + }, + { + "type": "list", + "bbox": [ + 0.228, + 0.593, + 0.787, + 0.757 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Our experimental results indicate that our method achieves state-of-the-art performance on synthetic, semi-synthetic, and real-world datasets, including a 0.61dB improvement in PSNR on the GoPro dataset compared to previous studies [33]. Moreover, our method produces more distinct subjective results, as demonstrated in Figure 1." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.128 + ], + "angle": 0, + "content": "Z. Sun et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.145, + 0.389, + 0.161 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.177, + 0.527, + 0.194 + ], + "angle": 0, + "content": "2.1 Event-based Motion Deblurring" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.203, + 0.789, + 0.459 + ], + "angle": 0, + "content": "The utilization of event information to assist in deblurring tasks has become increasingly common in recent research [9, 12, 25, 31, 39, 43, 49]. The high temporal resolution and low latency of event cameras not only capture edge positions but also embed moving temporal information within the event stream. This information is crucial to reconstruct the latent sharp image. Many methods have been attempted to explore the relationship between event information, blurred images, and sharp images. One such method [25] includes the Event-based Double Integral (EDI) model, which leverages the principles of event generation to formulate an optimization problem that establishes a correlation between a blurry image and a latent sharp image. In recent studies, the event stream is typically converted into event frames [30, 38], which simplifies the processing of event data within traditional CNN frameworks. [9] combines the motion feature extracted from event data with the image feature, using them as the input for the decoder. [50] employs a Multi-Layer Perceptron (MLP) to predict the fused feature value from event and image features. [33] integrates them through a cross-modal attention method, which generates a query matrix from the image features, but derives key and value matrices from the event features." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.46, + 0.79, + 0.551 + ], + "angle": 0, + "content": "Confronting the prevalent issue of integrating image and event features is the problem of inter-modal inconsistencies, which traditional fusion techniques cannot easily mitigate. To address this, our approach seeks to align these two modalities, optimizing the use of event data. Through this alignment, our goal is to minimize the inconsistencies spawned when merging differing modalities, thus amplifying the efficacy of the deblurring process." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.573, + 0.442, + 0.589 + ], + "angle": 0, + "content": "2.2 Event Representation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.789, + 0.81 + ], + "angle": 0, + "content": "Event data, characterized by its spatial sparsity and strong temporal continuity due to its asynchronous generation, necessitates a custom-designed event representation to deliver the required information. Methods [11, 26, 41] that process events individually maximize the use of high temporal resolution, as suggested by [26], which proposed an adaptive spiking neuron model to manage the fluctuating event input. The effectiveness of these algorithms, however, can be significantly hampered by the noise within the event due to independent processing. A Time Surface [1, 14, 22, 51] is created using the timestamp of the last event at each pixel, which prominently reveals the rich temporal information contained within the event. Furthermore, the event frame [5, 7, 21, 23, 45] transforms the original 4D event data into a 2D frame by accumulating events on the number or the polarity pixel-wise, facilitating extraction with conventional image processing techniques. In [45], the authors utilize ternary data for their event frames to allocate high blur region of the image." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.81, + 0.789, + 0.84 + ], + "angle": 0, + "content": "However, existing event frame approaches result in abundant light intensity change information being lost in the squeezing process, making it impossible to" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.322, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "Motion Aware Event Representation-driven Image Deblurring" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.224 + ], + "angle": 0, + "content": "distinguish the order of event arrivals and lacking the ability to detect multiple motion patterns. Some techniques attempt to recover some of this lost data by dividing the exposure time into smaller segments, though this essentially equates to sampling the original data. Consequently, we design a motion-aware event representation specifically tailored for motion deblurring." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.25, + 0.331, + 0.266 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.285, + 0.421, + 0.3 + ], + "angle": 0, + "content": "3.1 Overall Framework" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.314, + 0.789, + 0.542 + ], + "angle": 0, + "content": "The overall framework of our proposed method is depicted in Figure 2. We adopt a classical encoder-decoder architecture for our approach. Initially, the blurred image is inputted into the image branch, where we use the baseline conv block mentioned in [3] as our fundamental block, to extract relevant features. Simultaneously, the corresponding event frames generated by our deviation accumulation are fed into the Recurrent Motion Extraction (RME) module, which facilitates the extraction of features covering motion in local and global time scales. Following each block of the two branches, the Feature Alignment and Fusion (FAF) module is employed to eliminate the inconsistencies between the two modalities. By performing down-sampling operations, we are able to obtain both shallow and semantic information. In the decoder, skip connections are utilized to incorporate the extracted features for image reconstruction. The architecture of the conv block and the decoder is shown in the supplementary materials. The entire training process is conducted in an end-to-end manner, with the PSNR Loss function [4] employed to optimize the network's parameters." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.567, + 0.646, + 0.582 + ], + "angle": 0, + "content": "3.2 Deviation Accumulation Event Representation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.596, + 0.788, + 0.642 + ], + "angle": 0, + "content": "Event cameras are bio-inspired sensors which asynchronously output signals. They trigger an event while the logarithm of the intensity change exceeds the preset threshold \\( c \\) at any pixel, which is formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.323, + 0.655, + 0.786, + 0.695 + ], + "angle": 0, + "content": "\\[\np = \\left\\{ \\begin{array}{l l} + 1 & \\text {i f} \\log \\left(I _ {x y} (t)\\right) - \\log \\left(I _ {x y} (t - \\Delta t)\\right) \\geq c, \\\\ - 1 & \\text {i f} \\log \\left(I _ {x y} (t)\\right) - \\log \\left(I _ {x y} (t - \\Delta t)\\right) \\leq - c, \\end{array} \\right. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.706, + 0.788, + 0.798 + ], + "angle": 0, + "content": "where \\( p \\) is the event's polarity, \\( I_{xy}(t) \\) and \\( I_{xy}(t - \\Delta t) \\) represent the intensity at the pixel coordinate \\( (x,y) \\) at time \\( t \\) and \\( t - \\Delta t \\), respectively. An event is a quaternion value \\( (x,y,t,p) \\) where \\( x,y \\) denotes the pixel coordinate, \\( t \\) represents the time of change, the polarity \\( p \\in \\{+1, -1\\} \\) indicates the increase and decrease. For a given blurry image \\( B \\), it can be formulated as the average of latent intensity images \\( I(t) \\) during the exposure time [25]:" + }, + { + "type": "equation", + "bbox": [ + 0.438, + 0.81, + 0.786, + 0.845 + ], + "angle": 0, + "content": "\\[\nB = \\frac {1}{T} \\int_ {0} ^ {T} I (t) d t, \\tag {2}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.128 + ], + "angle": 0, + "content": "Z. Sun et al." + }, + { + "type": "image", + "bbox": [ + 0.233, + 0.145, + 0.773, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.332, + 0.789, + 0.457 + ], + "angle": 0, + "content": "Fig. 2: Overview of the architecture of our proposed method for image deblurring. The blurred image and related event frames within the exposure time are fed into the corresponding network branches respectively. The core module of the event branch: (a) Recurrent Motion Extraction (RME) module processes event frames from longer time scales to shorter, merging old features into new features through weighted sum, which results in perceiving various motion patterns under different time ranges. The image branch extracts the features with conv block and fuses them with event features using (b) Feature Alignment and Fusion (FAF) module. FAF module eliminates the inter-modal inconsistencies in a bi-directional way ensuring better information interaction." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.487, + 0.789, + 0.52 + ], + "angle": 0, + "content": "where \\(T\\) is the exposure time. For simplicity of understanding, we now only consider the pixel at \\((x,y)\\) then combining Equations (1) and (2), we have" + }, + { + "type": "equation", + "bbox": [ + 0.333, + 0.527, + 0.787, + 0.601 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} B _ {x y} = \\frac {1}{T} \\int_ {0} ^ {T} \\exp (\\log (I _ {x y} (t))) d t (3) \\\\ = \\frac {1}{T} \\int_ {0} ^ {T} e x p (\\log (I _ {x y} (0)) + c \\int_ {0} ^ {t} p (s) d s) d t. (4) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.608, + 0.789, + 0.67 + ], + "angle": 0, + "content": "The internal part of the exponential operation is the logarithmic value of the light intensity of the pixel at time \\( t \\), while the external part of the exponential operation is the integration and averaging of each moment within the exposure time. Based on the event data, we can discretize Equation (4):" + }, + { + "type": "equation", + "bbox": [ + 0.352, + 0.681, + 0.787, + 0.77 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} B _ {x y} = \\frac {1}{N} \\sum_ {t ^ {\\prime} \\in \\mathcal {T}} \\exp \\left(\\log \\left(I _ {x y} (0)\\right) + c \\sum_ {t = 0} ^ {t ^ {\\prime}} p _ {t}\\right) (5) \\\\ = \\frac {I _ {x y} (0)}{N} \\sum_ {t ^ {\\prime} \\in \\mathcal {T}} e x p \\left(c \\sum_ {t = 0} ^ {t ^ {\\prime}} p _ {t}\\right), (6) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.81 + ], + "angle": 0, + "content": "where \\(\\mathcal{T} = \\{t_1, t_2, \\dots, t_N\\}\\) indicates the set of time events occurs, \\(N\\) is the number of events, and \\(p_t\\) means the polarity of the event at time \\(t\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.811, + 0.788, + 0.84 + ], + "angle": 0, + "content": "The aforementioned derivation prompts us to consider not only the polarity of the event but also the impact of the alteration in the absolute value after" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.322, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "Motion Aware Event Representation-driven Image Deblurring" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.147, + 0.542, + 0.163 + ], + "angle": 0, + "content": "Algorithm 1 Deviation Accumulation (DA)" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.167, + 0.786, + 0.209 + ], + "angle": 0, + "content": "Input: taking pixel at \\((x,y)\\) for an example, the events stream \\((x,y,t_i,p_i)\\) in which \\(i\\in [1,N]\\), deviation from the initial moment \\(D\\), accumulation of deviation \\(R\\), Output: the value of DA event representation \\(V\\)." + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.21, + 0.338, + 0.222 + ], + "angle": 0, + "content": "1: \\(D = 0,R = 0\\)" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.224, + 0.473, + 0.237 + ], + "angle": 0, + "content": "2: Counting the number of events \\(N\\);" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.238, + 0.378, + 0.251 + ], + "angle": 0, + "content": "3: for \\((i = 1\\) to \\(N)\\) do" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.252, + 0.338, + 0.265 + ], + "angle": 0, + "content": "4: \\(D = D + p_{i}\\)" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.266, + 0.338, + 0.277 + ], + "angle": 0, + "content": "5: \\(R = R + D\\)" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.279, + 0.301, + 0.291 + ], + "angle": 0, + "content": "6: end for" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.293, + 0.31, + 0.306 + ], + "angle": 0, + "content": "7: \\( V = R / N \\)" + }, + { + "type": "list", + "bbox": [ + 0.223, + 0.21, + 0.473, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.341, + 0.787, + 0.538 + ], + "angle": 0, + "content": "each accumulation when considering the problem of deblurring. However, since event data cannot provide \\( I_{xy}(0) \\), the absolute light intensity is not accessible. But we can indirectly capture this information through the deviation from the initial moment. Therefore, we design a new event representation method for the deblurring problem presented in Algorithm 1. First, we accumulate all the events in the exposure time, so we get a 2D matrix named frequency matrix representing the number of the event occurs at each pixel position. We deal with the events considering their polarities subsequently, recording the deviation at each pixel by an additional tensor. Upon the arrival of the subsequent event, we update the deviation according to polarity and then add the deviation to the corresponding position. In the end, by dividing the cumulative matrix by the frequency matrix, we obtain the average deviation change over the exposure period as the new event representation named Deviation Accumulation (DA)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.539, + 0.787, + 0.765 + ], + "angle": 0, + "content": "Traditional polarity accumulation methods directly accumulating event polarities lead to identical outputs whenever the number of positive and negative events within a given time frame \\(([0, T])\\) remains unchanged. Nevertheless, by considering the deviation brought by historical events, our method can generate distinct values for events arriving in different orders, thereby improving the ability to perceive motion patterns and providing more precise motion information for the deblurring procedure. For example, as illustrated in Figure 3, (a) shows event signals generated by a motion pattern which represents that the light intensity at that pixel has risen twice and then fallen twice while (d) shows another pattern. Facing these two patterns, (b) and (e) depict the changing process of the outputs of traditional method and our method after each event is accumulated. After processing all four events, traditional polarity accumulation gets the same output of 0 for both patterns, but our DA representation yields 1 and 0.5 respectively which contains richer motion information. More examples of motion patterns our method can distinguish are shown in supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.766, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Overall, the phenomenon that polarity accumulation methods represent some different motion patterns as the same value introduces ambiguity into the deblurring process. Our method has a better perception of the motion patterns and offers a more distinct motion edge than conventional approaches from the visual effect as shown in Figure 3 (c) and (f). Inspired by [33], directly segmenting" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.128 + ], + "angle": 0, + "content": "Z. Sun et al." + }, + { + "type": "image", + "bbox": [ + 0.264, + 0.149, + 0.412, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.418, + 0.146, + 0.596, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.264, + 0.258, + 0.412, + 0.361 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.419, + 0.256, + 0.594, + 0.361 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.619, + 0.148, + 0.738, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.619, + 0.257, + 0.738, + 0.361 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.373, + 0.788, + 0.43 + ], + "angle": 0, + "content": "Fig. 3: (a) and (d) are event signals of two different motion patterns. (b) shows the Polarity Accumulation (PA) can't distinguish (a) and (d) as it gets the same output of 0 after processing all events while (e) shows our Deviation Accumulation (DA) can. DA perceives more motion information than PA as shown in (c) and (f)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.455, + 0.788, + 0.562 + ], + "angle": 0, + "content": "the entire exposure time averagely is not suitable for motion deblurring, which focuses on capturing rapid motion but falls short in adequately modeling the continuous movement throughout the exposure. In this paper, we symmetrically accumulate events along the \\( T / 2 \\) axis using the DA method. Taking \\( [0,T / 2] \\) as an example, we accumulate events on different time scales. The longest accumulating range is \\( T / 2 \\), followed by \\( T / 3 \\), and the shortest is \\( T / 6 \\). A detailed schematic is provided in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.58, + 0.574, + 0.594 + ], + "angle": 0, + "content": "3.3 Recurrent Motion Extraction Module" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.601, + 0.788, + 0.721 + ], + "angle": 0, + "content": "The goal of the Recurrent Motion Extraction (RME) module is to cater to both global slow motion and local fast motion in different time scales. Based on our observations of the motion process and event representations, the focus of the event representations varies in different time scales. In shorter time intervals, the global slow motion is not significant in the event representation due to the smaller number of events triggered. On the contrary slow motion can be clearly recorded in longer intervals, but at the same time, the violent moving part will produce edge spreading artifact." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.722, + 0.788, + 0.798 + ], + "angle": 0, + "content": "To enable the network to accurately model the various in-homogeneous motions, we devised an RME module that incorporates a recurrent mechanism. As shown in Figure 2(a), the event features are fed into the processing module from the longest time scales to the shortest and we use \\( e1 \\), \\( e2 \\), and \\( e3 \\) to denote them. Formally, the RME module can be presented as" + }, + { + "type": "equation", + "bbox": [ + 0.315, + 0.804, + 0.787, + 0.845 + ], + "angle": 0, + "content": "\\[\ne _ {i} ^ {\\prime} = \\left\\{ \\begin{array}{l l} f _ {e x} \\left(e _ {i}\\right) & \\text {i f} i = 1, \\\\ \\left(1 - \\alpha\\right) \\times f _ {e x} \\left(e _ {i}\\right) + \\alpha \\times f _ {r e} \\left(e _ {i - 1} ^ {\\prime}\\right) & \\text {i f} i = 2, 3, \\end{array} \\right. \\tag {7}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.321, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "Motion Aware Event Representation-driven Image Deblurring" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.253 + ], + "angle": 0, + "content": "where \\( f_{ex}(\\cdot) \\) and \\( f_{re}(\\cdot) \\) are two baseline conv blocks, \\( f_{ex}(\\cdot) \\) is used to extract information from the current representation, \\( f_{re}(\\cdot) \\) is used to retain useful information from previous features, \\( \\alpha \\) is the learnable weight as a trade-off between two features. By sharing the parameters of \\( f_{ex}(\\cdot) \\) and \\( f_{re}(\\cdot) \\) for different time scales, we force the network to be able to perceive motions at varying time scales. Each input feature is symmetric on the time axis so that the priori information about the motion direction is simultaneously fed into the network." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.275, + 0.582, + 0.29 + ], + "angle": 0, + "content": "3.4 Feature Alignment and Fusion Module" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.3, + 0.789, + 0.573 + ], + "angle": 0, + "content": "Although the information captured by the event camera exhibits a high correlation with the frame camera during the equivalent exposure period, there is still an inevitable gap originating from the modal dissimilarities between them. This phenomenon is frequently mentioned in multi-modal learning [18,42], but in event-guided image restoration tasks people usually [9,33,45] fuse features from two modalities directly using convolutions or attention mechanisms. Therefore, we propose an effective Feature Alignment and Fusion (FAF) module to facilitate enhanced feature fusion in a more coherent environment. Concretely, as shown in Figure 2(b), this module consists of a feature alignment module and a multi-head attention module. We use the operations of multiplication and addition successively to merge the raw information of one modality into the changing process of the other modality, which fills the gap of inter-modal inconsistency. The aligned features are fed into the attention module. In each fusion process, attention plays a bi-directional role, we not only generate the Q matrix through the frame features and the K and V matrix through the event features but also use the event features to perform look-ups on the frame features. The weights of the attention module are shared so that we can fuse the motion information under a more consistent space. Therefore, FAF module can be expressed as" + }, + { + "type": "equation", + "bbox": [ + 0.228, + 0.583, + 0.785, + 0.6 + ], + "angle": 0, + "content": "\\[\nx ^ {\\prime} = \\operatorname {C o n v} _ {2} (x) + (\\operatorname {C o n v} _ {1} (x) \\times e + e), \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.229, + 0.605, + 0.785, + 0.621 + ], + "angle": 0, + "content": "\\[\ne ^ {\\prime} = \\operatorname {C o n v} _ {2} (e) + (\\operatorname {C o n v} _ {1} (e) \\times x + x), \\tag {9}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.229, + 0.625, + 0.785, + 0.641 + ], + "angle": 0, + "content": "\\[\nF = \\operatorname {C o n v} _ {3} (\\operatorname {C o n c a t} ([ \\text {A t t e n t i o n} (Q _ {x ^ {\\prime}}, K _ {e ^ {\\prime}}, V _ {e ^ {\\prime}}), \\text {A t t e n t i o n} (Q _ {e ^ {\\prime}}, K _ {x ^ {\\prime}}, V _ {x ^ {\\prime}}) ])), \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.654, + 0.788, + 0.701 + ], + "angle": 0, + "content": "where \\( x' \\) and \\( e' \\) are aligned image features \\( x \\) and event feature \\( e \\) respectively, Attention(\\cdot) denotes the multi-head attention operation, Concat(\\cdot) denotes the concatenation operation along the channel dimension." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.723, + 0.506, + 0.74 + ], + "angle": 0, + "content": "4 Experiments and Analysis" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.755, + 0.448, + 0.77 + ], + "angle": 0, + "content": "4.1 Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.784, + 0.809 + ], + "angle": 0, + "content": "Datasets. We use three different datasets containing synthetic, semi-synthetic, and real-world frames and events to evaluate the proposed method." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.786, + 0.84 + ], + "angle": 0, + "content": "GoPro: We train the network on the GoPro dataset which is the benchmark dataset for the image motion deblurring [24]. The blurry image is averaged from" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.127 + ], + "angle": 0, + "content": "Z. Sun et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.187 + ], + "angle": 0, + "content": "Table 1: The quantitative results on GoPro, HS-ERGB, and REBlur test datasets. The best values are highlighted in bold. HINet+ and UFPNet+ are event-enhanced versions of HINet and UFPNet by concatenating our DA to their input." + }, + { + "type": "table", + "bbox": [ + 0.229, + 0.2, + 0.772, + 0.424 + ], + "angle": 0, + "content": "
MethodInputGoProHS-ERGBREBlurParams(M)
PSNRSSIMPSNRSSIMPSNRSSIM
LEVS [10]F20.840.54722.130.555--18.2
EDI [25]F+E29.060.94323.930.70436.520.9640.5
SRN [35]F30.260.934--35.100.96110.2
EVDI [49]F+E30.400.90625.130.707--0.4
HINet [4]F32.710.95927.320.80735.580.96588.7
Restormer [47]F32.920.96127.550.80835.500.95926.1
MSDI-Net [17]F33.280.96427.460.80936.140.968241.3
NAFNet [3]F33.710.96727.640.81136.150.96967.8
UFPNet [6]F34.060.96827.640.80936.110.96880.3
HINet+ [4]F+E34.630.96827.660.80837.920.97688.7
UFPNet+ [6]F+E35.220.97227.680.80937.970.97680.3
EFNet [33]F+E35.460.97226.680.80038.120.9758.5
OursF+E36.070.97627.930.81238.470.97813.9
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.463, + 0.785, + 0.553 + ], + "angle": 0, + "content": "adjacent sharp frames, by doing so the dataset contains 3214 pairs of blurry and sharp images. According to the standard division, we adopt 2103 pairs for training and 1111 pairs for testing. For the event data, we use the GoPro raw event dataset provided by [33] which synthesizes events with a random threshold \\( c \\) following the Gaussian distribution by ESIM simulator [27]. Following the DA method, the raw event data is preprocessed into 6 event frames for each image." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.553, + 0.787, + 0.629 + ], + "angle": 0, + "content": "HS-ERGB: The HS-ERGB dataset [37] consists of sharp videos and real-world events, we use [50] released normal blur version which synthesizes blurry frames by averaging 49 interpolating images. Each blurry frame in the dataset has a known exposure time, allowing us to generate the DA representation for each frame within its respective exposure duration." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.787, + 0.704 + ], + "angle": 0, + "content": "REBlur: The REBlur dataset collects a number of sequences of real-world event data, corresponding with blurry images and reliable ground-truth sharp images [33]. It contains 12 kinds of linear and nonlinear motions. Following the original setting, we use 486 pairs of blurry-sharp images with associated events for training and 983 pairs for testing." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.789, + 0.84 + ], + "angle": 0, + "content": "Implementation details. We train the end-to-end network without pre-training on a single Tesla V100 GPU. We extract patches of size \\(256 \\times 256\\) for training images and corresponding event frames, and the batch size is set to 2 by default. The AdamW optimizer is employed (\\(\\beta_{1} = 0.9\\) and \\(\\beta_{2} = 0.99\\)) with the initial learning rate 0.0001 with the cosine annealing schedule where the \\(T_{max}\\) is 400K iteration. For data augmentation, we perform horizontal and vertical flips, the random noise and the hot pixels are added into event frames in order to simulate the real-world situation. For HS-ERGB and REBlur datasets, We fine-tune the model trained on GoPro with the training set of them. The fine-tuning process" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.322, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "Motion Aware Event Representation-driven Image Deblurring" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.261, + 0.147, + 0.432, + 0.279 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.262, + 0.28, + 0.43, + 0.291 + ], + "angle": 0, + "content": "(a) Blur image in GoPro testset" + }, + { + "type": "image", + "bbox": [ + 0.436, + 0.147, + 0.51, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.442, + 0.207, + 0.498, + 0.218 + ], + "angle": 0, + "content": "(b) HINet" + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.147, + 0.588, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.516, + 0.207, + 0.581, + 0.218 + ], + "angle": 0, + "content": "(c) UFPNet" + }, + { + "type": "image", + "bbox": [ + 0.591, + 0.147, + 0.665, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.596, + 0.207, + 0.653, + 0.218 + ], + "angle": 0, + "content": "(d) EFNet" + }, + { + "type": "image", + "bbox": [ + 0.668, + 0.147, + 0.743, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.671, + 0.207, + 0.739, + 0.218 + ], + "angle": 0, + "content": "(e) NAFNet" + }, + { + "type": "image_caption", + "bbox": [ + 0.262, + 0.279, + 0.43, + 0.291 + ], + "angle": 0, + "content": "(a) Blur image in GoPro testset" + }, + { + "type": "image", + "bbox": [ + 0.436, + 0.222, + 0.51, + 0.279 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.442, + 0.28, + 0.501, + 0.291 + ], + "angle": 0, + "content": "(f) HINet+" + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.222, + 0.588, + 0.279 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.515, + 0.28, + 0.585, + 0.291 + ], + "angle": 0, + "content": "(g) UFPNet+" + }, + { + "type": "image", + "bbox": [ + 0.591, + 0.222, + 0.664, + 0.279 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.605, + 0.28, + 0.652, + 0.291 + ], + "angle": 0, + "content": "(h) Ours" + }, + { + "type": "image", + "bbox": [ + 0.669, + 0.222, + 0.744, + 0.279 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.688, + 0.28, + 0.725, + 0.291 + ], + "angle": 0, + "content": "(i) GT" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.304, + 0.785, + 0.331 + ], + "angle": 0, + "content": "Fig. 4: Qualitative comparisons on the GoPro dataset. The notation is the same as in Table 1. Zoom in for better view." + }, + { + "type": "image", + "bbox": [ + 0.262, + 0.349, + 0.433, + 0.48 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.263, + 0.481, + 0.434, + 0.493 + ], + "angle": 0, + "content": "(a) Blur image in REBlur testset" + }, + { + "type": "image", + "bbox": [ + 0.436, + 0.349, + 0.511, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.442, + 0.409, + 0.498, + 0.42 + ], + "angle": 0, + "content": "(b) HINet" + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.349, + 0.588, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.517, + 0.409, + 0.582, + 0.42 + ], + "angle": 0, + "content": "(c) UFPNet" + }, + { + "type": "image", + "bbox": [ + 0.591, + 0.349, + 0.666, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.597, + 0.409, + 0.654, + 0.42 + ], + "angle": 0, + "content": "(d) EFNet" + }, + { + "type": "image", + "bbox": [ + 0.669, + 0.349, + 0.743, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.672, + 0.409, + 0.739, + 0.42 + ], + "angle": 0, + "content": "(e) NAFNet" + }, + { + "type": "image_caption", + "bbox": [ + 0.262, + 0.499, + 0.787, + 0.534 + ], + "angle": 0, + "content": ": Qualitative comparisons on the REBlur dataset. The notation is the same as Table 1. Zoom in for better view." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.564, + 0.785, + 0.595 + ], + "angle": 0, + "content": "consisted of 4000 iterations while keeping the other configurations consistent with the previous experiments." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.618, + 0.62, + 0.634 + ], + "angle": 0, + "content": "4.2 Comparison with State-of-the-Art Methods" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Table 1 shows the comparison results of our method and other SOTA methods on GoPro [24], HS-ERGB [50] and REBlur [33]. We compare our network with not only event-based methods, but also with some SOTA image-only deblurring networks enhancing with event input. The results illustrate that: (1) Our method outperforms all other methods in synthetic, semi-synthetic and real-world datasets, which indicates that with the proper utilization of events the designed approach is able to handle blur with different distributions. Specifically, the average gains of our method over the second-best one are 0.61dB, 0.25dB and 0.35dB in terms of PSNR on GoPro, HS-ERGB and REBlur datasets, respectively. (2) Through the enhancement of our proposed event representation, the HINet+ and the UFPNet+ increase 1.92dB and 1.16dB in PSNR on the benchmark GoPro dataset compared with their image-only version, respectively. Taking HINet as an example, with the same operation of [33], HINet with DA" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.127 + ], + "angle": 0, + "content": "Z. Sun et al." + }, + { + "type": "image", + "bbox": [ + 0.268, + 0.144, + 0.344, + 0.203 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.283, + 0.204, + 0.328, + 0.216 + ], + "angle": 0, + "content": "(a) GT" + }, + { + "type": "image", + "bbox": [ + 0.346, + 0.144, + 0.422, + 0.203 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.352, + 0.204, + 0.417, + 0.216 + ], + "angle": 0, + "content": "(b) Blurry" + }, + { + "type": "image", + "bbox": [ + 0.425, + 0.144, + 0.501, + 0.203 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.428, + 0.204, + 0.5, + 0.216 + ], + "angle": 0, + "content": "(c) UFPNet" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.144, + 0.58, + 0.203 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.505, + 0.204, + 0.578, + 0.216 + ], + "angle": 0, + "content": "(d) NAFNet" + }, + { + "type": "image", + "bbox": [ + 0.582, + 0.144, + 0.658, + 0.203 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.589, + 0.204, + 0.652, + 0.216 + ], + "angle": 0, + "content": "(e) EFNet" + }, + { + "type": "image", + "bbox": [ + 0.66, + 0.144, + 0.736, + 0.203 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.673, + 0.204, + 0.725, + 0.216 + ], + "angle": 0, + "content": "(f) Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.305, + 0.228, + 0.696, + 0.242 + ], + "angle": 0, + "content": "Fig. 6: Qualitative comparisons on the HS-ERGB dataset." + }, + { + "type": "image", + "bbox": [ + 0.269, + 0.257, + 0.383, + 0.347 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.269, + 0.349, + 0.382, + 0.435 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.299, + 0.436, + 0.352, + 0.447 + ], + "angle": 0, + "content": "(a) Blurry" + }, + { + "type": "image", + "bbox": [ + 0.385, + 0.258, + 0.501, + 0.347 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.386, + 0.349, + 0.501, + 0.434 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.405, + 0.436, + 0.481, + 0.447 + ], + "angle": 0, + "content": "(b) SCER [33]" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.258, + 0.618, + 0.347 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.349, + 0.618, + 0.435 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.524, + 0.436, + 0.596, + 0.447 + ], + "angle": 0, + "content": "(c) DA (Ours)" + }, + { + "type": "image", + "bbox": [ + 0.621, + 0.258, + 0.735, + 0.347 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.621, + 0.349, + 0.735, + 0.435 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.659, + 0.436, + 0.697, + 0.447 + ], + "angle": 0, + "content": "(d) GT" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.459, + 0.785, + 0.487 + ], + "angle": 0, + "content": "Fig. 7: Visual comparisons on GoPro dataset with different event representation. Compared to SCER representation, our DA representation reduces edge artifacts." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.518, + 0.785, + 0.547 + ], + "angle": 0, + "content": "event representation exceeds 0.94dB over the result in [33], which presents the fact that DA extracts more conducive information for motion deblurring." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.549, + 0.785, + 0.639 + ], + "angle": 0, + "content": "We also show the visual comparison results to verify the effectiveness of our method in Figure 4, Figure 5 and Figure 6 with representative samples from each dataset. As illustrated in the figures, our model reconstructs better details and has minor spatial distortions compared with other methods. This indicates that our event representation perceives more motion information and the modules effectively embed it to reconstruct a high-quality image with enhanced details." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.663, + 0.401, + 0.677 + ], + "angle": 0, + "content": "4.3 Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.785, + 0.719 + ], + "angle": 0, + "content": "We conduct plenty ablation experiments on GoPro dataset to verify the effectiveness of our event representation and proposed method." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Effectiveness of Event Representation. To demonstrate the adaptability of our event representation for motion deblurring tasks, as shown in Table 2, we replace our DA representation with several other types of representations and trained the network accordingly. The results reveal certain shortcomings of the alternative representations. Firstly, the stack representation simply accumulates all events on their polarities which leaves out the temporal information. Secondly, the SBT approach [40], which subdivides the entire exposure time equally, proves to be inadequate for modeling the varying degrees of blur. Thirdly, the SCER" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.322, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Motion Aware Event Representation-driven Image Deblurring" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.339, + 0.145, + 0.663, + 0.159 + ], + "angle": 0, + "content": "Table 2: Ablation study of each module effects." + }, + { + "type": "table", + "bbox": [ + 0.242, + 0.172, + 0.758, + 0.34 + ], + "angle": 0, + "content": "
Event RepresentationRMEFusionPSNRSSIM
Stack33.240.953
SBT [40]✓FAF(Ours)35.630.973
SCER [33]35.810.975
Concat.34.020.967
Add34.740.970
Attention35.750.974
ACA(Ours)✓Only Align Event+Attention35.780.975
Only Align Image+Attention35.740.974
35.420.973
ACA(Ours)✗FAF(Ours)36.070.976
✓
" + }, + { + "type": "image", + "bbox": [ + 0.249, + 0.354, + 0.416, + 0.428 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.293, + 0.429, + 0.375, + 0.44 + ], + "angle": 0, + "content": "(a) w/o RME" + }, + { + "type": "image", + "bbox": [ + 0.418, + 0.354, + 0.586, + 0.428 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.461, + 0.429, + 0.545, + 0.44 + ], + "angle": 0, + "content": "(b) w/o RME" + }, + { + "type": "image", + "bbox": [ + 0.588, + 0.354, + 0.755, + 0.428 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.631, + 0.429, + 0.712, + 0.44 + ], + "angle": 0, + "content": "(c) w/o RME" + }, + { + "type": "image", + "bbox": [ + 0.25, + 0.441, + 0.416, + 0.514 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.296, + 0.515, + 0.371, + 0.527 + ], + "angle": 0, + "content": "(d) w/RME" + }, + { + "type": "image", + "bbox": [ + 0.418, + 0.441, + 0.585, + 0.514 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.465, + 0.515, + 0.54, + 0.527 + ], + "angle": 0, + "content": "(e) w/RME" + }, + { + "type": "image", + "bbox": [ + 0.587, + 0.441, + 0.754, + 0.514 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.634, + 0.515, + 0.708, + 0.527 + ], + "angle": 0, + "content": "(f) w/RME" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.539, + 0.788, + 0.581 + ], + "angle": 0, + "content": "Fig. 8: Visual comparison results of w/ and w/o RME module. All of these scenarios involve movement in different directions and speeds, RME module can successfully capture motion features in these complex motion scenarios. Zoom in for better view." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.612, + 0.788, + 0.734 + ], + "angle": 0, + "content": "method [33], which directly accumulates event polarities, is found to be incapable of capturing the influence of absolute intensity changes. This limitation is illustrated in Figure 3. In contrast, our DA representation not only possesses a solid physical foundation for event generation but also perceives more motion patterns. As a result, it achieves a 0.26dB improvement in performance compared to the best alternative representation. We also provide the visual comparison on the image reconstruction results in Figure 7. Our DA representation reduces motion-generated edge artifacts, resulting in a better deblurring effect." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Effectiveness of Recurrent Motion Extraction Module. To assess the RME module's contribution, we replace it with a baseline conv block and adjust the parameters for fair comparison. In this approach, we concatenate event frames of different time scales and feed them into the conv block. The findings, as presented in the last two rows of Table 2, clearly indicate that the contribution of \\(0.65\\mathrm{dB}\\) increment in PSNR stems from the network structure design, rather than an increase in parameter quantity. This shows the importance of incorpo" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.127 + ], + "angle": 0, + "content": "Z. Sun et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.207 + ], + "angle": 0, + "content": "rating the module for effectively leveraging motion information in multi-timescales and improving performance in motion deblurring tasks. We also present the visual comparisons when removing the RME module in Figure 8, the RME module adeptly captures both fast local and slow global motion characteristics." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.208, + 0.788, + 0.359 + ], + "angle": 0, + "content": "Effectiveness of Feature Alignment and Fusion Module. To illustrate the effectiveness of our FAF module, we compare it with several common fusion methods, including concatenation, addition, and cross-modality attention. As shown in Table 2, it is arduous to utilize the motion information contained in event feature through the direct concatenation or addition. While attention mechanisms can capture relative parts in the two features, the inherent inconsistencies within the two modalities still lead to a 0.32dB performance gap when compared to our FAF module. We also use a t-SNE visualization in Figure 9 to demonstrate the effectiveness of our feature alignment, where two originally in-congruent modal features are mixed up together after the FAF module." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.36, + 0.489, + 0.631 + ], + "angle": 0, + "content": "Furthermore, we also verify the importance of bi-directional alignment by conducting experiments solely on aligning event or image features. Since both events and frames provide crucial information for image restoration, such as edge motion information and textual information, respectively, it is unreasonable to let one modality dominate the alignment process. As shown in Table 2, the bi-directional alignment outperforms the highest score achieved by one-way correction by 0.29dB. This demonstrates the necessity of fully utilizing the mutual information between the two modalities in order to achieve better image reconstruction results." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.375, + 0.771, + 0.546 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.546, + 0.788, + 0.629 + ], + "angle": 0, + "content": "Fig. 9: t-SNE visualization of original and aligned features. The distributions of the aligned two features overlap more compared to the original features. This indicates the effectiveness of FAF module to reduce the inter-modal inconsistency." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.656, + 0.36, + 0.672 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In this paper, we propose an event-based image deblurring method. Firstly, we rethink the limit of traditional event preprocessing method and propose a new deviation accumulation (DA) technique. With deviation, the new approach can distinguish the order of events and perceive more motion patterns. Moreover, we design an image deblurring network equipped with recurrent motion extraction (RME) module and feature alignment and fusion (FAF) module to fit our proposed event representation. The RME module can better perceive motions in different time scales, and the FAF module promotes better integration of the two modalities. Finally, extensive experimental results over several datasets demonstrate the effectiveness of the proposed algorithm." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.322, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Motion Aware Event Representation-driven Image Deblurring" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.177 + ], + "angle": 0, + "content": "Acknowledgements. This work was supported by National Natural Science Foundation of China (NSFC) under Grants 62225207 and 62276243." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.203, + 0.323, + 0.219 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.238, + 0.785, + 0.279 + ], + "angle": 0, + "content": "1. Benosman, R., Clercq, C., Lagorce, X., Ieng, S.H., Bartolozzi, C.: Event-based visual flow. IEEE transactions on neural networks and learning systems 25(2), 407-417 (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.281, + 0.785, + 0.308 + ], + "angle": 0, + "content": "2. Chang, M., Feng, H., Xu, Z., Li, Q.: Low-light image restoration with short-and long-exposure raw pairs. IEEE Transactions on Multimedia 24, 702-714 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.309, + 0.785, + 0.35 + ], + "angle": 0, + "content": "3. Chen, L., Chu, X., Zhang, X., Sun, J.: Simple baselines for image restoration. In: Proceedings of the European conference on computer vision (ECCV). pp. 17-33. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.351, + 0.785, + 0.392 + ], + "angle": 0, + "content": "4. Chen, L., Lu, X., Zhang, J., Chu, X., Chen, C.: Hinet: Half instance normalization network for image restoration. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 182-192 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.393, + 0.785, + 0.434 + ], + "angle": 0, + "content": "5. Cook, M., Gugelmann, L., Jug, F., Krautz, C., Steger, A.: Interacting maps for fast visual interpretation. In: The 2011 International Joint Conference on Neural Networks. pp. 770-776. IEEE (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.435, + 0.785, + 0.49 + ], + "angle": 0, + "content": "6. Fang, Z., Wu, F., Dong, W., Li, X., Wu, J., Shi, G.: Self-supervised non-uniform kernel estimation with flow-based motion prior for blind image deblurring. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18105-18114 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.491, + 0.785, + 0.532 + ], + "angle": 0, + "content": "7. Gehrig, D., Rebecq, H., Gallego, G., Scaramuzza, D.: Eklt: Asynchronous photometric feature tracking using events and frames. International Journal of Computer Vision 128(3), 601-618 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.533, + 0.785, + 0.574 + ], + "angle": 0, + "content": "8. Ji, S.W., Lee, J., Kim, S.W., Hong, J.P., Baek, S.J., Jung, S.W., Ko, S.J.: Xydeblur: divide and conquer for single image deblurring. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17421-17430 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.575, + 0.785, + 0.616 + ], + "angle": 0, + "content": "9. Jiang, Z., Zhang, Y., Zou, D., Ren, J., Lv, J., Liu, Y.: Learning event-based motion deblurring. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3320-3329 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.617, + 0.785, + 0.658 + ], + "angle": 0, + "content": "10. Jin, M., Meishvili, G., Favaro, P.: Learning to extract a video sequence from a single motion-blurred image. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 6334-6342 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.659, + 0.785, + 0.687 + ], + "angle": 0, + "content": "1. Kim, H., Handa, A., Benosman, R., Ieng, S.H., Davison, A.J.: Simultaneous mosaicing and tracking with an event camera. J. Solid State Circ 43, 566-576 (2008)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.688, + 0.785, + 0.728 + ], + "angle": 0, + "content": "2. Kim, T., Lee, J., Wang, L., Yoon, K.J.: Event-guided deblurring of unknown exposure time videos. In: Proceedings of the European conference on computer vision (ECCV). pp. 519-538. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.729, + 0.785, + 0.756 + ], + "angle": 0, + "content": "3. Krishnan, D., Tay, T., Fergus, R.: Blind deconvolution using a normalized sparsity measure. In: CVPR 2011. pp. 233-240. IEEE (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.757, + 0.785, + 0.799 + ], + "angle": 0, + "content": "4. Lagorce, X., Orchard, G., Galluppi, F., Shi, B.E., Benosman, R.B.: Hots: a hierarchy of event-based time-surfaces for pattern recognition. IEEE transactions on pattern analysis and machine intelligence 39(7), 1346-1359 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "5. Levin, A., Weiss, Y., Durand, F., Freeman, W.T.: Understanding and evaluating blind deconvolution algorithms. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 1964-1971. IEEE (2009)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.238, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.127 + ], + "angle": 0, + "content": "Z. Sun et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.203 + ], + "angle": 0, + "content": "16. Li, D., Shi, X., Zhang, Y., Cheung, K.C., See, S., Wang, X., Qin, H., Li, H.: A simple baseline for video restoration with grouped spatial-temporal shift. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9822-9832 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.204, + 0.787, + 0.245 + ], + "angle": 0, + "content": "17. Li, D., Zhang, Y., Cheung, K.C., Wang, X., Qin, H., Li, H.: Learning degradation representations for image deblurring. In: Proceedings of the European conference on computer vision (ECCV). pp. 736-753. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.246, + 0.787, + 0.3 + ], + "angle": 0, + "content": "18. Li, Y., Yu, A.W., Meng, T., Caine, B., Ngiam, J., Peng, D., Shen, J., Lu, Y., Zhou, D., Le, Q.V., et al.: Deepfusion: Lidar-camera deep fusion for multi-modal 3d object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17182-17191 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.301, + 0.787, + 0.342 + ], + "angle": 0, + "content": "19. Liang, M., Yang, B., Wang, S., Urtasun, R.: Deep continuous fusion for multi-sensor 3d object detection. In: Proceedings of the European conference on computer vision (ECCV). pp. 641-656 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.343, + 0.787, + 0.384 + ], + "angle": 0, + "content": "20. Liang, X., Qian, Y., Guo, Q., Cheng, H., Liang, J.: Af: An association-based fusion method for multi-modal classification. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(12), 9236-9254 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.787, + 0.411 + ], + "angle": 0, + "content": "21. Liu, M., Delbruck, T.: Adaptive time-slice block-matching optical flow algorithm for dynamic vision sensors. BMVC (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.412, + 0.787, + 0.467 + ], + "angle": 0, + "content": "22. Manderscheid, J., Sironi, A., Bourdis, N., Migliore, D., Lepetit, V.: Speed invariant time surface for learning to detect corner points with event-based cameras. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10245-10254 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.468, + 0.787, + 0.522 + ], + "angle": 0, + "content": "23. Maqueda, A.I., Loquercio, A., Gallego, G., García, N., Scaramuzza, D.: Event-based vision meets deep learning on steering prediction for self-driving cars. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5419-5427 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.523, + 0.787, + 0.563 + ], + "angle": 0, + "content": "24. Nah, S., Hyun Kim, T., Mu Lee, K.: Deep multi-scale convolutional neural network for dynamic scene deblurring. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3883-3891 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.565, + 0.787, + 0.619 + ], + "angle": 0, + "content": "25. Pan, L., Scheerlinck, C., Yu, X., Hartley, R., Liu, M., Dai, Y.: Bringing a blurry frame alive at high frame-rate with an event camera. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6820-6829 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.62, + 0.787, + 0.674 + ], + "angle": 0, + "content": "26. Paredes-Vallés, F., Scheper, K.Y., De Croon, G.C.: Unsupervised learning of a hierarchical spiking neural network for optical flow estimation: From events to global motion perception. IEEE transactions on pattern analysis and machine intelligence 42(8), 2051-2064 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.675, + 0.787, + 0.702 + ], + "angle": 0, + "content": "27. Rebecq, H., Gehrig, D., Scaramuzza, D.: Esim: an open event camera simulator. In: Conference on robot learning. pp. 969-982. PMLR (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.703, + 0.787, + 0.73 + ], + "angle": 0, + "content": "28. Ren, D., Shang, W., Yang, Y., Zuo, W.: Aggregating long-term sharp features via hybrid transformers for video deblurring. arXiv preprint arXiv:2309.07054 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.731, + 0.787, + 0.771 + ], + "angle": 0, + "content": "29. Ren, S., Du, Y., Lv, J., Han, G., He, S.: Learning from the master: Distilling cross-modal advanced knowledge for lip reading. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13325-13333 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.772, + 0.787, + 0.813 + ], + "angle": 0, + "content": "30. Shang, W., Ren, D., Zou, D., Ren, J.S., Luo, P., Zuo, W.: Bringing events into video deblurring with non-consecutively blurry frames. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4531-4540 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.814, + 0.787, + 0.84 + ], + "angle": 0, + "content": "31. Song, C., Bajaj, C., Huang, Q.: Deblursr: Event-based motion deblurring under the spiking representation. arXiv preprint arXiv:2303.08977 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.322, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Motion Aware Event Representation-driven Image Deblurring" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "32. Su, S., Delbracio, M., Wang, J., Sapiro, G., Heidrich, W., Wang, O.: Deep video deblurring for hand-held cameras. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1279-1288 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.788, + 0.245 + ], + "angle": 0, + "content": "33. Sun, L., Sakaridis, C., Liang, J., Jiang, Q., Yang, K., Sun, P., Ye, Y., Wang, K., Gool, L.V.: Event-based fusion for motion deblurring with cross-modal attention. In: Proceedings of the European conference on computer vision (ECCV). pp. 412-428. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.246, + 0.788, + 0.286 + ], + "angle": 0, + "content": "34. Sun, L., Cho, S., Wang, J., Hays, J.: Edge-based blur kernel estimation using patch priors. In: IEEE international conference on computational photography (ICCP). pp. 1-8. IEEE (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.287, + 0.788, + 0.328 + ], + "angle": 0, + "content": "35. Tao, X., Gao, H., Shen, X., Wang, J., Jia, J.: Scale-recurrent network for deep image deblurring. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 8174-8182 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.328, + 0.788, + 0.37 + ], + "angle": 0, + "content": "36. Tian, Y., Krishnan, D., Isola, P.: Contrastive multiview coding. In: Proceedings of the European conference on computer vision (ECCV). pp. 776-794. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.371, + 0.788, + 0.424 + ], + "angle": 0, + "content": "37. Tulyakov, S., Gehrig, D., Georgoulis, S., Erbach, J., Gehrig, M., Li, Y., Scaramuzza, D.: Time lens: Event-based video frame interpolation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 16155-16164 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.425, + 0.788, + 0.467 + ], + "angle": 0, + "content": "38. Vitoria, P., Georgoulis, S., Tulyakov, S., Bochicchio, A., Erbach, J., Li, Y.: Event-based image deblurring with dynamic motion awareness. In: Proceedings of the European conference on computer vision (ECCV). pp. 95-112. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.468, + 0.788, + 0.509 + ], + "angle": 0, + "content": "39. Wang, B., He, J., Yu, L., Xia, G.S., Yang, W.: Event enhanced high-quality image recovery. In: Proceedings of the European conference on computer vision (ECCV). pp. 155-171. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.509, + 0.788, + 0.563 + ], + "angle": 0, + "content": "40. Wang, L., Ho, Y.S., Yoon, K.J., et al.: Event-based high dynamic range image and very high frame rate video generation using conditional generative adversarial networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10081-10090 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.564, + 0.788, + 0.605 + ], + "angle": 0, + "content": "41. Weikersdorfer, D., Conradt, J.: Event-based particle filtering for robot self-localization. In: 2012 IEEE International Conference on Robotics and Biomimetics (ROBIO). pp. 866-870. IEEE (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.606, + 0.788, + 0.633 + ], + "angle": 0, + "content": "42. Xia, W., Li, X., Deng, A., Xiong, H., Dou, D., Hu, D.: Robust cross-modal knowledge distillation for unconstrained videos. arXiv preprint arXiv:2304.07775 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.634, + 0.788, + 0.674 + ], + "angle": 0, + "content": "43. Xu, F., Yu, L., Wang, B., Yang, W., Xia, G.S., Jia, X., Qiao, Z., Liu, J.: Motion deblurring with real events. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2583-2592 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.675, + 0.788, + 0.715 + ], + "angle": 0, + "content": "44. Xu, L., Jia, J.: Two-phase kernel estimation for robust motion deblurring. In: Proceedings of the European conference on computer vision (ECCV). pp. 157-170. Springer (2010)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.716, + 0.788, + 0.757 + ], + "angle": 0, + "content": "45. Yang, D., Yamac, M.: Motion aware double attention network for dynamic scene deblurring. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1113-1123 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.758, + 0.788, + 0.785 + ], + "angle": 0, + "content": "46. Yuan, L., Sun, J., Quan, L., Shum, H.Y.: Image deblurring with blurred/noisy image pairs. In: ACM SIGGRAPH 2007 papers, pp. 1-es (2007)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.786, + 0.788, + 0.84 + ], + "angle": 0, + "content": "47. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H.: Restormer: Efficient transformer for high-resolution image restoration. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5728-5739 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.127 + ], + "angle": 0, + "content": "Z. Sun et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.175 + ], + "angle": 0, + "content": "48. Zhang, S., Zhen, A., Stevenson, R.L.: Deep motion blur removal using noisy/blurry image pairs. Journal of Electronic Imaging 30(3), 033022-033022 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.177, + 0.787, + 0.217 + ], + "angle": 0, + "content": "49. Zhang, X., Yu, L.: Unifying motion deblurring and frame interpolation with events. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17765-17774 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.218, + 0.787, + 0.259 + ], + "angle": 0, + "content": "50. Zhang, X., Yu, L., Yang, W., Liu, J., Xia, G.S.: Generalizing event-based motion deblurring in real-world scenarios. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10734-10744 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.26, + 0.787, + 0.301 + ], + "angle": 0, + "content": "51. Zhou, Y., Gallego, G., Rebecq, H., Kneip, L., Li, H., Scaramuzza, D.: Semi-dense 3d reconstruction with a stereo event camera. In: Proceedings of the European conference on computer vision (ECCV). pp. 235-251 (2018)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.301 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Motion Aware Event Representation-driven Image Deblurring/a9105a02-b537-456a-9778-80b745c423ab_origin.pdf b/2024/Motion Aware Event Representation-driven Image Deblurring/a9105a02-b537-456a-9778-80b745c423ab_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..1f4f976dedb51ea832a9e9942ad3ed111f0a48a6 --- /dev/null +++ b/2024/Motion Aware Event Representation-driven Image Deblurring/a9105a02-b537-456a-9778-80b745c423ab_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1c1fb906670531866b90c9b0e4be252cf62060042ee9805762d0c2c32d8cd87d +size 2934585 diff --git a/2024/Motion Aware Event Representation-driven Image Deblurring/full.md b/2024/Motion Aware Event Representation-driven Image Deblurring/full.md new file mode 100644 index 0000000000000000000000000000000000000000..18701821cf7f170b9b94312c336edb9d8825fb5b --- /dev/null +++ b/2024/Motion Aware Event Representation-driven Image Deblurring/full.md @@ -0,0 +1,389 @@ +# Motion Aware Event Representation-driven Image Deblurring + +Zhijing Sun, Xueyang Fu, Longzhuo Huang, Aiping Liu, and Zheng-Jun Zha* + +School of Information Science and Technology and MoE Key Laboratory of Brain-inspired Intelligent Perception and Cognition, University of Science and Technology of China, Hefei, 230026, China {sunzhijing, hlz_0}@mail.ustc.edu.cn {xyfu, aipingl, zhazj}@ustc.edu.cn + +Abstract. Traditional image deblurring struggles with high-quality reconstruction due to limited motion data from single blurred images. Excitingly, the high-temporal resolution of event cameras records motion more precisely in a different modality, transforming image deblurring. However, many event camera-based methods, which only care about the final value of the polarity accumulation, ignore the influence of the absolute intensity change where events generate so fall short in perceiving motion patterns and effectively aiding image reconstruction. To overcome this, in this work, we propose a new event preprocessing technique that accumulates the deviation from the initial moment each time the event is updated. This process can distinguish the order of events to improve the perception of object motion patterns. To complement our proposed event representation, we create a recurrent module designed to meticulously extract motion features across local and global time scales. To further facilitate the event feature and image feature integration, which assists in image reconstruction, we develop a bi-directional feature alignment and fusion module. This module works to lessen inter-modal inconsistencies. Our approach has been thoroughly tested through rigorous experiments carried out on several datasets with different distributions. These trials have delivered promising results, with our method achieving top-tier performance in both quantitative and qualitative assessments. Code is available at https://github.com/ZhijingS/DA_event_deblur. + +Keywords: Image deblurring $\cdot$ Event camera $\cdot$ Event representation + +# 1 Introduction + +Frame-based cameras often produce blurred images due to factors like camera shake or object motion, given their configured exposure time. Such blur, a common form of image degradation, occurs across various scenes. While photographers typically aim for clear and focused photos, clear inputs are equally critical for numerous computer vision tasks. These tasks include super-resolution, target + +![](images/5bfd42e747bf5dbc6e4d8cc6284f047ed99a601db36a9a27c56293f41bf9dedb.jpg) +(a) GT (PSNR |SSIM) + +![](images/17adc58bc8c7075408d48df65f7dadc8389c5148ab985309ad083c9396351c0f.jpg) +(b) Blurry (28.62dB |0.819) + +![](images/9296ea44d1b5bdf6a1adb5c182b71bd016571a194d8379b57fef5dee10f83353.jpg) +(c) HINet [4] (29.14dB |0.837) +Fig. 1: A deblurring example on HS-ERGB [37] dataset via recent state-of-the-art methods. Compared to other methods, our approach yields a more distinct result owing to its enhanced motion perception capability. + +![](images/13044559a5ade1896c43bb1813ecb568a3239ac67268cd28c59ff07c03370df3.jpg) +(d) NAFNet [3] (30.41dB |0.855) + +![](images/473f75e245093e693cdddd4d784fb33482225b41569de5edb3f7f85fa30dcde8.jpg) +(e) EFNet [33] (30.76dB |0.856) + +![](images/0ad319576978f72a3e13af2d0e95f04fc6e4cb45953156a234d058cded602fea.jpg) +(f) Ours (31.01dB [0.858] + +detection, and semantic segmentation, to name a few. Hence, image deblurring serves as a fundamental task in the realm of low-level computer vision, reinforcing the image quality and supporting the proper function of various algorithms. + +Traditional image-based techniques often attempt to derive a clear image from a single blurred one. They do this by leveraging natural image priors or making assumptions about blur operations [6, 8, 13, 15, 34, 44]. However, in capturing a single blurred image, much motion information is lost. This lack of motion information during exposure creates undesirable artifacts in the deblurred results, especially in complex motion scenarios, as illustrated in Figures 1c and 1d. To counter this challenge, recent methodologies use more temporal information, such as differently exposed frames [2, 46, 48] and neighboring frames [16, 28, 32]. However, these approaches have their limitations. For instance, the use of differently exposed frames is constrained by the low temporal resolution of frame-based cameras, improving the reconstruction effect primarily in slow-motion scenes. The method that uses adjacent frames to estimate the current frame's motion information can lead to accumulative errors. In summary, traditional deblurring techniques encounter hurdles due to the inability of frame-based cameras to capture motion information during exposure. + +As opposed to frame-based cameras, event cameras, inspired by biological systems, offer a promising solution to this issue. These cameras' microsecond-level temporal resolution $(\mu \mathrm{s})$ and asynchronous output architecture allow the recording of fast motions. This capability enables event streams to store pixel intensity changes and provide accurate motion information to deblur the corresponding frame. By using event frames as additional input, some methods attempt to guide the deblurring process through these event cameras [9,25,39,43,49]. One method [45] subdivides the event stream into ternary 2D data and generates a mask to localize high-blur regions, thereby enhancing the network's ability to distinguish different blur levels. Another method [33] symmetrically accumulates event polarities and fuses event and image features using an attention mechanism, enabling the integration of motion information into image reconstruction. However, these methods primarily employ the event stream in a manner akin to polarity accumulation. In essence, they consider only the final accumulation result of relative intensity change between subsequent moments, neglecting the ab + +solute intensity change caused by historical events. Consequently, this approach fails to differentiate between various motion patterns, injecting unnecessary ambiguity into the deblurring process, and leading to edge artifacts as displayed in Figure 1e. + +Given the limited perception of motion during exposure offered by polarity accumulation, we posit that the deviation from the initial moment provides more precise motion information for image deblurring. Each time the event generates, the absolute light intensity changes. The deviation from the initial moment enables us to consider the impact of changes in absolute light intensity, even when the absolute light intensity at the initial moment is unknown. Thus, the deviation, to a degree, implies the order of arrival of the events, offering different information for varying motion patterns. As a result, we propose the Deviation Accumulation (DA) method to enhance motion perception. Specifically, every time a new event arrives, our method accumulates the deviation relative to the initial moment, thus encompassing more comprehensive motion information. Furthermore, we find that longer time scales for event accumulation can capture global motion patterns, while shorter scales precisely depict local fast-motion patterns. To better perceive motion at both global and local time scales, we have designed a Recurrent Motion Extraction (RME) module that works with corresponding event representations. Moreover, in line with multimodal learning [18-20, 29, 36, 42], we present an effective Feature Alignment and Fusion (FAF) module. This module serves to address the issue that simple fusion methods struggle with the disparate nature of event and frame modalities. As a result, these two modalities align, ensuring better information interaction. We integrate these modules into a single image deblurring network. This network extracts the motion features within the exposure time from the event stream and combines them with the texture information supplied by the image to complete the motion blur removal, as illustrated in Figure 1f. + +The key contributions of this paper are as follows: + +- We introduce an event preprocessing method designed specifically for the deblurring problem. The deviation accumulation is capable of responding distinctively to various motion patterns, thereby adding accurate motion information for image deblurring tasks. +- We develop a network equipped with recurrent motion extraction modules, designed to better extract motion features on local and global time scales. This design enables the network to handle multiple composite motion scenarios effectively. +- We propose a feature alignment and fusion module to mitigate the impacts of inter-modal inconsistencies and ensure that motion information more effectively guides the image deblurring process. + +Our experimental results indicate that our method achieves state-of-the-art performance on synthetic, semi-synthetic, and real-world datasets, including a 0.61dB improvement in PSNR on the GoPro dataset compared to previous studies [33]. Moreover, our method produces more distinct subjective results, as demonstrated in Figure 1. + +# 2 Related Work + +# 2.1 Event-based Motion Deblurring + +The utilization of event information to assist in deblurring tasks has become increasingly common in recent research [9, 12, 25, 31, 39, 43, 49]. The high temporal resolution and low latency of event cameras not only capture edge positions but also embed moving temporal information within the event stream. This information is crucial to reconstruct the latent sharp image. Many methods have been attempted to explore the relationship between event information, blurred images, and sharp images. One such method [25] includes the Event-based Double Integral (EDI) model, which leverages the principles of event generation to formulate an optimization problem that establishes a correlation between a blurry image and a latent sharp image. In recent studies, the event stream is typically converted into event frames [30, 38], which simplifies the processing of event data within traditional CNN frameworks. [9] combines the motion feature extracted from event data with the image feature, using them as the input for the decoder. [50] employs a Multi-Layer Perceptron (MLP) to predict the fused feature value from event and image features. [33] integrates them through a cross-modal attention method, which generates a query matrix from the image features, but derives key and value matrices from the event features. + +Confronting the prevalent issue of integrating image and event features is the problem of inter-modal inconsistencies, which traditional fusion techniques cannot easily mitigate. To address this, our approach seeks to align these two modalities, optimizing the use of event data. Through this alignment, our goal is to minimize the inconsistencies spawned when merging differing modalities, thus amplifying the efficacy of the deblurring process. + +# 2.2 Event Representation + +Event data, characterized by its spatial sparsity and strong temporal continuity due to its asynchronous generation, necessitates a custom-designed event representation to deliver the required information. Methods [11, 26, 41] that process events individually maximize the use of high temporal resolution, as suggested by [26], which proposed an adaptive spiking neuron model to manage the fluctuating event input. The effectiveness of these algorithms, however, can be significantly hampered by the noise within the event due to independent processing. A Time Surface [1, 14, 22, 51] is created using the timestamp of the last event at each pixel, which prominently reveals the rich temporal information contained within the event. Furthermore, the event frame [5, 7, 21, 23, 45] transforms the original 4D event data into a 2D frame by accumulating events on the number or the polarity pixel-wise, facilitating extraction with conventional image processing techniques. In [45], the authors utilize ternary data for their event frames to allocate high blur region of the image. + +However, existing event frame approaches result in abundant light intensity change information being lost in the squeezing process, making it impossible to + +distinguish the order of event arrivals and lacking the ability to detect multiple motion patterns. Some techniques attempt to recover some of this lost data by dividing the exposure time into smaller segments, though this essentially equates to sampling the original data. Consequently, we design a motion-aware event representation specifically tailored for motion deblurring. + +# 3 Method + +# 3.1 Overall Framework + +The overall framework of our proposed method is depicted in Figure 2. We adopt a classical encoder-decoder architecture for our approach. Initially, the blurred image is inputted into the image branch, where we use the baseline conv block mentioned in [3] as our fundamental block, to extract relevant features. Simultaneously, the corresponding event frames generated by our deviation accumulation are fed into the Recurrent Motion Extraction (RME) module, which facilitates the extraction of features covering motion in local and global time scales. Following each block of the two branches, the Feature Alignment and Fusion (FAF) module is employed to eliminate the inconsistencies between the two modalities. By performing down-sampling operations, we are able to obtain both shallow and semantic information. In the decoder, skip connections are utilized to incorporate the extracted features for image reconstruction. The architecture of the conv block and the decoder is shown in the supplementary materials. The entire training process is conducted in an end-to-end manner, with the PSNR Loss function [4] employed to optimize the network's parameters. + +# 3.2 Deviation Accumulation Event Representation + +Event cameras are bio-inspired sensors which asynchronously output signals. They trigger an event while the logarithm of the intensity change exceeds the preset threshold $c$ at any pixel, which is formulated as + +$$ +p = \left\{ \begin{array}{l l} + 1 & \text {i f} \log \left(I _ {x y} (t)\right) - \log \left(I _ {x y} (t - \Delta t)\right) \geq c, \\ - 1 & \text {i f} \log \left(I _ {x y} (t)\right) - \log \left(I _ {x y} (t - \Delta t)\right) \leq - c, \end{array} \right. \tag {1} +$$ + +where $p$ is the event's polarity, $I_{xy}(t)$ and $I_{xy}(t - \Delta t)$ represent the intensity at the pixel coordinate $(x,y)$ at time $t$ and $t - \Delta t$ , respectively. An event is a quaternion value $(x,y,t,p)$ where $x,y$ denotes the pixel coordinate, $t$ represents the time of change, the polarity $p \in \{+1, -1\}$ indicates the increase and decrease. For a given blurry image $B$ , it can be formulated as the average of latent intensity images $I(t)$ during the exposure time [25]: + +$$ +B = \frac {1}{T} \int_ {0} ^ {T} I (t) d t, \tag {2} +$$ + +![](images/727fbf1276279060979dc2c666d494d80bcb46f3f42f97cec3fb0bcc253b654c.jpg) +Fig. 2: Overview of the architecture of our proposed method for image deblurring. The blurred image and related event frames within the exposure time are fed into the corresponding network branches respectively. The core module of the event branch: (a) Recurrent Motion Extraction (RME) module processes event frames from longer time scales to shorter, merging old features into new features through weighted sum, which results in perceiving various motion patterns under different time ranges. The image branch extracts the features with conv block and fuses them with event features using (b) Feature Alignment and Fusion (FAF) module. FAF module eliminates the inter-modal inconsistencies in a bi-directional way ensuring better information interaction. + +where $T$ is the exposure time. For simplicity of understanding, we now only consider the pixel at $(x,y)$ then combining Equations (1) and (2), we have + +$$ +\begin{array}{l} B _ {x y} = \frac {1}{T} \int_ {0} ^ {T} \exp (\log (I _ {x y} (t))) d t (3) \\ = \frac {1}{T} \int_ {0} ^ {T} e x p (\log (I _ {x y} (0)) + c \int_ {0} ^ {t} p (s) d s) d t. (4) \\ \end{array} +$$ + +The internal part of the exponential operation is the logarithmic value of the light intensity of the pixel at time $t$ , while the external part of the exponential operation is the integration and averaging of each moment within the exposure time. Based on the event data, we can discretize Equation (4): + +$$ +\begin{array}{l} B _ {x y} = \frac {1}{N} \sum_ {t ^ {\prime} \in \mathcal {T}} \exp \left(\log \left(I _ {x y} (0)\right) + c \sum_ {t = 0} ^ {t ^ {\prime}} p _ {t}\right) (5) \\ = \frac {I _ {x y} (0)}{N} \sum_ {t ^ {\prime} \in \mathcal {T}} e x p \left(c \sum_ {t = 0} ^ {t ^ {\prime}} p _ {t}\right), (6) \\ \end{array} +$$ + +where $\mathcal{T} = \{t_1, t_2, \dots, t_N\}$ indicates the set of time events occurs, $N$ is the number of events, and $p_t$ means the polarity of the event at time $t$ . + +The aforementioned derivation prompts us to consider not only the polarity of the event but also the impact of the alteration in the absolute value after + +# Algorithm 1 Deviation Accumulation (DA) + +Input: taking pixel at $(x,y)$ for an example, the events stream $(x,y,t_i,p_i)$ in which $i\in [1,N]$ , deviation from the initial moment $D$ , accumulation of deviation $R$ , Output: the value of DA event representation $V$ . + +1: $D = 0,R = 0$ +2: Counting the number of events $N$ ; +3: for $(i = 1$ to $N)$ do +4: $D = D + p_{i}$ +5: $R = R + D$ +6: end for +7: $V = R / N$ + +each accumulation when considering the problem of deblurring. However, since event data cannot provide $I_{xy}(0)$ , the absolute light intensity is not accessible. But we can indirectly capture this information through the deviation from the initial moment. Therefore, we design a new event representation method for the deblurring problem presented in Algorithm 1. First, we accumulate all the events in the exposure time, so we get a 2D matrix named frequency matrix representing the number of the event occurs at each pixel position. We deal with the events considering their polarities subsequently, recording the deviation at each pixel by an additional tensor. Upon the arrival of the subsequent event, we update the deviation according to polarity and then add the deviation to the corresponding position. In the end, by dividing the cumulative matrix by the frequency matrix, we obtain the average deviation change over the exposure period as the new event representation named Deviation Accumulation (DA). + +Traditional polarity accumulation methods directly accumulating event polarities lead to identical outputs whenever the number of positive and negative events within a given time frame $([0, T])$ remains unchanged. Nevertheless, by considering the deviation brought by historical events, our method can generate distinct values for events arriving in different orders, thereby improving the ability to perceive motion patterns and providing more precise motion information for the deblurring procedure. For example, as illustrated in Figure 3, (a) shows event signals generated by a motion pattern which represents that the light intensity at that pixel has risen twice and then fallen twice while (d) shows another pattern. Facing these two patterns, (b) and (e) depict the changing process of the outputs of traditional method and our method after each event is accumulated. After processing all four events, traditional polarity accumulation gets the same output of 0 for both patterns, but our DA representation yields 1 and 0.5 respectively which contains richer motion information. More examples of motion patterns our method can distinguish are shown in supplementary materials. + +Overall, the phenomenon that polarity accumulation methods represent some different motion patterns as the same value introduces ambiguity into the deblurring process. Our method has a better perception of the motion patterns and offers a more distinct motion edge than conventional approaches from the visual effect as shown in Figure 3 (c) and (f). Inspired by [33], directly segmenting + +![](images/f27df0ad0618d5db90f55bedf00ddfb7b27b5a54a5d9f563739f621712895c21.jpg) + +![](images/42187f67118f8b940e6d4101d46ceb7e90f1c0c1d5e25fc0efcf2e6b161825e5.jpg) + +![](images/2594bfdcceba4383dab92ecf16d109c444ee6f9000430d2f4182e0ca07f17411.jpg) +Fig. 3: (a) and (d) are event signals of two different motion patterns. (b) shows the Polarity Accumulation (PA) can't distinguish (a) and (d) as it gets the same output of 0 after processing all events while (e) shows our Deviation Accumulation (DA) can. DA perceives more motion information than PA as shown in (c) and (f). + +![](images/1c9c90660ce6ee749ed03a3071e895c9dbd79e1abc3052138dd0644cf4112f44.jpg) + +![](images/16ebf709d3c72c1877985cdb2c25867915be52e3966f58db11e80140a702ae40.jpg) + +![](images/a6efd72a2e849957ba7e7139a202dbda424f44cc76c8a7f95b59c21f5a3f3d94.jpg) + +the entire exposure time averagely is not suitable for motion deblurring, which focuses on capturing rapid motion but falls short in adequately modeling the continuous movement throughout the exposure. In this paper, we symmetrically accumulate events along the $T / 2$ axis using the DA method. Taking $[0,T / 2]$ as an example, we accumulate events on different time scales. The longest accumulating range is $T / 2$ , followed by $T / 3$ , and the shortest is $T / 6$ . A detailed schematic is provided in the supplementary materials. + +# 3.3 Recurrent Motion Extraction Module + +The goal of the Recurrent Motion Extraction (RME) module is to cater to both global slow motion and local fast motion in different time scales. Based on our observations of the motion process and event representations, the focus of the event representations varies in different time scales. In shorter time intervals, the global slow motion is not significant in the event representation due to the smaller number of events triggered. On the contrary slow motion can be clearly recorded in longer intervals, but at the same time, the violent moving part will produce edge spreading artifact. + +To enable the network to accurately model the various in-homogeneous motions, we devised an RME module that incorporates a recurrent mechanism. As shown in Figure 2(a), the event features are fed into the processing module from the longest time scales to the shortest and we use $e1$ , $e2$ , and $e3$ to denote them. Formally, the RME module can be presented as + +$$ +e _ {i} ^ {\prime} = \left\{ \begin{array}{l l} f _ {e x} \left(e _ {i}\right) & \text {i f} i = 1, \\ \left(1 - \alpha\right) \times f _ {e x} \left(e _ {i}\right) + \alpha \times f _ {r e} \left(e _ {i - 1} ^ {\prime}\right) & \text {i f} i = 2, 3, \end{array} \right. \tag {7} +$$ + +where $f_{ex}(\cdot)$ and $f_{re}(\cdot)$ are two baseline conv blocks, $f_{ex}(\cdot)$ is used to extract information from the current representation, $f_{re}(\cdot)$ is used to retain useful information from previous features, $\alpha$ is the learnable weight as a trade-off between two features. By sharing the parameters of $f_{ex}(\cdot)$ and $f_{re}(\cdot)$ for different time scales, we force the network to be able to perceive motions at varying time scales. Each input feature is symmetric on the time axis so that the priori information about the motion direction is simultaneously fed into the network. + +# 3.4 Feature Alignment and Fusion Module + +Although the information captured by the event camera exhibits a high correlation with the frame camera during the equivalent exposure period, there is still an inevitable gap originating from the modal dissimilarities between them. This phenomenon is frequently mentioned in multi-modal learning [18,42], but in event-guided image restoration tasks people usually [9,33,45] fuse features from two modalities directly using convolutions or attention mechanisms. Therefore, we propose an effective Feature Alignment and Fusion (FAF) module to facilitate enhanced feature fusion in a more coherent environment. Concretely, as shown in Figure 2(b), this module consists of a feature alignment module and a multi-head attention module. We use the operations of multiplication and addition successively to merge the raw information of one modality into the changing process of the other modality, which fills the gap of inter-modal inconsistency. The aligned features are fed into the attention module. In each fusion process, attention plays a bi-directional role, we not only generate the Q matrix through the frame features and the K and V matrix through the event features but also use the event features to perform look-ups on the frame features. The weights of the attention module are shared so that we can fuse the motion information under a more consistent space. Therefore, FAF module can be expressed as + +$$ +x ^ {\prime} = \operatorname {C o n v} _ {2} (x) + (\operatorname {C o n v} _ {1} (x) \times e + e), \tag {8} +$$ + +$$ +e ^ {\prime} = \operatorname {C o n v} _ {2} (e) + (\operatorname {C o n v} _ {1} (e) \times x + x), \tag {9} +$$ + +$$ +F = \operatorname {C o n v} _ {3} (\operatorname {C o n c a t} ([ \text {A t t e n t i o n} (Q _ {x ^ {\prime}}, K _ {e ^ {\prime}}, V _ {e ^ {\prime}}), \text {A t t e n t i o n} (Q _ {e ^ {\prime}}, K _ {x ^ {\prime}}, V _ {x ^ {\prime}}) ])), \tag {10} +$$ + +where $x'$ and $e'$ are aligned image features $x$ and event feature $e$ respectively, Attention(\cdot) denotes the multi-head attention operation, Concat(\cdot) denotes the concatenation operation along the channel dimension. + +# 4 Experiments and Analysis + +# 4.1 Experimental Settings + +Datasets. We use three different datasets containing synthetic, semi-synthetic, and real-world frames and events to evaluate the proposed method. + +GoPro: We train the network on the GoPro dataset which is the benchmark dataset for the image motion deblurring [24]. The blurry image is averaged from + +Table 1: The quantitative results on GoPro, HS-ERGB, and REBlur test datasets. The best values are highlighted in bold. HINet+ and UFPNet+ are event-enhanced versions of HINet and UFPNet by concatenating our DA to their input. + +
MethodInputGoProHS-ERGBREBlurParams(M)
PSNRSSIMPSNRSSIMPSNRSSIM
LEVS [10]F20.840.54722.130.555--18.2
EDI [25]F+E29.060.94323.930.70436.520.9640.5
SRN [35]F30.260.934--35.100.96110.2
EVDI [49]F+E30.400.90625.130.707--0.4
HINet [4]F32.710.95927.320.80735.580.96588.7
Restormer [47]F32.920.96127.550.80835.500.95926.1
MSDI-Net [17]F33.280.96427.460.80936.140.968241.3
NAFNet [3]F33.710.96727.640.81136.150.96967.8
UFPNet [6]F34.060.96827.640.80936.110.96880.3
HINet+ [4]F+E34.630.96827.660.80837.920.97688.7
UFPNet+ [6]F+E35.220.97227.680.80937.970.97680.3
EFNet [33]F+E35.460.97226.680.80038.120.9758.5
OursF+E36.070.97627.930.81238.470.97813.9
+ +adjacent sharp frames, by doing so the dataset contains 3214 pairs of blurry and sharp images. According to the standard division, we adopt 2103 pairs for training and 1111 pairs for testing. For the event data, we use the GoPro raw event dataset provided by [33] which synthesizes events with a random threshold $c$ following the Gaussian distribution by ESIM simulator [27]. Following the DA method, the raw event data is preprocessed into 6 event frames for each image. + +HS-ERGB: The HS-ERGB dataset [37] consists of sharp videos and real-world events, we use [50] released normal blur version which synthesizes blurry frames by averaging 49 interpolating images. Each blurry frame in the dataset has a known exposure time, allowing us to generate the DA representation for each frame within its respective exposure duration. + +REBlur: The REBlur dataset collects a number of sequences of real-world event data, corresponding with blurry images and reliable ground-truth sharp images [33]. It contains 12 kinds of linear and nonlinear motions. Following the original setting, we use 486 pairs of blurry-sharp images with associated events for training and 983 pairs for testing. + +Implementation details. We train the end-to-end network without pre-training on a single Tesla V100 GPU. We extract patches of size $256 \times 256$ for training images and corresponding event frames, and the batch size is set to 2 by default. The AdamW optimizer is employed ( $\beta_{1} = 0.9$ and $\beta_{2} = 0.99$ ) with the initial learning rate 0.0001 with the cosine annealing schedule where the $T_{max}$ is 400K iteration. For data augmentation, we perform horizontal and vertical flips, the random noise and the hot pixels are added into event frames in order to simulate the real-world situation. For HS-ERGB and REBlur datasets, We fine-tune the model trained on GoPro with the training set of them. The fine-tuning process + +![](images/20a836d4e920f402f1279f0912bc458bdbdbfc258f2ce5277dd485ce5c876056.jpg) +(a) Blur image in GoPro testset + +(a) Blur image in GoPro testset + +![](images/090bccd521fd0515526e1fdddda2a1d27d3b2e0539b66ab88db218c349e85371.jpg) +(b) HINet + +![](images/de240c96888191bad7fd23d6f5d1fd42c49aab6b15b4ad351ec42ba2462276a4.jpg) +(c) UFPNet + +![](images/855a3626dcaa19b3a6f81c7a38abba6894d8210cbb0351136d26db20c5ded06a.jpg) +(d) EFNet + +![](images/9d59b2a9ca86a9e8467a6a4624297134ae54fe4303e5a9a48ebb08b3760aa938.jpg) +(e) NAFNet + +![](images/33624c427a9def027a5e47ac4beec07fe892844f016eb3e4c139b6b2dff08fd6.jpg) +(f) HINet+ + +![](images/b844597b0b2628d9b494fbb49079e9538da6c357ce669b0189408d786989910e.jpg) +(g) UFPNet+ + +![](images/525fd323a7e502680d99c7d455846726cc7a4848aa71c023a1ac82cea2cdfb8e.jpg) +(h) Ours + +![](images/1352bdc01951e52d0cda0812c9706632ddf092d3b3b441a507c212605ee64399.jpg) +(i) GT + +![](images/0843f6c080552a51c92468fb751bc0e4305eec693b7bb36b6c1b2a4b492e83be.jpg) +Fig. 4: Qualitative comparisons on the GoPro dataset. The notation is the same as in Table 1. Zoom in for better view. +(a) Blur image in REBlur testset +: Qualitative comparisons on the REBlur dataset. The notation is the same as Table 1. Zoom in for better view. + +![](images/e3739c1a5a39ccec854698002b1ed5041ff11d891789b8f9a04cefc4b762dc8d.jpg) +(b) HINet + +![](images/1cee75ff1bd3522ee52b1985f367cea90277b0673d9543de97be08f469dad2dc.jpg) +(c) UFPNet + +![](images/49d03350932205f4c7456d7b8004e01f93a025b15a3769694659b63a66aebda9.jpg) +(d) EFNet + +![](images/b60a52f299b5b271275e91e328b201333ba9fc08023bb258ff4899f4394d7338.jpg) +(e) NAFNet + +consisted of 4000 iterations while keeping the other configurations consistent with the previous experiments. + +# 4.2 Comparison with State-of-the-Art Methods + +Table 1 shows the comparison results of our method and other SOTA methods on GoPro [24], HS-ERGB [50] and REBlur [33]. We compare our network with not only event-based methods, but also with some SOTA image-only deblurring networks enhancing with event input. The results illustrate that: (1) Our method outperforms all other methods in synthetic, semi-synthetic and real-world datasets, which indicates that with the proper utilization of events the designed approach is able to handle blur with different distributions. Specifically, the average gains of our method over the second-best one are 0.61dB, 0.25dB and 0.35dB in terms of PSNR on GoPro, HS-ERGB and REBlur datasets, respectively. (2) Through the enhancement of our proposed event representation, the HINet+ and the UFPNet+ increase 1.92dB and 1.16dB in PSNR on the benchmark GoPro dataset compared with their image-only version, respectively. Taking HINet as an example, with the same operation of [33], HINet with DA + +![](images/734fd1b1a209722f46554a3beff3f560d60b9d4c52ce867556462ddd29643728.jpg) +(a) GT + +![](images/23665851adff685957098865bc95f18a700e064978f64467d82b3f335ad0e2c5.jpg) +(b) Blurry + +![](images/e436f01102b7a8802bc8f0f4fda16ca5e7ee47f048512df78a52a0737fd0df72.jpg) +(c) UFPNet + +![](images/116209ed0274009ad194fa824ab167736e9ae06fc2d09cdee0ae569b63bfb553.jpg) +(d) NAFNet + +![](images/75b923889f3b670525027bcd31c23cc3db0dd31392f76380b22b20f16a83888f.jpg) +(e) EFNet + +![](images/1e7c2eabaea83f858b8621cc979ffad7fca20e9e2d345effe15da8aa95aa0aa0.jpg) +(f) Ours + +![](images/41e0651326b76cccb5d5710c1442995461b95c4ddbe1fcacc18b94d134bc4605.jpg) +Fig. 6: Qualitative comparisons on the HS-ERGB dataset. + +![](images/a720aa58896079b0e0b39e7458067a7fbec6479614a71146e05ad3f99b7d9700.jpg) +(a) Blurry + +![](images/68e51ec54743ea9c8e64884178afd1a439dff5366a8d45a80122461dcc428e57.jpg) + +![](images/aca0016fe3ce7794d334f49c234da809ebe1c6ab4418653c1f8df20c1c5fd89a.jpg) +(b) SCER [33] + +![](images/c83f0d2d8eb537da6533e4b9f86be810fee0e31af3b33665379c2553ef6c7061.jpg) + +![](images/e006b54c00ea90658069b0fd4689d9fe01929f4c59460730601fbe9ea5ab731a.jpg) +(c) DA (Ours) + +![](images/2355dc71134c668061af47e87b50aa1f1aa24daa40fa40843e9198ed91743256.jpg) + +![](images/f1a387eda2a305341be0a7284808fbc25adaa1a60547ed3ebee3eb97710f9f44.jpg) +(d) GT +Fig. 7: Visual comparisons on GoPro dataset with different event representation. Compared to SCER representation, our DA representation reduces edge artifacts. + +event representation exceeds 0.94dB over the result in [33], which presents the fact that DA extracts more conducive information for motion deblurring. + +We also show the visual comparison results to verify the effectiveness of our method in Figure 4, Figure 5 and Figure 6 with representative samples from each dataset. As illustrated in the figures, our model reconstructs better details and has minor spatial distortions compared with other methods. This indicates that our event representation perceives more motion information and the modules effectively embed it to reconstruct a high-quality image with enhanced details. + +# 4.3 Ablation Studies + +We conduct plenty ablation experiments on GoPro dataset to verify the effectiveness of our event representation and proposed method. + +Effectiveness of Event Representation. To demonstrate the adaptability of our event representation for motion deblurring tasks, as shown in Table 2, we replace our DA representation with several other types of representations and trained the network accordingly. The results reveal certain shortcomings of the alternative representations. Firstly, the stack representation simply accumulates all events on their polarities which leaves out the temporal information. Secondly, the SBT approach [40], which subdivides the entire exposure time equally, proves to be inadequate for modeling the varying degrees of blur. Thirdly, the SCER + +Table 2: Ablation study of each module effects. + +
Event RepresentationRMEFusionPSNRSSIM
Stack33.240.953
SBT [40]✓FAF(Ours)35.630.973
SCER [33]35.810.975
Concat.34.020.967
Add34.740.970
Attention35.750.974
ACA(Ours)✓Only Align Event+Attention35.780.975
Only Align Image+Attention35.740.974
35.420.973
ACA(Ours)✗FAF(Ours)36.070.976
✓
+ +![](images/81f7908c63a90b433e90594dc13fa912cb37f5504f5f993a8d50ce93a37c0d8f.jpg) +(a) w/o RME + +![](images/11be934ee6cb38b10dfb35c3fc2d91777a02b74c7898ac526a9795643b474ccb.jpg) +(b) w/o RME + +![](images/1f18276d5b5d8ae61f7cddeb2945d238a52e8902dde8319fdaa11eaef552ba1d.jpg) +(c) w/o RME + +![](images/7815bdf963277e3713ad059dbd509e05af04828114043170576ce6d39efadbb0.jpg) +(d) w/RME +Fig. 8: Visual comparison results of w/ and w/o RME module. All of these scenarios involve movement in different directions and speeds, RME module can successfully capture motion features in these complex motion scenarios. Zoom in for better view. + +![](images/7a552d1a9227850a52be4560f38c65a17ba9c1bf2257581d64e68c82f2f56e03.jpg) +(e) w/RME + +![](images/a03e2927b922c9edc7b8ca0699305e75b5d34138e48b931449aba8642488c107.jpg) +(f) w/RME + +method [33], which directly accumulates event polarities, is found to be incapable of capturing the influence of absolute intensity changes. This limitation is illustrated in Figure 3. In contrast, our DA representation not only possesses a solid physical foundation for event generation but also perceives more motion patterns. As a result, it achieves a 0.26dB improvement in performance compared to the best alternative representation. We also provide the visual comparison on the image reconstruction results in Figure 7. Our DA representation reduces motion-generated edge artifacts, resulting in a better deblurring effect. + +Effectiveness of Recurrent Motion Extraction Module. To assess the RME module's contribution, we replace it with a baseline conv block and adjust the parameters for fair comparison. In this approach, we concatenate event frames of different time scales and feed them into the conv block. The findings, as presented in the last two rows of Table 2, clearly indicate that the contribution of $0.65\mathrm{dB}$ increment in PSNR stems from the network structure design, rather than an increase in parameter quantity. This shows the importance of incorpo + +rating the module for effectively leveraging motion information in multi-timescales and improving performance in motion deblurring tasks. We also present the visual comparisons when removing the RME module in Figure 8, the RME module adeptly captures both fast local and slow global motion characteristics. + +Effectiveness of Feature Alignment and Fusion Module. To illustrate the effectiveness of our FAF module, we compare it with several common fusion methods, including concatenation, addition, and cross-modality attention. As shown in Table 2, it is arduous to utilize the motion information contained in event feature through the direct concatenation or addition. While attention mechanisms can capture relative parts in the two features, the inherent inconsistencies within the two modalities still lead to a 0.32dB performance gap when compared to our FAF module. We also use a t-SNE visualization in Figure 9 to demonstrate the effectiveness of our feature alignment, where two originally in-congruent modal features are mixed up together after the FAF module. + +Furthermore, we also verify the importance of bi-directional alignment by conducting experiments solely on aligning event or image features. Since both events and frames provide crucial information for image restoration, such as edge motion information and textual information, respectively, it is unreasonable to let one modality dominate the alignment process. As shown in Table 2, the bi-directional alignment outperforms the highest score achieved by one-way correction by 0.29dB. This demonstrates the necessity of fully utilizing the mutual information between the two modalities in order to achieve better image reconstruction results. + +![](images/5fb4c0b4afc0a3f1a56e00ab48428f673c8a69bd8eea8b47eb91ccc2480c19a4.jpg) +Fig. 9: t-SNE visualization of original and aligned features. The distributions of the aligned two features overlap more compared to the original features. This indicates the effectiveness of FAF module to reduce the inter-modal inconsistency. + +# 5 Conclusion + +In this paper, we propose an event-based image deblurring method. Firstly, we rethink the limit of traditional event preprocessing method and propose a new deviation accumulation (DA) technique. With deviation, the new approach can distinguish the order of events and perceive more motion patterns. Moreover, we design an image deblurring network equipped with recurrent motion extraction (RME) module and feature alignment and fusion (FAF) module to fit our proposed event representation. The RME module can better perceive motions in different time scales, and the FAF module promotes better integration of the two modalities. Finally, extensive experimental results over several datasets demonstrate the effectiveness of the proposed algorithm. + +Acknowledgements. This work was supported by National Natural Science Foundation of China (NSFC) under Grants 62225207 and 62276243. + +# References + +1. Benosman, R., Clercq, C., Lagorce, X., Ieng, S.H., Bartolozzi, C.: Event-based visual flow. IEEE transactions on neural networks and learning systems 25(2), 407-417 (2013) +2. Chang, M., Feng, H., Xu, Z., Li, Q.: Low-light image restoration with short-and long-exposure raw pairs. IEEE Transactions on Multimedia 24, 702-714 (2021) +3. Chen, L., Chu, X., Zhang, X., Sun, J.: Simple baselines for image restoration. In: Proceedings of the European conference on computer vision (ECCV). pp. 17-33. Springer (2022) +4. Chen, L., Lu, X., Zhang, J., Chu, X., Chen, C.: Hinet: Half instance normalization network for image restoration. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 182-192 (2021) +5. Cook, M., Gugelmann, L., Jug, F., Krautz, C., Steger, A.: Interacting maps for fast visual interpretation. In: The 2011 International Joint Conference on Neural Networks. pp. 770-776. IEEE (2011) +6. Fang, Z., Wu, F., Dong, W., Li, X., Wu, J., Shi, G.: Self-supervised non-uniform kernel estimation with flow-based motion prior for blind image deblurring. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18105-18114 (2023) +7. Gehrig, D., Rebecq, H., Gallego, G., Scaramuzza, D.: Eklt: Asynchronous photometric feature tracking using events and frames. International Journal of Computer Vision 128(3), 601-618 (2020) +8. Ji, S.W., Lee, J., Kim, S.W., Hong, J.P., Baek, S.J., Jung, S.W., Ko, S.J.: Xydeblur: divide and conquer for single image deblurring. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17421-17430 (2022) +9. Jiang, Z., Zhang, Y., Zou, D., Ren, J., Lv, J., Liu, Y.: Learning event-based motion deblurring. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3320-3329 (2020) +10. Jin, M., Meishvili, G., Favaro, P.: Learning to extract a video sequence from a single motion-blurred image. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 6334-6342 (2018) +1. Kim, H., Handa, A., Benosman, R., Ieng, S.H., Davison, A.J.: Simultaneous mosaicing and tracking with an event camera. J. Solid State Circ 43, 566-576 (2008) +2. Kim, T., Lee, J., Wang, L., Yoon, K.J.: Event-guided deblurring of unknown exposure time videos. In: Proceedings of the European conference on computer vision (ECCV). pp. 519-538. Springer (2022) +3. Krishnan, D., Tay, T., Fergus, R.: Blind deconvolution using a normalized sparsity measure. In: CVPR 2011. pp. 233-240. IEEE (2011) +4. Lagorce, X., Orchard, G., Galluppi, F., Shi, B.E., Benosman, R.B.: Hots: a hierarchy of event-based time-surfaces for pattern recognition. IEEE transactions on pattern analysis and machine intelligence 39(7), 1346-1359 (2016) +5. Levin, A., Weiss, Y., Durand, F., Freeman, W.T.: Understanding and evaluating blind deconvolution algorithms. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 1964-1971. IEEE (2009) + +16. Li, D., Shi, X., Zhang, Y., Cheung, K.C., See, S., Wang, X., Qin, H., Li, H.: A simple baseline for video restoration with grouped spatial-temporal shift. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9822-9832 (2023) +17. Li, D., Zhang, Y., Cheung, K.C., Wang, X., Qin, H., Li, H.: Learning degradation representations for image deblurring. In: Proceedings of the European conference on computer vision (ECCV). pp. 736-753. Springer (2022) +18. Li, Y., Yu, A.W., Meng, T., Caine, B., Ngiam, J., Peng, D., Shen, J., Lu, Y., Zhou, D., Le, Q.V., et al.: Deepfusion: Lidar-camera deep fusion for multi-modal 3d object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17182-17191 (2022) +19. Liang, M., Yang, B., Wang, S., Urtasun, R.: Deep continuous fusion for multi-sensor 3d object detection. In: Proceedings of the European conference on computer vision (ECCV). pp. 641-656 (2018) +20. Liang, X., Qian, Y., Guo, Q., Cheng, H., Liang, J.: Af: An association-based fusion method for multi-modal classification. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(12), 9236-9254 (2021) +21. Liu, M., Delbruck, T.: Adaptive time-slice block-matching optical flow algorithm for dynamic vision sensors. BMVC (2018) +22. Manderscheid, J., Sironi, A., Bourdis, N., Migliore, D., Lepetit, V.: Speed invariant time surface for learning to detect corner points with event-based cameras. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10245-10254 (2019) +23. Maqueda, A.I., Loquercio, A., Gallego, G., García, N., Scaramuzza, D.: Event-based vision meets deep learning on steering prediction for self-driving cars. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5419-5427 (2018) +24. Nah, S., Hyun Kim, T., Mu Lee, K.: Deep multi-scale convolutional neural network for dynamic scene deblurring. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3883-3891 (2017) +25. Pan, L., Scheerlinck, C., Yu, X., Hartley, R., Liu, M., Dai, Y.: Bringing a blurry frame alive at high frame-rate with an event camera. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6820-6829 (2019) +26. Paredes-Vallés, F., Scheper, K.Y., De Croon, G.C.: Unsupervised learning of a hierarchical spiking neural network for optical flow estimation: From events to global motion perception. IEEE transactions on pattern analysis and machine intelligence 42(8), 2051-2064 (2019) +27. Rebecq, H., Gehrig, D., Scaramuzza, D.: Esim: an open event camera simulator. In: Conference on robot learning. pp. 969-982. PMLR (2018) +28. Ren, D., Shang, W., Yang, Y., Zuo, W.: Aggregating long-term sharp features via hybrid transformers for video deblurring. arXiv preprint arXiv:2309.07054 (2023) +29. Ren, S., Du, Y., Lv, J., Han, G., He, S.: Learning from the master: Distilling cross-modal advanced knowledge for lip reading. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13325-13333 (2021) +30. Shang, W., Ren, D., Zou, D., Ren, J.S., Luo, P., Zuo, W.: Bringing events into video deblurring with non-consecutively blurry frames. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4531-4540 (2021) +31. Song, C., Bajaj, C., Huang, Q.: Deblursr: Event-based motion deblurring under the spiking representation. arXiv preprint arXiv:2303.08977 (2023) + +32. Su, S., Delbracio, M., Wang, J., Sapiro, G., Heidrich, W., Wang, O.: Deep video deblurring for hand-held cameras. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1279-1288 (2017) +33. Sun, L., Sakaridis, C., Liang, J., Jiang, Q., Yang, K., Sun, P., Ye, Y., Wang, K., Gool, L.V.: Event-based fusion for motion deblurring with cross-modal attention. In: Proceedings of the European conference on computer vision (ECCV). pp. 412-428. Springer (2022) +34. Sun, L., Cho, S., Wang, J., Hays, J.: Edge-based blur kernel estimation using patch priors. In: IEEE international conference on computational photography (ICCP). pp. 1-8. IEEE (2013) +35. Tao, X., Gao, H., Shen, X., Wang, J., Jia, J.: Scale-recurrent network for deep image deblurring. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 8174-8182 (2018) +36. Tian, Y., Krishnan, D., Isola, P.: Contrastive multiview coding. In: Proceedings of the European conference on computer vision (ECCV). pp. 776-794. Springer (2020) +37. Tulyakov, S., Gehrig, D., Georgoulis, S., Erbach, J., Gehrig, M., Li, Y., Scaramuzza, D.: Time lens: Event-based video frame interpolation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 16155-16164 (2021) +38. Vitoria, P., Georgoulis, S., Tulyakov, S., Bochicchio, A., Erbach, J., Li, Y.: Event-based image deblurring with dynamic motion awareness. In: Proceedings of the European conference on computer vision (ECCV). pp. 95-112. Springer (2022) +39. Wang, B., He, J., Yu, L., Xia, G.S., Yang, W.: Event enhanced high-quality image recovery. In: Proceedings of the European conference on computer vision (ECCV). pp. 155-171. Springer (2020) +40. Wang, L., Ho, Y.S., Yoon, K.J., et al.: Event-based high dynamic range image and very high frame rate video generation using conditional generative adversarial networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10081-10090 (2019) +41. Weikersdorfer, D., Conradt, J.: Event-based particle filtering for robot self-localization. In: 2012 IEEE International Conference on Robotics and Biomimetics (ROBIO). pp. 866-870. IEEE (2012) +42. Xia, W., Li, X., Deng, A., Xiong, H., Dou, D., Hu, D.: Robust cross-modal knowledge distillation for unconstrained videos. arXiv preprint arXiv:2304.07775 (2023) +43. Xu, F., Yu, L., Wang, B., Yang, W., Xia, G.S., Jia, X., Qiao, Z., Liu, J.: Motion deblurring with real events. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2583-2592 (2021) +44. Xu, L., Jia, J.: Two-phase kernel estimation for robust motion deblurring. In: Proceedings of the European conference on computer vision (ECCV). pp. 157-170. Springer (2010) +45. Yang, D., Yamac, M.: Motion aware double attention network for dynamic scene deblurring. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1113-1123 (2022) +46. Yuan, L., Sun, J., Quan, L., Shum, H.Y.: Image deblurring with blurred/noisy image pairs. In: ACM SIGGRAPH 2007 papers, pp. 1-es (2007) +47. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H.: Restormer: Efficient transformer for high-resolution image restoration. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5728-5739 (2022) + +48. Zhang, S., Zhen, A., Stevenson, R.L.: Deep motion blur removal using noisy/blurry image pairs. Journal of Electronic Imaging 30(3), 033022-033022 (2021) +49. Zhang, X., Yu, L.: Unifying motion deblurring and frame interpolation with events. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17765-17774 (2022) +50. Zhang, X., Yu, L., Yang, W., Liu, J., Xia, G.S.: Generalizing event-based motion deblurring in real-world scenarios. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10734-10744 (2023) +51. Zhou, Y., Gallego, G., Rebecq, H., Kneip, L., Li, H., Scaramuzza, D.: Semi-dense 3d reconstruction with a stereo event camera. In: Proceedings of the European conference on computer vision (ECCV). pp. 235-251 (2018) \ No newline at end of file diff --git a/2024/Motion Aware Event Representation-driven Image Deblurring/images.zip b/2024/Motion Aware Event Representation-driven Image Deblurring/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..9e2a4f2b9a99baa6449c609f26a5cd104bdba634 --- /dev/null +++ b/2024/Motion Aware Event Representation-driven Image Deblurring/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5ac24081eab05afd3713baedd54f8320d0a6b03b3369bcecf44e17e4236c44e5 +size 462060 diff --git a/2024/Motion Aware Event Representation-driven Image Deblurring/layout.json b/2024/Motion Aware Event Representation-driven Image Deblurring/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..24fc1ceb3c8d1a3ef1b96ecf87b5ed20bb29a6a6 --- /dev/null +++ b/2024/Motion Aware Event Representation-driven Image Deblurring/layout.json @@ -0,0 +1,11094 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 156, + 111, + 459, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 156, + 111, + 459, + 148 + ], + "spans": [ + { + "bbox": [ + 156, + 111, + 459, + 148 + ], + "type": "text", + "content": "Motion Aware Event Representation-driven Image Deblurring" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 137, + 168, + 476, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 168, + 476, + 181 + ], + "spans": [ + { + "bbox": [ + 137, + 168, + 476, + 181 + ], + "type": "text", + "content": "Zhijing Sun, Xueyang Fu, Longzhuo Huang, Aiping Liu, and Zheng-Jun Zha*" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 153, + 190, + 461, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 153, + 190, + 461, + 246 + ], + "spans": [ + { + "bbox": [ + 153, + 190, + 461, + 246 + ], + "type": "text", + "content": "School of Information Science and Technology and MoE Key Laboratory of Brain-inspired Intelligent Perception and Cognition, University of Science and Technology of China, Hefei, 230026, China {sunzhijing, hlz_0}@mail.ustc.edu.cn {xyfu, aipingl, zhazj}@ustc.edu.cn" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 276, + 452, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 276, + 452, + 518 + ], + "spans": [ + { + "bbox": [ + 160, + 276, + 452, + 518 + ], + "type": "text", + "content": "Abstract. Traditional image deblurring struggles with high-quality reconstruction due to limited motion data from single blurred images. Excitingly, the high-temporal resolution of event cameras records motion more precisely in a different modality, transforming image deblurring. However, many event camera-based methods, which only care about the final value of the polarity accumulation, ignore the influence of the absolute intensity change where events generate so fall short in perceiving motion patterns and effectively aiding image reconstruction. To overcome this, in this work, we propose a new event preprocessing technique that accumulates the deviation from the initial moment each time the event is updated. This process can distinguish the order of events to improve the perception of object motion patterns. To complement our proposed event representation, we create a recurrent module designed to meticulously extract motion features across local and global time scales. To further facilitate the event feature and image feature integration, which assists in image reconstruction, we develop a bi-directional feature alignment and fusion module. This module works to lessen inter-modal inconsistencies. Our approach has been thoroughly tested through rigorous experiments carried out on several datasets with different distributions. These trials have delivered promising results, with our method achieving top-tier performance in both quantitative and qualitative assessments. Code is available at https://github.com/ZhijingS/DA_event_deblur." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 528, + 443, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 528, + 443, + 540 + ], + "spans": [ + { + "bbox": [ + 160, + 528, + 443, + 540 + ], + "type": "text", + "content": "Keywords: Image deblurring " + }, + { + "bbox": [ + 160, + 528, + 443, + 540 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 528, + 443, + 540 + ], + "type": "text", + "content": " Event camera " + }, + { + "bbox": [ + 160, + 528, + 443, + 540 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 528, + 443, + 540 + ], + "type": "text", + "content": " Event representation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 561, + 229, + 573 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 561, + 229, + 573 + ], + "spans": [ + { + "bbox": [ + 133, + 561, + 229, + 573 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 586, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 586, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 586, + 482, + 647 + ], + "type": "text", + "content": "Frame-based cameras often produce blurred images due to factors like camera shake or object motion, given their configured exposure time. Such blur, a common form of image degradation, occurs across various scenes. While photographers typically aim for clear and focused photos, clear inputs are equally critical for numerous computer vision tasks. These tasks include super-resolution, target" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 142, + 654, + 238, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 654, + 238, + 665 + ], + "spans": [ + { + "bbox": [ + 142, + 654, + 238, + 665 + ], + "type": "text", + "content": "*Corresponding author" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 191, + 171 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 191, + 171 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 191, + 171 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 191, + 171 + ], + "type": "image", + "image_path": "5bfd42e747bf5dbc6e4d8cc6284f047ed99a601db36a9a27c56293f41bf9dedb.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 148, + 172, + 175, + 198 + ], + "lines": [ + { + "bbox": [ + 148, + 172, + 175, + 198 + ], + "spans": [ + { + "bbox": [ + 148, + 172, + 175, + 198 + ], + "type": "text", + "content": "(a) GT (PSNR |SSIM)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 192, + 114, + 249, + 171 + ], + "blocks": [ + { + "bbox": [ + 192, + 114, + 249, + 171 + ], + "lines": [ + { + "bbox": [ + 192, + 114, + 249, + 171 + ], + "spans": [ + { + "bbox": [ + 192, + 114, + 249, + 171 + ], + "type": "image", + "image_path": "17adc58bc8c7075408d48df65f7dadc8389c5148ab985309ad083c9396351c0f.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 201, + 172, + 239, + 198 + ], + "lines": [ + { + "bbox": [ + 201, + 172, + 239, + 198 + ], + "spans": [ + { + "bbox": [ + 201, + 172, + 239, + 198 + ], + "type": "text", + "content": "(b) Blurry (28.62dB |0.819)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 250, + 114, + 306, + 171 + ], + "blocks": [ + { + "bbox": [ + 250, + 114, + 306, + 171 + ], + "lines": [ + { + "bbox": [ + 250, + 114, + 306, + 171 + ], + "spans": [ + { + "bbox": [ + 250, + 114, + 306, + 171 + ], + "type": "image", + "image_path": "9296ea44d1b5bdf6a1adb5c182b71bd016571a194d8379b57fef5dee10f83353.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 254, + 172, + 304, + 198 + ], + "lines": [ + { + "bbox": [ + 254, + 172, + 304, + 198 + ], + "spans": [ + { + "bbox": [ + 254, + 172, + 304, + 198 + ], + "type": "text", + "content": "(c) HINet [4] (29.14dB |0.837)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 206, + 482, + 240 + ], + "lines": [ + { + "bbox": [ + 130, + 206, + 482, + 240 + ], + "spans": [ + { + "bbox": [ + 130, + 206, + 482, + 240 + ], + "type": "text", + "content": "Fig. 1: A deblurring example on HS-ERGB [37] dataset via recent state-of-the-art methods. Compared to other methods, our approach yields a more distinct result owing to its enhanced motion perception capability." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 307, + 114, + 365, + 171 + ], + "blocks": [ + { + "bbox": [ + 307, + 114, + 365, + 171 + ], + "lines": [ + { + "bbox": [ + 307, + 114, + 365, + 171 + ], + "spans": [ + { + "bbox": [ + 307, + 114, + 365, + 171 + ], + "type": "image", + "image_path": "13044559a5ade1896c43bb1813ecb568a3239ac67268cd28c59ff07c03370df3.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 308, + 172, + 365, + 198 + ], + "lines": [ + { + "bbox": [ + 308, + 172, + 365, + 198 + ], + "spans": [ + { + "bbox": [ + 308, + 172, + 365, + 198 + ], + "type": "text", + "content": "(d) NAFNet [3] (30.41dB |0.855)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 366, + 114, + 423, + 171 + ], + "blocks": [ + { + "bbox": [ + 366, + 114, + 423, + 171 + ], + "lines": [ + { + "bbox": [ + 366, + 114, + 423, + 171 + ], + "spans": [ + { + "bbox": [ + 366, + 114, + 423, + 171 + ], + "type": "image", + "image_path": "473f75e245093e693cdddd4d784fb33482225b41569de5edb3f7f85fa30dcde8.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 367, + 172, + 421, + 198 + ], + "lines": [ + { + "bbox": [ + 367, + 172, + 421, + 198 + ], + "spans": [ + { + "bbox": [ + 367, + 172, + 421, + 198 + ], + "type": "text", + "content": "(e) EFNet [33] (30.76dB |0.856)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 425, + 114, + 481, + 171 + ], + "blocks": [ + { + "bbox": [ + 425, + 114, + 481, + 171 + ], + "lines": [ + { + "bbox": [ + 425, + 114, + 481, + 171 + ], + "spans": [ + { + "bbox": [ + 425, + 114, + 481, + 171 + ], + "type": "image", + "image_path": "0ad319576978f72a3e13af2d0e95f04fc6e4cb45953156a234d058cded602fea.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 436, + 172, + 468, + 198 + ], + "lines": [ + { + "bbox": [ + 436, + 172, + 468, + 198 + ], + "spans": [ + { + "bbox": [ + 436, + 172, + 468, + 198 + ], + "type": "text", + "content": "(f) Ours (31.01dB [0.858]" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 267, + 480, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 267, + 480, + 304 + ], + "spans": [ + { + "bbox": [ + 130, + 267, + 480, + 304 + ], + "type": "text", + "content": "detection, and semantic segmentation, to name a few. Hence, image deblurring serves as a fundamental task in the realm of low-level computer vision, reinforcing the image quality and supporting the proper function of various algorithms." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 305, + 482, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 305, + 482, + 484 + ], + "spans": [ + { + "bbox": [ + 130, + 305, + 482, + 484 + ], + "type": "text", + "content": "Traditional image-based techniques often attempt to derive a clear image from a single blurred one. They do this by leveraging natural image priors or making assumptions about blur operations [6, 8, 13, 15, 34, 44]. However, in capturing a single blurred image, much motion information is lost. This lack of motion information during exposure creates undesirable artifacts in the deblurred results, especially in complex motion scenarios, as illustrated in Figures 1c and 1d. To counter this challenge, recent methodologies use more temporal information, such as differently exposed frames [2, 46, 48] and neighboring frames [16, 28, 32]. However, these approaches have their limitations. For instance, the use of differently exposed frames is constrained by the low temporal resolution of frame-based cameras, improving the reconstruction effect primarily in slow-motion scenes. The method that uses adjacent frames to estimate the current frame's motion information can lead to accumulative errors. In summary, traditional deblurring techniques encounter hurdles due to the inability of frame-based cameras to capture motion information during exposure." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": "As opposed to frame-based cameras, event cameras, inspired by biological systems, offer a promising solution to this issue. These cameras' microsecond-level temporal resolution " + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "inline_equation", + "content": "(\\mu \\mathrm{s})" + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": " and asynchronous output architecture allow the recording of fast motions. This capability enables event streams to store pixel intensity changes and provide accurate motion information to deblur the corresponding frame. By using event frames as additional input, some methods attempt to guide the deblurring process through these event cameras [9,25,39,43,49]. One method [45] subdivides the event stream into ternary 2D data and generates a mask to localize high-blur regions, thereby enhancing the network's ability to distinguish different blur levels. Another method [33] symmetrically accumulates event polarities and fuses event and image features using an attention mechanism, enabling the integration of motion information into image reconstruction. However, these methods primarily employ the event stream in a manner akin to polarity accumulation. In essence, they consider only the final accumulation result of relative intensity change between subsequent moments, neglecting the ab" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "text", + "content": "Z. Sun et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": "solute intensity change caused by historical events. Consequently, this approach fails to differentiate between various motion patterns, injecting unnecessary ambiguity into the deblurring process, and leading to edge artifacts as displayed in Figure 1e." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 164, + 482, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 164, + 482, + 450 + ], + "spans": [ + { + "bbox": [ + 130, + 164, + 482, + 450 + ], + "type": "text", + "content": "Given the limited perception of motion during exposure offered by polarity accumulation, we posit that the deviation from the initial moment provides more precise motion information for image deblurring. Each time the event generates, the absolute light intensity changes. The deviation from the initial moment enables us to consider the impact of changes in absolute light intensity, even when the absolute light intensity at the initial moment is unknown. Thus, the deviation, to a degree, implies the order of arrival of the events, offering different information for varying motion patterns. As a result, we propose the Deviation Accumulation (DA) method to enhance motion perception. Specifically, every time a new event arrives, our method accumulates the deviation relative to the initial moment, thus encompassing more comprehensive motion information. Furthermore, we find that longer time scales for event accumulation can capture global motion patterns, while shorter scales precisely depict local fast-motion patterns. To better perceive motion at both global and local time scales, we have designed a Recurrent Motion Extraction (RME) module that works with corresponding event representations. Moreover, in line with multimodal learning [18-20, 29, 36, 42], we present an effective Feature Alignment and Fusion (FAF) module. This module serves to address the issue that simple fusion methods struggle with the disparate nature of event and frame modalities. As a result, these two modalities align, ensuring better information interaction. We integrate these modules into a single image deblurring network. This network extracts the motion features within the exposure time from the event stream and combines them with the texture information supplied by the image to complete the motion blur removal, as illustrated in Figure 1f." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 146, + 451, + 368, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 451, + 368, + 462 + ], + "spans": [ + { + "bbox": [ + 146, + 451, + 368, + 462 + ], + "type": "text", + "content": "The key contributions of this paper are as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 139, + 469, + 481, + 599 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 139, + 469, + 481, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 469, + 481, + 515 + ], + "spans": [ + { + "bbox": [ + 139, + 469, + 481, + 515 + ], + "type": "text", + "content": "- We introduce an event preprocessing method designed specifically for the deblurring problem. The deviation accumulation is capable of responding distinctively to various motion patterns, thereby adding accurate motion information for image deblurring tasks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 139, + 516, + 481, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 516, + 481, + 563 + ], + "spans": [ + { + "bbox": [ + 139, + 516, + 481, + 563 + ], + "type": "text", + "content": "- We develop a network equipped with recurrent motion extraction modules, designed to better extract motion features on local and global time scales. This design enables the network to handle multiple composite motion scenarios effectively." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 139, + 563, + 481, + 599 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 563, + 481, + 599 + ], + "spans": [ + { + "bbox": [ + 139, + 563, + 481, + 599 + ], + "type": "text", + "content": "- We propose a feature alignment and fusion module to mitigate the impacts of inter-modal inconsistencies and ensure that motion information more effectively guides the image deblurring process." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": "Our experimental results indicate that our method achieves state-of-the-art performance on synthetic, semi-synthetic, and real-world datasets, including a 0.61dB improvement in PSNR on the GoPro dataset compared to previous studies [33]. Moreover, our method produces more distinct subjective results, as demonstrated in Figure 1." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 197, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 197, + 91, + 447, + 102 + ], + "type": "text", + "content": "Motion Aware Event Representation-driven Image Deblurring" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 238, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 238, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 238, + 127 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 322, + 153 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 322, + 153 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 322, + 153 + ], + "type": "text", + "content": "2.1 Event-based Motion Deblurring" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 160, + 482, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 160, + 482, + 363 + ], + "spans": [ + { + "bbox": [ + 130, + 160, + 482, + 363 + ], + "type": "text", + "content": "The utilization of event information to assist in deblurring tasks has become increasingly common in recent research [9, 12, 25, 31, 39, 43, 49]. The high temporal resolution and low latency of event cameras not only capture edge positions but also embed moving temporal information within the event stream. This information is crucial to reconstruct the latent sharp image. Many methods have been attempted to explore the relationship between event information, blurred images, and sharp images. One such method [25] includes the Event-based Double Integral (EDI) model, which leverages the principles of event generation to formulate an optimization problem that establishes a correlation between a blurry image and a latent sharp image. In recent studies, the event stream is typically converted into event frames [30, 38], which simplifies the processing of event data within traditional CNN frameworks. [9] combines the motion feature extracted from event data with the image feature, using them as the input for the decoder. [50] employs a Multi-Layer Perceptron (MLP) to predict the fused feature value from event and image features. [33] integrates them through a cross-modal attention method, which generates a query matrix from the image features, but derives key and value matrices from the event features." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 364, + 483, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 364, + 483, + 436 + ], + "spans": [ + { + "bbox": [ + 130, + 364, + 483, + 436 + ], + "type": "text", + "content": "Confronting the prevalent issue of integrating image and event features is the problem of inter-modal inconsistencies, which traditional fusion techniques cannot easily mitigate. To address this, our approach seeks to align these two modalities, optimizing the use of event data. Through this alignment, our goal is to minimize the inconsistencies spawned when merging differing modalities, thus amplifying the efficacy of the deblurring process." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 453, + 270, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 453, + 270, + 466 + ], + "spans": [ + { + "bbox": [ + 132, + 453, + 270, + 466 + ], + "type": "text", + "content": "2.2 Event Representation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 474, + 482, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 482, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 482, + 641 + ], + "type": "text", + "content": "Event data, characterized by its spatial sparsity and strong temporal continuity due to its asynchronous generation, necessitates a custom-designed event representation to deliver the required information. Methods [11, 26, 41] that process events individually maximize the use of high temporal resolution, as suggested by [26], which proposed an adaptive spiking neuron model to manage the fluctuating event input. The effectiveness of these algorithms, however, can be significantly hampered by the noise within the event due to independent processing. A Time Surface [1, 14, 22, 51] is created using the timestamp of the last event at each pixel, which prominently reveals the rich temporal information contained within the event. Furthermore, the event frame [5, 7, 21, 23, 45] transforms the original 4D event data into a 2D frame by accumulating events on the number or the polarity pixel-wise, facilitating extraction with conventional image processing techniques. In [45], the authors utilize ternary data for their event frames to allocate high blur region of the image." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 641, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 641, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 641, + 482, + 665 + ], + "type": "text", + "content": "However, existing event frame approaches result in abundant light intensity change information being lost in the squeezing process, making it impossible to" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "text", + "content": "Z. Sun et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": "distinguish the order of event arrivals and lacking the ability to detect multiple motion patterns. Some techniques attempt to recover some of this lost data by dividing the exposure time into smaller segments, though this essentially equates to sampling the original data. Consequently, we design a motion-aware event representation specifically tailored for motion deblurring." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 198, + 202, + 210 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 198, + 202, + 210 + ], + "spans": [ + { + "bbox": [ + 132, + 198, + 202, + 210 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 225, + 257, + 237 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 225, + 257, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 225, + 257, + 237 + ], + "type": "text", + "content": "3.1 Overall Framework" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 248, + 482, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 248, + 482, + 429 + ], + "spans": [ + { + "bbox": [ + 130, + 248, + 482, + 429 + ], + "type": "text", + "content": "The overall framework of our proposed method is depicted in Figure 2. We adopt a classical encoder-decoder architecture for our approach. Initially, the blurred image is inputted into the image branch, where we use the baseline conv block mentioned in [3] as our fundamental block, to extract relevant features. Simultaneously, the corresponding event frames generated by our deviation accumulation are fed into the Recurrent Motion Extraction (RME) module, which facilitates the extraction of features covering motion in local and global time scales. Following each block of the two branches, the Feature Alignment and Fusion (FAF) module is employed to eliminate the inconsistencies between the two modalities. By performing down-sampling operations, we are able to obtain both shallow and semantic information. In the decoder, skip connections are utilized to incorporate the extracted features for image reconstruction. The architecture of the conv block and the decoder is shown in the supplementary materials. The entire training process is conducted in an end-to-end manner, with the PSNR Loss function [4] employed to optimize the network's parameters." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 449, + 395, + 460 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 449, + 395, + 460 + ], + "spans": [ + { + "bbox": [ + 132, + 449, + 395, + 460 + ], + "type": "text", + "content": "3.2 Deviation Accumulation Event Representation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 472, + 482, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 472, + 482, + 508 + ], + "spans": [ + { + "bbox": [ + 130, + 472, + 482, + 508 + ], + "type": "text", + "content": "Event cameras are bio-inspired sensors which asynchronously output signals. They trigger an event while the logarithm of the intensity change exceeds the preset threshold " + }, + { + "bbox": [ + 130, + 472, + 482, + 508 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 472, + 482, + 508 + ], + "type": "text", + "content": " at any pixel, which is formulated as" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 197, + 518, + 481, + 550 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 518, + 481, + 550 + ], + "spans": [ + { + "bbox": [ + 197, + 518, + 481, + 550 + ], + "type": "interline_equation", + "content": "p = \\left\\{ \\begin{array}{l l} + 1 & \\text {i f} \\log \\left(I _ {x y} (t)\\right) - \\log \\left(I _ {x y} (t - \\Delta t)\\right) \\geq c, \\\\ - 1 & \\text {i f} \\log \\left(I _ {x y} (t)\\right) - \\log \\left(I _ {x y} (t - \\Delta t)\\right) \\leq - c, \\end{array} \\right. \\tag {1}", + "image_path": "a39a14f4f2903d152e5e4a1209bcaebb40e4e911dff5b5de11724b4a8f21f904.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 559, + 482, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 559, + 482, + 632 + ], + "spans": [ + { + "bbox": [ + 130, + 559, + 482, + 632 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 559, + 482, + 632 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 130, + 559, + 482, + 632 + ], + "type": "text", + "content": " is the event's polarity, " + }, + { + "bbox": [ + 130, + 559, + 482, + 632 + ], + "type": "inline_equation", + "content": "I_{xy}(t)" + }, + { + "bbox": [ + 130, + 559, + 482, + 632 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 559, + 482, + 632 + ], + "type": "inline_equation", + "content": "I_{xy}(t - \\Delta t)" + }, + { + "bbox": [ + 130, + 559, + 482, + 632 + ], + "type": "text", + "content": " represent the intensity at the pixel coordinate " + }, + { + "bbox": [ + 130, + 559, + 482, + 632 + ], + "type": "inline_equation", + "content": "(x,y)" + }, + { + "bbox": [ + 130, + 559, + 482, + 632 + ], + "type": "text", + "content": " at time " + }, + { + "bbox": [ + 130, + 559, + 482, + 632 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 559, + 482, + 632 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 559, + 482, + 632 + ], + "type": "inline_equation", + "content": "t - \\Delta t" + }, + { + "bbox": [ + 130, + 559, + 482, + 632 + ], + "type": "text", + "content": ", respectively. An event is a quaternion value " + }, + { + "bbox": [ + 130, + 559, + 482, + 632 + ], + "type": "inline_equation", + "content": "(x,y,t,p)" + }, + { + "bbox": [ + 130, + 559, + 482, + 632 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 130, + 559, + 482, + 632 + ], + "type": "inline_equation", + "content": "x,y" + }, + { + "bbox": [ + 130, + 559, + 482, + 632 + ], + "type": "text", + "content": " denotes the pixel coordinate, " + }, + { + "bbox": [ + 130, + 559, + 482, + 632 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 559, + 482, + 632 + ], + "type": "text", + "content": " represents the time of change, the polarity " + }, + { + "bbox": [ + 130, + 559, + 482, + 632 + ], + "type": "inline_equation", + "content": "p \\in \\{+1, -1\\}" + }, + { + "bbox": [ + 130, + 559, + 482, + 632 + ], + "type": "text", + "content": " indicates the increase and decrease. For a given blurry image " + }, + { + "bbox": [ + 130, + 559, + 482, + 632 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 130, + 559, + 482, + 632 + ], + "type": "text", + "content": ", it can be formulated as the average of latent intensity images " + }, + { + "bbox": [ + 130, + 559, + 482, + 632 + ], + "type": "inline_equation", + "content": "I(t)" + }, + { + "bbox": [ + 130, + 559, + 482, + 632 + ], + "type": "text", + "content": " during the exposure time [25]:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 268, + 641, + 481, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 268, + 641, + 481, + 669 + ], + "spans": [ + { + "bbox": [ + 268, + 641, + 481, + 669 + ], + "type": "interline_equation", + "content": "B = \\frac {1}{T} \\int_ {0} ^ {T} I (t) d t, \\tag {2}", + "image_path": "2a0e90c4450543a4125dbbe1eb26d14539710c5b72e45f4366861d092e0b34de.jpg" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 197, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 197, + 91, + 448, + 103 + ], + "type": "text", + "content": "Motion Aware Event Representation-driven Image Deblurring" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 142, + 114, + 473, + 257 + ], + "blocks": [ + { + "bbox": [ + 142, + 114, + 473, + 257 + ], + "lines": [ + { + "bbox": [ + 142, + 114, + 473, + 257 + ], + "spans": [ + { + "bbox": [ + 142, + 114, + 473, + 257 + ], + "type": "image", + "image_path": "727fbf1276279060979dc2c666d494d80bcb46f3f42f97cec3fb0bcc253b654c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 262, + 482, + 361 + ], + "lines": [ + { + "bbox": [ + 131, + 262, + 482, + 361 + ], + "spans": [ + { + "bbox": [ + 131, + 262, + 482, + 361 + ], + "type": "text", + "content": "Fig. 2: Overview of the architecture of our proposed method for image deblurring. The blurred image and related event frames within the exposure time are fed into the corresponding network branches respectively. The core module of the event branch: (a) Recurrent Motion Extraction (RME) module processes event frames from longer time scales to shorter, merging old features into new features through weighted sum, which results in perceiving various motion patterns under different time ranges. The image branch extracts the features with conv block and fuses them with event features using (b) Feature Alignment and Fusion (FAF) module. FAF module eliminates the inter-modal inconsistencies in a bi-directional way ensuring better information interaction." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 385, + 482, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 385, + 482, + 411 + ], + "spans": [ + { + "bbox": [ + 131, + 385, + 482, + 411 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 385, + 482, + 411 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 131, + 385, + 482, + 411 + ], + "type": "text", + "content": " is the exposure time. For simplicity of understanding, we now only consider the pixel at " + }, + { + "bbox": [ + 131, + 385, + 482, + 411 + ], + "type": "inline_equation", + "content": "(x,y)" + }, + { + "bbox": [ + 131, + 385, + 482, + 411 + ], + "type": "text", + "content": " then combining Equations (1) and (2), we have" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 203, + 417, + 481, + 475 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 203, + 417, + 481, + 475 + ], + "spans": [ + { + "bbox": [ + 203, + 417, + 481, + 475 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} B _ {x y} = \\frac {1}{T} \\int_ {0} ^ {T} \\exp (\\log (I _ {x y} (t))) d t (3) \\\\ = \\frac {1}{T} \\int_ {0} ^ {T} e x p (\\log (I _ {x y} (0)) + c \\int_ {0} ^ {t} p (s) d s) d t. (4) \\\\ \\end{array}", + "image_path": "fbc32aed71a76cb334d2c61b1cbfc7cce72bb199cfbd713fa1676563e04d57d2.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 481, + 482, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 481, + 482, + 530 + ], + "spans": [ + { + "bbox": [ + 131, + 481, + 482, + 530 + ], + "type": "text", + "content": "The internal part of the exponential operation is the logarithmic value of the light intensity of the pixel at time " + }, + { + "bbox": [ + 131, + 481, + 482, + 530 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 131, + 481, + 482, + 530 + ], + "type": "text", + "content": ", while the external part of the exponential operation is the integration and averaging of each moment within the exposure time. Based on the event data, we can discretize Equation (4):" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 215, + 539, + 481, + 609 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 215, + 539, + 481, + 609 + ], + "spans": [ + { + "bbox": [ + 215, + 539, + 481, + 609 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} B _ {x y} = \\frac {1}{N} \\sum_ {t ^ {\\prime} \\in \\mathcal {T}} \\exp \\left(\\log \\left(I _ {x y} (0)\\right) + c \\sum_ {t = 0} ^ {t ^ {\\prime}} p _ {t}\\right) (5) \\\\ = \\frac {I _ {x y} (0)}{N} \\sum_ {t ^ {\\prime} \\in \\mathcal {T}} e x p \\left(c \\sum_ {t = 0} ^ {t ^ {\\prime}} p _ {t}\\right), (6) \\\\ \\end{array}", + "image_path": "74b032a3deb582e89d36e70caad99ded0464be8cc699f109011f37d2498596ff.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 617, + 482, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 641 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 617, + 482, + 641 + ], + "type": "inline_equation", + "content": "\\mathcal{T} = \\{t_1, t_2, \\dots, t_N\\}" + }, + { + "bbox": [ + 130, + 617, + 482, + 641 + ], + "type": "text", + "content": " indicates the set of time events occurs, " + }, + { + "bbox": [ + 130, + 617, + 482, + 641 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 617, + 482, + 641 + ], + "type": "text", + "content": " is the number of events, and " + }, + { + "bbox": [ + 130, + 617, + 482, + 641 + ], + "type": "inline_equation", + "content": "p_t" + }, + { + "bbox": [ + 130, + 617, + 482, + 641 + ], + "type": "text", + "content": " means the polarity of the event at time " + }, + { + "bbox": [ + 130, + 617, + 482, + 641 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 617, + 482, + 641 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 642, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 642, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 642, + 482, + 665 + ], + "type": "text", + "content": "The aforementioned derivation prompts us to consider not only the polarity of the event but also the impact of the alteration in the absolute value after" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "text", + "content": "Z. Sun et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 116, + 331, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 331, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 331, + 129 + ], + "type": "text", + "content": "Algorithm 1 Deviation Accumulation (DA)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 132, + 481, + 165 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 132, + 481, + 165 + ], + "spans": [ + { + "bbox": [ + 132, + 132, + 481, + 165 + ], + "type": "text", + "content": "Input: taking pixel at " + }, + { + "bbox": [ + 132, + 132, + 481, + 165 + ], + "type": "inline_equation", + "content": "(x,y)" + }, + { + "bbox": [ + 132, + 132, + 481, + 165 + ], + "type": "text", + "content": " for an example, the events stream " + }, + { + "bbox": [ + 132, + 132, + 481, + 165 + ], + "type": "inline_equation", + "content": "(x,y,t_i,p_i)" + }, + { + "bbox": [ + 132, + 132, + 481, + 165 + ], + "type": "text", + "content": " in which " + }, + { + "bbox": [ + 132, + 132, + 481, + 165 + ], + "type": "inline_equation", + "content": "i\\in [1,N]" + }, + { + "bbox": [ + 132, + 132, + 481, + 165 + ], + "type": "text", + "content": ", deviation from the initial moment " + }, + { + "bbox": [ + 132, + 132, + 481, + 165 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 132, + 132, + 481, + 165 + ], + "type": "text", + "content": ", accumulation of deviation " + }, + { + "bbox": [ + 132, + 132, + 481, + 165 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 132, + 132, + 481, + 165 + ], + "type": "text", + "content": ", Output: the value of DA event representation " + }, + { + "bbox": [ + 132, + 132, + 481, + 165 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 132, + 132, + 481, + 165 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 136, + 166, + 289, + 242 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 136, + 166, + 206, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 166, + 206, + 175 + ], + "spans": [ + { + "bbox": [ + 136, + 166, + 206, + 175 + ], + "type": "text", + "content": "1: " + }, + { + "bbox": [ + 136, + 166, + 206, + 175 + ], + "type": "inline_equation", + "content": "D = 0,R = 0" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 136, + 177, + 289, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 177, + 289, + 187 + ], + "spans": [ + { + "bbox": [ + 136, + 177, + 289, + 187 + ], + "type": "text", + "content": "2: Counting the number of events " + }, + { + "bbox": [ + 136, + 177, + 289, + 187 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 136, + 177, + 289, + 187 + ], + "type": "text", + "content": ";" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 136, + 188, + 231, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 188, + 231, + 198 + ], + "spans": [ + { + "bbox": [ + 136, + 188, + 231, + 198 + ], + "type": "text", + "content": "3: for " + }, + { + "bbox": [ + 136, + 188, + 231, + 198 + ], + "type": "inline_equation", + "content": "(i = 1" + }, + { + "bbox": [ + 136, + 188, + 231, + 198 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 136, + 188, + 231, + 198 + ], + "type": "inline_equation", + "content": "N)" + }, + { + "bbox": [ + 136, + 188, + 231, + 198 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 136, + 199, + 206, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 199, + 206, + 209 + ], + "spans": [ + { + "bbox": [ + 136, + 199, + 206, + 209 + ], + "type": "text", + "content": "4: " + }, + { + "bbox": [ + 136, + 199, + 206, + 209 + ], + "type": "inline_equation", + "content": "D = D + p_{i}" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 136, + 210, + 206, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 210, + 206, + 219 + ], + "spans": [ + { + "bbox": [ + 136, + 210, + 206, + 219 + ], + "type": "text", + "content": "5: " + }, + { + "bbox": [ + 136, + 210, + 206, + 219 + ], + "type": "inline_equation", + "content": "R = R + D" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 136, + 220, + 184, + 230 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 220, + 184, + 230 + ], + "spans": [ + { + "bbox": [ + 136, + 220, + 184, + 230 + ], + "type": "text", + "content": "6: end for" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 136, + 232, + 189, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 232, + 189, + 242 + ], + "spans": [ + { + "bbox": [ + 136, + 232, + 189, + 242 + ], + "type": "text", + "content": "7: " + }, + { + "bbox": [ + 136, + 232, + 189, + 242 + ], + "type": "inline_equation", + "content": "V = R / N" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 130, + 270, + 481, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 270, + 481, + 426 + ], + "spans": [ + { + "bbox": [ + 130, + 270, + 481, + 426 + ], + "type": "text", + "content": "each accumulation when considering the problem of deblurring. However, since event data cannot provide " + }, + { + "bbox": [ + 130, + 270, + 481, + 426 + ], + "type": "inline_equation", + "content": "I_{xy}(0)" + }, + { + "bbox": [ + 130, + 270, + 481, + 426 + ], + "type": "text", + "content": ", the absolute light intensity is not accessible. But we can indirectly capture this information through the deviation from the initial moment. Therefore, we design a new event representation method for the deblurring problem presented in Algorithm 1. First, we accumulate all the events in the exposure time, so we get a 2D matrix named frequency matrix representing the number of the event occurs at each pixel position. We deal with the events considering their polarities subsequently, recording the deviation at each pixel by an additional tensor. Upon the arrival of the subsequent event, we update the deviation according to polarity and then add the deviation to the corresponding position. In the end, by dividing the cumulative matrix by the frequency matrix, we obtain the average deviation change over the exposure period as the new event representation named Deviation Accumulation (DA)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 426, + 481, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 481, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 481, + 605 + ], + "type": "text", + "content": "Traditional polarity accumulation methods directly accumulating event polarities lead to identical outputs whenever the number of positive and negative events within a given time frame " + }, + { + "bbox": [ + 130, + 426, + 481, + 605 + ], + "type": "inline_equation", + "content": "([0, T])" + }, + { + "bbox": [ + 130, + 426, + 481, + 605 + ], + "type": "text", + "content": " remains unchanged. Nevertheless, by considering the deviation brought by historical events, our method can generate distinct values for events arriving in different orders, thereby improving the ability to perceive motion patterns and providing more precise motion information for the deblurring procedure. For example, as illustrated in Figure 3, (a) shows event signals generated by a motion pattern which represents that the light intensity at that pixel has risen twice and then fallen twice while (d) shows another pattern. Facing these two patterns, (b) and (e) depict the changing process of the outputs of traditional method and our method after each event is accumulated. After processing all four events, traditional polarity accumulation gets the same output of 0 for both patterns, but our DA representation yields 1 and 0.5 respectively which contains richer motion information. More examples of motion patterns our method can distinguish are shown in supplementary materials." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 606, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 606, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 606, + 481, + 666 + ], + "type": "text", + "content": "Overall, the phenomenon that polarity accumulation methods represent some different motion patterns as the same value introduces ambiguity into the deblurring process. Our method has a better perception of the motion patterns and offers a more distinct motion edge than conventional approaches from the visual effect as shown in Figure 3 (c) and (f). Inspired by [33], directly segmenting" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 197, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 197, + 91, + 448, + 103 + ], + "type": "text", + "content": "Motion Aware Event Representation-driven Image Deblurring" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 161, + 118, + 252, + 202 + ], + "blocks": [ + { + "bbox": [ + 161, + 118, + 252, + 202 + ], + "lines": [ + { + "bbox": [ + 161, + 118, + 252, + 202 + ], + "spans": [ + { + "bbox": [ + 161, + 118, + 252, + 202 + ], + "type": "image", + "image_path": "f27df0ad0618d5db90f55bedf00ddfb7b27b5a54a5d9f563739f621712895c21.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 255, + 115, + 364, + 201 + ], + "blocks": [ + { + "bbox": [ + 255, + 115, + 364, + 201 + ], + "lines": [ + { + "bbox": [ + 255, + 115, + 364, + 201 + ], + "spans": [ + { + "bbox": [ + 255, + 115, + 364, + 201 + ], + "type": "image", + "image_path": "42187f67118f8b940e6d4101d46ceb7e90f1c0c1d5e25fc0efcf2e6b161825e5.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 161, + 204, + 252, + 285 + ], + "blocks": [ + { + "bbox": [ + 161, + 204, + 252, + 285 + ], + "lines": [ + { + "bbox": [ + 161, + 204, + 252, + 285 + ], + "spans": [ + { + "bbox": [ + 161, + 204, + 252, + 285 + ], + "type": "image", + "image_path": "2594bfdcceba4383dab92ecf16d109c444ee6f9000430d2f4182e0ca07f17411.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 295, + 482, + 340 + ], + "lines": [ + { + "bbox": [ + 130, + 295, + 482, + 340 + ], + "spans": [ + { + "bbox": [ + 130, + 295, + 482, + 340 + ], + "type": "text", + "content": "Fig. 3: (a) and (d) are event signals of two different motion patterns. (b) shows the Polarity Accumulation (PA) can't distinguish (a) and (d) as it gets the same output of 0 after processing all events while (e) shows our Deviation Accumulation (DA) can. DA perceives more motion information than PA as shown in (c) and (f)." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 256, + 202, + 363, + 285 + ], + "blocks": [ + { + "bbox": [ + 256, + 202, + 363, + 285 + ], + "lines": [ + { + "bbox": [ + 256, + 202, + 363, + 285 + ], + "spans": [ + { + "bbox": [ + 256, + 202, + 363, + 285 + ], + "type": "image", + "image_path": "1c9c90660ce6ee749ed03a3071e895c9dbd79e1abc3052138dd0644cf4112f44.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 378, + 117, + 451, + 202 + ], + "blocks": [ + { + "bbox": [ + 378, + 117, + 451, + 202 + ], + "lines": [ + { + "bbox": [ + 378, + 117, + 451, + 202 + ], + "spans": [ + { + "bbox": [ + 378, + 117, + 451, + 202 + ], + "type": "image", + "image_path": "16ebf709d3c72c1877985cdb2c25867915be52e3966f58db11e80140a702ae40.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 378, + 203, + 451, + 285 + ], + "blocks": [ + { + "bbox": [ + 378, + 203, + 451, + 285 + ], + "lines": [ + { + "bbox": [ + 378, + 203, + 451, + 285 + ], + "spans": [ + { + "bbox": [ + 378, + 203, + 451, + 285 + ], + "type": "image", + "image_path": "a6efd72a2e849957ba7e7139a202dbda424f44cc76c8a7f95b59c21f5a3f3d94.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 360, + 482, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 360, + 482, + 445 + ], + "spans": [ + { + "bbox": [ + 130, + 360, + 482, + 445 + ], + "type": "text", + "content": "the entire exposure time averagely is not suitable for motion deblurring, which focuses on capturing rapid motion but falls short in adequately modeling the continuous movement throughout the exposure. In this paper, we symmetrically accumulate events along the " + }, + { + "bbox": [ + 130, + 360, + 482, + 445 + ], + "type": "inline_equation", + "content": "T / 2" + }, + { + "bbox": [ + 130, + 360, + 482, + 445 + ], + "type": "text", + "content": " axis using the DA method. Taking " + }, + { + "bbox": [ + 130, + 360, + 482, + 445 + ], + "type": "inline_equation", + "content": "[0,T / 2]" + }, + { + "bbox": [ + 130, + 360, + 482, + 445 + ], + "type": "text", + "content": " as an example, we accumulate events on different time scales. The longest accumulating range is " + }, + { + "bbox": [ + 130, + 360, + 482, + 445 + ], + "type": "inline_equation", + "content": "T / 2" + }, + { + "bbox": [ + 130, + 360, + 482, + 445 + ], + "type": "text", + "content": ", followed by " + }, + { + "bbox": [ + 130, + 360, + 482, + 445 + ], + "type": "inline_equation", + "content": "T / 3" + }, + { + "bbox": [ + 130, + 360, + 482, + 445 + ], + "type": "text", + "content": ", and the shortest is " + }, + { + "bbox": [ + 130, + 360, + 482, + 445 + ], + "type": "inline_equation", + "content": "T / 6" + }, + { + "bbox": [ + 130, + 360, + 482, + 445 + ], + "type": "text", + "content": ". A detailed schematic is provided in the supplementary materials." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 459, + 351, + 470 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 459, + 351, + 470 + ], + "spans": [ + { + "bbox": [ + 131, + 459, + 351, + 470 + ], + "type": "text", + "content": "3.3 Recurrent Motion Extraction Module" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 475, + 482, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 475, + 482, + 571 + ], + "spans": [ + { + "bbox": [ + 130, + 475, + 482, + 571 + ], + "type": "text", + "content": "The goal of the Recurrent Motion Extraction (RME) module is to cater to both global slow motion and local fast motion in different time scales. Based on our observations of the motion process and event representations, the focus of the event representations varies in different time scales. In shorter time intervals, the global slow motion is not significant in the event representation due to the smaller number of events triggered. On the contrary slow motion can be clearly recorded in longer intervals, but at the same time, the violent moving part will produce edge spreading artifact." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 571, + 482, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 571, + 482, + 632 + ], + "spans": [ + { + "bbox": [ + 130, + 571, + 482, + 632 + ], + "type": "text", + "content": "To enable the network to accurately model the various in-homogeneous motions, we devised an RME module that incorporates a recurrent mechanism. As shown in Figure 2(a), the event features are fed into the processing module from the longest time scales to the shortest and we use " + }, + { + "bbox": [ + 130, + 571, + 482, + 632 + ], + "type": "inline_equation", + "content": "e1" + }, + { + "bbox": [ + 130, + 571, + 482, + 632 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 571, + 482, + 632 + ], + "type": "inline_equation", + "content": "e2" + }, + { + "bbox": [ + 130, + 571, + 482, + 632 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 571, + 482, + 632 + ], + "type": "inline_equation", + "content": "e3" + }, + { + "bbox": [ + 130, + 571, + 482, + 632 + ], + "type": "text", + "content": " to denote them. Formally, the RME module can be presented as" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 192, + 636, + 481, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 192, + 636, + 481, + 669 + ], + "spans": [ + { + "bbox": [ + 192, + 636, + 481, + 669 + ], + "type": "interline_equation", + "content": "e _ {i} ^ {\\prime} = \\left\\{ \\begin{array}{l l} f _ {e x} \\left(e _ {i}\\right) & \\text {i f} i = 1, \\\\ \\left(1 - \\alpha\\right) \\times f _ {e x} \\left(e _ {i}\\right) + \\alpha \\times f _ {r e} \\left(e _ {i - 1} ^ {\\prime}\\right) & \\text {i f} i = 2, 3, \\end{array} \\right. \\tag {7}", + "image_path": "c9a039b12008746d81e8373977fe42b9b1a33d08f05e2c907c50b83cbde0fd5f.jpg" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "text", + "content": "Z. Sun et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "inline_equation", + "content": "f_{ex}(\\cdot)" + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "inline_equation", + "content": "f_{re}(\\cdot)" + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": " are two baseline conv blocks, " + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "inline_equation", + "content": "f_{ex}(\\cdot)" + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": " is used to extract information from the current representation, " + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "inline_equation", + "content": "f_{re}(\\cdot)" + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": " is used to retain useful information from previous features, " + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": " is the learnable weight as a trade-off between two features. By sharing the parameters of " + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "inline_equation", + "content": "f_{ex}(\\cdot)" + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "inline_equation", + "content": "f_{re}(\\cdot)" + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": " for different time scales, we force the network to be able to perceive motions at varying time scales. Each input feature is symmetric on the time axis so that the priori information about the motion direction is simultaneously fed into the network." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 217, + 356, + 229 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 217, + 356, + 229 + ], + "spans": [ + { + "bbox": [ + 131, + 217, + 356, + 229 + ], + "type": "text", + "content": "3.4 Feature Alignment and Fusion Module" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 237, + 482, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 237, + 482, + 453 + ], + "spans": [ + { + "bbox": [ + 130, + 237, + 482, + 453 + ], + "type": "text", + "content": "Although the information captured by the event camera exhibits a high correlation with the frame camera during the equivalent exposure period, there is still an inevitable gap originating from the modal dissimilarities between them. This phenomenon is frequently mentioned in multi-modal learning [18,42], but in event-guided image restoration tasks people usually [9,33,45] fuse features from two modalities directly using convolutions or attention mechanisms. Therefore, we propose an effective Feature Alignment and Fusion (FAF) module to facilitate enhanced feature fusion in a more coherent environment. Concretely, as shown in Figure 2(b), this module consists of a feature alignment module and a multi-head attention module. We use the operations of multiplication and addition successively to merge the raw information of one modality into the changing process of the other modality, which fills the gap of inter-modal inconsistency. The aligned features are fed into the attention module. In each fusion process, attention plays a bi-directional role, we not only generate the Q matrix through the frame features and the K and V matrix through the event features but also use the event features to perform look-ups on the frame features. The weights of the attention module are shared so that we can fuse the motion information under a more consistent space. Therefore, FAF module can be expressed as" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 139, + 461, + 480, + 475 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 461, + 480, + 475 + ], + "spans": [ + { + "bbox": [ + 139, + 461, + 480, + 475 + ], + "type": "interline_equation", + "content": "x ^ {\\prime} = \\operatorname {C o n v} _ {2} (x) + (\\operatorname {C o n v} _ {1} (x) \\times e + e), \\tag {8}", + "image_path": "0d285aa2d0086ce503ef811b79473ffc2f14233a495d7e9cc7bf425efbed1fae.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 140, + 479, + 480, + 491 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 479, + 480, + 491 + ], + "spans": [ + { + "bbox": [ + 140, + 479, + 480, + 491 + ], + "type": "interline_equation", + "content": "e ^ {\\prime} = \\operatorname {C o n v} _ {2} (e) + (\\operatorname {C o n v} _ {1} (e) \\times x + x), \\tag {9}", + "image_path": "3e8762240c64d4aae50549086e9e80407f826ae815463aba648db7ec978fe2ce.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 140, + 495, + 480, + 507 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 495, + 480, + 507 + ], + "spans": [ + { + "bbox": [ + 140, + 495, + 480, + 507 + ], + "type": "interline_equation", + "content": "F = \\operatorname {C o n v} _ {3} (\\operatorname {C o n c a t} ([ \\text {A t t e n t i o n} (Q _ {x ^ {\\prime}}, K _ {e ^ {\\prime}}, V _ {e ^ {\\prime}}), \\text {A t t e n t i o n} (Q _ {e ^ {\\prime}}, K _ {x ^ {\\prime}}, V _ {x ^ {\\prime}}) ])), \\tag {10}", + "image_path": "fc07502b57d630960f7fcd23bad02d9cc0d3cd9594d1575b92d5abea5ce75789.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 517, + 482, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 517, + 482, + 555 + ], + "spans": [ + { + "bbox": [ + 131, + 517, + 482, + 555 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 517, + 482, + 555 + ], + "type": "inline_equation", + "content": "x'" + }, + { + "bbox": [ + 131, + 517, + 482, + 555 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 517, + 482, + 555 + ], + "type": "inline_equation", + "content": "e'" + }, + { + "bbox": [ + 131, + 517, + 482, + 555 + ], + "type": "text", + "content": " are aligned image features " + }, + { + "bbox": [ + 131, + 517, + 482, + 555 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 131, + 517, + 482, + 555 + ], + "type": "text", + "content": " and event feature " + }, + { + "bbox": [ + 131, + 517, + 482, + 555 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 131, + 517, + 482, + 555 + ], + "type": "text", + "content": " respectively, Attention(\\cdot) denotes the multi-head attention operation, Concat(\\cdot) denotes the concatenation operation along the channel dimension." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 572, + 309, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 572, + 309, + 586 + ], + "spans": [ + { + "bbox": [ + 131, + 572, + 309, + 586 + ], + "type": "text", + "content": "4 Experiments and Analysis" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 597, + 274, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 597, + 274, + 609 + ], + "spans": [ + { + "bbox": [ + 131, + 597, + 274, + 609 + ], + "type": "text", + "content": "4.1 Experimental Settings" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 617, + 479, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 479, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 479, + 640 + ], + "type": "text", + "content": "Datasets. We use three different datasets containing synthetic, semi-synthetic, and real-world frames and events to evaluate the proposed method." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "GoPro: We train the network on the GoPro dataset which is the benchmark dataset for the image motion deblurring [24]. The blurry image is averaged from" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 196, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 196, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 196, + 91, + 448, + 103 + ], + "type": "text", + "content": "Motion Aware Event Representation-driven Image Deblurring" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 140, + 158, + 472, + 335 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 1: The quantitative results on GoPro, HS-ERGB, and REBlur test datasets. The best values are highlighted in bold. HINet+ and UFPNet+ are event-enhanced versions of HINet and UFPNet by concatenating our DA to their input." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 140, + 158, + 472, + 335 + ], + "lines": [ + { + "bbox": [ + 140, + 158, + 472, + 335 + ], + "spans": [ + { + "bbox": [ + 140, + 158, + 472, + 335 + ], + "type": "table", + "html": "
MethodInputGoProHS-ERGBREBlurParams(M)
PSNRSSIMPSNRSSIMPSNRSSIM
LEVS [10]F20.840.54722.130.555--18.2
EDI [25]F+E29.060.94323.930.70436.520.9640.5
SRN [35]F30.260.934--35.100.96110.2
EVDI [49]F+E30.400.90625.130.707--0.4
HINet [4]F32.710.95927.320.80735.580.96588.7
Restormer [47]F32.920.96127.550.80835.500.95926.1
MSDI-Net [17]F33.280.96427.460.80936.140.968241.3
NAFNet [3]F33.710.96727.640.81136.150.96967.8
UFPNet [6]F34.060.96827.640.80936.110.96880.3
HINet+ [4]F+E34.630.96827.660.80837.920.97688.7
UFPNet+ [6]F+E35.220.97227.680.80937.970.97680.3
EFNet [33]F+E35.460.97226.680.80038.120.9758.5
OursF+E36.070.97627.930.81238.470.97813.9
", + "image_path": "c910256561819228f874ee586a3a0c12b3b6df3b608f13ac97ba3cd6c9d9c00b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 366, + 480, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 366, + 480, + 437 + ], + "spans": [ + { + "bbox": [ + 130, + 366, + 480, + 437 + ], + "type": "text", + "content": "adjacent sharp frames, by doing so the dataset contains 3214 pairs of blurry and sharp images. According to the standard division, we adopt 2103 pairs for training and 1111 pairs for testing. For the event data, we use the GoPro raw event dataset provided by [33] which synthesizes events with a random threshold " + }, + { + "bbox": [ + 130, + 366, + 480, + 437 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 366, + 480, + 437 + ], + "type": "text", + "content": " following the Gaussian distribution by ESIM simulator [27]. Following the DA method, the raw event data is preprocessed into 6 event frames for each image." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 437, + 481, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 481, + 498 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 481, + 498 + ], + "type": "text", + "content": "HS-ERGB: The HS-ERGB dataset [37] consists of sharp videos and real-world events, we use [50] released normal blur version which synthesizes blurry frames by averaging 49 interpolating images. Each blurry frame in the dataset has a known exposure time, allowing us to generate the DA representation for each frame within its respective exposure duration." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 498, + 481, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 481, + 557 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 481, + 557 + ], + "type": "text", + "content": "REBlur: The REBlur dataset collects a number of sequences of real-world event data, corresponding with blurry images and reliable ground-truth sharp images [33]. It contains 12 kinds of linear and nonlinear motions. Following the original setting, we use 486 pairs of blurry-sharp images with associated events for training and 983 pairs for testing." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "text", + "content": "Implementation details. We train the end-to-end network without pre-training on a single Tesla V100 GPU. We extract patches of size " + }, + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "text", + "content": " for training images and corresponding event frames, and the batch size is set to 2 by default. The AdamW optimizer is employed (" + }, + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\beta_{1} = 0.9" + }, + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\beta_{2} = 0.99" + }, + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "text", + "content": ") with the initial learning rate 0.0001 with the cosine annealing schedule where the " + }, + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "inline_equation", + "content": "T_{max}" + }, + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "text", + "content": " is 400K iteration. For data augmentation, we perform horizontal and vertical flips, the random noise and the hot pixels are added into event frames in order to simulate the real-world situation. For HS-ERGB and REBlur datasets, We fine-tune the model trained on GoPro with the training set of them. The fine-tuning process" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "Z. Sun et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 159, + 116, + 264, + 220 + ], + "blocks": [ + { + "bbox": [ + 159, + 116, + 264, + 220 + ], + "lines": [ + { + "bbox": [ + 159, + 116, + 264, + 220 + ], + "spans": [ + { + "bbox": [ + 159, + 116, + 264, + 220 + ], + "type": "image", + "image_path": "20a836d4e920f402f1279f0912bc458bdbdbfc258f2ce5277dd485ce5c876056.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 160, + 220, + 263, + 230 + ], + "lines": [ + { + "bbox": [ + 160, + 220, + 263, + 230 + ], + "spans": [ + { + "bbox": [ + 160, + 220, + 263, + 230 + ], + "type": "text", + "content": "(a) Blur image in GoPro testset" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 221, + 263, + 230 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 221, + 263, + 230 + ], + "spans": [ + { + "bbox": [ + 160, + 221, + 263, + 230 + ], + "type": "text", + "content": "(a) Blur image in GoPro testset" + } + ] + } + ], + "index": 3, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 266, + 116, + 312, + 162 + ], + "blocks": [ + { + "bbox": [ + 266, + 116, + 312, + 162 + ], + "lines": [ + { + "bbox": [ + 266, + 116, + 312, + 162 + ], + "spans": [ + { + "bbox": [ + 266, + 116, + 312, + 162 + ], + "type": "image", + "image_path": "090bccd521fd0515526e1fdddda2a1d27d3b2e0539b66ab88db218c349e85371.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 270, + 163, + 304, + 172 + ], + "lines": [ + { + "bbox": [ + 270, + 163, + 304, + 172 + ], + "spans": [ + { + "bbox": [ + 270, + 163, + 304, + 172 + ], + "type": "text", + "content": "(b) HINet" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 313, + 116, + 359, + 162 + ], + "blocks": [ + { + "bbox": [ + 313, + 116, + 359, + 162 + ], + "lines": [ + { + "bbox": [ + 313, + 116, + 359, + 162 + ], + "spans": [ + { + "bbox": [ + 313, + 116, + 359, + 162 + ], + "type": "image", + "image_path": "de240c96888191bad7fd23d6f5d1fd42c49aab6b15b4ad351ec42ba2462276a4.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 315, + 163, + 355, + 172 + ], + "lines": [ + { + "bbox": [ + 315, + 163, + 355, + 172 + ], + "spans": [ + { + "bbox": [ + 315, + 163, + 355, + 172 + ], + "type": "text", + "content": "(c) UFPNet" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 361, + 116, + 406, + 162 + ], + "blocks": [ + { + "bbox": [ + 361, + 116, + 406, + 162 + ], + "lines": [ + { + "bbox": [ + 361, + 116, + 406, + 162 + ], + "spans": [ + { + "bbox": [ + 361, + 116, + 406, + 162 + ], + "type": "image", + "image_path": "855a3626dcaa19b3a6f81c7a38abba6894d8210cbb0351136d26db20c5ded06a.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 364, + 163, + 399, + 172 + ], + "lines": [ + { + "bbox": [ + 364, + 163, + 399, + 172 + ], + "spans": [ + { + "bbox": [ + 364, + 163, + 399, + 172 + ], + "type": "text", + "content": "(d) EFNet" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 408, + 116, + 454, + 162 + ], + "blocks": [ + { + "bbox": [ + 408, + 116, + 454, + 162 + ], + "lines": [ + { + "bbox": [ + 408, + 116, + 454, + 162 + ], + "spans": [ + { + "bbox": [ + 408, + 116, + 454, + 162 + ], + "type": "image", + "image_path": "9d59b2a9ca86a9e8467a6a4624297134ae54fe4303e5a9a48ebb08b3760aa938.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 410, + 163, + 452, + 172 + ], + "lines": [ + { + "bbox": [ + 410, + 163, + 452, + 172 + ], + "spans": [ + { + "bbox": [ + 410, + 163, + 452, + 172 + ], + "type": "text", + "content": "(e) NAFNet" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 266, + 175, + 312, + 220 + ], + "blocks": [ + { + "bbox": [ + 266, + 175, + 312, + 220 + ], + "lines": [ + { + "bbox": [ + 266, + 175, + 312, + 220 + ], + "spans": [ + { + "bbox": [ + 266, + 175, + 312, + 220 + ], + "type": "image", + "image_path": "33624c427a9def027a5e47ac4beec07fe892844f016eb3e4c139b6b2dff08fd6.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 270, + 221, + 306, + 230 + ], + "lines": [ + { + "bbox": [ + 270, + 221, + 306, + 230 + ], + "spans": [ + { + "bbox": [ + 270, + 221, + 306, + 230 + ], + "type": "text", + "content": "(f) HINet+" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 313, + 175, + 359, + 220 + ], + "blocks": [ + { + "bbox": [ + 313, + 175, + 359, + 220 + ], + "lines": [ + { + "bbox": [ + 313, + 175, + 359, + 220 + ], + "spans": [ + { + "bbox": [ + 313, + 175, + 359, + 220 + ], + "type": "image", + "image_path": "b844597b0b2628d9b494fbb49079e9538da6c357ce669b0189408d786989910e.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 315, + 221, + 358, + 230 + ], + "lines": [ + { + "bbox": [ + 315, + 221, + 358, + 230 + ], + "spans": [ + { + "bbox": [ + 315, + 221, + 358, + 230 + ], + "type": "text", + "content": "(g) UFPNet+" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 361, + 175, + 406, + 220 + ], + "blocks": [ + { + "bbox": [ + 361, + 175, + 406, + 220 + ], + "lines": [ + { + "bbox": [ + 361, + 175, + 406, + 220 + ], + "spans": [ + { + "bbox": [ + 361, + 175, + 406, + 220 + ], + "type": "image", + "image_path": "525fd323a7e502680d99c7d455846726cc7a4848aa71c023a1ac82cea2cdfb8e.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 370, + 221, + 399, + 230 + ], + "lines": [ + { + "bbox": [ + 370, + 221, + 399, + 230 + ], + "spans": [ + { + "bbox": [ + 370, + 221, + 399, + 230 + ], + "type": "text", + "content": "(h) Ours" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 409, + 175, + 455, + 220 + ], + "blocks": [ + { + "bbox": [ + 409, + 175, + 455, + 220 + ], + "lines": [ + { + "bbox": [ + 409, + 175, + 455, + 220 + ], + "spans": [ + { + "bbox": [ + 409, + 175, + 455, + 220 + ], + "type": "image", + "image_path": "1352bdc01951e52d0cda0812c9706632ddf092d3b3b441a507c212605ee64399.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 421, + 221, + 443, + 230 + ], + "lines": [ + { + "bbox": [ + 421, + 221, + 443, + 230 + ], + "spans": [ + { + "bbox": [ + 421, + 221, + 443, + 230 + ], + "type": "text", + "content": "(i) GT" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 160, + 276, + 264, + 380 + ], + "blocks": [ + { + "bbox": [ + 131, + 240, + 480, + 262 + ], + "lines": [ + { + "bbox": [ + 131, + 240, + 480, + 262 + ], + "spans": [ + { + "bbox": [ + 131, + 240, + 480, + 262 + ], + "type": "text", + "content": "Fig. 4: Qualitative comparisons on the GoPro dataset. The notation is the same as in Table 1. Zoom in for better view." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 160, + 276, + 264, + 380 + ], + "lines": [ + { + "bbox": [ + 160, + 276, + 264, + 380 + ], + "spans": [ + { + "bbox": [ + 160, + 276, + 264, + 380 + ], + "type": "image", + "image_path": "0843f6c080552a51c92468fb751bc0e4305eec693b7bb36b6c1b2a4b492e83be.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 160, + 380, + 265, + 390 + ], + "lines": [ + { + "bbox": [ + 160, + 380, + 265, + 390 + ], + "spans": [ + { + "bbox": [ + 160, + 380, + 265, + 390 + ], + "type": "text", + "content": "(a) Blur image in REBlur testset" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 160, + 395, + 481, + 422 + ], + "lines": [ + { + "bbox": [ + 160, + 395, + 481, + 422 + ], + "spans": [ + { + "bbox": [ + 160, + 395, + 481, + 422 + ], + "type": "text", + "content": ": Qualitative comparisons on the REBlur dataset. The notation is the same as Table 1. Zoom in for better view." + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 266, + 276, + 312, + 322 + ], + "blocks": [ + { + "bbox": [ + 266, + 276, + 312, + 322 + ], + "lines": [ + { + "bbox": [ + 266, + 276, + 312, + 322 + ], + "spans": [ + { + "bbox": [ + 266, + 276, + 312, + 322 + ], + "type": "image", + "image_path": "e3739c1a5a39ccec854698002b1ed5041ff11d891789b8f9a04cefc4b762dc8d.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 270, + 323, + 304, + 332 + ], + "lines": [ + { + "bbox": [ + 270, + 323, + 304, + 332 + ], + "spans": [ + { + "bbox": [ + 270, + 323, + 304, + 332 + ], + "type": "text", + "content": "(b) HINet" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 315, + 276, + 359, + 322 + ], + "blocks": [ + { + "bbox": [ + 315, + 276, + 359, + 322 + ], + "lines": [ + { + "bbox": [ + 315, + 276, + 359, + 322 + ], + "spans": [ + { + "bbox": [ + 315, + 276, + 359, + 322 + ], + "type": "image", + "image_path": "1cee75ff1bd3522ee52b1985f367cea90277b0673d9543de97be08f469dad2dc.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 316, + 323, + 356, + 332 + ], + "lines": [ + { + "bbox": [ + 316, + 323, + 356, + 332 + ], + "spans": [ + { + "bbox": [ + 316, + 323, + 356, + 332 + ], + "type": "text", + "content": "(c) UFPNet" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 361, + 276, + 407, + 322 + ], + "blocks": [ + { + "bbox": [ + 361, + 276, + 407, + 322 + ], + "lines": [ + { + "bbox": [ + 361, + 276, + 407, + 322 + ], + "spans": [ + { + "bbox": [ + 361, + 276, + 407, + 322 + ], + "type": "image", + "image_path": "49d03350932205f4c7456d7b8004e01f93a025b15a3769694659b63a66aebda9.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 365, + 323, + 400, + 332 + ], + "lines": [ + { + "bbox": [ + 365, + 323, + 400, + 332 + ], + "spans": [ + { + "bbox": [ + 365, + 323, + 400, + 332 + ], + "type": "text", + "content": "(d) EFNet" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 409, + 276, + 454, + 322 + ], + "blocks": [ + { + "bbox": [ + 409, + 276, + 454, + 322 + ], + "lines": [ + { + "bbox": [ + 409, + 276, + 454, + 322 + ], + "spans": [ + { + "bbox": [ + 409, + 276, + 454, + 322 + ], + "type": "image", + "image_path": "b60a52f299b5b271275e91e328b201333ba9fc08023bb258ff4899f4394d7338.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 411, + 323, + 452, + 332 + ], + "lines": [ + { + "bbox": [ + 411, + 323, + 452, + 332 + ], + "spans": [ + { + "bbox": [ + 411, + 323, + 452, + 332 + ], + "type": "text", + "content": "(e) NAFNet" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "bbox": [ + 131, + 446, + 480, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 446, + 480, + 471 + ], + "spans": [ + { + "bbox": [ + 131, + 446, + 480, + 471 + ], + "type": "text", + "content": "consisted of 4000 iterations while keeping the other configurations consistent with the previous experiments." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 131, + 489, + 379, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 489, + 379, + 502 + ], + "spans": [ + { + "bbox": [ + 131, + 489, + 379, + 502 + ], + "type": "text", + "content": "4.2 Comparison with State-of-the-Art Methods" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": "Table 1 shows the comparison results of our method and other SOTA methods on GoPro [24], HS-ERGB [50] and REBlur [33]. We compare our network with not only event-based methods, but also with some SOTA image-only deblurring networks enhancing with event input. The results illustrate that: (1) Our method outperforms all other methods in synthetic, semi-synthetic and real-world datasets, which indicates that with the proper utilization of events the designed approach is able to handle blur with different distributions. Specifically, the average gains of our method over the second-best one are 0.61dB, 0.25dB and 0.35dB in terms of PSNR on GoPro, HS-ERGB and REBlur datasets, respectively. (2) Through the enhancement of our proposed event representation, the HINet+ and the UFPNet+ increase 1.92dB and 1.16dB in PSNR on the benchmark GoPro dataset compared with their image-only version, respectively. Taking HINet as an example, with the same operation of [33], HINet with DA" + } + ] + } + ], + "index": 35 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 197, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 197, + 91, + 448, + 102 + ], + "type": "text", + "content": "Motion Aware Event Representation-driven Image Deblurring" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 164, + 114, + 210, + 160 + ], + "blocks": [ + { + "bbox": [ + 164, + 114, + 210, + 160 + ], + "lines": [ + { + "bbox": [ + 164, + 114, + 210, + 160 + ], + "spans": [ + { + "bbox": [ + 164, + 114, + 210, + 160 + ], + "type": "image", + "image_path": "734fd1b1a209722f46554a3beff3f560d60b9d4c52ce867556462ddd29643728.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 173, + 161, + 200, + 171 + ], + "lines": [ + { + "bbox": [ + 173, + 161, + 200, + 171 + ], + "spans": [ + { + "bbox": [ + 173, + 161, + 200, + 171 + ], + "type": "text", + "content": "(a) GT" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 211, + 114, + 258, + 160 + ], + "blocks": [ + { + "bbox": [ + 211, + 114, + 258, + 160 + ], + "lines": [ + { + "bbox": [ + 211, + 114, + 258, + 160 + ], + "spans": [ + { + "bbox": [ + 211, + 114, + 258, + 160 + ], + "type": "image", + "image_path": "23665851adff685957098865bc95f18a700e064978f64467d82b3f335ad0e2c5.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 215, + 161, + 255, + 171 + ], + "lines": [ + { + "bbox": [ + 215, + 161, + 255, + 171 + ], + "spans": [ + { + "bbox": [ + 215, + 161, + 255, + 171 + ], + "type": "text", + "content": "(b) Blurry" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 260, + 114, + 306, + 160 + ], + "blocks": [ + { + "bbox": [ + 260, + 114, + 306, + 160 + ], + "lines": [ + { + "bbox": [ + 260, + 114, + 306, + 160 + ], + "spans": [ + { + "bbox": [ + 260, + 114, + 306, + 160 + ], + "type": "image", + "image_path": "e436f01102b7a8802bc8f0f4fda16ca5e7ee47f048512df78a52a0737fd0df72.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 261, + 161, + 306, + 171 + ], + "lines": [ + { + "bbox": [ + 261, + 161, + 306, + 171 + ], + "spans": [ + { + "bbox": [ + 261, + 161, + 306, + 171 + ], + "type": "text", + "content": "(c) UFPNet" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 307, + 114, + 354, + 160 + ], + "blocks": [ + { + "bbox": [ + 307, + 114, + 354, + 160 + ], + "lines": [ + { + "bbox": [ + 307, + 114, + 354, + 160 + ], + "spans": [ + { + "bbox": [ + 307, + 114, + 354, + 160 + ], + "type": "image", + "image_path": "116209ed0274009ad194fa824ab167736e9ae06fc2d09cdee0ae569b63bfb553.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 309, + 161, + 353, + 171 + ], + "lines": [ + { + "bbox": [ + 309, + 161, + 353, + 171 + ], + "spans": [ + { + "bbox": [ + 309, + 161, + 353, + 171 + ], + "type": "text", + "content": "(d) NAFNet" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 356, + 114, + 402, + 160 + ], + "blocks": [ + { + "bbox": [ + 356, + 114, + 402, + 160 + ], + "lines": [ + { + "bbox": [ + 356, + 114, + 402, + 160 + ], + "spans": [ + { + "bbox": [ + 356, + 114, + 402, + 160 + ], + "type": "image", + "image_path": "75b923889f3b670525027bcd31c23cc3db0dd31392f76380b22b20f16a83888f.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 360, + 161, + 399, + 171 + ], + "lines": [ + { + "bbox": [ + 360, + 161, + 399, + 171 + ], + "spans": [ + { + "bbox": [ + 360, + 161, + 399, + 171 + ], + "type": "text", + "content": "(e) EFNet" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 403, + 114, + 450, + 160 + ], + "blocks": [ + { + "bbox": [ + 403, + 114, + 450, + 160 + ], + "lines": [ + { + "bbox": [ + 403, + 114, + 450, + 160 + ], + "spans": [ + { + "bbox": [ + 403, + 114, + 450, + 160 + ], + "type": "image", + "image_path": "1e7c2eabaea83f858b8621cc979ffad7fca20e9e2d345effe15da8aa95aa0aa0.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 411, + 161, + 443, + 171 + ], + "lines": [ + { + "bbox": [ + 411, + 161, + 443, + 171 + ], + "spans": [ + { + "bbox": [ + 411, + 161, + 443, + 171 + ], + "type": "text", + "content": "(f) Ours" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 164, + 203, + 234, + 274 + ], + "blocks": [ + { + "bbox": [ + 186, + 180, + 425, + 191 + ], + "lines": [ + { + "bbox": [ + 186, + 180, + 425, + 191 + ], + "spans": [ + { + "bbox": [ + 186, + 180, + 425, + 191 + ], + "type": "text", + "content": "Fig. 6: Qualitative comparisons on the HS-ERGB dataset." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 164, + 203, + 234, + 274 + ], + "lines": [ + { + "bbox": [ + 164, + 203, + 234, + 274 + ], + "spans": [ + { + "bbox": [ + 164, + 203, + 234, + 274 + ], + "type": "image", + "image_path": "41e0651326b76cccb5d5710c1442995461b95c4ddbe1fcacc18b94d134bc4605.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 164, + 276, + 233, + 344 + ], + "blocks": [ + { + "bbox": [ + 164, + 276, + 233, + 344 + ], + "lines": [ + { + "bbox": [ + 164, + 276, + 233, + 344 + ], + "spans": [ + { + "bbox": [ + 164, + 276, + 233, + 344 + ], + "type": "image", + "image_path": "a720aa58896079b0e0b39e7458067a7fbec6479614a71146e05ad3f99b7d9700.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 182, + 345, + 215, + 354 + ], + "lines": [ + { + "bbox": [ + 182, + 345, + 215, + 354 + ], + "spans": [ + { + "bbox": [ + 182, + 345, + 215, + 354 + ], + "type": "text", + "content": "(a) Blurry" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 235, + 204, + 306, + 274 + ], + "blocks": [ + { + "bbox": [ + 235, + 204, + 306, + 274 + ], + "lines": [ + { + "bbox": [ + 235, + 204, + 306, + 274 + ], + "spans": [ + { + "bbox": [ + 235, + 204, + 306, + 274 + ], + "type": "image", + "image_path": "68e51ec54743ea9c8e64884178afd1a439dff5366a8d45a80122461dcc428e57.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 236, + 276, + 306, + 343 + ], + "blocks": [ + { + "bbox": [ + 236, + 276, + 306, + 343 + ], + "lines": [ + { + "bbox": [ + 236, + 276, + 306, + 343 + ], + "spans": [ + { + "bbox": [ + 236, + 276, + 306, + 343 + ], + "type": "image", + "image_path": "aca0016fe3ce7794d334f49c234da809ebe1c6ab4418653c1f8df20c1c5fd89a.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 247, + 345, + 294, + 354 + ], + "lines": [ + { + "bbox": [ + 247, + 345, + 294, + 354 + ], + "spans": [ + { + "bbox": [ + 247, + 345, + 294, + 354 + ], + "type": "text", + "content": "(b) SCER [33]" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 307, + 204, + 378, + 274 + ], + "blocks": [ + { + "bbox": [ + 307, + 204, + 378, + 274 + ], + "lines": [ + { + "bbox": [ + 307, + 204, + 378, + 274 + ], + "spans": [ + { + "bbox": [ + 307, + 204, + 378, + 274 + ], + "type": "image", + "image_path": "c83f0d2d8eb537da6533e4b9f86be810fee0e31af3b33665379c2553ef6c7061.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 308, + 276, + 378, + 344 + ], + "blocks": [ + { + "bbox": [ + 308, + 276, + 378, + 344 + ], + "lines": [ + { + "bbox": [ + 308, + 276, + 378, + 344 + ], + "spans": [ + { + "bbox": [ + 308, + 276, + 378, + 344 + ], + "type": "image", + "image_path": "e006b54c00ea90658069b0fd4689d9fe01929f4c59460730601fbe9ea5ab731a.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 320, + 345, + 364, + 354 + ], + "lines": [ + { + "bbox": [ + 320, + 345, + 364, + 354 + ], + "spans": [ + { + "bbox": [ + 320, + 345, + 364, + 354 + ], + "type": "text", + "content": "(c) DA (Ours)" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 380, + 204, + 449, + 274 + ], + "blocks": [ + { + "bbox": [ + 380, + 204, + 449, + 274 + ], + "lines": [ + { + "bbox": [ + 380, + 204, + 449, + 274 + ], + "spans": [ + { + "bbox": [ + 380, + 204, + 449, + 274 + ], + "type": "image", + "image_path": "2355dc71134c668061af47e87b50aa1f1aa24daa40fa40843e9198ed91743256.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 380, + 276, + 449, + 344 + ], + "blocks": [ + { + "bbox": [ + 380, + 276, + 449, + 344 + ], + "lines": [ + { + "bbox": [ + 380, + 276, + 449, + 344 + ], + "spans": [ + { + "bbox": [ + 380, + 276, + 449, + 344 + ], + "type": "image", + "image_path": "f1a387eda2a305341be0a7284808fbc25adaa1a60547ed3ebee3eb97710f9f44.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 403, + 345, + 426, + 354 + ], + "lines": [ + { + "bbox": [ + 403, + 345, + 426, + 354 + ], + "spans": [ + { + "bbox": [ + 403, + 345, + 426, + 354 + ], + "type": "text", + "content": "(d) GT" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 363, + 480, + 385 + ], + "lines": [ + { + "bbox": [ + 131, + 363, + 480, + 385 + ], + "spans": [ + { + "bbox": [ + 131, + 363, + 480, + 385 + ], + "type": "text", + "content": "Fig. 7: Visual comparisons on GoPro dataset with different event representation. Compared to SCER representation, our DA representation reduces edge artifacts." + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "bbox": [ + 130, + 410, + 480, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 410, + 480, + 433 + ], + "spans": [ + { + "bbox": [ + 130, + 410, + 480, + 433 + ], + "type": "text", + "content": "event representation exceeds 0.94dB over the result in [33], which presents the fact that DA extracts more conducive information for motion deblurring." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 130, + 434, + 480, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 434, + 480, + 506 + ], + "spans": [ + { + "bbox": [ + 130, + 434, + 480, + 506 + ], + "type": "text", + "content": "We also show the visual comparison results to verify the effectiveness of our method in Figure 4, Figure 5 and Figure 6 with representative samples from each dataset. As illustrated in the figures, our model reconstructs better details and has minor spatial distortions compared with other methods. This indicates that our event representation perceives more motion information and the modules effectively embed it to reconstruct a high-quality image with enhanced details." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 131, + 525, + 245, + 536 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 525, + 245, + 536 + ], + "spans": [ + { + "bbox": [ + 131, + 525, + 245, + 536 + ], + "type": "text", + "content": "4.3 Ablation Studies" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 130, + 545, + 480, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 480, + 569 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 480, + 569 + ], + "type": "text", + "content": "We conduct plenty ablation experiments on GoPro dataset to verify the effectiveness of our event representation and proposed method." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 130, + 570, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 481, + 665 + ], + "type": "text", + "content": "Effectiveness of Event Representation. To demonstrate the adaptability of our event representation for motion deblurring tasks, as shown in Table 2, we replace our DA representation with several other types of representations and trained the network accordingly. The results reveal certain shortcomings of the alternative representations. Firstly, the stack representation simply accumulates all events on their polarities which leaves out the temporal information. Secondly, the SBT approach [40], which subdivides the entire exposure time equally, proves to be inadequate for modeling the varying degrees of blur. Thirdly, the SCER" + } + ] + } + ], + "index": 32 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "Z. Sun et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 148, + 136, + 463, + 269 + ], + "blocks": [ + { + "bbox": [ + 207, + 114, + 405, + 125 + ], + "lines": [ + { + "bbox": [ + 207, + 114, + 405, + 125 + ], + "spans": [ + { + "bbox": [ + 207, + 114, + 405, + 125 + ], + "type": "text", + "content": "Table 2: Ablation study of each module effects." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 148, + 136, + 463, + 269 + ], + "lines": [ + { + "bbox": [ + 148, + 136, + 463, + 269 + ], + "spans": [ + { + "bbox": [ + 148, + 136, + 463, + 269 + ], + "type": "table", + "html": "
Event RepresentationRMEFusionPSNRSSIM
Stack33.240.953
SBT [40]✓FAF(Ours)35.630.973
SCER [33]35.810.975
Concat.34.020.967
Add34.740.970
Attention35.750.974
ACA(Ours)✓Only Align Event+Attention35.780.975
Only Align Image+Attention35.740.974
35.420.973
ACA(Ours)✗FAF(Ours)36.070.976
✓
", + "image_path": "5d434a7baeecb04e0bb393f3d63371deeecd9117414eb116d1158d1296fd0d7d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 152, + 280, + 254, + 338 + ], + "blocks": [ + { + "bbox": [ + 152, + 280, + 254, + 338 + ], + "lines": [ + { + "bbox": [ + 152, + 280, + 254, + 338 + ], + "spans": [ + { + "bbox": [ + 152, + 280, + 254, + 338 + ], + "type": "image", + "image_path": "81f7908c63a90b433e90594dc13fa912cb37f5504f5f993a8d50ce93a37c0d8f.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 179, + 339, + 229, + 348 + ], + "lines": [ + { + "bbox": [ + 179, + 339, + 229, + 348 + ], + "spans": [ + { + "bbox": [ + 179, + 339, + 229, + 348 + ], + "type": "text", + "content": "(a) w/o RME" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 255, + 280, + 358, + 338 + ], + "blocks": [ + { + "bbox": [ + 255, + 280, + 358, + 338 + ], + "lines": [ + { + "bbox": [ + 255, + 280, + 358, + 338 + ], + "spans": [ + { + "bbox": [ + 255, + 280, + 358, + 338 + ], + "type": "image", + "image_path": "11be934ee6cb38b10dfb35c3fc2d91777a02b74c7898ac526a9795643b474ccb.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 282, + 339, + 333, + 348 + ], + "lines": [ + { + "bbox": [ + 282, + 339, + 333, + 348 + ], + "spans": [ + { + "bbox": [ + 282, + 339, + 333, + 348 + ], + "type": "text", + "content": "(b) w/o RME" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 359, + 280, + 462, + 338 + ], + "blocks": [ + { + "bbox": [ + 359, + 280, + 462, + 338 + ], + "lines": [ + { + "bbox": [ + 359, + 280, + 462, + 338 + ], + "spans": [ + { + "bbox": [ + 359, + 280, + 462, + 338 + ], + "type": "image", + "image_path": "1f18276d5b5d8ae61f7cddeb2945d238a52e8902dde8319fdaa11eaef552ba1d.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 386, + 339, + 435, + 348 + ], + "lines": [ + { + "bbox": [ + 386, + 339, + 435, + 348 + ], + "spans": [ + { + "bbox": [ + 386, + 339, + 435, + 348 + ], + "type": "text", + "content": "(c) w/o RME" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 153, + 349, + 254, + 407 + ], + "blocks": [ + { + "bbox": [ + 153, + 349, + 254, + 407 + ], + "lines": [ + { + "bbox": [ + 153, + 349, + 254, + 407 + ], + "spans": [ + { + "bbox": [ + 153, + 349, + 254, + 407 + ], + "type": "image", + "image_path": "7815bdf963277e3713ad059dbd509e05af04828114043170576ce6d39efadbb0.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 181, + 407, + 227, + 417 + ], + "lines": [ + { + "bbox": [ + 181, + 407, + 227, + 417 + ], + "spans": [ + { + "bbox": [ + 181, + 407, + 227, + 417 + ], + "type": "text", + "content": "(d) w/RME" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 426, + 482, + 460 + ], + "lines": [ + { + "bbox": [ + 131, + 426, + 482, + 460 + ], + "spans": [ + { + "bbox": [ + 131, + 426, + 482, + 460 + ], + "type": "text", + "content": "Fig. 8: Visual comparison results of w/ and w/o RME module. All of these scenarios involve movement in different directions and speeds, RME module can successfully capture motion features in these complex motion scenarios. Zoom in for better view." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 255, + 349, + 358, + 407 + ], + "blocks": [ + { + "bbox": [ + 255, + 349, + 358, + 407 + ], + "lines": [ + { + "bbox": [ + 255, + 349, + 358, + 407 + ], + "spans": [ + { + "bbox": [ + 255, + 349, + 358, + 407 + ], + "type": "image", + "image_path": "7a552d1a9227850a52be4560f38c65a17ba9c1bf2257581d64e68c82f2f56e03.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 284, + 407, + 330, + 417 + ], + "lines": [ + { + "bbox": [ + 284, + 407, + 330, + 417 + ], + "spans": [ + { + "bbox": [ + 284, + 407, + 330, + 417 + ], + "type": "text", + "content": "(e) w/RME" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 359, + 349, + 461, + 407 + ], + "blocks": [ + { + "bbox": [ + 359, + 349, + 461, + 407 + ], + "lines": [ + { + "bbox": [ + 359, + 349, + 461, + 407 + ], + "spans": [ + { + "bbox": [ + 359, + 349, + 461, + 407 + ], + "type": "image", + "image_path": "a03e2927b922c9edc7b8ca0699305e75b5d34138e48b931449aba8642488c107.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 388, + 407, + 433, + 417 + ], + "lines": [ + { + "bbox": [ + 388, + 407, + 433, + 417 + ], + "spans": [ + { + "bbox": [ + 388, + 407, + 433, + 417 + ], + "type": "text", + "content": "(f) w/RME" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 484, + 482, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 484, + 482, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 484, + 482, + 581 + ], + "type": "text", + "content": "method [33], which directly accumulates event polarities, is found to be incapable of capturing the influence of absolute intensity changes. This limitation is illustrated in Figure 3. In contrast, our DA representation not only possesses a solid physical foundation for event generation but also perceives more motion patterns. As a result, it achieves a 0.26dB improvement in performance compared to the best alternative representation. We also provide the visual comparison on the image reconstruction results in Figure 7. Our DA representation reduces motion-generated edge artifacts, resulting in a better deblurring effect." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "Effectiveness of Recurrent Motion Extraction Module. To assess the RME module's contribution, we replace it with a baseline conv block and adjust the parameters for fair comparison. In this approach, we concatenate event frames of different time scales and feed them into the conv block. The findings, as presented in the last two rows of Table 2, clearly indicate that the contribution of " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "0.65\\mathrm{dB}" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": " increment in PSNR stems from the network structure design, rather than an increase in parameter quantity. This shows the importance of incorpo" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 197, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 197, + 91, + 447, + 102 + ], + "type": "text", + "content": "Motion Aware Event Representation-driven Image Deblurring" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "content": "rating the module for effectively leveraging motion information in multi-timescales and improving performance in motion deblurring tasks. We also present the visual comparisons when removing the RME module in Figure 8, the RME module adeptly captures both fast local and slow global motion characteristics." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 164, + 482, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 164, + 482, + 284 + ], + "spans": [ + { + "bbox": [ + 130, + 164, + 482, + 284 + ], + "type": "text", + "content": "Effectiveness of Feature Alignment and Fusion Module. To illustrate the effectiveness of our FAF module, we compare it with several common fusion methods, including concatenation, addition, and cross-modality attention. As shown in Table 2, it is arduous to utilize the motion information contained in event feature through the direct concatenation or addition. While attention mechanisms can capture relative parts in the two features, the inherent inconsistencies within the two modalities still lead to a 0.32dB performance gap when compared to our FAF module. We also use a t-SNE visualization in Figure 9 to demonstrate the effectiveness of our feature alignment, where two originally in-congruent modal features are mixed up together after the FAF module." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 285, + 299, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 285, + 299, + 499 + ], + "spans": [ + { + "bbox": [ + 130, + 285, + 299, + 499 + ], + "type": "text", + "content": "Furthermore, we also verify the importance of bi-directional alignment by conducting experiments solely on aligning event or image features. Since both events and frames provide crucial information for image restoration, such as edge motion information and textual information, respectively, it is unreasonable to let one modality dominate the alignment process. As shown in Table 2, the bi-directional alignment outperforms the highest score achieved by one-way correction by 0.29dB. This demonstrates the necessity of fully utilizing the mutual information between the two modalities in order to achieve better image reconstruction results." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 316, + 297, + 471, + 432 + ], + "blocks": [ + { + "bbox": [ + 316, + 297, + 471, + 432 + ], + "lines": [ + { + "bbox": [ + 316, + 297, + 471, + 432 + ], + "spans": [ + { + "bbox": [ + 316, + 297, + 471, + 432 + ], + "type": "image", + "image_path": "5fb4c0b4afc0a3f1a56e00ab48428f673c8a69bd8eea8b47eb91ccc2480c19a4.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 432, + 482, + 498 + ], + "lines": [ + { + "bbox": [ + 304, + 432, + 482, + 498 + ], + "spans": [ + { + "bbox": [ + 304, + 432, + 482, + 498 + ], + "type": "text", + "content": "Fig. 9: t-SNE visualization of original and aligned features. The distributions of the aligned two features overlap more compared to the original features. This indicates the effectiveness of FAF module to reduce the inter-modal inconsistency." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 519, + 220, + 532 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 519, + 220, + 532 + ], + "spans": [ + { + "bbox": [ + 132, + 519, + 220, + 532 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "In this paper, we propose an event-based image deblurring method. Firstly, we rethink the limit of traditional event preprocessing method and propose a new deviation accumulation (DA) technique. With deviation, the new approach can distinguish the order of events and perceive more motion patterns. Moreover, we design an image deblurring network equipped with recurrent motion extraction (RME) module and feature alignment and fusion (FAF) module to fit our proposed event representation. The RME module can better perceive motions in different time scales, and the FAF module promotes better integration of the two modalities. Finally, extensive experimental results over several datasets demonstrate the effectiveness of the proposed algorithm." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "Z. Sun et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 140 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 140 + ], + "type": "text", + "content": "Acknowledgements. This work was supported by National Natural Science Foundation of China (NSFC) under Grants 62225207 and 62276243." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 160, + 197, + 173 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 160, + 197, + 173 + ], + "spans": [ + { + "bbox": [ + 132, + 160, + 197, + 173 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 188, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 138, + 188, + 480, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 188, + 480, + 220 + ], + "spans": [ + { + "bbox": [ + 138, + 188, + 480, + 220 + ], + "type": "text", + "content": "1. Benosman, R., Clercq, C., Lagorce, X., Ieng, S.H., Bartolozzi, C.: Event-based visual flow. IEEE transactions on neural networks and learning systems 25(2), 407-417 (2013)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 222, + 480, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 222, + 480, + 243 + ], + "spans": [ + { + "bbox": [ + 138, + 222, + 480, + 243 + ], + "type": "text", + "content": "2. Chang, M., Feng, H., Xu, Z., Li, Q.: Low-light image restoration with short-and long-exposure raw pairs. IEEE Transactions on Multimedia 24, 702-714 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 244, + 480, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 244, + 480, + 277 + ], + "spans": [ + { + "bbox": [ + 138, + 244, + 480, + 277 + ], + "type": "text", + "content": "3. Chen, L., Chu, X., Zhang, X., Sun, J.: Simple baselines for image restoration. In: Proceedings of the European conference on computer vision (ECCV). pp. 17-33. Springer (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 277, + 480, + 310 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 277, + 480, + 310 + ], + "spans": [ + { + "bbox": [ + 138, + 277, + 480, + 310 + ], + "type": "text", + "content": "4. Chen, L., Lu, X., Zhang, J., Chu, X., Chen, C.: Hinet: Half instance normalization network for image restoration. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 182-192 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 311, + 480, + 343 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 311, + 480, + 343 + ], + "spans": [ + { + "bbox": [ + 138, + 311, + 480, + 343 + ], + "type": "text", + "content": "5. Cook, M., Gugelmann, L., Jug, F., Krautz, C., Steger, A.: Interacting maps for fast visual interpretation. In: The 2011 International Joint Conference on Neural Networks. pp. 770-776. IEEE (2011)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 344, + 480, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 344, + 480, + 388 + ], + "spans": [ + { + "bbox": [ + 138, + 344, + 480, + 388 + ], + "type": "text", + "content": "6. Fang, Z., Wu, F., Dong, W., Li, X., Wu, J., Shi, G.: Self-supervised non-uniform kernel estimation with flow-based motion prior for blind image deblurring. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18105-18114 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 388, + 480, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 388, + 480, + 421 + ], + "spans": [ + { + "bbox": [ + 138, + 388, + 480, + 421 + ], + "type": "text", + "content": "7. Gehrig, D., Rebecq, H., Gallego, G., Scaramuzza, D.: Eklt: Asynchronous photometric feature tracking using events and frames. International Journal of Computer Vision 128(3), 601-618 (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 422, + 480, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 422, + 480, + 454 + ], + "spans": [ + { + "bbox": [ + 138, + 422, + 480, + 454 + ], + "type": "text", + "content": "8. Ji, S.W., Lee, J., Kim, S.W., Hong, J.P., Baek, S.J., Jung, S.W., Ko, S.J.: Xydeblur: divide and conquer for single image deblurring. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17421-17430 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 455, + 480, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 455, + 480, + 487 + ], + "spans": [ + { + "bbox": [ + 138, + 455, + 480, + 487 + ], + "type": "text", + "content": "9. Jiang, Z., Zhang, Y., Zou, D., Ren, J., Lv, J., Liu, Y.: Learning event-based motion deblurring. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3320-3329 (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 488, + 480, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 488, + 480, + 521 + ], + "spans": [ + { + "bbox": [ + 138, + 488, + 480, + 521 + ], + "type": "text", + "content": "10. Jin, M., Meishvili, G., Favaro, P.: Learning to extract a video sequence from a single motion-blurred image. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 6334-6342 (2018)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 521, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 521, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 138, + 521, + 480, + 544 + ], + "type": "text", + "content": "1. Kim, H., Handa, A., Benosman, R., Ieng, S.H., Davison, A.J.: Simultaneous mosaicing and tracking with an event camera. J. Solid State Circ 43, 566-576 (2008)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 544, + 480, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 544, + 480, + 576 + ], + "spans": [ + { + "bbox": [ + 138, + 544, + 480, + 576 + ], + "type": "text", + "content": "2. Kim, T., Lee, J., Wang, L., Yoon, K.J.: Event-guided deblurring of unknown exposure time videos. In: Proceedings of the European conference on computer vision (ECCV). pp. 519-538. Springer (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 577, + 480, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 577, + 480, + 598 + ], + "spans": [ + { + "bbox": [ + 138, + 577, + 480, + 598 + ], + "type": "text", + "content": "3. Krishnan, D., Tay, T., Fergus, R.: Blind deconvolution using a normalized sparsity measure. In: CVPR 2011. pp. 233-240. IEEE (2011)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 599, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 599, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 599, + 480, + 632 + ], + "type": "text", + "content": "4. Lagorce, X., Orchard, G., Galluppi, F., Shi, B.E., Benosman, R.B.: Hots: a hierarchy of event-based time-surfaces for pattern recognition. IEEE transactions on pattern analysis and machine intelligence 39(7), 1346-1359 (2016)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "type": "text", + "content": "5. Levin, A., Weiss, Y., Durand, F., Freeman, W.T.: Understanding and evaluating blind deconvolution algorithms. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 1964-1971. IEEE (2009)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 197, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 197, + 91, + 447, + 102 + ], + "type": "text", + "content": "Motion Aware Event Representation-driven Image Deblurring" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "type": "text", + "content": "16. Li, D., Shi, X., Zhang, Y., Cheung, K.C., See, S., Wang, X., Qin, H., Li, H.: A simple baseline for video restoration with grouped spatial-temporal shift. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9822-9832 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 161, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 161, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 133, + 161, + 481, + 194 + ], + "type": "text", + "content": "17. Li, D., Zhang, Y., Cheung, K.C., Wang, X., Qin, H., Li, H.: Learning degradation representations for image deblurring. In: Proceedings of the European conference on computer vision (ECCV). pp. 736-753. Springer (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 194, + 481, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 194, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 133, + 194, + 481, + 237 + ], + "type": "text", + "content": "18. Li, Y., Yu, A.W., Meng, T., Caine, B., Ngiam, J., Peng, D., Shen, J., Lu, Y., Zhou, D., Le, Q.V., et al.: Deepfusion: Lidar-camera deep fusion for multi-modal 3d object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17182-17191 (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 238, + 481, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 238, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 133, + 238, + 481, + 270 + ], + "type": "text", + "content": "19. Liang, M., Yang, B., Wang, S., Urtasun, R.: Deep continuous fusion for multi-sensor 3d object detection. In: Proceedings of the European conference on computer vision (ECCV). pp. 641-656 (2018)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 271, + 481, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 481, + 304 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 481, + 304 + ], + "type": "text", + "content": "20. Liang, X., Qian, Y., Guo, Q., Cheng, H., Liang, J.: Af: An association-based fusion method for multi-modal classification. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(12), 9236-9254 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 304, + 481, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 481, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 481, + 325 + ], + "type": "text", + "content": "21. Liu, M., Delbruck, T.: Adaptive time-slice block-matching optical flow algorithm for dynamic vision sensors. BMVC (2018)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 326, + 481, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 481, + 369 + ], + "type": "text", + "content": "22. Manderscheid, J., Sironi, A., Bourdis, N., Migliore, D., Lepetit, V.: Speed invariant time surface for learning to detect corner points with event-based cameras. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10245-10254 (2019)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 370, + 481, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 370, + 481, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 370, + 481, + 413 + ], + "type": "text", + "content": "23. Maqueda, A.I., Loquercio, A., Gallego, G., García, N., Scaramuzza, D.: Event-based vision meets deep learning on steering prediction for self-driving cars. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5419-5427 (2018)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 414, + 481, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 414, + 481, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 414, + 481, + 445 + ], + "type": "text", + "content": "24. Nah, S., Hyun Kim, T., Mu Lee, K.: Deep multi-scale convolutional neural network for dynamic scene deblurring. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3883-3891 (2017)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 447, + 481, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 447, + 481, + 490 + ], + "spans": [ + { + "bbox": [ + 132, + 447, + 481, + 490 + ], + "type": "text", + "content": "25. Pan, L., Scheerlinck, C., Yu, X., Hartley, R., Liu, M., Dai, Y.: Bringing a blurry frame alive at high frame-rate with an event camera. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6820-6829 (2019)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 491, + 481, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 491, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 491, + 481, + 533 + ], + "type": "text", + "content": "26. Paredes-Vallés, F., Scheper, K.Y., De Croon, G.C.: Unsupervised learning of a hierarchical spiking neural network for optical flow estimation: From events to global motion perception. IEEE transactions on pattern analysis and machine intelligence 42(8), 2051-2064 (2019)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 534, + 481, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 481, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 481, + 555 + ], + "type": "text", + "content": "27. Rebecq, H., Gehrig, D., Scaramuzza, D.: Esim: an open event camera simulator. In: Conference on robot learning. pp. 969-982. PMLR (2018)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 556, + 481, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 556, + 481, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 556, + 481, + 578 + ], + "type": "text", + "content": "28. Ren, D., Shang, W., Yang, Y., Zuo, W.: Aggregating long-term sharp features via hybrid transformers for video deblurring. arXiv preprint arXiv:2309.07054 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 578, + 481, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 481, + 610 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 481, + 610 + ], + "type": "text", + "content": "29. Ren, S., Du, Y., Lv, J., Han, G., He, S.: Learning from the master: Distilling cross-modal advanced knowledge for lip reading. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13325-13333 (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 611, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 481, + 643 + ], + "type": "text", + "content": "30. Shang, W., Ren, D., Zou, D., Ren, J.S., Luo, P., Zuo, W.: Bringing events into video deblurring with non-consecutively blurry frames. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4531-4540 (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 644, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 644, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 644, + 481, + 665 + ], + "type": "text", + "content": "31. Song, C., Bajaj, C., Huang, Q.: Deblursr: Event-based motion deblurring under the spiking representation. arXiv preprint arXiv:2303.08977 (2023)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "Z. Sun et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "32. Su, S., Delbracio, M., Wang, J., Sapiro, G., Heidrich, W., Wang, O.: Deep video deblurring for hand-held cameras. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1279-1288 (2017)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 482, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 482, + 194 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 482, + 194 + ], + "type": "text", + "content": "33. Sun, L., Sakaridis, C., Liang, J., Jiang, Q., Yang, K., Sun, P., Ye, Y., Wang, K., Gool, L.V.: Event-based fusion for motion deblurring with cross-modal attention. In: Proceedings of the European conference on computer vision (ECCV). pp. 412-428. Springer (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 194, + 482, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 194, + 482, + 226 + ], + "spans": [ + { + "bbox": [ + 130, + 194, + 482, + 226 + ], + "type": "text", + "content": "34. Sun, L., Cho, S., Wang, J., Hays, J.: Edge-based blur kernel estimation using patch priors. In: IEEE international conference on computational photography (ICCP). pp. 1-8. IEEE (2013)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 227, + 482, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 227, + 482, + 259 + ], + "spans": [ + { + "bbox": [ + 130, + 227, + 482, + 259 + ], + "type": "text", + "content": "35. Tao, X., Gao, H., Shen, X., Wang, J., Jia, J.: Scale-recurrent network for deep image deblurring. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 8174-8182 (2018)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 259, + 482, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 259, + 482, + 293 + ], + "spans": [ + { + "bbox": [ + 130, + 259, + 482, + 293 + ], + "type": "text", + "content": "36. Tian, Y., Krishnan, D., Isola, P.: Contrastive multiview coding. In: Proceedings of the European conference on computer vision (ECCV). pp. 776-794. Springer (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 293, + 482, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 293, + 482, + 335 + ], + "spans": [ + { + "bbox": [ + 130, + 293, + 482, + 335 + ], + "type": "text", + "content": "37. Tulyakov, S., Gehrig, D., Georgoulis, S., Erbach, J., Gehrig, M., Li, Y., Scaramuzza, D.: Time lens: Event-based video frame interpolation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 16155-16164 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 336, + 482, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 336, + 482, + 369 + ], + "spans": [ + { + "bbox": [ + 130, + 336, + 482, + 369 + ], + "type": "text", + "content": "38. Vitoria, P., Georgoulis, S., Tulyakov, S., Bochicchio, A., Erbach, J., Li, Y.: Event-based image deblurring with dynamic motion awareness. In: Proceedings of the European conference on computer vision (ECCV). pp. 95-112. Springer (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 370, + 482, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 370, + 482, + 403 + ], + "spans": [ + { + "bbox": [ + 130, + 370, + 482, + 403 + ], + "type": "text", + "content": "39. Wang, B., He, J., Yu, L., Xia, G.S., Yang, W.: Event enhanced high-quality image recovery. In: Proceedings of the European conference on computer vision (ECCV). pp. 155-171. Springer (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 403, + 482, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 403, + 482, + 445 + ], + "spans": [ + { + "bbox": [ + 130, + 403, + 482, + 445 + ], + "type": "text", + "content": "40. Wang, L., Ho, Y.S., Yoon, K.J., et al.: Event-based high dynamic range image and very high frame rate video generation using conditional generative adversarial networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10081-10090 (2019)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 446, + 482, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 446, + 482, + 479 + ], + "spans": [ + { + "bbox": [ + 130, + 446, + 482, + 479 + ], + "type": "text", + "content": "41. Weikersdorfer, D., Conradt, J.: Event-based particle filtering for robot self-localization. In: 2012 IEEE International Conference on Robotics and Biomimetics (ROBIO). pp. 866-870. IEEE (2012)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 479, + 482, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 479, + 482, + 501 + ], + "spans": [ + { + "bbox": [ + 130, + 479, + 482, + 501 + ], + "type": "text", + "content": "42. Xia, W., Li, X., Deng, A., Xiong, H., Dou, D., Hu, D.: Robust cross-modal knowledge distillation for unconstrained videos. arXiv preprint arXiv:2304.07775 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 502, + 482, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 502, + 482, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 502, + 482, + 533 + ], + "type": "text", + "content": "43. Xu, F., Yu, L., Wang, B., Yang, W., Xia, G.S., Jia, X., Qiao, Z., Liu, J.: Motion deblurring with real events. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2583-2592 (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 534, + 482, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 534, + 482, + 566 + ], + "spans": [ + { + "bbox": [ + 130, + 534, + 482, + 566 + ], + "type": "text", + "content": "44. Xu, L., Jia, J.: Two-phase kernel estimation for robust motion deblurring. In: Proceedings of the European conference on computer vision (ECCV). pp. 157-170. Springer (2010)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 567, + 482, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 567, + 482, + 599 + ], + "spans": [ + { + "bbox": [ + 130, + 567, + 482, + 599 + ], + "type": "text", + "content": "45. Yang, D., Yamac, M.: Motion aware double attention network for dynamic scene deblurring. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1113-1123 (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 600, + 482, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 600, + 482, + 621 + ], + "spans": [ + { + "bbox": [ + 130, + 600, + 482, + 621 + ], + "type": "text", + "content": "46. Yuan, L., Sun, J., Quan, L., Shum, H.Y.: Image deblurring with blurred/noisy image pairs. In: ACM SIGGRAPH 2007 papers, pp. 1-es (2007)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 622, + 482, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 622, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 622, + 482, + 665 + ], + "type": "text", + "content": "47. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H.: Restormer: Efficient transformer for high-resolution image restoration. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5728-5739 (2022)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 197, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 197, + 91, + 447, + 102 + ], + "type": "text", + "content": "Motion Aware Event Representation-driven Image Deblurring" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 238 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 138 + ], + "type": "text", + "content": "48. Zhang, S., Zhen, A., Stevenson, R.L.: Deep motion blur removal using noisy/blurry image pairs. Journal of Electronic Imaging 30(3), 033022-033022 (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 140, + 481, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 140, + 481, + 171 + ], + "spans": [ + { + "bbox": [ + 133, + 140, + 481, + 171 + ], + "type": "text", + "content": "49. Zhang, X., Yu, L.: Unifying motion deblurring and frame interpolation with events. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17765-17774 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 481, + 205 + ], + "type": "text", + "content": "50. Zhang, X., Yu, L., Yang, W., Liu, J., Xia, G.S.: Generalizing event-based motion deblurring in real-world scenarios. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10734-10744 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 205, + 481, + 238 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 481, + 238 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 481, + 238 + ], + "type": "text", + "content": "51. Zhou, Y., Gallego, G., Rebecq, H., Kneip, L., Li, H., Scaramuzza, D.: Semi-dense 3d reconstruction with a stereo event camera. In: Proceedings of the European conference on computer vision (ECCV). pp. 235-251 (2018)" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "Z. Sun et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Motion Keyframe Interpolation for Any Human Skeleton using Point Cloud-based Human Motion Data Homogenisation/391ac41a-5023-4e3a-b209-fdaecfa62b6a_content_list.json b/2024/Motion Keyframe Interpolation for Any Human Skeleton using Point Cloud-based Human Motion Data Homogenisation/391ac41a-5023-4e3a-b209-fdaecfa62b6a_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..c2dd108f6cbc2735654b5f20cc7ed81eb0fdaaa3 --- /dev/null +++ b/2024/Motion Keyframe Interpolation for Any Human Skeleton using Point Cloud-based Human Motion Data Homogenisation/391ac41a-5023-4e3a-b209-fdaecfa62b6a_content_list.json @@ -0,0 +1,1685 @@ +[ + { + "type": "text", + "text": "Motion Keyframe Interpolation for Any Human Skeleton via Temporally Consistent Point Cloud Sampling and Reconstruction", + "text_level": 1, + "bbox": [ + 225, + 140, + 779, + 209 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Clinton Mo $^{1}$ , Kun Hu $^{1,\\star}$ , Chengjiang Long $^{2}$ , Dong Yuan $^{1}$ , and Zhiyong Wang $^{1}$", + "bbox": [ + 215, + 233, + 782, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 The University of Sydney, Darlington NSW 2008, Australia \nclmo6615@uni.sydney.edu.au, {kun.hu, zhiyong.wang, dong.yuan}@sydney.edu.au", + "bbox": [ + 294, + 261, + 705, + 303 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Meta Reality Labs, Burlingame, CA, USA", + "bbox": [ + 351, + 303, + 648, + 318 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "clong1@meta.com", + "bbox": [ + 442, + 319, + 560, + 330 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. In the character animation field, modern supervised keyframe interpolation models have demonstrated exceptional performance in constructing natural human motions from sparse pose definitions. As supervised models, large motion datasets are necessary to facilitate the learning process; however, since motion is represented with fixed hierarchical skeletons, such datasets are incompatible for skeletons outside the datasets' native configurations. Consequently, the expected availability of a motion dataset for desired skeletons severely hinders the feasibility of learned interpolation in practice. To combat this limitation, we propose Point Cloud-based Motion Representation Learning (PC-MRL), an unsupervised approach to enabling cross-compatibility between skeletons for motion interpolation learning. PC-MRL consists of a skeleton obfuscation strategy using temporal point cloud sampling, and an unsupervised skeleton reconstruction method from point clouds. We devise a temporal point-wise K-nearest neighbors loss for unsupervised learning. Moreover, we propose First-frame Offset Quaternion (FOQ) and Rest Pose Augmentation (RPA) strategies to overcome necessary limitations of our unsupervised point cloud-to-skeletal motion process. Comprehensive experiments demonstrate the effectiveness of PC-MRL in motion interpolation for desired skeletons without supervision from native datasets.", + "bbox": [ + 259, + 366, + 743, + 643 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: 3D point clouds $\\cdot$ Human body motion $\\cdot$ Dataset creation", + "bbox": [ + 261, + 656, + 733, + 672 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 695, + 375, + 710 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D character animation workflows largely rely on the concepts of keyframing with interpolation, often referred to as the pose-to-pose principle [13]. By defining key poses at correct timings, algorithms can be employed to generate intermediate poses, and thereby eliminating the need for defining each frame individually [5, 42]. Although this keyframe-based workflow is less costly than frame-by-frame production, human motion is often complex, and natural depictions", + "bbox": [ + 212, + 726, + 787, + 818 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding author.", + "bbox": [ + 218, + 825, + 385, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d2218860b0ef14c222532445ddae55cd3415e56b8bcf0d28f405c5ce4aab6c0c.jpg", + "image_caption": [ + "(a) Linear interpolation" + ], + "image_footnote": [], + "bbox": [ + 267, + 150, + 390, + 257 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e3bfa326343bc12db7c8e9e70f1524810ef7753ef624ffd77c7e09b16cfd69c3.jpg", + "image_caption": [ + "(b) PC-MRL (Our method)" + ], + "image_footnote": [], + "bbox": [ + 444, + 150, + 560, + 257 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/44d2a55e7203e40d15ac1c42294ae3da0256fcc101b9cd092f1004ab6aa6faa2.jpg", + "image_caption": [ + "(c) Ground truth" + ], + "image_footnote": [], + "bbox": [ + 617, + 150, + 735, + 257 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ee25c25af7037e509bc3473cb4d65ecefc982ec85197badae4aac3be6468f142.jpg", + "image_caption": [ + "(a) Hierarchical skeleton" + ], + "image_footnote": [], + "bbox": [ + 295, + 320, + 359, + 417 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/61f3a519e8e4097725bad99f609add4adf90d3642355878e3fb74294765cbaf4.jpg", + "image_caption": [ + "Fig. 1: Motion interpolation of walking keyframes conducted by our approach.", + "(b) Skeletal motion" + ], + "image_footnote": [], + "bbox": [ + 444, + 318, + 555, + 417 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/f908a2abf1d811ef7f297dd1286a831a856acd18723f2f794f65c2e56c3db689.jpg", + "image_caption": [ + "(c) Point cloud representation", + "Fig. 2: Our method leverages point cloud representations (c) to obfuscate the hierarchical dependencies of skeletal motion. Joint colours correspond to ordinal indices in original motion data representation. Point clouds are coloured by temporal index." + ], + "image_footnote": [], + "bbox": [ + 617, + 319, + 733, + 417 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "demand a significant number of keyframes, especially for realistic motions that adhere to physical properties and constraints. Although motion capture (MoCap) is frequently employed as an alternative for achieving realistic human motions, MoCap often nonetheless necessitates subsequent keyframing to correct unwanted motion elements and address other imperfections.", + "bbox": [ + 212, + 521, + 784, + 597 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In recent years, the interpolation process has been widely studied, particularly with the advent of machine learning methods for sequential data. Compared to conventional methods such as linear interpolation (LERP), machine learning methods have demonstrated the capability to derive visually natural motions from sparser keyframe sets [12, 18, 34, 37, 40], which we can observe in Fig. 1. Data-driven interpolation approaches require large motion capture datasets to learn the necessary motion features for effective synthesis of transitions between keyframes. In the field of 3D motions, this aspect significantly restricts the feasibility of learned interpolation methods in animation practice, as motion/pose data are tied to specific skeletal configurations and are not cross-compatible. That is, the dataset's native skeleton, i.e. source skeleton, is structurally different and incompatible with the desired skeleton, i.e. target skeleton.", + "bbox": [ + 212, + 598, + 785, + 777 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Therefore, we propose a novel point cloud-based motion representation learning (PC-MRL) approach, in which skeletal hierarchies are obfuscated by sampling into the point cloud medium, and reconstructed into skeletal motion data using an unsupervised neural network. Unlike raw pose and motion data, the", + "bbox": [ + 212, + 779, + 785, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "C. Mo et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "point cloud format is a non-hierarchical representation, and effectively obscures any skeletal configuration as displayed in Fig. 2. PC-MRL samples a set of points around the rest position of skeletons using a combination of uniform and normal distributions, and geometrically parents each point to an associated bone in order to effectively represent the associated skeletons' motion data in a temporally consistent manner. Subsequently, for the reconstruction of motion data into a given desired skeleton, we propose an unsupervised learning scheme, featuring a K-nearest neighbors (KNN) loss between cross-skeleton point clouds to optimize a transformer neural network. The network processes the point cloud-based motion representations and generates the corresponding target skeletal motion representations. By obscuring both source and predicted target motion data into the point cloud space, our strategy allows for a direct comparison, utilizing KNN to minimize the visual difference between the two point clouds.", + "bbox": [ + 212, + 146, + 787, + 343 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The unsupervised nature of our target skeleton reconstruction scheme creates two main disparities between the expected and predicted motion features. Firstly, the exact roll rotation axis of bones cannot be represented in the point cloud format. To this end, we introduce First-frame Offset Quaternion (FOQ) representations to incorporate relative roll values, which are obtainable from temporally consistent point clouds. This approach is agnostic to absolute roll rotations, providing a standardisation mechanism for the resulting motion data within the context of motion interpolation learning. Second, KNN-based objectives learn skeletal representations in a geometrically optimal manner, which does not always reflect the desired skeletal behaviour. This disparity is commonly observed with smaller skeletal features, such as shoulder and hip bones. As such, we augment training motions using Rest Pose Augmentation (RPA) to increase the range of skeletal behaviours that our interpolation model learns. Extensive experiments demonstrate the effectiveness of our proposed method, achieving performance targets near the level of direct dataset supervision.", + "bbox": [ + 212, + 343, + 787, + 568 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In summary, this paper presents the following key contributions:", + "bbox": [ + 238, + 569, + 702, + 583 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. We propose a novel method for achieving unsupervised human motion reconstruction from point cloud data, enabling skeleton-agnostic motion interpolation learning for the first time.", + "2. To train motion interpolation models, we formulate first-frame offset quaternions to represent bone rotations with relative roll data, as well as a rest position augmentation strategy to address skeletal configuration variability.", + "3. We perform comprehensive experiments to demonstrate our method's effectiveness towards learning motion interpolation, despite the absence of directly compatible datasets in the training process." + ], + "bbox": [ + 220, + 592, + 782, + 724 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related works", + "text_level": 1, + "bbox": [ + 215, + 750, + 392, + 763 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We explore existing research on motion interpolation approaches, and motion data modelling methods in general. In addition, our proposed approach bears resemblance to unsupervised approaches for the motion re-targeting research problem, and as such, we will also analyse existing methods in this field.", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Motion Keyframe Interpolation for Any Human Skeleton", + "bbox": [ + 351, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Data-Driven Motion Modelling Methods", + "text_level": 1, + "bbox": [ + 215, + 146, + 596, + 161 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Contemporary research in motion modelling predominantly explore learned data-driven methods, using neural networks to capture, model, and establish correlations between the motion features of various skeletal configurations [29, 35, 41, 51, 55]. Motion features alone have demonstrated considerable efficacy in cross-skeleton imitation using a variety of strategies, including task-driven objectives [19, 45], diffusion-based generation [25, 49, 52], and latent feature consistency [1-3, 7, 46]. Latent consistency in particular learns inter-skeleton correlations without the need for paired datasets, which reduces its barrier of entry.", + "bbox": [ + 212, + 172, + 785, + 292 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Likewise, motion interpolation strategies have evolved through the use of learning based approaches. The inherent numerical precision of motion data has traditionally posed a challenge for the adoption of the learning approaches, due to their approximate nature [21,27]. This challenge is further amplified by the temporally sparse distribution of keyframes. Recently, a number of recurrent neural network based methods [17, 18, 43, 53] and transformer-based networks [12, 34, 37] have demonstrated encouraging interpolation performance for the transition between keyframes. Additionally, various other motion modelling methods are capable for keyframe based interpolation, albeit with limited intermediate frame representation capabilities [19, 25, 45].", + "bbox": [ + 212, + 292, + 787, + 443 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "A major constraint inherent in all data-driven motion modelling methods is their dependency on datasets featuring specific skeletal configurations. To address this, skeleton-free training strategies have been explored for elementary tasks involving 3D characters. These include vertex-based pose transfer [7,8,31,47], and point cloud-based human shape reconstruction [23,48]. Notably, in these methods, 3D mesh and point cloud coordinates serve as general 3D data representations. Concurrently, other research has significantly minimized the required volume of training data, bringing it down to single example sequences. This reduction has been achieved through the use of patch matching techniques [28] and imitation learning within physics-based simulations [30,39]. Reinforcement learning has facilitated the generation of goal-driven motion without pre-existing datasets [32,38,50]. However, these simulation-based methods have typically been either exceedingly complex to implement in practice, or yield results that are too inflexible and/or prone to errors for animation workflows.", + "bbox": [ + 212, + 444, + 787, + 655 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Motion Re-Targeting without Supervision", + "text_level": 1, + "bbox": [ + 215, + 678, + 606, + 694 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Conventional approaches in motion re-targeting have been a cornerstone of 3D animation for decades, primarily dedicated to converting motion capture data into 3D skeletal motions [4, 15]. Typically, raw motion capture data is optically recorded, yielding primarily spatial information. In contrast, motion data is largely rotational in nature [5], often represented in $SO(3)$ space. Therefore, a principal objective of motion re-targeting solutions has been to effectively bridge these two distinct types of representations. In [14], extensive manually adjustable constraints were introduced for flexible re-targeting between skeletons of identical topology (i.e., same hierarchy, different bone lengths and rest poses). Key", + "bbox": [ + 212, + 704, + 785, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "C. Mo et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b2e7a16acdf66b05482e71dd89d4c86ffdc781ecfb2d07b2bd8b9030fd37b84b.jpg", + "image_caption": [ + "Fig. 3: Illustration of point cloud-based motion representation learning (PC-MRL) pipeline for keyframe interpolation to any human skeleton. Our unsupervised point cloud-to-pose decoding approach (I) first adapts motion features from existing datasets to new skeletons. We then employ the adapted motion data to supervise motion interpolation models (II), using FOQ and rest pose augmentations to address remaining discrepancies between the adapted and expected target skeleton motion patterns." + ], + "image_footnote": [], + "bbox": [ + 223, + 152, + 772, + 280 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "practices in this approach, including end-effector matching, joint contacts, and inter-skeleton point correlations, are still relevant to current animation pipelines. Point correlations using Inverse Kinematics (IK) have also been proposed as an intermediate representation to facilitate the adaptation between topologically distinct skeletons [9, 36], albeit with hierarchical pairing limitations.", + "bbox": [ + 212, + 414, + 782, + 489 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In response, morphologically-independent motion control systems were proposed using IK joint chains and joint-wise constraints [20,26]. Such systems often expect highly standardised motion skeleton features, and are generally impractical for re-targeting motions produced for different control schemes. Additionally, the reliance on Euler angles for extensive classical re-targeting constraints poses a variety of issues regarding rotational freedom [11], as well as compatibility with animation workflows [33], which predominantly use quaternion-based systems.", + "bbox": [ + 212, + 489, + 784, + 595 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Despite these advancements, the challenge of arbitrary skeleton re-targeting from motion datasets still remains largely unresolved, limiting any widespread application of data-driven pipelines. Our paper aims to effectively address this pervasive limitation.", + "bbox": [ + 212, + 595, + 785, + 656 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Methodology", + "text_level": 1, + "bbox": [ + 214, + 681, + 380, + 698 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As illustrated in Fig. 3, our proposed method consists of two main components: point cloud-to-motion reconstruction and motion keyframe interpolation.", + "bbox": [ + 212, + 714, + 782, + 744 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Quaternion-based Motion Representations", + "text_level": 1, + "bbox": [ + 214, + 768, + 609, + 782 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We first declare a set of notations for representing skeletons. A skeleton $S$ is defined by a tree of bones $\\mathcal{B}$ . A bone $\\mathbf{v}_n \\in \\mathcal{B}$ is defined by its bone parent $\\mathbf{v}_m \\in \\mathcal{B}$ , and a head position $\\mathbf{H}_n \\in \\mathbb{R}^3$ which denotes the bone's base offset from", + "bbox": [ + 212, + 794, + 785, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Motion Keyframe Interpolation for Any Human Skeleton", + "bbox": [ + 351, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "its parent head $\\mathbf{H}_m$ . In addition, a tail position $\\mathbf{T}_n \\in \\mathbb{R}^3$ helps to define the end point of the bone's line segment representation, starting from its head position, of $\\mathbf{v}_n$ . The root bone $\\mathbf{v}_0$ has no parent, and thus no head position $\\mathbf{H}_0$ . The set of $\\mathbf{H}_n$ and $\\mathbf{T}_n$ for each bone $\\mathbf{v}_n \\in \\mathcal{B}$ make up the rest position of the skeleton.", + "bbox": [ + 212, + 146, + 782, + 205 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Next, we define the properties of motion data. At any given frame $t$ , a skeleton's pose position can be represented as global 3D positions $\\mathbf{p}_{n,t} \\in \\mathbb{R}^3$ and global quaternion rotations $\\mathbf{q}_{n,t} \\in \\mathbb{R}^4$ for each bone $\\mathbf{v}_n \\in \\mathcal{B}$ . For a given motion sequence $M$ , only the root bone is given a 3D position in pose position, i.e. $\\mathbf{p}_{0,t}$ as all other pose positions are derived using Forwards Kinematics (FK [11,24]):", + "bbox": [ + 212, + 207, + 784, + 282 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {p} _ {n, t} = \\mathbf {p} _ {m, t} + Q R \\left(\\mathbf {H} _ {n}, \\mathbf {q} _ {m, t}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 398, + 292, + 784, + 306 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $QR$ is the quaternion rotation function, and $\\mathbf{v}_m$ is the parent joint. In summary, a motion sequence $M$ of $|M|$ frames is defined by:", + "bbox": [ + 212, + 314, + 782, + 345 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nM = \\left\\{\\mathbf {p} _ {0, t} \\cup \\left\\{\\mathbf {q} _ {n, t} \\mid \\mathbf {v} _ {n} \\in \\mathcal {B} \\right\\} \\mid 0 \\leq t < | M | \\right\\} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 348, + 354, + 784, + 369 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Point Cloud Obfuscation and Skeleton Reconstruction", + "text_level": 1, + "bbox": [ + 214, + 388, + 707, + 402 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our approach is based on the hypothesis that 3D motion sequences can be visually represented by a more universal format: 3D point cloud sequences. By geometrically sampling the volume around a skeleton, point cloud data becomes a non-hierarchical representation of its pose(s), obfuscating the original skeletal configuration. Due to this decoupling between the skeleton configuration and motion representation, a successful motion data reconstruction from point cloud data would be the first step to enabling learnable cross-skeleton motion features. Formally, for a given source skeleton $S_A$ with an associated motion sequence $M_A$ , and any humanoid skeletal configuration $S_B$ , the point cloud obfuscation and reconstruction module is to learn a function that can adequately perform $F_{S_A \\to S_B}(M_A) = \\hat{M}_B$ , where $\\hat{M}_B$ is a visually similar motion adaptation of $M_A$ on skeleton $S_B$ , which is an estimation of $M_B$ .", + "bbox": [ + 212, + 410, + 784, + 592 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Motion Data Obfuscation with Point Clouds To sample the point cloud $\\mathcal{X}$ , we first generate a set of points $u\\in \\mathcal{X}$ along the line segments defined by the global head and tail positions for each bone $\\mathbf{v}_n$ . In the bone-local space to $\\mathbf{v}_n$ , this would mean sampling $u$ by a uniformly distributed factor $\\alpha \\sim \\mathcal{U}_{[0,1]}$ . Specifically, $\\alpha = 0$ aligns with the head position of the bone, and $\\alpha = 1$ indicates the tail position. By further sampling $u$ via a normal distribution, our sampling strategy can capture the surrounding volume around the skeleton. Formally, to sample $u$ from its associated bone $\\mathbf{v}_n$ within a standard deviation of $\\sigma$ , we have:", + "bbox": [ + 212, + 611, + 784, + 731 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {H} _ {x} \\sim \\mathcal {N} (\\alpha \\mathbf {T} _ {n}, \\sigma),\n$$\n", + "text_format": "latex", + "bbox": [ + 408, + 753, + 535, + 768 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {p} _ {u, t} = \\mathbf {p} _ {n, t} + Q R (\\mathbf {H} _ {x}, \\mathbf {q} _ {n, t}) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 403, + 763, + 784, + 787 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Notably, this sampling process as in Eq. 3 is associated with the skeletal configuration only, by which the sampled points are temporally consistent for further processing with the motion sequence. Specifically, each point maintains a", + "bbox": [ + 212, + 794, + 782, + 839 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "C. Mo et al.", + "bbox": [ + 271, + 114, + 354, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "constant position from its associated bone in the bone-local space. The lateral distance of each point from their parent bone segment allows point clouds to react to the bone's rotational roll axis.", + "bbox": [ + 212, + 146, + 782, + 190 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In a pure point cloud representation, it is possible to remove the relationships between bones. However, to clearly distinguish symmetrical body features in different skeletons, we introduce and embed body group associations for the points during our reconstruction process. Specifically, every bone is categorised by one of these body groups: [spine, left arm, right arm, left leg, right leg]. Each sampled point is assigned q group attribute in line with the associated bone. The attribute is represented by a one-hot feature vector $\\mathbf{g}_u$ . To this end, we characterize $u = \\{\\mathbf{p}_{u,t}, \\mathbf{g}_u\\} \\in \\mathcal{X}$ .", + "bbox": [ + 212, + 191, + 784, + 313 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Skeleton Reconstruction Given a sampled point cloud representation $\\mathcal{X}_A$ from $S_A$ and $M_A$ as input, we aim to train a motion adaptation function $F_{S_A\\to S_B}$ which is tasked with producing a visually similar skeletal motion $\\hat{M}_B$ for $S_B$ using a temporal and set-based neural network. As shown in Fig. 3, to construct $F_{S_A\\rightarrow S_B}$ , we employ a Point Transformer [54] to learn embeddings for the unordered point cloud sets frame-wisely, and a standard temporal transformer to process the sequence and decode $\\mathcal{X}_A$ into $\\hat{M}_B$ .", + "bbox": [ + 212, + 338, + 782, + 443 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To optimize $F_{S_A \\to S_B}$ , our objective function maximises the similarity between the sampled point clouds of both $M_A$ and $\\hat{M}_B$ . $\\mathcal{X}_A$ can be viewed as the ground truth to $\\mathcal{X}_B$ , which enables an unsupervised strategy for $F_{S_A \\to S_B}$ . In detail, we first sample a point cloud $\\mathcal{X}_B$ from $S_B$ based on $\\hat{M}_B$ derived by $F_{S_A \\to S_B}$ . Next, a temporally consistent KNN-Loss objective is devised to minimise the $\\ell_2$ distance between any given point of $\\mathcal{X}_B$ from its $K$ nearest neighbouring points in $\\mathcal{X}_A$ . The $K$ nearest neighbours are determined based on inter-point distance throughout the entire motion sequence, instead of a frame-wise setting, as shown in Fig. 4. The points with differing body groups are excluded from the neighbour search. Mathematically, we have:", + "bbox": [ + 212, + 445, + 784, + 595 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\delta \\left(u _ {A}, u _ {B}\\right) = \\left\\{ \\begin{array}{l l} \\sum_ {t} \\left\\| \\mathbf {p} _ {u _ {A}, t} - \\mathbf {p} _ {u _ {B}, t} \\right\\| _ {2}, & \\text {i f} \\mathbf {g} _ {u _ {A}} = \\mathbf {g} _ {u _ {B}}, \\\\ \\inf , & \\text {o t h e r w i s e}, \\end{array} \\right. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 316, + 608, + 784, + 647 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {K N N}} \\left(\\mathcal {X} _ {A}, \\mathcal {X} _ {B}\\right) = \\sum_ {u _ {A} \\in \\mathcal {X} _ {A}, u _ {B} \\in N _ {k} \\left(u _ {A}, \\mathcal {X} _ {B}\\right)} \\delta \\left(u _ {A}, u _ {B}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 316, + 651, + 784, + 685 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $N_{k}(u_{A},\\mathcal{X}_{B})$ is the set containing points that are the k-nearest neighbors of $u_{A}$ in $\\mathcal{X}_B$ , based on $\\delta (u_A,u_B)$ distance.", + "bbox": [ + 212, + 696, + 782, + 727 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In addition to the KNN-Loss, we introduce an optional end-effector loss in the skeleton space for the hand, foot, and head bones, which provides direct positional guidance for matched end-effector pairs $E_{AB}$ :", + "bbox": [ + 212, + 728, + 782, + 773 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {e n d}} = \\sum_ {(\\mathbf {v} _ {A}, \\mathbf {v} _ {B}) \\in E _ {A B}} \\frac {1}{| M _ {A} |} \\sum_ {t} \\left\\| \\mathbf {p} _ {\\mathbf {v} _ {A}, t} - \\mathbf {p} _ {\\mathbf {v} _ {B}, t} \\right\\| _ {2}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 334, + 785, + 784, + 821 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Motion Keyframe Interpolation for Any Human Skeleton", + "bbox": [ + 351, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3a14d93f1d2874a63ea21ea43eb030712e6913a3b9a1c82718ed631ffec26b07.jpg", + "image_caption": [ + "Fig. 4: Visualisation of our temporally consistent KNN objective, processed independently per body group. For each point from $\\mathcal{X}_A$ (a), the total $\\ell_2$ distance throughout the sequence is measured between each point (b) from $\\mathcal{X}_B$ . The $\\ell_2$ distances from the K-nearest points are summed to produce KNN-Loss (c)." + ], + "image_footnote": [], + "bbox": [ + 225, + 146, + 776, + 280 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.3 First-Frame Offset Quaternions", + "text_level": 1, + "bbox": [ + 215, + 378, + 521, + 393 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Point clouds lack absolute roll axis information for each bone. As such, we introduce a First-frame Offset Quaternion (FOQ) transform to incorporate the available relative roll data for motion modelling. We formulate all quaternions of a motion sequence relative to their values from the initial frame. In detail, for any quaternion sequence $Q = \\{\\mathbf{q}_0, \\mathbf{q}_1, \\mathbf{q}_2, \\ldots\\}$ , we can simply obtain its FOQ transform as $\\{\\mathbf{q}_0 \\times \\mathbf{q}_0^{-1}, \\mathbf{q}_1 \\times \\mathbf{q}_0^{-1}, \\mathbf{q}_2 \\times \\mathbf{q}_0^{-1}, \\ldots\\}$ , where $\\mathbf{q}_0^{-1}$ is the quaternion conjugate of $\\mathbf{q}_0$ . An FOQ sequence can easily be converted back to a quaternion sequence as long as any $\\mathbf{q}_t \\in Q$ is known.", + "bbox": [ + 212, + 405, + 787, + 526 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Crucially, FOQ is a roll-invariant rotation representation, that is, FOQ remains identical regardless of a bone's absolute roll position in the original data. Trivially, to adjust the rotational roll of a bone $\\mathbf{v}_n$ , we perform the quaternion multiplication for each frame: $\\mathbf{q}_{n,t} \\times \\mathbf{r}$ , where $\\mathbf{r}$ is a roll quaternion. As a quaternion on the roll axis of $\\mathbf{v}_n$ , $\\mathbf{r}$ is constrained within $\\mathbf{r} = \\alpha + \\beta (x\\mathbf{i} + y\\mathbf{j} + z\\mathbf{k})$ , where $x,y,z$ are the 3D XYZ values of $\\mathbf{T}_n$ , and $\\alpha, \\beta$ are scalars that control the roll magnitude. Note that $\\mathbf{r}$ is constant throughout the sequence. Therefore, we can deduce the roll-invariant nature of any constant $\\mathbf{r}$ :", + "bbox": [ + 212, + 527, + 787, + 647 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\left(\\mathbf {q} _ {n, t} \\times \\mathbf {r}\\right) \\times \\left(\\mathbf {q} _ {0, t} \\times \\mathbf {r}\\right) ^ {- 1} = \\mathbf {q} _ {n, t} \\times \\mathbf {r} \\times \\mathbf {r} ^ {- 1} \\times \\mathbf {q} _ {0, t} ^ {- 1} \\\\ = \\mathbf {q} _ {n, t} \\mathbf {q} _ {0, t} ^ {- 1} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 326, + 657, + 671, + 699 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.4 Motion Interpolation Transformer and Rest Pose Augmentation", + "text_level": 1, + "bbox": [ + 214, + 720, + 785, + 738 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To leverage our point cloud-based representation learning for cross-skeleton interpolation, a transformer model - CITL [34] is adapted, allowing for an efficient interpolation pipeline with existing datasets to alternative skeletons. We introduce two key modifications to CITL and its training strategy, focusing on enhancing motion quality. First, quaternion predictions are substituted with FOQ predictions. This is inspired by the RNN-based method TGcomplete [18], which predicts", + "bbox": [ + 212, + 750, + 787, + 842 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "C. Mo et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 7 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 IK target alignment with maximum real quaternion component" + ], + "code_body": "1: Input: $\\mathbf{p_{IK}} = 3D$ IK target unit vector \n2: $\\mathbf{T} =$ resting tail position of associated bone \n3: Output: $\\mathbf{q}_{\\mathrm{max}} =$ output quaternion, to be applied on original bone quaternion \n4: $\\eta \\leftarrow \\frac{\\mathbf{T}}{|\\mathbf{T}|}$ $\\triangleright$ 3D unit coordinates $[\\eta_x,\\eta_y,\\eta_z]$ \n5: $\\omega \\leftarrow \\frac{\\mathbf{p_{IK}}}{|\\mathbf{p_{IK}}|}$ $\\triangleright$ 3D unit coordinates $[\\omega_x,\\omega_y,\\omega_z]$ \n6: $\\mathbf{q}_{\\mathrm{min}} \\gets \\frac{0 + (\\eta_x + \\omega_x)\\mathbf{i} + (\\eta_y + \\omega_y)\\mathbf{j} + (\\eta_z + \\omega_z)\\mathbf{k}}{2}$ \n7: $\\mathbf{q}_{\\mathrm{max}} \\gets \\mathbf{q}_{\\mathrm{min}} \\times (0 + \\eta_x\\mathbf{i} + \\eta_y\\mathbf{j} + \\eta_z\\mathbf{k})$", + "bbox": [ + 218, + 164, + 787, + 273 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "quaternion offsets from the preceding frame rather than raw quaternions, thereby achieving greater generalisability and superior quality in interpolation.", + "bbox": [ + 212, + 301, + 782, + 332 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "For our model's CITL-based architecture, we substitute the original sinusoidal positional encoding scheme with learned relative positions [44]. While the original implementation relies on the continual nature of sinusoidal functions, we observe that relative positional embeddings with zero initialisation converges upon similarly continuous behaviour, and significantly lowers the complexity of positional attention. The lessened focus on positional relations allows the model to synthesise deeper behaviours within the pose space, improving overall generalisability. As a side effect, the model also accepts arbitrary length inputs beyond the training data length.", + "bbox": [ + 212, + 332, + 784, + 467 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Notably, the intended rest pose of a skeletal configuration is often geometrically sub-optimal for point cloud matching, and can manifest in multiple forms. To address this, we devise a rest position augmentation strategy, leveraging a per-bone Inverse Kinematics (IK) system. RPA broadens the range of intended rest poses that the learned interpolation model can accommodate. In detail, during each phase of the Forward Kinematics (FK) process, we augment the bone tails with a random additional offset. We re-align the augmented global tail position as closely as possible to its original location using a quaternion rotation with the maximum possible real component. Algorithm 1 describes the method by which this quaternion can be derived. Since the real component describes, in cosine, the amount of rotation around an axis, this minimizes numerical disturbance from RPA while altering the visual representation to the desired extent.", + "bbox": [ + 212, + 468, + 785, + 650 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experimental Results", + "text_level": 1, + "bbox": [ + 215, + 686, + 460, + 704 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Datasets", + "text_level": 1, + "bbox": [ + 215, + 715, + 333, + 729 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "To thoroughly evaluate the effectiveness of the proposed method, we conducted extensive experiments on the following widely used motion capture datasets:", + "bbox": [ + 212, + 739, + 782, + 771 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "- LaFAN1 [18] contains long, high quality motions in controllable video game character styles. The dataset was recorded on a large motion capture stage, and manually cleaned to production standards. The native skeleton of LaFAN1 contains 6 spinal bones, and 4 bones for each limb.", + "bbox": [ + 223, + 779, + 792, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Motion Keyframe Interpolation for Any Human Skeleton", + "bbox": [ + 351, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "- Human3.6M [6,22] comprises of 3.6 million motion frames. The motions captured in Human3.6M tend to be less dynamic, as they were recorded on a relatively compact $4 \\times 3 \\mathrm{~m}$ stage. Its native skeleton includes 5 spinal bones, 4 bones for each limb, and an additional finger bone for each hand.", + "bbox": [ + 223, + 145, + 785, + 205 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "- CMU Mocap [10] is a diverse motion dataset encompassing 113 categories. It features a mix of static and dynamic motions, captured on a $3 \\times 8 \\mathrm{~m}$ stage. The skeleton is identical to that of Human3.6M, with the exception of the pelvis, which is divided into two hip bones and a lower back bone.", + "bbox": [ + 225, + 205, + 785, + 267 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "For PC-MRL, we designated the CMU MoCap skeleton as $S_B$ . Its native dataset is exclusively utilized for testing purposes, while LaFAN1 and Human3.6M are used simultaneously for training. $^3$", + "bbox": [ + 215, + 271, + 785, + 318 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 Implementation Details", + "text_level": 1, + "bbox": [ + 215, + 335, + 455, + 351 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "In our experiments, point cloud obfuscation adopted a 256-point setting for the sampling process with $\\sigma = 0.05$ and each point was characterized by a 64-element vector. To maintain consistency, we ensure that all bone offsets and motion positions are defined in metre units. For skeletal reconstruction, 4 Point Transformer [54] layers was utilized to construct the neural network. Within each layer, the feature size was doubled, and the size of the point set was downsampled by a factor of 4 through a farthest-first traversal approach, eventually resulting in a singular feature vector. In the final step, two linear layers with ReLU activations were utilized to transform the resulted feature vector in a frame-wise manner, producing the output skeletal representation. During training, we set $K = 8$ for the KNN loss to optimize the network. Additionally, the use of unit quaternion values in practice relies on modulus division for constraint, a process that can potentially lead to exploding gradients. To address this, we introduced an objective function $\\mathcal{L}_q$ that measures the $\\ell_2$ norms' difference between the raw quaternion output and the expected unit quaternion. All experiments were trained and evaluated on a single NVIDIA GeForce RTX 3090 GPU.", + "bbox": [ + 212, + 357, + 787, + 599 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.3 Motion Interpolation Comparison against Supervised Methods", + "text_level": 1, + "bbox": [ + 215, + 617, + 779, + 633 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "To demonstrate the efficacy of our method, we produce and compare our method against the conventional LERP method, as well as three state-of-the-art interpolation models trained on the original CMU MoCap dataset. Specifically, we measure the performance of the RNN-based $\\mathrm{TG}_{\\mathrm{complete}}$ model [18], the BERT-based motion interpolation adaptation [12], and a transformer-based encoder-decoder approach [34]. We additionally provide results on a variant of our method trained without our RPA strategy. The experimental configuration for each state-of-the-art model is identical to their original implementations. To measure interpolation accuracy, we employ the standard $\\ell_2$ positional distance (L2P), $\\ell_2$ quaternion difference (L2Q), and NPSS [16] for visual similarity evaluation [18].", + "bbox": [ + 212, + 638, + 787, + 790 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "C. Mo et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 9 + }, + { + "type": "page_footnote", + "text": "3 The authors Clinton Mo, Kun Hu and Zhiyong Wang are signatories of the dataset licenses and produced all experimental results of the paper. Meta Inc. was not granted access to the datasets.", + "bbox": [ + 217, + 796, + 785, + 839 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/d6f4353838fb737cf53e19a26d670c71b8f1f903ef28e0ecfc73dc2930b5e69b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Motion categoryMethodL2P↓L2Q↓NPSS↓
515305153051530
BasketballLERP0.07870.40090.80330.15880.58340.96110.13150.61201.8202
TGcomplete [18]0.10500.44190.71040.18590.64390.93030.16000.65791.2136
BERT [12]0.08890.37370.68650.15850.55600.89920.16890.61771.4587
CITL [34]0.06250.29730.57080.14100.49320.80110.10470.40981.0959
PC-MRL w/o RPA0.07670.35340.67150.15050.54310.90490.13640.55961.3471
PC-MRL (Ours)0.07520.34430.66620.15610.54950.90880.13970.55721.3015
GolfLERP0.01890.12290.31870.04550.18700.39870.05870.39791.0312
TGcomplete [18]0.04730.26160.45460.06930.31910.52970.09540.50621.1534
BERT [12]0.02760.11590.29670.06120.20820.42580.10990.39060.9255
CITL [34]0.02010.10430.25890.04760.17660.35520.06360.25810.7799
PC-MRL w/out RPA0.02740.13040.36220.05550.20650.49230.09110.37591.2083
PC-MRL (Ours)0.02730.11300.27060.05980.20120.38440.10490.37160.8150
SwimmingLERP0.07310.36800.73740.13260.53270.94760.21470.92221.6972
TGcomplete [18]0.12920.57960.91410.18450.74201.11990.26600.97161.7698
BERT [12]0.10880.38960.75160.18190.55180.96220.28770.90741.6682
CITL [34]0.09050.38950.74750.14840.55350.95330.20840.83171.6122
PC-MRL w/o RPA0.07220.37540.74590.13450.54420.95940.19440.87311.6608
PC-MRL (Ours)0.07340.35950.71390.13550.53310.92480.19560.85911.6370
Walking & RunningLERP0.04550.25050.55270.11200.39970.62330.06550.26820.9071
TGcomplete [18]0.05400.17860.31030.11470.28660.38660.08950.26160.4265
BERT [12]0.05520.19980.35440.11110.31010.47620.09010.27600.6458
CITL [34]0.03280.10340.24270.09420.21020.33840.05260.15680.4080
PC-MRL w/o RPA0.03970.16570.34170.10580.28790.47220.06400.22880.6021
PC-MRL (Ours)0.03770.13920.28480.10640.27780.42810.06580.21290.5098
", + "bbox": [ + 223, + 143, + 782, + 388 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Table 1: Performance comparisons for various motion categories, measured in L2P, L2Q, and NPSS. All methods except PC-MRL are trained on the original CMU MoCap dataset. Each motion contains 128 frames with keyframes placed every 5, 15, or 30 frames. The top 2 results for each test are highlighted in purple and blue respectively.", + "bbox": [ + 215, + 388, + 787, + 444 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "The results listed in Table 1 clearly demonstrate our method's ability to approach the accuracy levels exhibited by directly supervised state-of-the-art methods. PC-MRL consistently outperforms both the LERP standard and RNN-based $\\mathrm{TG}_{\\mathrm{complete}}$ model, particularly in longer keyframe interval scenarios where the quality and depth of learned motion features is most critical and impactful. Likewise, the inclusion of RPA also tends to improve the long interval performance of PC-MRL, particularly when precise movements (i.e. golf) or deeper motion features (i.e. locomotion) are expected. Given this, we can observe that RPA definitively improves the overall consistency of the PC-MRL method in scenarios that may be unseen or less effectively represented by our point cloud-to-motion system.", + "bbox": [ + 212, + 489, + 787, + 655 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Native supervision, i.e. with CITL, unsurprisingly produces the most accurate interpolation model, with a notable exception of swimming motions, where our PC-MRL supervision produces even stronger results. We believe the abundance of low crawl motions, a similar motion to swimming in LaFAN1, provides more meaningful supervision over the original CMU dataset, which conversely has few similar motions once swimming motions are filtered out. On the other end of the spectrum, i.e. for high data scenarios such as walking and running locomotion from Fig. 5, all models including our PC-MRL approach observe the strongest improvements for learned methods over the naive LERP method. Both cases strongly support our method's capability to supervise highly intricate motion patterns, despite the incompleteness of the point cloud representation and reliance on relative rotations and geometric optimisation assumptions.", + "bbox": [ + 212, + 657, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Motion Keyframe Interpolation for Any Human Skeleton", + "bbox": [ + 351, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/5bf89b54c8e79c1caf8b33c94c9efab4976a017e1a19a022bbde137256c2059a.jpg", + "image_caption": [ + "Fig. 5: Walking keyframe interpolation examples produced by methods from Table 1. Close-up layered comparisons against the original motion are provided on the right." + ], + "image_footnote": [], + "bbox": [ + 243, + 155, + 759, + 411 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.4 Cross-Skeleton Motion Re-Targeting Experiments", + "text_level": 1, + "bbox": [ + 214, + 488, + 671, + 505 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We further conduct an experiment on motion re-targeting to directly benchmark our method against the sole existing state-of-the-art method for unpaired motion re-targeting with unrestricted skeletons, i.e. primal skeletons (PS) [1]. For direct motion re-targeting evaluation, we utilized the KNN-Loss and end-effector loss as metrics, owing to their effective measurement of visual similarity and non-sensitivity towards absolute roll axis correctness. Though absolute rolls are generally necessary for a complete motion re-targeting solution, this aspect is out of scope for our project as our main goal is motion interpolation.", + "bbox": [ + 212, + 513, + 787, + 635 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Table 2 indicates the superior performance of our method over the state-of-the-art PS method in terms of visual similarity. At low $\\sigma$ values, the higher relative $\\mathcal{L}_{\\mathrm{KNN}}$ of the primal skeleton method underscores its geometric deviations", + "bbox": [ + 212, + 636, + 787, + 683 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/1b9559e395566ecda0d7e8684a2e77835c2ba857c8b6aa16239b6973f57981d4.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Model\\( L_{\\text{KNN}}(\\mathcal{X}_A,\\mathcal{X}_B) \\downarrow \\)\\( L_{\\text{end}} \\downarrow \\)
\\( \\sigma = 0 \\)\\( \\sigma = 0.01 \\)\\( \\sigma = 0.05 \\)\\( \\sigma = 0.1 \\)\\( \\sigma = 0.2 \\)
Original motion1.00992.02345.73999.663616.92230
Primal skeleton [1]8.11098.03539.420612.829120.63329.6990
PC-MRL (Ours)3.87144.12316.727710.307817.52615.6258
", + "bbox": [ + 267, + 707, + 738, + 768 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Table 2: Average unpaired LaFAN1 to CMU motion re-targeting performance of the primal skeleton method and our point cloud method. $\\mathcal{L}_{\\mathrm{KNN}}$ is measured using 1024-point clouds at $K = 8$ . Independently sampled point clouds from the original LaFAN1 skeleton are provided as optimal $\\mathcal{L}_{\\mathrm{KNN}}$ reference. All values are scaled by $\\times 100$ .", + "bbox": [ + 212, + 767, + 787, + 824 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "C. Mo et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/4d71934c2946edcc88669cce7457ddcefb2a49921e2cbd04202c403bdc4878ed.jpg", + "image_caption": [ + "Fig. 6: Jumping motion re-targeting examples produced on the CMU skeleton by the primal skeleton method (middle row) and our method (bottom row). Sampled point cloud representations of each motion sequence are provided on the right." + ], + "image_footnote": [], + "bbox": [ + 233, + 146, + 767, + 277 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "from the original motion. In contrast, our method demonstrates significantly better visual adherence to the original geometry. As $\\sigma$ increases, our method approaches near-perfect alignment with the original motion. Figure 6 visually indicates this, showing our method follows the original motion closely, while the latent consistency-based primal skeleton method struggles to generalise and decode the skeleton's internal structures without directly supervised objectives, such as the end-effector positions provided during training. In addition, Fig 7 demonstrates that, like existing re-targeting approaches, our method is able to produce adequate results on disproportionate skeletons.", + "bbox": [ + 212, + 375, + 787, + 512 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Since our approach is largely focused on enabling non-native motion interpolation supervision, we additionally compare our PC-MRL method against a CITL model supervised by PS in Table 3. Due to the suboptimal re-targeting performance of PS, such interpolation models exhibit expectedly poor accuracy.", + "bbox": [ + 212, + 516, + 787, + 578 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/22c90d7f153bf66827f9df7e59a0f616eb3eed7b990ed10dbb5963987623cddf.jpg", + "image_caption": [ + "Fig. 7: Re-targeted falling motion from LaFAN1 (top) on the CMU skeleton (middle) and a short custom skeleton (bottom), performed by our point cloud-to-motion model." + ], + "image_footnote": [], + "bbox": [ + 290, + 616, + 718, + 762 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Motion Keyframe Interpolation for Any Human Skeleton", + "bbox": [ + 351, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/78c7a253c6c77c9a25a2144dec57a7944e36688644f4412cc77be868fa8bb19c.jpg", + "table_caption": [], + "table_footnote": [ + "Table 3: Average L2P, L2Q, and NPSS performance of CITL trained on native and PS-generated datasets, compared to our method." + ], + "table_body": "
MethodL2P↓L2Q↓NPSS↓
515305153051530
CITL - PS0.20430.63321.15240.23180.74281.14420.11220.42761.4622
CITL - Native0.04950.23060.48980.09790.36040.66090.11490.46841.1591
PC-MRL0.04940.23370.49600.10300.37710.68130.13700.54331.1760
", + "bbox": [ + 223, + 143, + 782, + 207 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Limitations", + "text_level": 1, + "bbox": [ + 215, + 273, + 362, + 289 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Due to the inherent limitations of the point cloud representation, certain constraints on the applicability of our point cloud-to-motion method are inevitable. Primarily, the reliance of our point cloud re-targeting method on the relative nature of First-frame Offset Quaternions (FOQ) for most motion learning tasks necessitates the presence of at least one known pose. In the absence of a known pose, relative quaternions are incapable of reconstructing absolute rotational values, which are essential for motion applications. Secondly, due to our method of sampling point clouds with a consistent deviation from each bone segment, the resultant representation inevitably obscures finer details. This includes elements like fingers and facial controls, to an extent that is beyond rectification.", + "bbox": [ + 212, + 304, + 787, + 455 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Our proposed objective function relies entirely on geometric segments, necessitating that all bones possess a non-zero length. This stipulation is deemed a reasonable prerequisite for generic human skeletons, as each bone is typically responsible for manipulating a discernible segment of the human body. Due to technical considerations, the skeletal configurations in each of our datasets included at least one bone of zero length, which we ignored in our experiments in favour of global quaternion rotations of all bones with non-zero length.", + "bbox": [ + 212, + 455, + 787, + 561 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "As a result of these limitations, we refrained from claiming state-of-the-art performance for general motion re-targeting tasks.", + "bbox": [ + 212, + 561, + 785, + 592 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 613, + 359, + 628 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "This paper introduces a novel learning-based motion interpolation method designed to enable cross-skeleton compatibility with existing motion datasets. Our proposed method PC-MRL employs a process of point cloud obfuscation and skeleton reconstruction. The point cloud space represents human motions in a non-hierarchical and skeleton-agnostic manner. It enables a KNN-based objective to be optimised without dataset supervision, guiding a neural network to generate high-quality motion features with any target skeleton. We address the rotational information loss of our point cloud format by presenting an offset quaternion strategy compatible with concurrent transformer-based models. Through extensive experiments, we have demonstrated the efficacy of PC-MRL in performing motion interpolation without relying on native motion data. Moreover, PC-MRL has achieved superior visual similarity metrics in the domain of motion re-targeting. We concluded our work by discussing the limitations of PC-MRL.", + "bbox": [ + 212, + 643, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "C. Mo et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Aberman, K., Li, P., Lischinski, D., Sorkine-Hornung, O., Cohen-Or, D., Chen, B.: Skeleton-aware networks for deep motion retargeting 39(4), 62–1 (2020) 4, 12", + "2. Aberman, K., Wu, R., Lischinski, D., Chen, B., Cohen-Or, D.: Learning character-agnostic motion for motion retargeting in 2d 38(4), 1-14 (jul 2019) 4", + "3. Annabi, L., Ma, Z., et al.: Unsupervised motion retargeting for human-robot imitation. In: Companion of the 2024 ACM/IEEE International Conference on Human-Robot Interactio. ACM (2024) 4", + "4. Bodenheimer, B., Rose, C., Rosenthal, S., Pella, J.: The process of motion capture: Dealing with the data. In: Computer Animation and Simulation. pp. 3-18. Springer (1997) 4", + "5. Burtnyk, N., Wein, M.: Computer-generated key-frame animation. Journal of the SMPTE 80(3), 149-153 (1971) 1, 4", + "6. Catalin Ionescu, Fuxin Li, C.S.: Latent structured models for human pose estimation. In: International Conference on Computer Vision (2011) 10", + "7. Chen, H., Tang, H., Timofte, R., Gool, L.V., Zhao, G.: Lart: Neural correspondence learning with latent regularization transformer for 3d motion transfer. Advances in Neural Information Processing Systems 36 (2024) 4", + "8. Chen, J., Li, C., Lee, G.H.: Weakly-supervised 3D pose transfer with keypoints. In: ICCV. pp. 15156-15165 (2023) 4", + "9. Choi, K.J., Ko, H.S.: Online motion retargetting. The Journal of Visualization and Computer Animation 11(5), 223-235 (2000) 5", + "0. CMU: Carnegie-mellon university motion capture database (2003), http://mocap.cs.cmu.edu/ 10", + "1. Dam, E.B., Koch, M., Lillholm, M.: Quaternions, interpolation and animation, vol. 2. CiteSeer (1998) 5, 6", + "2. Duan, Y., Lin, Y., Zou, Z., Yuan, Y., Qian, Z., Zhang, B.: A unified framework for real time motion completion. In: AAAI. vol. 36, pp. 4459-4467 (2022) 2, 4, 10, 11", + "3. Frank, T., Johnston, O.: Disney Animation: The Illusion of Life. Abbeville Publishing Group, New York (1981) 1", + "4. Gleicher, M.: Retargetting motion to new characters. In: Computer Graphics and Interactive Techniques. pp. 33-42 (1998) 4", + "5. Gleicher, M.: Animation from observation: Motion capture and motion editing 33(4), 51-54 (1999) 4", + "6. Gopalakrishnan, A., Mali, A., Kifer, D., Giles, L., Ororbia, A.G.: A neural temporal model for human motion prediction. In: CVPR. pp. 12116-12125 (2019) 10", + "7. Harvey, F.G., Pal, C.: Recurrent transition networks for character locomotion. In: SIGGRAPH Asia 2018 Technical Briefs, pp. 1-4 (2018) 4", + "8. Harvey, F.G., Yurick, M., Nowrouzezahrai, D., Pal, C.: Robust motion in betweening 39(4), 60-1 (2020) 2, 4, 8, 9, 10, 11", + "9. He, C., Saito, J., Zachary, J., Rushmeier, H., Zhou, Y.: Nemf: Neural motion fields for kinematic animation 35, 4244-4256 (2022) 4", + "20. Hecker, C., Raabe, B., Enslow, R.W., DeWeese, J., Maynard, J., van Prooijen, K.: Real-time motion retargeting to highly varied user-created morphologies 27(3), 1-11 (2008) 5", + "21. Hernandez, A., Gall, J., Moreno-Noguer, F.: Human motion prediction via spatiotemporal inpainting. In: ICCV. pp. 7134-7143 (2019) 4", + "22. Ionescu, C., Papava, D., Olaru, V., Sminchisescu, C.: Human3.6m: Large scale datasets and predictive methods for 3D human sensing in natural environments." + ], + "bbox": [ + 225, + 176, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Motion Keyframe Interpolation for Any Human Skeleton", + "bbox": [ + 351, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "IEEE Transactions on Pattern Analysis and Machine Intelligence 36(7), 1325-1339 (jul 2014) 10", + "23. Jiang, H., Cai, J., Zheng, J.: Skeleton-aware 3D human shape reconstruction from point clouds. In: ICCV. pp. 5431-5441 (2019) 4", + "24. Kantor, I.L., Solodovnikov, A.S., Shenitzer, A.: Hypercomplex numbers: an elementary introduction to algebras, vol. 302. Springer (1989) 6", + "25. Karunratanakul, K., Preechakul, K., Suwajanakorn, S., Tang, S.: Guided motion diffusion for controllable human motion synthesis. In: ICCV. pp. 2151-2162 (2023) 4", + "26. Kulpa, R., Multon, F., Arnaldi, B.: Morphology-independent representation of motions for interactive human-like animation. In: Eurographics (2005) 5", + "27. Lehrmann, A.M., Gehler, P.V., Nowozin, S.: Efficient nonlinear markov models for human motion. In: CVPR. pp. 1314-1321 (2014) 4", + "28. Li, P., Aberman, K., Zhang, Z., Hanocka, R., Sorkine-Hornung, O.: Ganimator: Neural motion synthesis from a single sequence 41(4), 1-12 (2022) 4", + "29. Li, W., Wang, Z., Mai, R., Ren, P., Zhang, Q., Zhou, Y., Xu, N., Zhuang, J., Xin, B., Gao, L., et al.: Modular design automation of the morphologies, controllers, and vision systems for intelligent robots: a survey. Visual Intelligence 1(1), 2 (2023) 4", + "30. Li, Z., Peng, X.B., Abbeel, P., Levine, S., Berseth, G., Sreenath, K.: Reinforcement learning for versatile, dynamic, and robust bipedal locomotion control. arXiv preprint arXiv:2401.16889 (2024) 4", + "31. Liao, Z., Yang, J., Saito, J., Pons-Moll, G., Zhou, Y.: Skeleton-free pose transfer for stylized 3D characters. In: ECCV. pp. 640-656. Springer (2022) 4", + "32. Makoviychuk, V., Wawrzyniak, L., Guo, Y., Lu, M., Storey, K., Macklin, M., Hoeller, D., Rudin, N., Allshire, A., Handa, A., et al.: Isaac gym: High performancegpu-based physics simulation for robot learning. arXiv preprint arXiv:2108.10470 (2021) 4", + "33. Merry, B., Marais, P., Gain, J.: Animation space: A truly linear framework for character animation. ACM Transactions on Graphics (TOG) 25(4), 1400-1423 (2006) 5", + "34. Mo, C.A., Hu, K., Long, C., Wang, Z.: Continuous intermediate token learning with implicit motion manifold for keyframe based motion interpolation. In: CVPR. pp. 13894-13903 (2023) 2, 4, 8, 10, 11", + "35. Mo, C.A., Hu, K., Mei, S., Chen, Z., Wang, Z.: Keyframe extraction from motion capture sequences with graph based deep reinforcement learning. In: ACM International Conference on Multimedia. pp. 5194-5202 (2021) 4", + "36. Monzani, J.S., Baerlocher, P., Boulic, R., Thalmann, D.: Using an intermediate skeleton and inverse kinematics for motion retargeting. In: Computer Graphics Forum. vol. 19, pp. 11-19. Wiley Online Library (2000) 5", + "37. Oreshkin, B.N., Valkanas, A., Harvey, F.G., Menard, L.S., Bocquelet, F., Coates, M.J.: Motion inbetweening via deep $\\delta$ -interpolator (2022) 2, 4", + "38. Peng, X.B., Abbeel, P., Levine, S., van de Panne, M.: Deepmimic: Example-guided deep reinforcement learning of physics-based character skills 37(4), 143:1-143:14 (Jul 2018) 4", + "39. Peng, X.B., Ma, Z., Abbeel, P., Levine, S., Kanazawa, A.: Amp: Adversarial motion priors for stylized physics-based character control 40(4) (Jul 2021) 4", + "40. Qin, J., Zheng, Y., Zhou, K.: Motion in-between via two-stage transformers 41(6), 1-16 (2022) 2", + "41. Reda, D., Won, J., Ye, Y., van de Panne, M., Winkler, A.: Physics-based motion retargeting from sparse inputs. ACM Computer Graphics and Interactive Techniques 6(3), 1-19 (2023) 4" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "C. Mo et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "42. Reeves, W.T.: Inbetweening for computer animation utilizing moving point constraints 15(3), 263-269 (1981) 1", + "43. Ren, T., Yu, J., Guo, S., Ma, Y., Ouyang, Y., Zeng, Z., Zhang, Y., Qin, Y.: Diverse motion in-betweening from sparse keyframes with dual posture stitching. IEEE Transactions on Visualization and Computer Graphics (2024) 4", + "44. Shaw, P., Uszkoreit, J., Vaswani, A.: Self-attention with relative position representations. In: NAACL. Association for Computational Linguistics (2018) 9", + "45. Tiwari, G., Antic, D., Lenssen, J.E., Sarafianos, N., Tung, T., Pons-Moll, G.: Posendf: Modeling human pose manifolds with neural distance fields. In: ECCV. pp. 572-589. Springer (2022) 4", + "46. Villegas, R., Yang, J., Ceylan, D., Lee, H.: Neural kinematic networks for unsupervised motion retargeting. In: CVPR. pp. 8639-8648 (2018) 4", + "47. Wang, J., Li, X., Liu, S., De Mello, S., Gallo, O., Wang, X., Kautz, J.: Zero-shot pose transfer for unrigged stylized 3D characters. In: CVPR. pp. 8704-8714 (2023) 4", + "48. Wang, K., Xie, J., Zhang, G., Liu, L., Yang, J.: Sequential 3D human pose and shape estimation from point clouds. In: CVPR. pp. 7275-7284 (2020) 4", + "49. Yuan, Y., Song, J., Iqbal, U., Vahdat, A., Kautz, J.: Physdiff: Physics-guided human motion diffusion model. In: ICCV. pp. 16010-16021 (2023) 4", + "50. Yuan, Y., Wei, S.E., Simon, T., Kitani, K., Saragih, J.: Simpoe: Simulated character control for 3D human pose estimation. In: CVPR. pp. 7159-7169 (2021) 4", + "51. Zhang, J., Weng, J., Kang, D., Zhao, F., Huang, S., Zhe, X., Bao, L., Shan, Y., Wang, J., Tu, Z.: Skinned motion retargeting with residual perception of motion semantics & geometry. In: CVPR. pp. 13864-13872 (2023) 4", + "52. Zhang, M., Li, H., Cai, Z., Ren, J., Yang, L., Liu, Z.: Finemogen: Fine-grained spatio-temporal motion generation and editing. Advances in Neural Information Processing Systems 36 (2024) 4", + "53. Zhang, X., van de Panne, M.: Data-driven autocompletion for keyframe animation. In: ACM SIGGRAPH Conference on Motion, Interaction and Games. pp. 1-11 (2018) 4", + "54. Zhao, H., Jiang, L., Jia, J., Torr, P.H., Koltun, V.: Point transformer. In: ICCV. pp. 16259-16268 (2021) 7, 10", + "55. Zhu, W., Yang, Z., Di, Z., Wu, W., Wang, Y., Loy, C.C.: Mocanet: Motion retargeting in-the-wild via canonicalization networks. In: AAAI. vol. 36, pp. 3617-3625 (2022) 4" + ], + "bbox": [ + 215, + 147, + 784, + 631 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Motion Keyframe Interpolation for Any Human Skeleton", + "bbox": [ + 351, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/Motion Keyframe Interpolation for Any Human Skeleton using Point Cloud-based Human Motion Data Homogenisation/391ac41a-5023-4e3a-b209-fdaecfa62b6a_model.json b/2024/Motion Keyframe Interpolation for Any Human Skeleton using Point Cloud-based Human Motion Data Homogenisation/391ac41a-5023-4e3a-b209-fdaecfa62b6a_model.json new file mode 100644 index 0000000000000000000000000000000000000000..8708b2b6a2d14eaa282a8992f05ce744cef308f4 --- /dev/null +++ b/2024/Motion Keyframe Interpolation for Any Human Skeleton using Point Cloud-based Human Motion Data Homogenisation/391ac41a-5023-4e3a-b209-fdaecfa62b6a_model.json @@ -0,0 +1,2379 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.227, + 0.141, + 0.78, + 0.21 + ], + "angle": 0, + "content": "Motion Keyframe Interpolation for Any Human Skeleton via Temporally Consistent Point Cloud Sampling and Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.234, + 0.784, + 0.253 + ], + "angle": 0, + "content": "Clinton Mo\\(^{1}\\), Kun Hu\\(^{1,\\star}\\), Chengjiang Long\\(^{2}\\), Dong Yuan\\(^{1}\\), and Zhiyong Wang\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.295, + 0.262, + 0.707, + 0.304 + ], + "angle": 0, + "content": "1 The University of Sydney, Darlington NSW 2008, Australia \nclmo6615@uni.sydney.edu.au, {kun.hu, zhiyong.wang, dong.yuan}@sydney.edu.au" + }, + { + "type": "text", + "bbox": [ + 0.352, + 0.304, + 0.65, + 0.319 + ], + "angle": 0, + "content": "\\(^{2}\\) Meta Reality Labs, Burlingame, CA, USA" + }, + { + "type": "text", + "bbox": [ + 0.443, + 0.32, + 0.562, + 0.332 + ], + "angle": 0, + "content": "clong1@meta.com" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.367, + 0.744, + 0.645 + ], + "angle": 0, + "content": "Abstract. In the character animation field, modern supervised keyframe interpolation models have demonstrated exceptional performance in constructing natural human motions from sparse pose definitions. As supervised models, large motion datasets are necessary to facilitate the learning process; however, since motion is represented with fixed hierarchical skeletons, such datasets are incompatible for skeletons outside the datasets' native configurations. Consequently, the expected availability of a motion dataset for desired skeletons severely hinders the feasibility of learned interpolation in practice. To combat this limitation, we propose Point Cloud-based Motion Representation Learning (PC-MRL), an unsupervised approach to enabling cross-compatibility between skeletons for motion interpolation learning. PC-MRL consists of a skeleton obfuscation strategy using temporal point cloud sampling, and an unsupervised skeleton reconstruction method from point clouds. We devise a temporal point-wise K-nearest neighbors loss for unsupervised learning. Moreover, we propose First-frame Offset Quaternion (FOQ) and Rest Pose Augmentation (RPA) strategies to overcome necessary limitations of our unsupervised point cloud-to-skeletal motion process. Comprehensive experiments demonstrate the effectiveness of PC-MRL in motion interpolation for desired skeletons without supervision from native datasets." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.657, + 0.735, + 0.673 + ], + "angle": 0, + "content": "Keywords: 3D point clouds \\(\\cdot\\) Human body motion \\(\\cdot\\) Dataset creation" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.696, + 0.377, + 0.712 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.727, + 0.788, + 0.819 + ], + "angle": 0, + "content": "3D character animation workflows largely rely on the concepts of keyframing with interpolation, often referred to as the pose-to-pose principle [13]. By defining key poses at correct timings, algorithms can be employed to generate intermediate poses, and thereby eliminating the need for defining each frame individually [5, 42]. Although this keyframe-based workflow is less costly than frame-by-frame production, human motion is often complex, and natural depictions" + }, + { + "type": "page_footnote", + "bbox": [ + 0.22, + 0.826, + 0.386, + 0.841 + ], + "angle": 0, + "content": "* Corresponding author." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.127 + ], + "angle": 0, + "content": "C. Mo et al." + }, + { + "type": "image", + "bbox": [ + 0.268, + 0.151, + 0.391, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.258, + 0.262, + 0.398, + 0.274 + ], + "angle": 0, + "content": "(a) Linear interpolation" + }, + { + "type": "image", + "bbox": [ + 0.445, + 0.151, + 0.561, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.42, + 0.261, + 0.583, + 0.275 + ], + "angle": 0, + "content": "(b) PC-MRL (Our method)" + }, + { + "type": "image", + "bbox": [ + 0.619, + 0.151, + 0.736, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.625, + 0.262, + 0.728, + 0.274 + ], + "angle": 0, + "content": "(c) Ground truth" + }, + { + "type": "image_caption", + "bbox": [ + 0.239, + 0.285, + 0.761, + 0.3 + ], + "angle": 0, + "content": "Fig. 1: Motion interpolation of walking keyframes conducted by our approach." + }, + { + "type": "image", + "bbox": [ + 0.296, + 0.321, + 0.361, + 0.419 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.256, + 0.425, + 0.4, + 0.438 + ], + "angle": 0, + "content": "(a) Hierarchical skeleton" + }, + { + "type": "image", + "bbox": [ + 0.446, + 0.319, + 0.557, + 0.418 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.445, + 0.425, + 0.56, + 0.438 + ], + "angle": 0, + "content": "(b) Skeletal motion" + }, + { + "type": "image", + "bbox": [ + 0.619, + 0.32, + 0.734, + 0.419 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.591, + 0.425, + 0.763, + 0.438 + ], + "angle": 0, + "content": "(c) Point cloud representation" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.449, + 0.788, + 0.492 + ], + "angle": 0, + "content": "Fig. 2: Our method leverages point cloud representations (c) to obfuscate the hierarchical dependencies of skeletal motion. Joint colours correspond to ordinal indices in original motion data representation. Point clouds are coloured by temporal index." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.522, + 0.785, + 0.598 + ], + "angle": 0, + "content": "demand a significant number of keyframes, especially for realistic motions that adhere to physical properties and constraints. Although motion capture (MoCap) is frequently employed as an alternative for achieving realistic human motions, MoCap often nonetheless necessitates subsequent keyframing to correct unwanted motion elements and address other imperfections." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.787, + 0.779 + ], + "angle": 0, + "content": "In recent years, the interpolation process has been widely studied, particularly with the advent of machine learning methods for sequential data. Compared to conventional methods such as linear interpolation (LERP), machine learning methods have demonstrated the capability to derive visually natural motions from sparser keyframe sets [12, 18, 34, 37, 40], which we can observe in Fig. 1. Data-driven interpolation approaches require large motion capture datasets to learn the necessary motion features for effective synthesis of transitions between keyframes. In the field of 3D motions, this aspect significantly restricts the feasibility of learned interpolation methods in animation practice, as motion/pose data are tied to specific skeletal configurations and are not cross-compatible. That is, the dataset's native skeleton, i.e. source skeleton, is structurally different and incompatible with the desired skeleton, i.e. target skeleton." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Therefore, we propose a novel point cloud-based motion representation learning (PC-MRL) approach, in which skeletal hierarchies are obfuscated by sampling into the point cloud medium, and reconstructed into skeletal motion data using an unsupervised neural network. Unlike raw pose and motion data, the" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.352, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Motion Keyframe Interpolation for Any Human Skeleton" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.344 + ], + "angle": 0, + "content": "point cloud format is a non-hierarchical representation, and effectively obscures any skeletal configuration as displayed in Fig. 2. PC-MRL samples a set of points around the rest position of skeletons using a combination of uniform and normal distributions, and geometrically parents each point to an associated bone in order to effectively represent the associated skeletons' motion data in a temporally consistent manner. Subsequently, for the reconstruction of motion data into a given desired skeleton, we propose an unsupervised learning scheme, featuring a K-nearest neighbors (KNN) loss between cross-skeleton point clouds to optimize a transformer neural network. The network processes the point cloud-based motion representations and generates the corresponding target skeletal motion representations. By obscuring both source and predicted target motion data into the point cloud space, our strategy allows for a direct comparison, utilizing KNN to minimize the visual difference between the two point clouds." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.344, + 0.788, + 0.569 + ], + "angle": 0, + "content": "The unsupervised nature of our target skeleton reconstruction scheme creates two main disparities between the expected and predicted motion features. Firstly, the exact roll rotation axis of bones cannot be represented in the point cloud format. To this end, we introduce First-frame Offset Quaternion (FOQ) representations to incorporate relative roll values, which are obtainable from temporally consistent point clouds. This approach is agnostic to absolute roll rotations, providing a standardisation mechanism for the resulting motion data within the context of motion interpolation learning. Second, KNN-based objectives learn skeletal representations in a geometrically optimal manner, which does not always reflect the desired skeletal behaviour. This disparity is commonly observed with smaller skeletal features, such as shoulder and hip bones. As such, we augment training motions using Rest Pose Augmentation (RPA) to increase the range of skeletal behaviours that our interpolation model learns. Extensive experiments demonstrate the effectiveness of our proposed method, achieving performance targets near the level of direct dataset supervision." + }, + { + "type": "text", + "bbox": [ + 0.239, + 0.57, + 0.704, + 0.584 + ], + "angle": 0, + "content": "In summary, this paper presents the following key contributions:" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.593, + 0.784, + 0.635 + ], + "angle": 0, + "content": "1. We propose a novel method for achieving unsupervised human motion reconstruction from point cloud data, enabling skeleton-agnostic motion interpolation learning for the first time." + }, + { + "type": "text", + "bbox": [ + 0.222, + 0.638, + 0.784, + 0.68 + ], + "angle": 0, + "content": "2. To train motion interpolation models, we formulate first-frame offset quaternions to represent bone rotations with relative roll data, as well as a rest position augmentation strategy to address skeletal configuration variability." + }, + { + "type": "text", + "bbox": [ + 0.222, + 0.683, + 0.784, + 0.725 + ], + "angle": 0, + "content": "3. We perform comprehensive experiments to demonstrate our method's effectiveness towards learning motion interpolation, despite the absence of directly compatible datasets in the training process." + }, + { + "type": "list", + "bbox": [ + 0.222, + 0.593, + 0.784, + 0.725 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.75, + 0.393, + 0.765 + ], + "angle": 0, + "content": "2 Related works" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.842 + ], + "angle": 0, + "content": "We explore existing research on motion interpolation approaches, and motion data modelling methods in general. In addition, our proposed approach bears resemblance to unsupervised approaches for the motion re-targeting research problem, and as such, we will also analyse existing methods in this field." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.127 + ], + "angle": 0, + "content": "C. Mo et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.597, + 0.162 + ], + "angle": 0, + "content": "2.1 Data-Driven Motion Modelling Methods" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.173, + 0.787, + 0.293 + ], + "angle": 0, + "content": "Contemporary research in motion modelling predominantly explore learned data-driven methods, using neural networks to capture, model, and establish correlations between the motion features of various skeletal configurations [29, 35, 41, 51, 55]. Motion features alone have demonstrated considerable efficacy in cross-skeleton imitation using a variety of strategies, including task-driven objectives [19, 45], diffusion-based generation [25, 49, 52], and latent feature consistency [1-3, 7, 46]. Latent consistency in particular learns inter-skeleton correlations without the need for paired datasets, which reduces its barrier of entry." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.294, + 0.789, + 0.444 + ], + "angle": 0, + "content": "Likewise, motion interpolation strategies have evolved through the use of learning based approaches. The inherent numerical precision of motion data has traditionally posed a challenge for the adoption of the learning approaches, due to their approximate nature [21,27]. This challenge is further amplified by the temporally sparse distribution of keyframes. Recently, a number of recurrent neural network based methods [17, 18, 43, 53] and transformer-based networks [12, 34, 37] have demonstrated encouraging interpolation performance for the transition between keyframes. Additionally, various other motion modelling methods are capable for keyframe based interpolation, albeit with limited intermediate frame representation capabilities [19, 25, 45]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.445, + 0.789, + 0.656 + ], + "angle": 0, + "content": "A major constraint inherent in all data-driven motion modelling methods is their dependency on datasets featuring specific skeletal configurations. To address this, skeleton-free training strategies have been explored for elementary tasks involving 3D characters. These include vertex-based pose transfer [7,8,31,47], and point cloud-based human shape reconstruction [23,48]. Notably, in these methods, 3D mesh and point cloud coordinates serve as general 3D data representations. Concurrently, other research has significantly minimized the required volume of training data, bringing it down to single example sequences. This reduction has been achieved through the use of patch matching techniques [28] and imitation learning within physics-based simulations [30,39]. Reinforcement learning has facilitated the generation of goal-driven motion without pre-existing datasets [32,38,50]. However, these simulation-based methods have typically been either exceedingly complex to implement in practice, or yield results that are too inflexible and/or prone to errors for animation workflows." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.679, + 0.607, + 0.695 + ], + "angle": 0, + "content": "2.2 Motion Re-Targeting without Supervision" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Conventional approaches in motion re-targeting have been a cornerstone of 3D animation for decades, primarily dedicated to converting motion capture data into 3D skeletal motions [4, 15]. Typically, raw motion capture data is optically recorded, yielding primarily spatial information. In contrast, motion data is largely rotational in nature [5], often represented in \\(SO(3)\\) space. Therefore, a principal objective of motion re-targeting solutions has been to effectively bridge these two distinct types of representations. In [14], extensive manually adjustable constraints were introduced for flexible re-targeting between skeletons of identical topology (i.e., same hierarchy, different bone lengths and rest poses). Key" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.352, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Motion Keyframe Interpolation for Any Human Skeleton" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.153, + 0.773, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.301, + 0.788, + 0.385 + ], + "angle": 0, + "content": "Fig. 3: Illustration of point cloud-based motion representation learning (PC-MRL) pipeline for keyframe interpolation to any human skeleton. Our unsupervised point cloud-to-pose decoding approach (I) first adapts motion features from existing datasets to new skeletons. We then employ the adapted motion data to supervise motion interpolation models (II), using FOQ and rest pose augmentations to address remaining discrepancies between the adapted and expected target skeleton motion patterns." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.415, + 0.784, + 0.49 + ], + "angle": 0, + "content": "practices in this approach, including end-effector matching, joint contacts, and inter-skeleton point correlations, are still relevant to current animation pipelines. Point correlations using Inverse Kinematics (IK) have also been proposed as an intermediate representation to facilitate the adaptation between topologically distinct skeletons [9, 36], albeit with hierarchical pairing limitations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.491, + 0.785, + 0.596 + ], + "angle": 0, + "content": "In response, morphologically-independent motion control systems were proposed using IK joint chains and joint-wise constraints [20,26]. Such systems often expect highly standardised motion skeleton features, and are generally impractical for re-targeting motions produced for different control schemes. Additionally, the reliance on Euler angles for extensive classical re-targeting constraints poses a variety of issues regarding rotational freedom [11], as well as compatibility with animation workflows [33], which predominantly use quaternion-based systems." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.597, + 0.787, + 0.657 + ], + "angle": 0, + "content": "Despite these advancements, the challenge of arbitrary skeleton re-targeting from motion datasets still remains largely unresolved, limiting any widespread application of data-driven pipelines. Our paper aims to effectively address this pervasive limitation." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.682, + 0.381, + 0.699 + ], + "angle": 0, + "content": "3 Methodology" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.715, + 0.784, + 0.745 + ], + "angle": 0, + "content": "As illustrated in Fig. 3, our proposed method consists of two main components: point cloud-to-motion reconstruction and motion keyframe interpolation." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.769, + 0.611, + 0.784 + ], + "angle": 0, + "content": "3.1 Quaternion-based Motion Representations" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.842 + ], + "angle": 0, + "content": "We first declare a set of notations for representing skeletons. A skeleton \\( S \\) is defined by a tree of bones \\( \\mathcal{B} \\). A bone \\( \\mathbf{v}_n \\in \\mathcal{B} \\) is defined by its bone parent \\( \\mathbf{v}_m \\in \\mathcal{B} \\), and a head position \\( \\mathbf{H}_n \\in \\mathbb{R}^3 \\) which denotes the bone's base offset from" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.128 + ], + "angle": 0, + "content": "C. Mo et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.207 + ], + "angle": 0, + "content": "its parent head \\(\\mathbf{H}_m\\). In addition, a tail position \\(\\mathbf{T}_n \\in \\mathbb{R}^3\\) helps to define the end point of the bone's line segment representation, starting from its head position, of \\(\\mathbf{v}_n\\). The root bone \\(\\mathbf{v}_0\\) has no parent, and thus no head position \\(\\mathbf{H}_0\\). The set of \\(\\mathbf{H}_n\\) and \\(\\mathbf{T}_n\\) for each bone \\(\\mathbf{v}_n \\in \\mathcal{B}\\) make up the rest position of the skeleton." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.208, + 0.785, + 0.284 + ], + "angle": 0, + "content": "Next, we define the properties of motion data. At any given frame \\(t\\), a skeleton's pose position can be represented as global 3D positions \\(\\mathbf{p}_{n,t} \\in \\mathbb{R}^3\\) and global quaternion rotations \\(\\mathbf{q}_{n,t} \\in \\mathbb{R}^4\\) for each bone \\(\\mathbf{v}_n \\in \\mathcal{B}\\). For a given motion sequence \\(M\\), only the root bone is given a 3D position in pose position, i.e. \\(\\mathbf{p}_{0,t}\\) as all other pose positions are derived using Forwards Kinematics (FK [11,24]):" + }, + { + "type": "equation", + "bbox": [ + 0.399, + 0.293, + 0.785, + 0.308 + ], + "angle": 0, + "content": "\\[\n\\mathbf {p} _ {n, t} = \\mathbf {p} _ {m, t} + Q R \\left(\\mathbf {H} _ {n}, \\mathbf {q} _ {m, t}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.315, + 0.784, + 0.346 + ], + "angle": 0, + "content": "where \\( QR \\) is the quaternion rotation function, and \\( \\mathbf{v}_m \\) is the parent joint. In summary, a motion sequence \\( M \\) of \\( |M| \\) frames is defined by:" + }, + { + "type": "equation", + "bbox": [ + 0.349, + 0.355, + 0.785, + 0.37 + ], + "angle": 0, + "content": "\\[\nM = \\left\\{\\mathbf {p} _ {0, t} \\cup \\left\\{\\mathbf {q} _ {n, t} \\mid \\mathbf {v} _ {n} \\in \\mathcal {B} \\right\\} \\mid 0 \\leq t < | M | \\right\\} \\tag {2}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.389, + 0.708, + 0.403 + ], + "angle": 0, + "content": "3.2 Point Cloud Obfuscation and Skeleton Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.411, + 0.785, + 0.593 + ], + "angle": 0, + "content": "Our approach is based on the hypothesis that 3D motion sequences can be visually represented by a more universal format: 3D point cloud sequences. By geometrically sampling the volume around a skeleton, point cloud data becomes a non-hierarchical representation of its pose(s), obfuscating the original skeletal configuration. Due to this decoupling between the skeleton configuration and motion representation, a successful motion data reconstruction from point cloud data would be the first step to enabling learnable cross-skeleton motion features. Formally, for a given source skeleton \\( S_A \\) with an associated motion sequence \\( M_A \\), and any humanoid skeletal configuration \\( S_B \\), the point cloud obfuscation and reconstruction module is to learn a function that can adequately perform \\( F_{S_A \\to S_B}(M_A) = \\hat{M}_B \\), where \\( \\hat{M}_B \\) is a visually similar motion adaptation of \\( M_A \\) on skeleton \\( S_B \\), which is an estimation of \\( M_B \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.612, + 0.785, + 0.732 + ], + "angle": 0, + "content": "Motion Data Obfuscation with Point Clouds To sample the point cloud \\(\\mathcal{X}\\), we first generate a set of points \\(u\\in \\mathcal{X}\\) along the line segments defined by the global head and tail positions for each bone \\(\\mathbf{v}_n\\). In the bone-local space to \\(\\mathbf{v}_n\\), this would mean sampling \\(u\\) by a uniformly distributed factor \\(\\alpha \\sim \\mathcal{U}_{[0,1]}\\). Specifically, \\(\\alpha = 0\\) aligns with the head position of the bone, and \\(\\alpha = 1\\) indicates the tail position. By further sampling \\(u\\) via a normal distribution, our sampling strategy can capture the surrounding volume around the skeleton. Formally, to sample \\(u\\) from its associated bone \\(\\mathbf{v}_n\\) within a standard deviation of \\(\\sigma\\), we have:" + }, + { + "type": "equation", + "bbox": [ + 0.41, + 0.754, + 0.537, + 0.769 + ], + "angle": 0, + "content": "\\[\n\\mathbf {H} _ {x} \\sim \\mathcal {N} (\\alpha \\mathbf {T} _ {n}, \\sigma),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.405, + 0.764, + 0.785, + 0.788 + ], + "angle": 0, + "content": "\\[\n\\mathbf {p} _ {u, t} = \\mathbf {p} _ {n, t} + Q R (\\mathbf {H} _ {x}, \\mathbf {q} _ {n, t}) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.784, + 0.84 + ], + "angle": 0, + "content": "Notably, this sampling process as in Eq. 3 is associated with the skeletal configuration only, by which the sampled points are temporally consistent for further processing with the motion sequence. Specifically, each point maintains a" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.352, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Motion Keyframe Interpolation for Any Human Skeleton" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.191 + ], + "angle": 0, + "content": "constant position from its associated bone in the bone-local space. The lateral distance of each point from their parent bone segment allows point clouds to react to the bone's rotational roll axis." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.193, + 0.785, + 0.314 + ], + "angle": 0, + "content": "In a pure point cloud representation, it is possible to remove the relationships between bones. However, to clearly distinguish symmetrical body features in different skeletons, we introduce and embed body group associations for the points during our reconstruction process. Specifically, every bone is categorised by one of these body groups: [spine, left arm, right arm, left leg, right leg]. Each sampled point is assigned q group attribute in line with the associated bone. The attribute is represented by a one-hot feature vector \\(\\mathbf{g}_u\\). To this end, we characterize \\(u = \\{\\mathbf{p}_{u,t}, \\mathbf{g}_u\\} \\in \\mathcal{X}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.339, + 0.784, + 0.444 + ], + "angle": 0, + "content": "Skeleton Reconstruction Given a sampled point cloud representation \\(\\mathcal{X}_A\\) from \\(S_A\\) and \\(M_A\\) as input, we aim to train a motion adaptation function \\(F_{S_A\\to S_B}\\) which is tasked with producing a visually similar skeletal motion \\(\\hat{M}_B\\) for \\(S_B\\) using a temporal and set-based neural network. As shown in Fig. 3, to construct \\(F_{S_A\\rightarrow S_B}\\), we employ a Point Transformer [54] to learn embeddings for the unordered point cloud sets frame-wisely, and a standard temporal transformer to process the sequence and decode \\(\\mathcal{X}_A\\) into \\(\\hat{M}_B\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.446, + 0.785, + 0.596 + ], + "angle": 0, + "content": "To optimize \\( F_{S_A \\to S_B} \\), our objective function maximises the similarity between the sampled point clouds of both \\( M_A \\) and \\( \\hat{M}_B \\). \\( \\mathcal{X}_A \\) can be viewed as the ground truth to \\( \\mathcal{X}_B \\), which enables an unsupervised strategy for \\( F_{S_A \\to S_B} \\). In detail, we first sample a point cloud \\( \\mathcal{X}_B \\) from \\( S_B \\) based on \\( \\hat{M}_B \\) derived by \\( F_{S_A \\to S_B} \\). Next, a temporally consistent KNN-Loss objective is devised to minimise the \\( \\ell_2 \\) distance between any given point of \\( \\mathcal{X}_B \\) from its \\( K \\) nearest neighbouring points in \\( \\mathcal{X}_A \\). The \\( K \\) nearest neighbours are determined based on inter-point distance throughout the entire motion sequence, instead of a frame-wise setting, as shown in Fig. 4. The points with differing body groups are excluded from the neighbour search. Mathematically, we have:" + }, + { + "type": "equation", + "bbox": [ + 0.318, + 0.609, + 0.785, + 0.648 + ], + "angle": 0, + "content": "\\[\n\\delta \\left(u _ {A}, u _ {B}\\right) = \\left\\{ \\begin{array}{l l} \\sum_ {t} \\left\\| \\mathbf {p} _ {u _ {A}, t} - \\mathbf {p} _ {u _ {B}, t} \\right\\| _ {2}, & \\text {i f} \\mathbf {g} _ {u _ {A}} = \\mathbf {g} _ {u _ {B}}, \\\\ \\inf , & \\text {o t h e r w i s e}, \\end{array} \\right. \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.318, + 0.652, + 0.785, + 0.686 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {K N N}} \\left(\\mathcal {X} _ {A}, \\mathcal {X} _ {B}\\right) = \\sum_ {u _ {A} \\in \\mathcal {X} _ {A}, u _ {B} \\in N _ {k} \\left(u _ {A}, \\mathcal {X} _ {B}\\right)} \\delta \\left(u _ {A}, u _ {B}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.698, + 0.784, + 0.728 + ], + "angle": 0, + "content": "where \\(N_{k}(u_{A},\\mathcal{X}_{B})\\) is the set containing points that are the k-nearest neighbors of \\(u_{A}\\) in \\(\\mathcal{X}_B\\), based on \\(\\delta (u_A,u_B)\\) distance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.729, + 0.784, + 0.774 + ], + "angle": 0, + "content": "In addition to the KNN-Loss, we introduce an optional end-effector loss in the skeleton space for the hand, foot, and head bones, which provides direct positional guidance for matched end-effector pairs \\( E_{AB} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.335, + 0.786, + 0.785, + 0.823 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {e n d}} = \\sum_ {(\\mathbf {v} _ {A}, \\mathbf {v} _ {B}) \\in E _ {A B}} \\frac {1}{| M _ {A} |} \\sum_ {t} \\left\\| \\mathbf {p} _ {\\mathbf {v} _ {A}, t} - \\mathbf {p} _ {\\mathbf {v} _ {B}, t} \\right\\| _ {2}. \\tag {6}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.127 + ], + "angle": 0, + "content": "C. Mo et al." + }, + { + "type": "image", + "bbox": [ + 0.227, + 0.147, + 0.777, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.292, + 0.788, + 0.349 + ], + "angle": 0, + "content": "Fig. 4: Visualisation of our temporally consistent KNN objective, processed independently per body group. For each point from \\(\\mathcal{X}_A\\) (a), the total \\(\\ell_2\\) distance throughout the sequence is measured between each point (b) from \\(\\mathcal{X}_B\\). The \\(\\ell_2\\) distances from the K-nearest points are summed to produce KNN-Loss (c)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.379, + 0.522, + 0.394 + ], + "angle": 0, + "content": "3.3 First-Frame Offset Quaternions" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.406, + 0.788, + 0.527 + ], + "angle": 0, + "content": "Point clouds lack absolute roll axis information for each bone. As such, we introduce a First-frame Offset Quaternion (FOQ) transform to incorporate the available relative roll data for motion modelling. We formulate all quaternions of a motion sequence relative to their values from the initial frame. In detail, for any quaternion sequence \\( Q = \\{\\mathbf{q}_0, \\mathbf{q}_1, \\mathbf{q}_2, \\ldots\\} \\), we can simply obtain its FOQ transform as \\( \\{\\mathbf{q}_0 \\times \\mathbf{q}_0^{-1}, \\mathbf{q}_1 \\times \\mathbf{q}_0^{-1}, \\mathbf{q}_2 \\times \\mathbf{q}_0^{-1}, \\ldots\\} \\), where \\( \\mathbf{q}_0^{-1} \\) is the quaternion conjugate of \\( \\mathbf{q}_0 \\). An FOQ sequence can easily be converted back to a quaternion sequence as long as any \\( \\mathbf{q}_t \\in Q \\) is known." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.528, + 0.788, + 0.648 + ], + "angle": 0, + "content": "Crucially, FOQ is a roll-invariant rotation representation, that is, FOQ remains identical regardless of a bone's absolute roll position in the original data. Trivially, to adjust the rotational roll of a bone \\(\\mathbf{v}_n\\), we perform the quaternion multiplication for each frame: \\(\\mathbf{q}_{n,t} \\times \\mathbf{r}\\), where \\(\\mathbf{r}\\) is a roll quaternion. As a quaternion on the roll axis of \\(\\mathbf{v}_n\\), \\(\\mathbf{r}\\) is constrained within \\(\\mathbf{r} = \\alpha + \\beta (x\\mathbf{i} + y\\mathbf{j} + z\\mathbf{k})\\), where \\(x,y,z\\) are the 3D XYZ values of \\(\\mathbf{T}_n\\), and \\(\\alpha, \\beta\\) are scalars that control the roll magnitude. Note that \\(\\mathbf{r}\\) is constant throughout the sequence. Therefore, we can deduce the roll-invariant nature of any constant \\(\\mathbf{r}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.328, + 0.659, + 0.673, + 0.7 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\left(\\mathbf {q} _ {n, t} \\times \\mathbf {r}\\right) \\times \\left(\\mathbf {q} _ {0, t} \\times \\mathbf {r}\\right) ^ {- 1} = \\mathbf {q} _ {n, t} \\times \\mathbf {r} \\times \\mathbf {r} ^ {- 1} \\times \\mathbf {q} _ {0, t} ^ {- 1} \\\\ = \\mathbf {q} _ {n, t} \\mathbf {q} _ {0, t} ^ {- 1} \\\\ \\end{array}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.722, + 0.787, + 0.739 + ], + "angle": 0, + "content": "3.4 Motion Interpolation Transformer and Rest Pose Augmentation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.843 + ], + "angle": 0, + "content": "To leverage our point cloud-based representation learning for cross-skeleton interpolation, a transformer model - CITL [34] is adapted, allowing for an efficient interpolation pipeline with existing datasets to alternative skeletons. We introduce two key modifications to CITL and its training strategy, focusing on enhancing motion quality. First, quaternion predictions are substituted with FOQ predictions. This is inspired by the RNN-based method TGcomplete [18], which predicts" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.352, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Motion Keyframe Interpolation for Any Human Skeleton" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "code_caption", + "bbox": [ + 0.218, + 0.146, + 0.779, + 0.163 + ], + "angle": 0, + "content": "Algorithm 1 IK target alignment with maximum real quaternion component" + }, + { + "type": "algorithm", + "bbox": [ + 0.22, + 0.165, + 0.788, + 0.274 + ], + "angle": 0, + "content": "1: Input: \\(\\mathbf{p_{IK}} = 3D\\) IK target unit vector \n2: \\(\\mathbf{T} =\\) resting tail position of associated bone \n3: Output: \\(\\mathbf{q}_{\\mathrm{max}} =\\) output quaternion, to be applied on original bone quaternion \n4: \\(\\eta \\leftarrow \\frac{\\mathbf{T}}{|\\mathbf{T}|}\\) \\(\\triangleright\\) 3D unit coordinates \\([\\eta_x,\\eta_y,\\eta_z]\\) \n5: \\(\\omega \\leftarrow \\frac{\\mathbf{p_{IK}}}{|\\mathbf{p_{IK}}|}\\) \\(\\triangleright\\) 3D unit coordinates \\([\\omega_x,\\omega_y,\\omega_z]\\) \n6: \\(\\mathbf{q}_{\\mathrm{min}} \\gets \\frac{0 + (\\eta_x + \\omega_x)\\mathbf{i} + (\\eta_y + \\omega_y)\\mathbf{j} + (\\eta_z + \\omega_z)\\mathbf{k}}{2}\\) \n7: \\(\\mathbf{q}_{\\mathrm{max}} \\gets \\mathbf{q}_{\\mathrm{min}} \\times (0 + \\eta_x\\mathbf{i} + \\eta_y\\mathbf{j} + \\eta_z\\mathbf{k})\\)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.303, + 0.784, + 0.333 + ], + "angle": 0, + "content": "quaternion offsets from the preceding frame rather than raw quaternions, thereby achieving greater generalisability and superior quality in interpolation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.333, + 0.785, + 0.468 + ], + "angle": 0, + "content": "For our model's CITL-based architecture, we substitute the original sinusoidal positional encoding scheme with learned relative positions [44]. While the original implementation relies on the continual nature of sinusoidal functions, we observe that relative positional embeddings with zero initialisation converges upon similarly continuous behaviour, and significantly lowers the complexity of positional attention. The lessened focus on positional relations allows the model to synthesise deeper behaviours within the pose space, improving overall generalisability. As a side effect, the model also accepts arbitrary length inputs beyond the training data length." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.469, + 0.787, + 0.651 + ], + "angle": 0, + "content": "Notably, the intended rest pose of a skeletal configuration is often geometrically sub-optimal for point cloud matching, and can manifest in multiple forms. To address this, we devise a rest position augmentation strategy, leveraging a per-bone Inverse Kinematics (IK) system. RPA broadens the range of intended rest poses that the learned interpolation model can accommodate. In detail, during each phase of the Forward Kinematics (FK) process, we augment the bone tails with a random additional offset. We re-align the augmented global tail position as closely as possible to its original location using a quaternion rotation with the maximum possible real component. Algorithm 1 describes the method by which this quaternion can be derived. Since the real component describes, in cosine, the amount of rotation around an axis, this minimizes numerical disturbance from RPA while altering the visual representation to the desired extent." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.687, + 0.461, + 0.705 + ], + "angle": 0, + "content": "4 Experimental Results" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.717, + 0.334, + 0.731 + ], + "angle": 0, + "content": "4.1 Datasets" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.741, + 0.784, + 0.772 + ], + "angle": 0, + "content": "To thoroughly evaluate the effectiveness of the proposed method, we conducted extensive experiments on the following widely used motion capture datasets:" + }, + { + "type": "text", + "bbox": [ + 0.224, + 0.78, + 0.794, + 0.84 + ], + "angle": 0, + "content": "- LaFAN1 [18] contains long, high quality motions in controllable video game character styles. The dataset was recorded on a large motion capture stage, and manually cleaned to production standards. The native skeleton of LaFAN1 contains 6 spinal bones, and 4 bones for each limb." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.127 + ], + "angle": 0, + "content": "C. Mo et al." + }, + { + "type": "text", + "bbox": [ + 0.224, + 0.146, + 0.787, + 0.207 + ], + "angle": 0, + "content": "- Human3.6M [6,22] comprises of 3.6 million motion frames. The motions captured in Human3.6M tend to be less dynamic, as they were recorded on a relatively compact \\(4 \\times 3 \\mathrm{~m}\\) stage. Its native skeleton includes 5 spinal bones, 4 bones for each limb, and an additional finger bone for each hand." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.207, + 0.787, + 0.268 + ], + "angle": 0, + "content": "- CMU Mocap [10] is a diverse motion dataset encompassing 113 categories. It features a mix of static and dynamic motions, captured on a \\(3 \\times 8 \\mathrm{~m}\\) stage. The skeleton is identical to that of Human3.6M, with the exception of the pelvis, which is divided into two hip bones and a lower back bone." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.272, + 0.787, + 0.319 + ], + "angle": 0, + "content": "For PC-MRL, we designated the CMU MoCap skeleton as \\( S_B \\). Its native dataset is exclusively utilized for testing purposes, while LaFAN1 and Human3.6M are used simultaneously for training.\\(^3\\)" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.337, + 0.456, + 0.352 + ], + "angle": 0, + "content": "4.2 Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.358, + 0.789, + 0.6 + ], + "angle": 0, + "content": "In our experiments, point cloud obfuscation adopted a 256-point setting for the sampling process with \\(\\sigma = 0.05\\) and each point was characterized by a 64-element vector. To maintain consistency, we ensure that all bone offsets and motion positions are defined in metre units. For skeletal reconstruction, 4 Point Transformer [54] layers was utilized to construct the neural network. Within each layer, the feature size was doubled, and the size of the point set was downsampled by a factor of 4 through a farthest-first traversal approach, eventually resulting in a singular feature vector. In the final step, two linear layers with ReLU activations were utilized to transform the resulted feature vector in a frame-wise manner, producing the output skeletal representation. During training, we set \\(K = 8\\) for the KNN loss to optimize the network. Additionally, the use of unit quaternion values in practice relies on modulus division for constraint, a process that can potentially lead to exploding gradients. To address this, we introduced an objective function \\(\\mathcal{L}_q\\) that measures the \\(\\ell_2\\) norms' difference between the raw quaternion output and the expected unit quaternion. All experiments were trained and evaluated on a single NVIDIA GeForce RTX 3090 GPU." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.618, + 0.78, + 0.634 + ], + "angle": 0, + "content": "4.3 Motion Interpolation Comparison against Supervised Methods" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.639, + 0.788, + 0.791 + ], + "angle": 0, + "content": "To demonstrate the efficacy of our method, we produce and compare our method against the conventional LERP method, as well as three state-of-the-art interpolation models trained on the original CMU MoCap dataset. Specifically, we measure the performance of the RNN-based \\(\\mathrm{TG}_{\\mathrm{complete}}\\) model [18], the BERT-based motion interpolation adaptation [12], and a transformer-based encoder-decoder approach [34]. We additionally provide results on a variant of our method trained without our RPA strategy. The experimental configuration for each state-of-the-art model is identical to their original implementations. To measure interpolation accuracy, we employ the standard \\(\\ell_2\\) positional distance (L2P), \\(\\ell_2\\) quaternion difference (L2Q), and NPSS [16] for visual similarity evaluation [18]." + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.797, + 0.787, + 0.84 + ], + "angle": 0, + "content": "3 The authors Clinton Mo, Kun Hu and Zhiyong Wang are signatories of the dataset licenses and produced all experimental results of the paper. Meta Inc. was not granted access to the datasets." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.352, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Motion Keyframe Interpolation for Any Human Skeleton" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table", + "bbox": [ + 0.224, + 0.144, + 0.784, + 0.389 + ], + "angle": 0, + "content": "
Motion categoryMethodL2P↓L2Q↓NPSS↓
515305153051530
BasketballLERP0.07870.40090.80330.15880.58340.96110.13150.61201.8202
TGcomplete [18]0.10500.44190.71040.18590.64390.93030.16000.65791.2136
BERT [12]0.08890.37370.68650.15850.55600.89920.16890.61771.4587
CITL [34]0.06250.29730.57080.14100.49320.80110.10470.40981.0959
PC-MRL w/o RPA0.07670.35340.67150.15050.54310.90490.13640.55961.3471
PC-MRL (Ours)0.07520.34430.66620.15610.54950.90880.13970.55721.3015
GolfLERP0.01890.12290.31870.04550.18700.39870.05870.39791.0312
TGcomplete [18]0.04730.26160.45460.06930.31910.52970.09540.50621.1534
BERT [12]0.02760.11590.29670.06120.20820.42580.10990.39060.9255
CITL [34]0.02010.10430.25890.04760.17660.35520.06360.25810.7799
PC-MRL w/out RPA0.02740.13040.36220.05550.20650.49230.09110.37591.2083
PC-MRL (Ours)0.02730.11300.27060.05980.20120.38440.10490.37160.8150
SwimmingLERP0.07310.36800.73740.13260.53270.94760.21470.92221.6972
TGcomplete [18]0.12920.57960.91410.18450.74201.11990.26600.97161.7698
BERT [12]0.10880.38960.75160.18190.55180.96220.28770.90741.6682
CITL [34]0.09050.38950.74750.14840.55350.95330.20840.83171.6122
PC-MRL w/o RPA0.07220.37540.74590.13450.54420.95940.19440.87311.6608
PC-MRL (Ours)0.07340.35950.71390.13550.53310.92480.19560.85911.6370
Walking & RunningLERP0.04550.25050.55270.11200.39970.62330.06550.26820.9071
TGcomplete [18]0.05400.17860.31030.11470.28660.38660.08950.26160.4265
BERT [12]0.05520.19980.35440.11110.31010.47620.09010.27600.6458
CITL [34]0.03280.10340.24270.09420.21020.33840.05260.15680.4080
PC-MRL w/o RPA0.03970.16570.34170.10580.28790.47220.06400.22880.6021
PC-MRL (Ours)0.03770.13920.28480.10640.27780.42810.06580.21290.5098
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.39, + 0.788, + 0.445 + ], + "angle": 0, + "content": "Table 1: Performance comparisons for various motion categories, measured in L2P, L2Q, and NPSS. All methods except PC-MRL are trained on the original CMU MoCap dataset. Each motion contains 128 frames with keyframes placed every 5, 15, or 30 frames. The top 2 results for each test are highlighted in purple and blue respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.491, + 0.788, + 0.656 + ], + "angle": 0, + "content": "The results listed in Table 1 clearly demonstrate our method's ability to approach the accuracy levels exhibited by directly supervised state-of-the-art methods. PC-MRL consistently outperforms both the LERP standard and RNN-based \\(\\mathrm{TG}_{\\mathrm{complete}}\\) model, particularly in longer keyframe interval scenarios where the quality and depth of learned motion features is most critical and impactful. Likewise, the inclusion of RPA also tends to improve the long interval performance of PC-MRL, particularly when precise movements (i.e. golf) or deeper motion features (i.e. locomotion) are expected. Given this, we can observe that RPA definitively improves the overall consistency of the PC-MRL method in scenarios that may be unseen or less effectively represented by our point cloud-to-motion system." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Native supervision, i.e. with CITL, unsurprisingly produces the most accurate interpolation model, with a notable exception of swimming motions, where our PC-MRL supervision produces even stronger results. We believe the abundance of low crawl motions, a similar motion to swimming in LaFAN1, provides more meaningful supervision over the original CMU dataset, which conversely has few similar motions once swimming motions are filtered out. On the other end of the spectrum, i.e. for high data scenarios such as walking and running locomotion from Fig. 5, all models including our PC-MRL approach observe the strongest improvements for learned methods over the naive LERP method. Both cases strongly support our method's capability to supervise highly intricate motion patterns, despite the incompleteness of the point cloud representation and reliance on relative rotations and geometric optimisation assumptions." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.127 + ], + "angle": 0, + "content": "C. Mo et al." + }, + { + "type": "image", + "bbox": [ + 0.244, + 0.156, + 0.761, + 0.412 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.431, + 0.785, + 0.46 + ], + "angle": 0, + "content": "Fig. 5: Walking keyframe interpolation examples produced by methods from Table 1. Close-up layered comparisons against the original motion are provided on the right." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.489, + 0.673, + 0.506 + ], + "angle": 0, + "content": "4.4 Cross-Skeleton Motion Re-Targeting Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.515, + 0.788, + 0.636 + ], + "angle": 0, + "content": "We further conduct an experiment on motion re-targeting to directly benchmark our method against the sole existing state-of-the-art method for unpaired motion re-targeting with unrestricted skeletons, i.e. primal skeletons (PS) [1]. For direct motion re-targeting evaluation, we utilized the KNN-Loss and end-effector loss as metrics, owing to their effective measurement of visual similarity and non-sensitivity towards absolute roll axis correctness. Though absolute rolls are generally necessary for a complete motion re-targeting solution, this aspect is out of scope for our project as our main goal is motion interpolation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.637, + 0.788, + 0.684 + ], + "angle": 0, + "content": "Table 2 indicates the superior performance of our method over the state-of-the-art PS method in terms of visual similarity. At low \\(\\sigma\\) values, the higher relative \\(\\mathcal{L}_{\\mathrm{KNN}}\\) of the primal skeleton method underscores its geometric deviations" + }, + { + "type": "table", + "bbox": [ + 0.268, + 0.708, + 0.74, + 0.769 + ], + "angle": 0, + "content": "
Model\\( L_{\\text{KNN}}(\\mathcal{X}_A,\\mathcal{X}_B) \\downarrow \\)\\( L_{\\text{end}} \\downarrow \\)
\\( \\sigma = 0 \\)\\( \\sigma = 0.01 \\)\\( \\sigma = 0.05 \\)\\( \\sigma = 0.1 \\)\\( \\sigma = 0.2 \\)
Original motion1.00992.02345.73999.663616.92230
Primal skeleton [1]8.11098.03539.420612.829120.63329.6990
PC-MRL (Ours)3.87144.12316.727710.307817.52615.6258
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.768, + 0.788, + 0.825 + ], + "angle": 0, + "content": "Table 2: Average unpaired LaFAN1 to CMU motion re-targeting performance of the primal skeleton method and our point cloud method. \\(\\mathcal{L}_{\\mathrm{KNN}}\\) is measured using 1024-point clouds at \\(K = 8\\). Independently sampled point clouds from the original LaFAN1 skeleton are provided as optimal \\(\\mathcal{L}_{\\mathrm{KNN}}\\) reference. All values are scaled by \\(\\times 100\\)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.352, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Motion Keyframe Interpolation for Any Human Skeleton" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.235, + 0.147, + 0.769, + 0.279 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.292, + 0.788, + 0.335 + ], + "angle": 0, + "content": "Fig. 6: Jumping motion re-targeting examples produced on the CMU skeleton by the primal skeleton method (middle row) and our method (bottom row). Sampled point cloud representations of each motion sequence are provided on the right." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.375, + 0.788, + 0.513 + ], + "angle": 0, + "content": "from the original motion. In contrast, our method demonstrates significantly better visual adherence to the original geometry. As \\(\\sigma\\) increases, our method approaches near-perfect alignment with the original motion. Figure 6 visually indicates this, showing our method follows the original motion closely, while the latent consistency-based primal skeleton method struggles to generalise and decode the skeleton's internal structures without directly supervised objectives, such as the end-effector positions provided during training. In addition, Fig 7 demonstrates that, like existing re-targeting approaches, our method is able to produce adequate results on disproportionate skeletons." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.517, + 0.788, + 0.579 + ], + "angle": 0, + "content": "Since our approach is largely focused on enabling non-native motion interpolation supervision, we additionally compare our PC-MRL method against a CITL model supervised by PS in Table 3. Due to the suboptimal re-targeting performance of PS, such interpolation models exhibit expectedly poor accuracy." + }, + { + "type": "image", + "bbox": [ + 0.292, + 0.617, + 0.72, + 0.763 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.773, + 0.788, + 0.803 + ], + "angle": 0, + "content": "Fig. 7: Re-targeted falling motion from LaFAN1 (top) on the CMU skeleton (middle) and a short custom skeleton (bottom), performed by our point cloud-to-motion model." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.127 + ], + "angle": 0, + "content": "C. Mo et al." + }, + { + "type": "table", + "bbox": [ + 0.225, + 0.144, + 0.783, + 0.208 + ], + "angle": 0, + "content": "
MethodL2P↓L2Q↓NPSS↓
515305153051530
CITL - PS0.20430.63321.15240.23180.74281.14420.11220.42761.4622
CITL - Native0.04950.23060.48980.09790.36040.66090.11490.46841.1591
PC-MRL0.04940.23370.49600.10300.37710.68130.13700.54331.1760
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.216, + 0.208, + 0.786, + 0.234 + ], + "angle": 0, + "content": "Table 3: Average L2P, L2Q, and NPSS performance of CITL trained on native and PS-generated datasets, compared to our method." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.274, + 0.364, + 0.29 + ], + "angle": 0, + "content": "5 Limitations" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.305, + 0.788, + 0.456 + ], + "angle": 0, + "content": "Due to the inherent limitations of the point cloud representation, certain constraints on the applicability of our point cloud-to-motion method are inevitable. Primarily, the reliance of our point cloud re-targeting method on the relative nature of First-frame Offset Quaternions (FOQ) for most motion learning tasks necessitates the presence of at least one known pose. In the absence of a known pose, relative quaternions are incapable of reconstructing absolute rotational values, which are essential for motion applications. Secondly, due to our method of sampling point clouds with a consistent deviation from each bone segment, the resultant representation inevitably obscures finer details. This includes elements like fingers and facial controls, to an extent that is beyond rectification." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.457, + 0.788, + 0.562 + ], + "angle": 0, + "content": "Our proposed objective function relies entirely on geometric segments, necessitating that all bones possess a non-zero length. This stipulation is deemed a reasonable prerequisite for generic human skeletons, as each bone is typically responsible for manipulating a discernible segment of the human body. Due to technical considerations, the skeletal configurations in each of our datasets included at least one bone of zero length, which we ignored in our experiments in favour of global quaternion rotations of all bones with non-zero length." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.562, + 0.787, + 0.593 + ], + "angle": 0, + "content": "As a result of these limitations, we refrained from claiming state-of-the-art performance for general motion re-targeting tasks." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.614, + 0.36, + 0.63 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.788, + 0.841 + ], + "angle": 0, + "content": "This paper introduces a novel learning-based motion interpolation method designed to enable cross-skeleton compatibility with existing motion datasets. Our proposed method PC-MRL employs a process of point cloud obfuscation and skeleton reconstruction. The point cloud space represents human motions in a non-hierarchical and skeleton-agnostic manner. It enables a KNN-based objective to be optimised without dataset supervision, guiding a neural network to generate high-quality motion features with any target skeleton. We address the rotational information loss of our point cloud format by presenting an offset quaternion strategy compatible with concurrent transformer-based models. Through extensive experiments, we have demonstrated the efficacy of PC-MRL in performing motion interpolation without relying on native motion data. Moreover, PC-MRL has achieved superior visual similarity metrics in the domain of motion re-targeting. We concluded our work by discussing the limitations of PC-MRL." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.352, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Motion Keyframe Interpolation for Any Human Skeleton" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.177, + 0.785, + 0.204 + ], + "angle": 0, + "content": "1. Aberman, K., Li, P., Lischinski, D., Sorkine-Hornung, O., Cohen-Or, D., Chen, B.: Skeleton-aware networks for deep motion retargeting 39(4), 62–1 (2020) 4, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.205, + 0.785, + 0.232 + ], + "angle": 0, + "content": "2. Aberman, K., Wu, R., Lischinski, D., Chen, B., Cohen-Or, D.: Learning character-agnostic motion for motion retargeting in 2d 38(4), 1-14 (jul 2019) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.233, + 0.785, + 0.272 + ], + "angle": 0, + "content": "3. Annabi, L., Ma, Z., et al.: Unsupervised motion retargeting for human-robot imitation. In: Companion of the 2024 ACM/IEEE International Conference on Human-Robot Interactio. ACM (2024) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.273, + 0.785, + 0.314 + ], + "angle": 0, + "content": "4. Bodenheimer, B., Rose, C., Rosenthal, S., Pella, J.: The process of motion capture: Dealing with the data. In: Computer Animation and Simulation. pp. 3-18. Springer (1997) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.315, + 0.785, + 0.342 + ], + "angle": 0, + "content": "5. Burtnyk, N., Wein, M.: Computer-generated key-frame animation. Journal of the SMPTE 80(3), 149-153 (1971) 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.343, + 0.785, + 0.37 + ], + "angle": 0, + "content": "6. Catalin Ionescu, Fuxin Li, C.S.: Latent structured models for human pose estimation. In: International Conference on Computer Vision (2011) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.371, + 0.785, + 0.411 + ], + "angle": 0, + "content": "7. Chen, H., Tang, H., Timofte, R., Gool, L.V., Zhao, G.: Lart: Neural correspondence learning with latent regularization transformer for 3d motion transfer. Advances in Neural Information Processing Systems 36 (2024) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.412, + 0.785, + 0.439 + ], + "angle": 0, + "content": "8. Chen, J., Li, C., Lee, G.H.: Weakly-supervised 3D pose transfer with keypoints. In: ICCV. pp. 15156-15165 (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.44, + 0.785, + 0.467 + ], + "angle": 0, + "content": "9. Choi, K.J., Ko, H.S.: Online motion retargetting. The Journal of Visualization and Computer Animation 11(5), 223-235 (2000) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.468, + 0.785, + 0.494 + ], + "angle": 0, + "content": "0. CMU: Carnegie-mellon university motion capture database (2003), http://mocap.cs.cmu.edu/ 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.495, + 0.785, + 0.522 + ], + "angle": 0, + "content": "1. Dam, E.B., Koch, M., Lillholm, M.: Quaternions, interpolation and animation, vol. 2. CiteSeer (1998) 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.523, + 0.785, + 0.55 + ], + "angle": 0, + "content": "2. Duan, Y., Lin, Y., Zou, Z., Yuan, Y., Qian, Z., Zhang, B.: A unified framework for real time motion completion. In: AAAI. vol. 36, pp. 4459-4467 (2022) 2, 4, 10, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.551, + 0.785, + 0.577 + ], + "angle": 0, + "content": "3. Frank, T., Johnston, O.: Disney Animation: The Illusion of Life. Abbeville Publishing Group, New York (1981) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.578, + 0.785, + 0.605 + ], + "angle": 0, + "content": "4. Gleicher, M.: Retargetting motion to new characters. In: Computer Graphics and Interactive Techniques. pp. 33-42 (1998) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.606, + 0.785, + 0.633 + ], + "angle": 0, + "content": "5. Gleicher, M.: Animation from observation: Motion capture and motion editing 33(4), 51-54 (1999) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.634, + 0.785, + 0.661 + ], + "angle": 0, + "content": "6. Gopalakrishnan, A., Mali, A., Kifer, D., Giles, L., Ororbia, A.G.: A neural temporal model for human motion prediction. In: CVPR. pp. 12116-12125 (2019) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.661, + 0.785, + 0.688 + ], + "angle": 0, + "content": "7. Harvey, F.G., Pal, C.: Recurrent transition networks for character locomotion. In: SIGGRAPH Asia 2018 Technical Briefs, pp. 1-4 (2018) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.689, + 0.785, + 0.716 + ], + "angle": 0, + "content": "8. Harvey, F.G., Yurick, M., Nowrouzezahrai, D., Pal, C.: Robust motion in betweening 39(4), 60-1 (2020) 2, 4, 8, 9, 10, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.717, + 0.785, + 0.743 + ], + "angle": 0, + "content": "9. He, C., Saito, J., Zachary, J., Rushmeier, H., Zhou, Y.: Nemf: Neural motion fields for kinematic animation 35, 4244-4256 (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.744, + 0.785, + 0.784 + ], + "angle": 0, + "content": "20. Hecker, C., Raabe, B., Enslow, R.W., DeWeese, J., Maynard, J., van Prooijen, K.: Real-time motion retargeting to highly varied user-created morphologies 27(3), 1-11 (2008) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.785, + 0.785, + 0.812 + ], + "angle": 0, + "content": "21. Hernandez, A., Gall, J., Moreno-Noguer, F.: Human motion prediction via spatiotemporal inpainting. In: ICCV. pp. 7134-7143 (2019) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "22. Ionescu, C., Papava, D., Olaru, V., Sminchisescu, C.: Human3.6m: Large scale datasets and predictive methods for 3D human sensing in natural environments." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.177, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.127 + ], + "angle": 0, + "content": "C. Mo et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.245, + 0.147, + 0.786, + 0.175 + ], + "angle": 0, + "content": "IEEE Transactions on Pattern Analysis and Machine Intelligence 36(7), 1325-1339 (jul 2014) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.176, + 0.785, + 0.202 + ], + "angle": 0, + "content": "23. Jiang, H., Cai, J., Zheng, J.: Skeleton-aware 3D human shape reconstruction from point clouds. In: ICCV. pp. 5431-5441 (2019) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.203, + 0.785, + 0.228 + ], + "angle": 0, + "content": "24. Kantor, I.L., Solodovnikov, A.S., Shenitzer, A.: Hypercomplex numbers: an elementary introduction to algebras, vol. 302. Springer (1989) 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.229, + 0.785, + 0.268 + ], + "angle": 0, + "content": "25. Karunratanakul, K., Preechakul, K., Suwajanakorn, S., Tang, S.: Guided motion diffusion for controllable human motion synthesis. In: ICCV. pp. 2151-2162 (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.269, + 0.785, + 0.297 + ], + "angle": 0, + "content": "26. Kulpa, R., Multon, F., Arnaldi, B.: Morphology-independent representation of motions for interactive human-like animation. In: Eurographics (2005) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.297, + 0.785, + 0.324 + ], + "angle": 0, + "content": "27. Lehrmann, A.M., Gehler, P.V., Nowozin, S.: Efficient nonlinear markov models for human motion. In: CVPR. pp. 1314-1321 (2014) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.325, + 0.785, + 0.351 + ], + "angle": 0, + "content": "28. Li, P., Aberman, K., Zhang, Z., Hanocka, R., Sorkine-Hornung, O.: Ganimator: Neural motion synthesis from a single sequence 41(4), 1-12 (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.352, + 0.785, + 0.392 + ], + "angle": 0, + "content": "29. Li, W., Wang, Z., Mai, R., Ren, P., Zhang, Q., Zhou, Y., Xu, N., Zhuang, J., Xin, B., Gao, L., et al.: Modular design automation of the morphologies, controllers, and vision systems for intelligent robots: a survey. Visual Intelligence 1(1), 2 (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.393, + 0.785, + 0.433 + ], + "angle": 0, + "content": "30. Li, Z., Peng, X.B., Abbeel, P., Levine, S., Berseth, G., Sreenath, K.: Reinforcement learning for versatile, dynamic, and robust bipedal locomotion control. arXiv preprint arXiv:2401.16889 (2024) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.434, + 0.785, + 0.459 + ], + "angle": 0, + "content": "31. Liao, Z., Yang, J., Saito, J., Pons-Moll, G., Zhou, Y.: Skeleton-free pose transfer for stylized 3D characters. In: ECCV. pp. 640-656. Springer (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.46, + 0.785, + 0.514 + ], + "angle": 0, + "content": "32. Makoviychuk, V., Wawrzyniak, L., Guo, Y., Lu, M., Storey, K., Macklin, M., Hoeller, D., Rudin, N., Allshire, A., Handa, A., et al.: Isaac gym: High performancegpu-based physics simulation for robot learning. arXiv preprint arXiv:2108.10470 (2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.515, + 0.785, + 0.555 + ], + "angle": 0, + "content": "33. Merry, B., Marais, P., Gain, J.: Animation space: A truly linear framework for character animation. ACM Transactions on Graphics (TOG) 25(4), 1400-1423 (2006) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.556, + 0.785, + 0.595 + ], + "angle": 0, + "content": "34. Mo, C.A., Hu, K., Long, C., Wang, Z.: Continuous intermediate token learning with implicit motion manifold for keyframe based motion interpolation. In: CVPR. pp. 13894-13903 (2023) 2, 4, 8, 10, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.597, + 0.785, + 0.637 + ], + "angle": 0, + "content": "35. Mo, C.A., Hu, K., Mei, S., Chen, Z., Wang, Z.: Keyframe extraction from motion capture sequences with graph based deep reinforcement learning. In: ACM International Conference on Multimedia. pp. 5194-5202 (2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.638, + 0.785, + 0.678 + ], + "angle": 0, + "content": "36. Monzani, J.S., Baerlocher, P., Boulic, R., Thalmann, D.: Using an intermediate skeleton and inverse kinematics for motion retargeting. In: Computer Graphics Forum. vol. 19, pp. 11-19. Wiley Online Library (2000) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.679, + 0.785, + 0.704 + ], + "angle": 0, + "content": "37. Oreshkin, B.N., Valkanas, A., Harvey, F.G., Menard, L.S., Bocquelet, F., Coates, M.J.: Motion inbetweening via deep \\(\\delta\\)-interpolator (2022) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.705, + 0.785, + 0.745 + ], + "angle": 0, + "content": "38. Peng, X.B., Abbeel, P., Levine, S., van de Panne, M.: Deepmimic: Example-guided deep reinforcement learning of physics-based character skills 37(4), 143:1-143:14 (Jul 2018) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.746, + 0.785, + 0.772 + ], + "angle": 0, + "content": "39. Peng, X.B., Ma, Z., Abbeel, P., Levine, S., Kanazawa, A.: Amp: Adversarial motion priors for stylized physics-based character control 40(4) (Jul 2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.773, + 0.785, + 0.799 + ], + "angle": 0, + "content": "40. Qin, J., Zheng, Y., Zhou, K.: Motion in-between via two-stage transformers 41(6), 1-16 (2022) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "41. Reda, D., Won, J., Ye, Y., van de Panne, M., Winkler, A.: Physics-based motion retargeting from sparse inputs. ACM Computer Graphics and Interactive Techniques 6(3), 1-19 (2023) 4" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.786, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.352, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Motion Keyframe Interpolation for Any Human Skeleton" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.175 + ], + "angle": 0, + "content": "42. Reeves, W.T.: Inbetweening for computer animation utilizing moving point constraints 15(3), 263-269 (1981) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.176, + 0.785, + 0.217 + ], + "angle": 0, + "content": "43. Ren, T., Yu, J., Guo, S., Ma, Y., Ouyang, Y., Zeng, Z., Zhang, Y., Qin, Y.: Diverse motion in-betweening from sparse keyframes with dual posture stitching. IEEE Transactions on Visualization and Computer Graphics (2024) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.218, + 0.785, + 0.244 + ], + "angle": 0, + "content": "44. Shaw, P., Uszkoreit, J., Vaswani, A.: Self-attention with relative position representations. In: NAACL. Association for Computational Linguistics (2018) 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.245, + 0.785, + 0.286 + ], + "angle": 0, + "content": "45. Tiwari, G., Antic, D., Lenssen, J.E., Sarafianos, N., Tung, T., Pons-Moll, G.: Posendf: Modeling human pose manifolds with neural distance fields. In: ECCV. pp. 572-589. Springer (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.785, + 0.314 + ], + "angle": 0, + "content": "46. Villegas, R., Yang, J., Ceylan, D., Lee, H.: Neural kinematic networks for unsupervised motion retargeting. In: CVPR. pp. 8639-8648 (2018) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.785, + 0.354 + ], + "angle": 0, + "content": "47. Wang, J., Li, X., Liu, S., De Mello, S., Gallo, O., Wang, X., Kautz, J.: Zero-shot pose transfer for unrigged stylized 3D characters. In: CVPR. pp. 8704-8714 (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.355, + 0.785, + 0.383 + ], + "angle": 0, + "content": "48. Wang, K., Xie, J., Zhang, G., Liu, L., Yang, J.: Sequential 3D human pose and shape estimation from point clouds. In: CVPR. pp. 7275-7284 (2020) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.785, + 0.411 + ], + "angle": 0, + "content": "49. Yuan, Y., Song, J., Iqbal, U., Vahdat, A., Kautz, J.: Physdiff: Physics-guided human motion diffusion model. In: ICCV. pp. 16010-16021 (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.411, + 0.785, + 0.438 + ], + "angle": 0, + "content": "50. Yuan, Y., Wei, S.E., Simon, T., Kitani, K., Saragih, J.: Simpoe: Simulated character control for 3D human pose estimation. In: CVPR. pp. 7159-7169 (2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.785, + 0.481 + ], + "angle": 0, + "content": "51. Zhang, J., Weng, J., Kang, D., Zhao, F., Huang, S., Zhe, X., Bao, L., Shan, Y., Wang, J., Tu, Z.: Skinned motion retargeting with residual perception of motion semantics & geometry. In: CVPR. pp. 13864-13872 (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.481, + 0.785, + 0.522 + ], + "angle": 0, + "content": "52. Zhang, M., Li, H., Cai, Z., Ren, J., Yang, L., Liu, Z.: Finemogen: Fine-grained spatio-temporal motion generation and editing. Advances in Neural Information Processing Systems 36 (2024) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.522, + 0.785, + 0.563 + ], + "angle": 0, + "content": "53. Zhang, X., van de Panne, M.: Data-driven autocompletion for keyframe animation. In: ACM SIGGRAPH Conference on Motion, Interaction and Games. pp. 1-11 (2018) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.564, + 0.785, + 0.591 + ], + "angle": 0, + "content": "54. Zhao, H., Jiang, L., Jia, J., Torr, P.H., Koltun, V.: Point transformer. In: ICCV. pp. 16259-16268 (2021) 7, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.591, + 0.785, + 0.632 + ], + "angle": 0, + "content": "55. Zhu, W., Yang, Z., Di, Z., Wu, W., Wang, Y., Loy, C.C.: Mocanet: Motion retargeting in-the-wild via canonicalization networks. In: AAAI. vol. 36, pp. 3617-3625 (2022) 4" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.632 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Motion Keyframe Interpolation for Any Human Skeleton using Point Cloud-based Human Motion Data Homogenisation/391ac41a-5023-4e3a-b209-fdaecfa62b6a_origin.pdf b/2024/Motion Keyframe Interpolation for Any Human Skeleton using Point Cloud-based Human Motion Data Homogenisation/391ac41a-5023-4e3a-b209-fdaecfa62b6a_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..82e3cae3d186952ebc617cc0de1e0d8086008a5e --- /dev/null +++ b/2024/Motion Keyframe Interpolation for Any Human Skeleton using Point Cloud-based Human Motion Data Homogenisation/391ac41a-5023-4e3a-b209-fdaecfa62b6a_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:40ab048744f6ef182f5ad14bedba5bcaaac72728af59e52c5d9bb5d176049b72 +size 11324956 diff --git a/2024/Motion Keyframe Interpolation for Any Human Skeleton using Point Cloud-based Human Motion Data Homogenisation/full.md b/2024/Motion Keyframe Interpolation for Any Human Skeleton using Point Cloud-based Human Motion Data Homogenisation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..f71a7859240070e165563bccc71bd9a3a1aa3c69 --- /dev/null +++ b/2024/Motion Keyframe Interpolation for Any Human Skeleton using Point Cloud-based Human Motion Data Homogenisation/full.md @@ -0,0 +1,303 @@ +# Motion Keyframe Interpolation for Any Human Skeleton via Temporally Consistent Point Cloud Sampling and Reconstruction + +Clinton Mo $^{1}$ , Kun Hu $^{1,\star}$ , Chengjiang Long $^{2}$ , Dong Yuan $^{1}$ , and Zhiyong Wang $^{1}$ + +1 The University of Sydney, Darlington NSW 2008, Australia +clmo6615@uni.sydney.edu.au, {kun.hu, zhiyong.wang, dong.yuan}@sydney.edu.au + +$^{2}$ Meta Reality Labs, Burlingame, CA, USA + +clong1@meta.com + +Abstract. In the character animation field, modern supervised keyframe interpolation models have demonstrated exceptional performance in constructing natural human motions from sparse pose definitions. As supervised models, large motion datasets are necessary to facilitate the learning process; however, since motion is represented with fixed hierarchical skeletons, such datasets are incompatible for skeletons outside the datasets' native configurations. Consequently, the expected availability of a motion dataset for desired skeletons severely hinders the feasibility of learned interpolation in practice. To combat this limitation, we propose Point Cloud-based Motion Representation Learning (PC-MRL), an unsupervised approach to enabling cross-compatibility between skeletons for motion interpolation learning. PC-MRL consists of a skeleton obfuscation strategy using temporal point cloud sampling, and an unsupervised skeleton reconstruction method from point clouds. We devise a temporal point-wise K-nearest neighbors loss for unsupervised learning. Moreover, we propose First-frame Offset Quaternion (FOQ) and Rest Pose Augmentation (RPA) strategies to overcome necessary limitations of our unsupervised point cloud-to-skeletal motion process. Comprehensive experiments demonstrate the effectiveness of PC-MRL in motion interpolation for desired skeletons without supervision from native datasets. + +Keywords: 3D point clouds $\cdot$ Human body motion $\cdot$ Dataset creation + +# 1 Introduction + +3D character animation workflows largely rely on the concepts of keyframing with interpolation, often referred to as the pose-to-pose principle [13]. By defining key poses at correct timings, algorithms can be employed to generate intermediate poses, and thereby eliminating the need for defining each frame individually [5, 42]. Although this keyframe-based workflow is less costly than frame-by-frame production, human motion is often complex, and natural depictions + +![](images/d2218860b0ef14c222532445ddae55cd3415e56b8bcf0d28f405c5ce4aab6c0c.jpg) +(a) Linear interpolation + +![](images/e3bfa326343bc12db7c8e9e70f1524810ef7753ef624ffd77c7e09b16cfd69c3.jpg) +(b) PC-MRL (Our method) + +![](images/44d2a55e7203e40d15ac1c42294ae3da0256fcc101b9cd092f1004ab6aa6faa2.jpg) +(c) Ground truth + +![](images/ee25c25af7037e509bc3473cb4d65ecefc982ec85197badae4aac3be6468f142.jpg) +(a) Hierarchical skeleton + +![](images/61f3a519e8e4097725bad99f609add4adf90d3642355878e3fb74294765cbaf4.jpg) +Fig. 1: Motion interpolation of walking keyframes conducted by our approach. +(b) Skeletal motion + +![](images/f908a2abf1d811ef7f297dd1286a831a856acd18723f2f794f65c2e56c3db689.jpg) +(c) Point cloud representation +Fig. 2: Our method leverages point cloud representations (c) to obfuscate the hierarchical dependencies of skeletal motion. Joint colours correspond to ordinal indices in original motion data representation. Point clouds are coloured by temporal index. + +demand a significant number of keyframes, especially for realistic motions that adhere to physical properties and constraints. Although motion capture (MoCap) is frequently employed as an alternative for achieving realistic human motions, MoCap often nonetheless necessitates subsequent keyframing to correct unwanted motion elements and address other imperfections. + +In recent years, the interpolation process has been widely studied, particularly with the advent of machine learning methods for sequential data. Compared to conventional methods such as linear interpolation (LERP), machine learning methods have demonstrated the capability to derive visually natural motions from sparser keyframe sets [12, 18, 34, 37, 40], which we can observe in Fig. 1. Data-driven interpolation approaches require large motion capture datasets to learn the necessary motion features for effective synthesis of transitions between keyframes. In the field of 3D motions, this aspect significantly restricts the feasibility of learned interpolation methods in animation practice, as motion/pose data are tied to specific skeletal configurations and are not cross-compatible. That is, the dataset's native skeleton, i.e. source skeleton, is structurally different and incompatible with the desired skeleton, i.e. target skeleton. + +Therefore, we propose a novel point cloud-based motion representation learning (PC-MRL) approach, in which skeletal hierarchies are obfuscated by sampling into the point cloud medium, and reconstructed into skeletal motion data using an unsupervised neural network. Unlike raw pose and motion data, the + +point cloud format is a non-hierarchical representation, and effectively obscures any skeletal configuration as displayed in Fig. 2. PC-MRL samples a set of points around the rest position of skeletons using a combination of uniform and normal distributions, and geometrically parents each point to an associated bone in order to effectively represent the associated skeletons' motion data in a temporally consistent manner. Subsequently, for the reconstruction of motion data into a given desired skeleton, we propose an unsupervised learning scheme, featuring a K-nearest neighbors (KNN) loss between cross-skeleton point clouds to optimize a transformer neural network. The network processes the point cloud-based motion representations and generates the corresponding target skeletal motion representations. By obscuring both source and predicted target motion data into the point cloud space, our strategy allows for a direct comparison, utilizing KNN to minimize the visual difference between the two point clouds. + +The unsupervised nature of our target skeleton reconstruction scheme creates two main disparities between the expected and predicted motion features. Firstly, the exact roll rotation axis of bones cannot be represented in the point cloud format. To this end, we introduce First-frame Offset Quaternion (FOQ) representations to incorporate relative roll values, which are obtainable from temporally consistent point clouds. This approach is agnostic to absolute roll rotations, providing a standardisation mechanism for the resulting motion data within the context of motion interpolation learning. Second, KNN-based objectives learn skeletal representations in a geometrically optimal manner, which does not always reflect the desired skeletal behaviour. This disparity is commonly observed with smaller skeletal features, such as shoulder and hip bones. As such, we augment training motions using Rest Pose Augmentation (RPA) to increase the range of skeletal behaviours that our interpolation model learns. Extensive experiments demonstrate the effectiveness of our proposed method, achieving performance targets near the level of direct dataset supervision. + +In summary, this paper presents the following key contributions: + +1. We propose a novel method for achieving unsupervised human motion reconstruction from point cloud data, enabling skeleton-agnostic motion interpolation learning for the first time. +2. To train motion interpolation models, we formulate first-frame offset quaternions to represent bone rotations with relative roll data, as well as a rest position augmentation strategy to address skeletal configuration variability. +3. We perform comprehensive experiments to demonstrate our method's effectiveness towards learning motion interpolation, despite the absence of directly compatible datasets in the training process. + +# 2 Related works + +We explore existing research on motion interpolation approaches, and motion data modelling methods in general. In addition, our proposed approach bears resemblance to unsupervised approaches for the motion re-targeting research problem, and as such, we will also analyse existing methods in this field. + +# 2.1 Data-Driven Motion Modelling Methods + +Contemporary research in motion modelling predominantly explore learned data-driven methods, using neural networks to capture, model, and establish correlations between the motion features of various skeletal configurations [29, 35, 41, 51, 55]. Motion features alone have demonstrated considerable efficacy in cross-skeleton imitation using a variety of strategies, including task-driven objectives [19, 45], diffusion-based generation [25, 49, 52], and latent feature consistency [1-3, 7, 46]. Latent consistency in particular learns inter-skeleton correlations without the need for paired datasets, which reduces its barrier of entry. + +Likewise, motion interpolation strategies have evolved through the use of learning based approaches. The inherent numerical precision of motion data has traditionally posed a challenge for the adoption of the learning approaches, due to their approximate nature [21,27]. This challenge is further amplified by the temporally sparse distribution of keyframes. Recently, a number of recurrent neural network based methods [17, 18, 43, 53] and transformer-based networks [12, 34, 37] have demonstrated encouraging interpolation performance for the transition between keyframes. Additionally, various other motion modelling methods are capable for keyframe based interpolation, albeit with limited intermediate frame representation capabilities [19, 25, 45]. + +A major constraint inherent in all data-driven motion modelling methods is their dependency on datasets featuring specific skeletal configurations. To address this, skeleton-free training strategies have been explored for elementary tasks involving 3D characters. These include vertex-based pose transfer [7,8,31,47], and point cloud-based human shape reconstruction [23,48]. Notably, in these methods, 3D mesh and point cloud coordinates serve as general 3D data representations. Concurrently, other research has significantly minimized the required volume of training data, bringing it down to single example sequences. This reduction has been achieved through the use of patch matching techniques [28] and imitation learning within physics-based simulations [30,39]. Reinforcement learning has facilitated the generation of goal-driven motion without pre-existing datasets [32,38,50]. However, these simulation-based methods have typically been either exceedingly complex to implement in practice, or yield results that are too inflexible and/or prone to errors for animation workflows. + +# 2.2 Motion Re-Targeting without Supervision + +Conventional approaches in motion re-targeting have been a cornerstone of 3D animation for decades, primarily dedicated to converting motion capture data into 3D skeletal motions [4, 15]. Typically, raw motion capture data is optically recorded, yielding primarily spatial information. In contrast, motion data is largely rotational in nature [5], often represented in $SO(3)$ space. Therefore, a principal objective of motion re-targeting solutions has been to effectively bridge these two distinct types of representations. In [14], extensive manually adjustable constraints were introduced for flexible re-targeting between skeletons of identical topology (i.e., same hierarchy, different bone lengths and rest poses). Key + +![](images/b2e7a16acdf66b05482e71dd89d4c86ffdc781ecfb2d07b2bd8b9030fd37b84b.jpg) +Fig. 3: Illustration of point cloud-based motion representation learning (PC-MRL) pipeline for keyframe interpolation to any human skeleton. Our unsupervised point cloud-to-pose decoding approach (I) first adapts motion features from existing datasets to new skeletons. We then employ the adapted motion data to supervise motion interpolation models (II), using FOQ and rest pose augmentations to address remaining discrepancies between the adapted and expected target skeleton motion patterns. + +practices in this approach, including end-effector matching, joint contacts, and inter-skeleton point correlations, are still relevant to current animation pipelines. Point correlations using Inverse Kinematics (IK) have also been proposed as an intermediate representation to facilitate the adaptation between topologically distinct skeletons [9, 36], albeit with hierarchical pairing limitations. + +In response, morphologically-independent motion control systems were proposed using IK joint chains and joint-wise constraints [20,26]. Such systems often expect highly standardised motion skeleton features, and are generally impractical for re-targeting motions produced for different control schemes. Additionally, the reliance on Euler angles for extensive classical re-targeting constraints poses a variety of issues regarding rotational freedom [11], as well as compatibility with animation workflows [33], which predominantly use quaternion-based systems. + +Despite these advancements, the challenge of arbitrary skeleton re-targeting from motion datasets still remains largely unresolved, limiting any widespread application of data-driven pipelines. Our paper aims to effectively address this pervasive limitation. + +# 3 Methodology + +As illustrated in Fig. 3, our proposed method consists of two main components: point cloud-to-motion reconstruction and motion keyframe interpolation. + +# 3.1 Quaternion-based Motion Representations + +We first declare a set of notations for representing skeletons. A skeleton $S$ is defined by a tree of bones $\mathcal{B}$ . A bone $\mathbf{v}_n \in \mathcal{B}$ is defined by its bone parent $\mathbf{v}_m \in \mathcal{B}$ , and a head position $\mathbf{H}_n \in \mathbb{R}^3$ which denotes the bone's base offset from + +its parent head $\mathbf{H}_m$ . In addition, a tail position $\mathbf{T}_n \in \mathbb{R}^3$ helps to define the end point of the bone's line segment representation, starting from its head position, of $\mathbf{v}_n$ . The root bone $\mathbf{v}_0$ has no parent, and thus no head position $\mathbf{H}_0$ . The set of $\mathbf{H}_n$ and $\mathbf{T}_n$ for each bone $\mathbf{v}_n \in \mathcal{B}$ make up the rest position of the skeleton. + +Next, we define the properties of motion data. At any given frame $t$ , a skeleton's pose position can be represented as global 3D positions $\mathbf{p}_{n,t} \in \mathbb{R}^3$ and global quaternion rotations $\mathbf{q}_{n,t} \in \mathbb{R}^4$ for each bone $\mathbf{v}_n \in \mathcal{B}$ . For a given motion sequence $M$ , only the root bone is given a 3D position in pose position, i.e. $\mathbf{p}_{0,t}$ as all other pose positions are derived using Forwards Kinematics (FK [11,24]): + +$$ +\mathbf {p} _ {n, t} = \mathbf {p} _ {m, t} + Q R \left(\mathbf {H} _ {n}, \mathbf {q} _ {m, t}\right), \tag {1} +$$ + +where $QR$ is the quaternion rotation function, and $\mathbf{v}_m$ is the parent joint. In summary, a motion sequence $M$ of $|M|$ frames is defined by: + +$$ +M = \left\{\mathbf {p} _ {0, t} \cup \left\{\mathbf {q} _ {n, t} \mid \mathbf {v} _ {n} \in \mathcal {B} \right\} \mid 0 \leq t < | M | \right\} \tag {2} +$$ + +# 3.2 Point Cloud Obfuscation and Skeleton Reconstruction + +Our approach is based on the hypothesis that 3D motion sequences can be visually represented by a more universal format: 3D point cloud sequences. By geometrically sampling the volume around a skeleton, point cloud data becomes a non-hierarchical representation of its pose(s), obfuscating the original skeletal configuration. Due to this decoupling between the skeleton configuration and motion representation, a successful motion data reconstruction from point cloud data would be the first step to enabling learnable cross-skeleton motion features. Formally, for a given source skeleton $S_A$ with an associated motion sequence $M_A$ , and any humanoid skeletal configuration $S_B$ , the point cloud obfuscation and reconstruction module is to learn a function that can adequately perform $F_{S_A \to S_B}(M_A) = \hat{M}_B$ , where $\hat{M}_B$ is a visually similar motion adaptation of $M_A$ on skeleton $S_B$ , which is an estimation of $M_B$ . + +Motion Data Obfuscation with Point Clouds To sample the point cloud $\mathcal{X}$ , we first generate a set of points $u\in \mathcal{X}$ along the line segments defined by the global head and tail positions for each bone $\mathbf{v}_n$ . In the bone-local space to $\mathbf{v}_n$ , this would mean sampling $u$ by a uniformly distributed factor $\alpha \sim \mathcal{U}_{[0,1]}$ . Specifically, $\alpha = 0$ aligns with the head position of the bone, and $\alpha = 1$ indicates the tail position. By further sampling $u$ via a normal distribution, our sampling strategy can capture the surrounding volume around the skeleton. Formally, to sample $u$ from its associated bone $\mathbf{v}_n$ within a standard deviation of $\sigma$ , we have: + +$$ +\mathbf {H} _ {x} \sim \mathcal {N} (\alpha \mathbf {T} _ {n}, \sigma), +$$ + +$$ +\mathbf {p} _ {u, t} = \mathbf {p} _ {n, t} + Q R (\mathbf {H} _ {x}, \mathbf {q} _ {n, t}) \tag {3} +$$ + +Notably, this sampling process as in Eq. 3 is associated with the skeletal configuration only, by which the sampled points are temporally consistent for further processing with the motion sequence. Specifically, each point maintains a + +constant position from its associated bone in the bone-local space. The lateral distance of each point from their parent bone segment allows point clouds to react to the bone's rotational roll axis. + +In a pure point cloud representation, it is possible to remove the relationships between bones. However, to clearly distinguish symmetrical body features in different skeletons, we introduce and embed body group associations for the points during our reconstruction process. Specifically, every bone is categorised by one of these body groups: [spine, left arm, right arm, left leg, right leg]. Each sampled point is assigned q group attribute in line with the associated bone. The attribute is represented by a one-hot feature vector $\mathbf{g}_u$ . To this end, we characterize $u = \{\mathbf{p}_{u,t}, \mathbf{g}_u\} \in \mathcal{X}$ . + +Skeleton Reconstruction Given a sampled point cloud representation $\mathcal{X}_A$ from $S_A$ and $M_A$ as input, we aim to train a motion adaptation function $F_{S_A\to S_B}$ which is tasked with producing a visually similar skeletal motion $\hat{M}_B$ for $S_B$ using a temporal and set-based neural network. As shown in Fig. 3, to construct $F_{S_A\rightarrow S_B}$ , we employ a Point Transformer [54] to learn embeddings for the unordered point cloud sets frame-wisely, and a standard temporal transformer to process the sequence and decode $\mathcal{X}_A$ into $\hat{M}_B$ . + +To optimize $F_{S_A \to S_B}$ , our objective function maximises the similarity between the sampled point clouds of both $M_A$ and $\hat{M}_B$ . $\mathcal{X}_A$ can be viewed as the ground truth to $\mathcal{X}_B$ , which enables an unsupervised strategy for $F_{S_A \to S_B}$ . In detail, we first sample a point cloud $\mathcal{X}_B$ from $S_B$ based on $\hat{M}_B$ derived by $F_{S_A \to S_B}$ . Next, a temporally consistent KNN-Loss objective is devised to minimise the $\ell_2$ distance between any given point of $\mathcal{X}_B$ from its $K$ nearest neighbouring points in $\mathcal{X}_A$ . The $K$ nearest neighbours are determined based on inter-point distance throughout the entire motion sequence, instead of a frame-wise setting, as shown in Fig. 4. The points with differing body groups are excluded from the neighbour search. Mathematically, we have: + +$$ +\delta \left(u _ {A}, u _ {B}\right) = \left\{ \begin{array}{l l} \sum_ {t} \left\| \mathbf {p} _ {u _ {A}, t} - \mathbf {p} _ {u _ {B}, t} \right\| _ {2}, & \text {i f} \mathbf {g} _ {u _ {A}} = \mathbf {g} _ {u _ {B}}, \\ \inf , & \text {o t h e r w i s e}, \end{array} \right. \tag {4} +$$ + +$$ +\mathcal {L} _ {\mathrm {K N N}} \left(\mathcal {X} _ {A}, \mathcal {X} _ {B}\right) = \sum_ {u _ {A} \in \mathcal {X} _ {A}, u _ {B} \in N _ {k} \left(u _ {A}, \mathcal {X} _ {B}\right)} \delta \left(u _ {A}, u _ {B}\right), \tag {5} +$$ + +where $N_{k}(u_{A},\mathcal{X}_{B})$ is the set containing points that are the k-nearest neighbors of $u_{A}$ in $\mathcal{X}_B$ , based on $\delta (u_A,u_B)$ distance. + +In addition to the KNN-Loss, we introduce an optional end-effector loss in the skeleton space for the hand, foot, and head bones, which provides direct positional guidance for matched end-effector pairs $E_{AB}$ : + +$$ +\mathcal {L} _ {\text {e n d}} = \sum_ {(\mathbf {v} _ {A}, \mathbf {v} _ {B}) \in E _ {A B}} \frac {1}{| M _ {A} |} \sum_ {t} \left\| \mathbf {p} _ {\mathbf {v} _ {A}, t} - \mathbf {p} _ {\mathbf {v} _ {B}, t} \right\| _ {2}. \tag {6} +$$ + +![](images/3a14d93f1d2874a63ea21ea43eb030712e6913a3b9a1c82718ed631ffec26b07.jpg) +Fig. 4: Visualisation of our temporally consistent KNN objective, processed independently per body group. For each point from $\mathcal{X}_A$ (a), the total $\ell_2$ distance throughout the sequence is measured between each point (b) from $\mathcal{X}_B$ . The $\ell_2$ distances from the K-nearest points are summed to produce KNN-Loss (c). + +# 3.3 First-Frame Offset Quaternions + +Point clouds lack absolute roll axis information for each bone. As such, we introduce a First-frame Offset Quaternion (FOQ) transform to incorporate the available relative roll data for motion modelling. We formulate all quaternions of a motion sequence relative to their values from the initial frame. In detail, for any quaternion sequence $Q = \{\mathbf{q}_0, \mathbf{q}_1, \mathbf{q}_2, \ldots\}$ , we can simply obtain its FOQ transform as $\{\mathbf{q}_0 \times \mathbf{q}_0^{-1}, \mathbf{q}_1 \times \mathbf{q}_0^{-1}, \mathbf{q}_2 \times \mathbf{q}_0^{-1}, \ldots\}$ , where $\mathbf{q}_0^{-1}$ is the quaternion conjugate of $\mathbf{q}_0$ . An FOQ sequence can easily be converted back to a quaternion sequence as long as any $\mathbf{q}_t \in Q$ is known. + +Crucially, FOQ is a roll-invariant rotation representation, that is, FOQ remains identical regardless of a bone's absolute roll position in the original data. Trivially, to adjust the rotational roll of a bone $\mathbf{v}_n$ , we perform the quaternion multiplication for each frame: $\mathbf{q}_{n,t} \times \mathbf{r}$ , where $\mathbf{r}$ is a roll quaternion. As a quaternion on the roll axis of $\mathbf{v}_n$ , $\mathbf{r}$ is constrained within $\mathbf{r} = \alpha + \beta (x\mathbf{i} + y\mathbf{j} + z\mathbf{k})$ , where $x,y,z$ are the 3D XYZ values of $\mathbf{T}_n$ , and $\alpha, \beta$ are scalars that control the roll magnitude. Note that $\mathbf{r}$ is constant throughout the sequence. Therefore, we can deduce the roll-invariant nature of any constant $\mathbf{r}$ : + +$$ +\begin{array}{l} \left(\mathbf {q} _ {n, t} \times \mathbf {r}\right) \times \left(\mathbf {q} _ {0, t} \times \mathbf {r}\right) ^ {- 1} = \mathbf {q} _ {n, t} \times \mathbf {r} \times \mathbf {r} ^ {- 1} \times \mathbf {q} _ {0, t} ^ {- 1} \\ = \mathbf {q} _ {n, t} \mathbf {q} _ {0, t} ^ {- 1} \\ \end{array} +$$ + +# 3.4 Motion Interpolation Transformer and Rest Pose Augmentation + +To leverage our point cloud-based representation learning for cross-skeleton interpolation, a transformer model - CITL [34] is adapted, allowing for an efficient interpolation pipeline with existing datasets to alternative skeletons. We introduce two key modifications to CITL and its training strategy, focusing on enhancing motion quality. First, quaternion predictions are substituted with FOQ predictions. This is inspired by the RNN-based method TGcomplete [18], which predicts + +Algorithm 1 IK target alignment with maximum real quaternion component +1: Input: $\mathbf{p_{IK}} = 3D$ IK target unit vector +2: $\mathbf{T} =$ resting tail position of associated bone +3: Output: $\mathbf{q}_{\mathrm{max}} =$ output quaternion, to be applied on original bone quaternion +4: $\eta \leftarrow \frac{\mathbf{T}}{|\mathbf{T}|}$ $\triangleright$ 3D unit coordinates $[\eta_x,\eta_y,\eta_z]$ +5: $\omega \leftarrow \frac{\mathbf{p_{IK}}}{|\mathbf{p_{IK}}|}$ $\triangleright$ 3D unit coordinates $[\omega_x,\omega_y,\omega_z]$ +6: $\mathbf{q}_{\mathrm{min}} \gets \frac{0 + (\eta_x + \omega_x)\mathbf{i} + (\eta_y + \omega_y)\mathbf{j} + (\eta_z + \omega_z)\mathbf{k}}{2}$ +7: $\mathbf{q}_{\mathrm{max}} \gets \mathbf{q}_{\mathrm{min}} \times (0 + \eta_x\mathbf{i} + \eta_y\mathbf{j} + \eta_z\mathbf{k})$ + +quaternion offsets from the preceding frame rather than raw quaternions, thereby achieving greater generalisability and superior quality in interpolation. + +For our model's CITL-based architecture, we substitute the original sinusoidal positional encoding scheme with learned relative positions [44]. While the original implementation relies on the continual nature of sinusoidal functions, we observe that relative positional embeddings with zero initialisation converges upon similarly continuous behaviour, and significantly lowers the complexity of positional attention. The lessened focus on positional relations allows the model to synthesise deeper behaviours within the pose space, improving overall generalisability. As a side effect, the model also accepts arbitrary length inputs beyond the training data length. + +Notably, the intended rest pose of a skeletal configuration is often geometrically sub-optimal for point cloud matching, and can manifest in multiple forms. To address this, we devise a rest position augmentation strategy, leveraging a per-bone Inverse Kinematics (IK) system. RPA broadens the range of intended rest poses that the learned interpolation model can accommodate. In detail, during each phase of the Forward Kinematics (FK) process, we augment the bone tails with a random additional offset. We re-align the augmented global tail position as closely as possible to its original location using a quaternion rotation with the maximum possible real component. Algorithm 1 describes the method by which this quaternion can be derived. Since the real component describes, in cosine, the amount of rotation around an axis, this minimizes numerical disturbance from RPA while altering the visual representation to the desired extent. + +# 4 Experimental Results + +# 4.1 Datasets + +To thoroughly evaluate the effectiveness of the proposed method, we conducted extensive experiments on the following widely used motion capture datasets: + +- LaFAN1 [18] contains long, high quality motions in controllable video game character styles. The dataset was recorded on a large motion capture stage, and manually cleaned to production standards. The native skeleton of LaFAN1 contains 6 spinal bones, and 4 bones for each limb. + +- Human3.6M [6,22] comprises of 3.6 million motion frames. The motions captured in Human3.6M tend to be less dynamic, as they were recorded on a relatively compact $4 \times 3 \mathrm{~m}$ stage. Its native skeleton includes 5 spinal bones, 4 bones for each limb, and an additional finger bone for each hand. + +- CMU Mocap [10] is a diverse motion dataset encompassing 113 categories. It features a mix of static and dynamic motions, captured on a $3 \times 8 \mathrm{~m}$ stage. The skeleton is identical to that of Human3.6M, with the exception of the pelvis, which is divided into two hip bones and a lower back bone. + +For PC-MRL, we designated the CMU MoCap skeleton as $S_B$ . Its native dataset is exclusively utilized for testing purposes, while LaFAN1 and Human3.6M are used simultaneously for training. $^3$ + +# 4.2 Implementation Details + +In our experiments, point cloud obfuscation adopted a 256-point setting for the sampling process with $\sigma = 0.05$ and each point was characterized by a 64-element vector. To maintain consistency, we ensure that all bone offsets and motion positions are defined in metre units. For skeletal reconstruction, 4 Point Transformer [54] layers was utilized to construct the neural network. Within each layer, the feature size was doubled, and the size of the point set was downsampled by a factor of 4 through a farthest-first traversal approach, eventually resulting in a singular feature vector. In the final step, two linear layers with ReLU activations were utilized to transform the resulted feature vector in a frame-wise manner, producing the output skeletal representation. During training, we set $K = 8$ for the KNN loss to optimize the network. Additionally, the use of unit quaternion values in practice relies on modulus division for constraint, a process that can potentially lead to exploding gradients. To address this, we introduced an objective function $\mathcal{L}_q$ that measures the $\ell_2$ norms' difference between the raw quaternion output and the expected unit quaternion. All experiments were trained and evaluated on a single NVIDIA GeForce RTX 3090 GPU. + +# 4.3 Motion Interpolation Comparison against Supervised Methods + +To demonstrate the efficacy of our method, we produce and compare our method against the conventional LERP method, as well as three state-of-the-art interpolation models trained on the original CMU MoCap dataset. Specifically, we measure the performance of the RNN-based $\mathrm{TG}_{\mathrm{complete}}$ model [18], the BERT-based motion interpolation adaptation [12], and a transformer-based encoder-decoder approach [34]. We additionally provide results on a variant of our method trained without our RPA strategy. The experimental configuration for each state-of-the-art model is identical to their original implementations. To measure interpolation accuracy, we employ the standard $\ell_2$ positional distance (L2P), $\ell_2$ quaternion difference (L2Q), and NPSS [16] for visual similarity evaluation [18]. + +
Motion categoryMethodL2P↓L2Q↓NPSS↓
515305153051530
BasketballLERP0.07870.40090.80330.15880.58340.96110.13150.61201.8202
TGcomplete [18]0.10500.44190.71040.18590.64390.93030.16000.65791.2136
BERT [12]0.08890.37370.68650.15850.55600.89920.16890.61771.4587
CITL [34]0.06250.29730.57080.14100.49320.80110.10470.40981.0959
PC-MRL w/o RPA0.07670.35340.67150.15050.54310.90490.13640.55961.3471
PC-MRL (Ours)0.07520.34430.66620.15610.54950.90880.13970.55721.3015
GolfLERP0.01890.12290.31870.04550.18700.39870.05870.39791.0312
TGcomplete [18]0.04730.26160.45460.06930.31910.52970.09540.50621.1534
BERT [12]0.02760.11590.29670.06120.20820.42580.10990.39060.9255
CITL [34]0.02010.10430.25890.04760.17660.35520.06360.25810.7799
PC-MRL w/out RPA0.02740.13040.36220.05550.20650.49230.09110.37591.2083
PC-MRL (Ours)0.02730.11300.27060.05980.20120.38440.10490.37160.8150
SwimmingLERP0.07310.36800.73740.13260.53270.94760.21470.92221.6972
TGcomplete [18]0.12920.57960.91410.18450.74201.11990.26600.97161.7698
BERT [12]0.10880.38960.75160.18190.55180.96220.28770.90741.6682
CITL [34]0.09050.38950.74750.14840.55350.95330.20840.83171.6122
PC-MRL w/o RPA0.07220.37540.74590.13450.54420.95940.19440.87311.6608
PC-MRL (Ours)0.07340.35950.71390.13550.53310.92480.19560.85911.6370
Walking & RunningLERP0.04550.25050.55270.11200.39970.62330.06550.26820.9071
TGcomplete [18]0.05400.17860.31030.11470.28660.38660.08950.26160.4265
BERT [12]0.05520.19980.35440.11110.31010.47620.09010.27600.6458
CITL [34]0.03280.10340.24270.09420.21020.33840.05260.15680.4080
PC-MRL w/o RPA0.03970.16570.34170.10580.28790.47220.06400.22880.6021
PC-MRL (Ours)0.03770.13920.28480.10640.27780.42810.06580.21290.5098
+ +Table 1: Performance comparisons for various motion categories, measured in L2P, L2Q, and NPSS. All methods except PC-MRL are trained on the original CMU MoCap dataset. Each motion contains 128 frames with keyframes placed every 5, 15, or 30 frames. The top 2 results for each test are highlighted in purple and blue respectively. + +The results listed in Table 1 clearly demonstrate our method's ability to approach the accuracy levels exhibited by directly supervised state-of-the-art methods. PC-MRL consistently outperforms both the LERP standard and RNN-based $\mathrm{TG}_{\mathrm{complete}}$ model, particularly in longer keyframe interval scenarios where the quality and depth of learned motion features is most critical and impactful. Likewise, the inclusion of RPA also tends to improve the long interval performance of PC-MRL, particularly when precise movements (i.e. golf) or deeper motion features (i.e. locomotion) are expected. Given this, we can observe that RPA definitively improves the overall consistency of the PC-MRL method in scenarios that may be unseen or less effectively represented by our point cloud-to-motion system. + +Native supervision, i.e. with CITL, unsurprisingly produces the most accurate interpolation model, with a notable exception of swimming motions, where our PC-MRL supervision produces even stronger results. We believe the abundance of low crawl motions, a similar motion to swimming in LaFAN1, provides more meaningful supervision over the original CMU dataset, which conversely has few similar motions once swimming motions are filtered out. On the other end of the spectrum, i.e. for high data scenarios such as walking and running locomotion from Fig. 5, all models including our PC-MRL approach observe the strongest improvements for learned methods over the naive LERP method. Both cases strongly support our method's capability to supervise highly intricate motion patterns, despite the incompleteness of the point cloud representation and reliance on relative rotations and geometric optimisation assumptions. + +![](images/5bf89b54c8e79c1caf8b33c94c9efab4976a017e1a19a022bbde137256c2059a.jpg) +Fig. 5: Walking keyframe interpolation examples produced by methods from Table 1. Close-up layered comparisons against the original motion are provided on the right. + +# 4.4 Cross-Skeleton Motion Re-Targeting Experiments + +We further conduct an experiment on motion re-targeting to directly benchmark our method against the sole existing state-of-the-art method for unpaired motion re-targeting with unrestricted skeletons, i.e. primal skeletons (PS) [1]. For direct motion re-targeting evaluation, we utilized the KNN-Loss and end-effector loss as metrics, owing to their effective measurement of visual similarity and non-sensitivity towards absolute roll axis correctness. Though absolute rolls are generally necessary for a complete motion re-targeting solution, this aspect is out of scope for our project as our main goal is motion interpolation. + +Table 2 indicates the superior performance of our method over the state-of-the-art PS method in terms of visual similarity. At low $\sigma$ values, the higher relative $\mathcal{L}_{\mathrm{KNN}}$ of the primal skeleton method underscores its geometric deviations + +
Model\( L_{\text{KNN}}(\mathcal{X}_A,\mathcal{X}_B) \downarrow \)\( L_{\text{end}} \downarrow \)
\( \sigma = 0 \)\( \sigma = 0.01 \)\( \sigma = 0.05 \)\( \sigma = 0.1 \)\( \sigma = 0.2 \)
Original motion1.00992.02345.73999.663616.92230
Primal skeleton [1]8.11098.03539.420612.829120.63329.6990
PC-MRL (Ours)3.87144.12316.727710.307817.52615.6258
+ +Table 2: Average unpaired LaFAN1 to CMU motion re-targeting performance of the primal skeleton method and our point cloud method. $\mathcal{L}_{\mathrm{KNN}}$ is measured using 1024-point clouds at $K = 8$ . Independently sampled point clouds from the original LaFAN1 skeleton are provided as optimal $\mathcal{L}_{\mathrm{KNN}}$ reference. All values are scaled by $\times 100$ . + +![](images/4d71934c2946edcc88669cce7457ddcefb2a49921e2cbd04202c403bdc4878ed.jpg) +Fig. 6: Jumping motion re-targeting examples produced on the CMU skeleton by the primal skeleton method (middle row) and our method (bottom row). Sampled point cloud representations of each motion sequence are provided on the right. + +from the original motion. In contrast, our method demonstrates significantly better visual adherence to the original geometry. As $\sigma$ increases, our method approaches near-perfect alignment with the original motion. Figure 6 visually indicates this, showing our method follows the original motion closely, while the latent consistency-based primal skeleton method struggles to generalise and decode the skeleton's internal structures without directly supervised objectives, such as the end-effector positions provided during training. In addition, Fig 7 demonstrates that, like existing re-targeting approaches, our method is able to produce adequate results on disproportionate skeletons. + +Since our approach is largely focused on enabling non-native motion interpolation supervision, we additionally compare our PC-MRL method against a CITL model supervised by PS in Table 3. Due to the suboptimal re-targeting performance of PS, such interpolation models exhibit expectedly poor accuracy. + +![](images/22c90d7f153bf66827f9df7e59a0f616eb3eed7b990ed10dbb5963987623cddf.jpg) +Fig. 7: Re-targeted falling motion from LaFAN1 (top) on the CMU skeleton (middle) and a short custom skeleton (bottom), performed by our point cloud-to-motion model. + +
MethodL2P↓L2Q↓NPSS↓
515305153051530
CITL - PS0.20430.63321.15240.23180.74281.14420.11220.42761.4622
CITL - Native0.04950.23060.48980.09790.36040.66090.11490.46841.1591
PC-MRL0.04940.23370.49600.10300.37710.68130.13700.54331.1760
+ +Table 3: Average L2P, L2Q, and NPSS performance of CITL trained on native and PS-generated datasets, compared to our method. + +# 5 Limitations + +Due to the inherent limitations of the point cloud representation, certain constraints on the applicability of our point cloud-to-motion method are inevitable. Primarily, the reliance of our point cloud re-targeting method on the relative nature of First-frame Offset Quaternions (FOQ) for most motion learning tasks necessitates the presence of at least one known pose. In the absence of a known pose, relative quaternions are incapable of reconstructing absolute rotational values, which are essential for motion applications. Secondly, due to our method of sampling point clouds with a consistent deviation from each bone segment, the resultant representation inevitably obscures finer details. This includes elements like fingers and facial controls, to an extent that is beyond rectification. + +Our proposed objective function relies entirely on geometric segments, necessitating that all bones possess a non-zero length. This stipulation is deemed a reasonable prerequisite for generic human skeletons, as each bone is typically responsible for manipulating a discernible segment of the human body. Due to technical considerations, the skeletal configurations in each of our datasets included at least one bone of zero length, which we ignored in our experiments in favour of global quaternion rotations of all bones with non-zero length. + +As a result of these limitations, we refrained from claiming state-of-the-art performance for general motion re-targeting tasks. + +# 6 Conclusion + +This paper introduces a novel learning-based motion interpolation method designed to enable cross-skeleton compatibility with existing motion datasets. Our proposed method PC-MRL employs a process of point cloud obfuscation and skeleton reconstruction. The point cloud space represents human motions in a non-hierarchical and skeleton-agnostic manner. It enables a KNN-based objective to be optimised without dataset supervision, guiding a neural network to generate high-quality motion features with any target skeleton. We address the rotational information loss of our point cloud format by presenting an offset quaternion strategy compatible with concurrent transformer-based models. Through extensive experiments, we have demonstrated the efficacy of PC-MRL in performing motion interpolation without relying on native motion data. Moreover, PC-MRL has achieved superior visual similarity metrics in the domain of motion re-targeting. We concluded our work by discussing the limitations of PC-MRL. + +# References + +1. Aberman, K., Li, P., Lischinski, D., Sorkine-Hornung, O., Cohen-Or, D., Chen, B.: Skeleton-aware networks for deep motion retargeting 39(4), 62–1 (2020) 4, 12 +2. Aberman, K., Wu, R., Lischinski, D., Chen, B., Cohen-Or, D.: Learning character-agnostic motion for motion retargeting in 2d 38(4), 1-14 (jul 2019) 4 +3. Annabi, L., Ma, Z., et al.: Unsupervised motion retargeting for human-robot imitation. In: Companion of the 2024 ACM/IEEE International Conference on Human-Robot Interactio. ACM (2024) 4 +4. Bodenheimer, B., Rose, C., Rosenthal, S., Pella, J.: The process of motion capture: Dealing with the data. In: Computer Animation and Simulation. pp. 3-18. Springer (1997) 4 +5. Burtnyk, N., Wein, M.: Computer-generated key-frame animation. Journal of the SMPTE 80(3), 149-153 (1971) 1, 4 +6. Catalin Ionescu, Fuxin Li, C.S.: Latent structured models for human pose estimation. In: International Conference on Computer Vision (2011) 10 +7. Chen, H., Tang, H., Timofte, R., Gool, L.V., Zhao, G.: Lart: Neural correspondence learning with latent regularization transformer for 3d motion transfer. Advances in Neural Information Processing Systems 36 (2024) 4 +8. Chen, J., Li, C., Lee, G.H.: Weakly-supervised 3D pose transfer with keypoints. In: ICCV. pp. 15156-15165 (2023) 4 +9. Choi, K.J., Ko, H.S.: Online motion retargetting. The Journal of Visualization and Computer Animation 11(5), 223-235 (2000) 5 +0. CMU: Carnegie-mellon university motion capture database (2003), http://mocap.cs.cmu.edu/ 10 +1. Dam, E.B., Koch, M., Lillholm, M.: Quaternions, interpolation and animation, vol. 2. CiteSeer (1998) 5, 6 +2. Duan, Y., Lin, Y., Zou, Z., Yuan, Y., Qian, Z., Zhang, B.: A unified framework for real time motion completion. In: AAAI. vol. 36, pp. 4459-4467 (2022) 2, 4, 10, 11 +3. Frank, T., Johnston, O.: Disney Animation: The Illusion of Life. Abbeville Publishing Group, New York (1981) 1 +4. Gleicher, M.: Retargetting motion to new characters. In: Computer Graphics and Interactive Techniques. pp. 33-42 (1998) 4 +5. Gleicher, M.: Animation from observation: Motion capture and motion editing 33(4), 51-54 (1999) 4 +6. Gopalakrishnan, A., Mali, A., Kifer, D., Giles, L., Ororbia, A.G.: A neural temporal model for human motion prediction. In: CVPR. pp. 12116-12125 (2019) 10 +7. Harvey, F.G., Pal, C.: Recurrent transition networks for character locomotion. In: SIGGRAPH Asia 2018 Technical Briefs, pp. 1-4 (2018) 4 +8. Harvey, F.G., Yurick, M., Nowrouzezahrai, D., Pal, C.: Robust motion in betweening 39(4), 60-1 (2020) 2, 4, 8, 9, 10, 11 +9. He, C., Saito, J., Zachary, J., Rushmeier, H., Zhou, Y.: Nemf: Neural motion fields for kinematic animation 35, 4244-4256 (2022) 4 +20. Hecker, C., Raabe, B., Enslow, R.W., DeWeese, J., Maynard, J., van Prooijen, K.: Real-time motion retargeting to highly varied user-created morphologies 27(3), 1-11 (2008) 5 +21. Hernandez, A., Gall, J., Moreno-Noguer, F.: Human motion prediction via spatiotemporal inpainting. In: ICCV. pp. 7134-7143 (2019) 4 +22. Ionescu, C., Papava, D., Olaru, V., Sminchisescu, C.: Human3.6m: Large scale datasets and predictive methods for 3D human sensing in natural environments. + +IEEE Transactions on Pattern Analysis and Machine Intelligence 36(7), 1325-1339 (jul 2014) 10 +23. Jiang, H., Cai, J., Zheng, J.: Skeleton-aware 3D human shape reconstruction from point clouds. In: ICCV. pp. 5431-5441 (2019) 4 +24. Kantor, I.L., Solodovnikov, A.S., Shenitzer, A.: Hypercomplex numbers: an elementary introduction to algebras, vol. 302. Springer (1989) 6 +25. Karunratanakul, K., Preechakul, K., Suwajanakorn, S., Tang, S.: Guided motion diffusion for controllable human motion synthesis. In: ICCV. pp. 2151-2162 (2023) 4 +26. Kulpa, R., Multon, F., Arnaldi, B.: Morphology-independent representation of motions for interactive human-like animation. In: Eurographics (2005) 5 +27. Lehrmann, A.M., Gehler, P.V., Nowozin, S.: Efficient nonlinear markov models for human motion. In: CVPR. pp. 1314-1321 (2014) 4 +28. Li, P., Aberman, K., Zhang, Z., Hanocka, R., Sorkine-Hornung, O.: Ganimator: Neural motion synthesis from a single sequence 41(4), 1-12 (2022) 4 +29. Li, W., Wang, Z., Mai, R., Ren, P., Zhang, Q., Zhou, Y., Xu, N., Zhuang, J., Xin, B., Gao, L., et al.: Modular design automation of the morphologies, controllers, and vision systems for intelligent robots: a survey. Visual Intelligence 1(1), 2 (2023) 4 +30. Li, Z., Peng, X.B., Abbeel, P., Levine, S., Berseth, G., Sreenath, K.: Reinforcement learning for versatile, dynamic, and robust bipedal locomotion control. arXiv preprint arXiv:2401.16889 (2024) 4 +31. Liao, Z., Yang, J., Saito, J., Pons-Moll, G., Zhou, Y.: Skeleton-free pose transfer for stylized 3D characters. In: ECCV. pp. 640-656. Springer (2022) 4 +32. Makoviychuk, V., Wawrzyniak, L., Guo, Y., Lu, M., Storey, K., Macklin, M., Hoeller, D., Rudin, N., Allshire, A., Handa, A., et al.: Isaac gym: High performancegpu-based physics simulation for robot learning. arXiv preprint arXiv:2108.10470 (2021) 4 +33. Merry, B., Marais, P., Gain, J.: Animation space: A truly linear framework for character animation. ACM Transactions on Graphics (TOG) 25(4), 1400-1423 (2006) 5 +34. Mo, C.A., Hu, K., Long, C., Wang, Z.: Continuous intermediate token learning with implicit motion manifold for keyframe based motion interpolation. In: CVPR. pp. 13894-13903 (2023) 2, 4, 8, 10, 11 +35. Mo, C.A., Hu, K., Mei, S., Chen, Z., Wang, Z.: Keyframe extraction from motion capture sequences with graph based deep reinforcement learning. In: ACM International Conference on Multimedia. pp. 5194-5202 (2021) 4 +36. Monzani, J.S., Baerlocher, P., Boulic, R., Thalmann, D.: Using an intermediate skeleton and inverse kinematics for motion retargeting. In: Computer Graphics Forum. vol. 19, pp. 11-19. Wiley Online Library (2000) 5 +37. Oreshkin, B.N., Valkanas, A., Harvey, F.G., Menard, L.S., Bocquelet, F., Coates, M.J.: Motion inbetweening via deep $\delta$ -interpolator (2022) 2, 4 +38. Peng, X.B., Abbeel, P., Levine, S., van de Panne, M.: Deepmimic: Example-guided deep reinforcement learning of physics-based character skills 37(4), 143:1-143:14 (Jul 2018) 4 +39. Peng, X.B., Ma, Z., Abbeel, P., Levine, S., Kanazawa, A.: Amp: Adversarial motion priors for stylized physics-based character control 40(4) (Jul 2021) 4 +40. Qin, J., Zheng, Y., Zhou, K.: Motion in-between via two-stage transformers 41(6), 1-16 (2022) 2 +41. Reda, D., Won, J., Ye, Y., van de Panne, M., Winkler, A.: Physics-based motion retargeting from sparse inputs. ACM Computer Graphics and Interactive Techniques 6(3), 1-19 (2023) 4 + +42. Reeves, W.T.: Inbetweening for computer animation utilizing moving point constraints 15(3), 263-269 (1981) 1 +43. Ren, T., Yu, J., Guo, S., Ma, Y., Ouyang, Y., Zeng, Z., Zhang, Y., Qin, Y.: Diverse motion in-betweening from sparse keyframes with dual posture stitching. IEEE Transactions on Visualization and Computer Graphics (2024) 4 +44. Shaw, P., Uszkoreit, J., Vaswani, A.: Self-attention with relative position representations. In: NAACL. Association for Computational Linguistics (2018) 9 +45. Tiwari, G., Antic, D., Lenssen, J.E., Sarafianos, N., Tung, T., Pons-Moll, G.: Posendf: Modeling human pose manifolds with neural distance fields. In: ECCV. pp. 572-589. Springer (2022) 4 +46. Villegas, R., Yang, J., Ceylan, D., Lee, H.: Neural kinematic networks for unsupervised motion retargeting. In: CVPR. pp. 8639-8648 (2018) 4 +47. Wang, J., Li, X., Liu, S., De Mello, S., Gallo, O., Wang, X., Kautz, J.: Zero-shot pose transfer for unrigged stylized 3D characters. In: CVPR. pp. 8704-8714 (2023) 4 +48. Wang, K., Xie, J., Zhang, G., Liu, L., Yang, J.: Sequential 3D human pose and shape estimation from point clouds. In: CVPR. pp. 7275-7284 (2020) 4 +49. Yuan, Y., Song, J., Iqbal, U., Vahdat, A., Kautz, J.: Physdiff: Physics-guided human motion diffusion model. In: ICCV. pp. 16010-16021 (2023) 4 +50. Yuan, Y., Wei, S.E., Simon, T., Kitani, K., Saragih, J.: Simpoe: Simulated character control for 3D human pose estimation. In: CVPR. pp. 7159-7169 (2021) 4 +51. Zhang, J., Weng, J., Kang, D., Zhao, F., Huang, S., Zhe, X., Bao, L., Shan, Y., Wang, J., Tu, Z.: Skinned motion retargeting with residual perception of motion semantics & geometry. In: CVPR. pp. 13864-13872 (2023) 4 +52. Zhang, M., Li, H., Cai, Z., Ren, J., Yang, L., Liu, Z.: Finemogen: Fine-grained spatio-temporal motion generation and editing. Advances in Neural Information Processing Systems 36 (2024) 4 +53. Zhang, X., van de Panne, M.: Data-driven autocompletion for keyframe animation. In: ACM SIGGRAPH Conference on Motion, Interaction and Games. pp. 1-11 (2018) 4 +54. Zhao, H., Jiang, L., Jia, J., Torr, P.H., Koltun, V.: Point transformer. In: ICCV. pp. 16259-16268 (2021) 7, 10 +55. Zhu, W., Yang, Z., Di, Z., Wu, W., Wang, Y., Loy, C.C.: Mocanet: Motion retargeting in-the-wild via canonicalization networks. In: AAAI. vol. 36, pp. 3617-3625 (2022) 4 \ No newline at end of file diff --git a/2024/Motion Keyframe Interpolation for Any Human Skeleton using Point Cloud-based Human Motion Data Homogenisation/images.zip b/2024/Motion Keyframe Interpolation for Any Human Skeleton using Point Cloud-based Human Motion Data Homogenisation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..1a0ea9668bb9a72f5fe9d82c665691c22b02006b --- /dev/null +++ b/2024/Motion Keyframe Interpolation for Any Human Skeleton using Point Cloud-based Human Motion Data Homogenisation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f33e8b8893dee3a93be5dac335e00446af47660fff3b3700a27aa5b6c6e85a33 +size 565290 diff --git a/2024/Motion Keyframe Interpolation for Any Human Skeleton using Point Cloud-based Human Motion Data Homogenisation/layout.json b/2024/Motion Keyframe Interpolation for Any Human Skeleton using Point Cloud-based Human Motion Data Homogenisation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..437355d4bde8eeb73b133bacfbd0b45de4db7808 --- /dev/null +++ b/2024/Motion Keyframe Interpolation for Any Human Skeleton using Point Cloud-based Human Motion Data Homogenisation/layout.json @@ -0,0 +1,9958 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 138, + 111, + 477, + 166 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 111, + 477, + 166 + ], + "spans": [ + { + "bbox": [ + 138, + 111, + 477, + 166 + ], + "type": "text", + "content": "Motion Keyframe Interpolation for Any Human Skeleton via Temporally Consistent Point Cloud Sampling and Reconstruction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 132, + 185, + 479, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 185, + 479, + 200 + ], + "spans": [ + { + "bbox": [ + 132, + 185, + 479, + 200 + ], + "type": "text", + "content": "Clinton Mo" + }, + { + "bbox": [ + 132, + 185, + 479, + 200 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 132, + 185, + 479, + 200 + ], + "type": "text", + "content": ", Kun Hu" + }, + { + "bbox": [ + 132, + 185, + 479, + 200 + ], + "type": "inline_equation", + "content": "^{1,\\star}" + }, + { + "bbox": [ + 132, + 185, + 479, + 200 + ], + "type": "text", + "content": ", Chengjiang Long" + }, + { + "bbox": [ + 132, + 185, + 479, + 200 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 132, + 185, + 479, + 200 + ], + "type": "text", + "content": ", Dong Yuan" + }, + { + "bbox": [ + 132, + 185, + 479, + 200 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 132, + 185, + 479, + 200 + ], + "type": "text", + "content": ", and Zhiyong Wang" + }, + { + "bbox": [ + 132, + 185, + 479, + 200 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 180, + 207, + 432, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 180, + 207, + 432, + 240 + ], + "spans": [ + { + "bbox": [ + 180, + 207, + 432, + 240 + ], + "type": "text", + "content": "1 The University of Sydney, Darlington NSW 2008, Australia \nclmo6615@uni.sydney.edu.au, {kun.hu, zhiyong.wang, dong.yuan}@sydney.edu.au" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 215, + 240, + 397, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 215, + 240, + 397, + 252 + ], + "spans": [ + { + "bbox": [ + 215, + 240, + 397, + 252 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 215, + 240, + 397, + 252 + ], + "type": "text", + "content": " Meta Reality Labs, Burlingame, CA, USA" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 271, + 253, + 343, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 271, + 253, + 343, + 262 + ], + "spans": [ + { + "bbox": [ + 271, + 253, + 343, + 262 + ], + "type": "text", + "content": "clong1@meta.com" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 159, + 290, + 455, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 290, + 455, + 510 + ], + "spans": [ + { + "bbox": [ + 159, + 290, + 455, + 510 + ], + "type": "text", + "content": "Abstract. In the character animation field, modern supervised keyframe interpolation models have demonstrated exceptional performance in constructing natural human motions from sparse pose definitions. As supervised models, large motion datasets are necessary to facilitate the learning process; however, since motion is represented with fixed hierarchical skeletons, such datasets are incompatible for skeletons outside the datasets' native configurations. Consequently, the expected availability of a motion dataset for desired skeletons severely hinders the feasibility of learned interpolation in practice. To combat this limitation, we propose Point Cloud-based Motion Representation Learning (PC-MRL), an unsupervised approach to enabling cross-compatibility between skeletons for motion interpolation learning. PC-MRL consists of a skeleton obfuscation strategy using temporal point cloud sampling, and an unsupervised skeleton reconstruction method from point clouds. We devise a temporal point-wise K-nearest neighbors loss for unsupervised learning. Moreover, we propose First-frame Offset Quaternion (FOQ) and Rest Pose Augmentation (RPA) strategies to overcome necessary limitations of our unsupervised point cloud-to-skeletal motion process. Comprehensive experiments demonstrate the effectiveness of PC-MRL in motion interpolation for desired skeletons without supervision from native datasets." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 160, + 520, + 449, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 520, + 449, + 533 + ], + "spans": [ + { + "bbox": [ + 160, + 520, + 449, + 533 + ], + "type": "text", + "content": "Keywords: 3D point clouds " + }, + { + "bbox": [ + 160, + 520, + 449, + 533 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 520, + 449, + 533 + ], + "type": "text", + "content": " Human body motion " + }, + { + "bbox": [ + 160, + 520, + 449, + 533 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 520, + 449, + 533 + ], + "type": "text", + "content": " Dataset creation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 551, + 230, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 551, + 230, + 563 + ], + "spans": [ + { + "bbox": [ + 132, + 551, + 230, + 563 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 575, + 482, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 575, + 482, + 648 + ], + "spans": [ + { + "bbox": [ + 130, + 575, + 482, + 648 + ], + "type": "text", + "content": "3D character animation workflows largely rely on the concepts of keyframing with interpolation, often referred to as the pose-to-pose principle [13]. By defining key poses at correct timings, algorithms can be employed to generate intermediate poses, and thereby eliminating the need for defining each frame individually [5, 42]. Although this keyframe-based workflow is less costly than frame-by-frame production, human motion is often complex, and natural depictions" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 134, + 654, + 236, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 654, + 236, + 666 + ], + "spans": [ + { + "bbox": [ + 134, + 654, + 236, + 666 + ], + "type": "text", + "content": "* Corresponding author." + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 164, + 119, + 239, + 204 + ], + "blocks": [ + { + "bbox": [ + 164, + 119, + 239, + 204 + ], + "lines": [ + { + "bbox": [ + 164, + 119, + 239, + 204 + ], + "spans": [ + { + "bbox": [ + 164, + 119, + 239, + 204 + ], + "type": "image", + "image_path": "d2218860b0ef14c222532445ddae55cd3415e56b8bcf0d28f405c5ce4aab6c0c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 157, + 207, + 243, + 217 + ], + "lines": [ + { + "bbox": [ + 157, + 207, + 243, + 217 + ], + "spans": [ + { + "bbox": [ + 157, + 207, + 243, + 217 + ], + "type": "text", + "content": "(a) Linear interpolation" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 272, + 119, + 343, + 204 + ], + "blocks": [ + { + "bbox": [ + 272, + 119, + 343, + 204 + ], + "lines": [ + { + "bbox": [ + 272, + 119, + 343, + 204 + ], + "spans": [ + { + "bbox": [ + 272, + 119, + 343, + 204 + ], + "type": "image", + "image_path": "e3bfa326343bc12db7c8e9e70f1524810ef7753ef624ffd77c7e09b16cfd69c3.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 257, + 206, + 356, + 217 + ], + "lines": [ + { + "bbox": [ + 257, + 206, + 356, + 217 + ], + "spans": [ + { + "bbox": [ + 257, + 206, + 356, + 217 + ], + "type": "text", + "content": "(b) PC-MRL (Our method)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 378, + 119, + 450, + 204 + ], + "blocks": [ + { + "bbox": [ + 378, + 119, + 450, + 204 + ], + "lines": [ + { + "bbox": [ + 378, + 119, + 450, + 204 + ], + "spans": [ + { + "bbox": [ + 378, + 119, + 450, + 204 + ], + "type": "image", + "image_path": "44d2a55e7203e40d15ac1c42294ae3da0256fcc101b9cd092f1004ab6aa6faa2.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 382, + 207, + 445, + 217 + ], + "lines": [ + { + "bbox": [ + 382, + 207, + 445, + 217 + ], + "spans": [ + { + "bbox": [ + 382, + 207, + 445, + 217 + ], + "type": "text", + "content": "(c) Ground truth" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 181, + 254, + 220, + 331 + ], + "blocks": [ + { + "bbox": [ + 181, + 254, + 220, + 331 + ], + "lines": [ + { + "bbox": [ + 181, + 254, + 220, + 331 + ], + "spans": [ + { + "bbox": [ + 181, + 254, + 220, + 331 + ], + "type": "image", + "image_path": "ee25c25af7037e509bc3473cb4d65ecefc982ec85197badae4aac3be6468f142.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 156, + 336, + 244, + 346 + ], + "lines": [ + { + "bbox": [ + 156, + 336, + 244, + 346 + ], + "spans": [ + { + "bbox": [ + 156, + 336, + 244, + 346 + ], + "type": "text", + "content": "(a) Hierarchical skeleton" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 272, + 252, + 340, + 331 + ], + "blocks": [ + { + "bbox": [ + 146, + 225, + 465, + 237 + ], + "lines": [ + { + "bbox": [ + 146, + 225, + 465, + 237 + ], + "spans": [ + { + "bbox": [ + 146, + 225, + 465, + 237 + ], + "type": "text", + "content": "Fig. 1: Motion interpolation of walking keyframes conducted by our approach." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 272, + 252, + 340, + 331 + ], + "lines": [ + { + "bbox": [ + 272, + 252, + 340, + 331 + ], + "spans": [ + { + "bbox": [ + 272, + 252, + 340, + 331 + ], + "type": "image", + "image_path": "61f3a519e8e4097725bad99f609add4adf90d3642355878e3fb74294765cbaf4.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 272, + 336, + 342, + 346 + ], + "lines": [ + { + "bbox": [ + 272, + 336, + 342, + 346 + ], + "spans": [ + { + "bbox": [ + 272, + 336, + 342, + 346 + ], + "type": "text", + "content": "(b) Skeletal motion" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 378, + 253, + 449, + 331 + ], + "blocks": [ + { + "bbox": [ + 378, + 253, + 449, + 331 + ], + "lines": [ + { + "bbox": [ + 378, + 253, + 449, + 331 + ], + "spans": [ + { + "bbox": [ + 378, + 253, + 449, + 331 + ], + "type": "image", + "image_path": "f908a2abf1d811ef7f297dd1286a831a856acd18723f2f794f65c2e56c3db689.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 361, + 336, + 466, + 346 + ], + "lines": [ + { + "bbox": [ + 361, + 336, + 466, + 346 + ], + "spans": [ + { + "bbox": [ + 361, + 336, + 466, + 346 + ], + "type": "text", + "content": "(c) Point cloud representation" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 355, + 482, + 389 + ], + "lines": [ + { + "bbox": [ + 130, + 355, + 482, + 389 + ], + "spans": [ + { + "bbox": [ + 130, + 355, + 482, + 389 + ], + "type": "text", + "content": "Fig. 2: Our method leverages point cloud representations (c) to obfuscate the hierarchical dependencies of skeletal motion. Joint colours correspond to ordinal indices in original motion data representation. Point clouds are coloured by temporal index." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 413, + 480, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 413, + 480, + 473 + ], + "spans": [ + { + "bbox": [ + 130, + 413, + 480, + 473 + ], + "type": "text", + "content": "demand a significant number of keyframes, especially for realistic motions that adhere to physical properties and constraints. Although motion capture (MoCap) is frequently employed as an alternative for achieving realistic human motions, MoCap often nonetheless necessitates subsequent keyframing to correct unwanted motion elements and address other imperfections." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 474, + 481, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 481, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 481, + 616 + ], + "type": "text", + "content": "In recent years, the interpolation process has been widely studied, particularly with the advent of machine learning methods for sequential data. Compared to conventional methods such as linear interpolation (LERP), machine learning methods have demonstrated the capability to derive visually natural motions from sparser keyframe sets [12, 18, 34, 37, 40], which we can observe in Fig. 1. Data-driven interpolation approaches require large motion capture datasets to learn the necessary motion features for effective synthesis of transitions between keyframes. In the field of 3D motions, this aspect significantly restricts the feasibility of learned interpolation methods in animation practice, as motion/pose data are tied to specific skeletal configurations and are not cross-compatible. That is, the dataset's native skeleton, i.e. source skeleton, is structurally different and incompatible with the desired skeleton, i.e. target skeleton." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "content": "Therefore, we propose a novel point cloud-based motion representation learning (PC-MRL) approach, in which skeletal hierarchies are obfuscated by sampling into the point cloud medium, and reconstructed into skeletal motion data using an unsupervised neural network. Unlike raw pose and motion data, the" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "C. Mo et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "text", + "content": "point cloud format is a non-hierarchical representation, and effectively obscures any skeletal configuration as displayed in Fig. 2. PC-MRL samples a set of points around the rest position of skeletons using a combination of uniform and normal distributions, and geometrically parents each point to an associated bone in order to effectively represent the associated skeletons' motion data in a temporally consistent manner. Subsequently, for the reconstruction of motion data into a given desired skeleton, we propose an unsupervised learning scheme, featuring a K-nearest neighbors (KNN) loss between cross-skeleton point clouds to optimize a transformer neural network. The network processes the point cloud-based motion representations and generates the corresponding target skeletal motion representations. By obscuring both source and predicted target motion data into the point cloud space, our strategy allows for a direct comparison, utilizing KNN to minimize the visual difference between the two point clouds." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 272, + 482, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 272, + 482, + 450 + ], + "spans": [ + { + "bbox": [ + 130, + 272, + 482, + 450 + ], + "type": "text", + "content": "The unsupervised nature of our target skeleton reconstruction scheme creates two main disparities between the expected and predicted motion features. Firstly, the exact roll rotation axis of bones cannot be represented in the point cloud format. To this end, we introduce First-frame Offset Quaternion (FOQ) representations to incorporate relative roll values, which are obtainable from temporally consistent point clouds. This approach is agnostic to absolute roll rotations, providing a standardisation mechanism for the resulting motion data within the context of motion interpolation learning. Second, KNN-based objectives learn skeletal representations in a geometrically optimal manner, which does not always reflect the desired skeletal behaviour. This disparity is commonly observed with smaller skeletal features, such as shoulder and hip bones. As such, we augment training motions using Rest Pose Augmentation (RPA) to increase the range of skeletal behaviours that our interpolation model learns. Extensive experiments demonstrate the effectiveness of our proposed method, achieving performance targets near the level of direct dataset supervision." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 146, + 451, + 430, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 451, + 430, + 462 + ], + "spans": [ + { + "bbox": [ + 146, + 451, + 430, + 462 + ], + "type": "text", + "content": "In summary, this paper presents the following key contributions:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 135, + 469, + 479, + 574 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 136, + 469, + 479, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 469, + 479, + 502 + ], + "spans": [ + { + "bbox": [ + 136, + 469, + 479, + 502 + ], + "type": "text", + "content": "1. We propose a novel method for achieving unsupervised human motion reconstruction from point cloud data, enabling skeleton-agnostic motion interpolation learning for the first time." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 135, + 505, + 479, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 505, + 479, + 538 + ], + "spans": [ + { + "bbox": [ + 135, + 505, + 479, + 538 + ], + "type": "text", + "content": "2. To train motion interpolation models, we formulate first-frame offset quaternions to represent bone rotations with relative roll data, as well as a rest position augmentation strategy to address skeletal configuration variability." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 135, + 540, + 479, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 540, + 479, + 574 + ], + "spans": [ + { + "bbox": [ + 135, + 540, + 479, + 574 + ], + "type": "text", + "content": "3. We perform comprehensive experiments to demonstrate our method's effectiveness towards learning motion interpolation, despite the absence of directly compatible datasets in the training process." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 594, + 240, + 605 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 594, + 240, + 605 + ], + "spans": [ + { + "bbox": [ + 132, + 594, + 240, + 605 + ], + "type": "text", + "content": "2 Related works" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "We explore existing research on motion interpolation approaches, and motion data modelling methods in general. In addition, our proposed approach bears resemblance to unsupervised approaches for the motion re-targeting research problem, and as such, we will also analyse existing methods in this field." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 215, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 215, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 215, + 91, + 447, + 102 + ], + "type": "text", + "content": "Motion Keyframe Interpolation for Any Human Skeleton" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 365, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 365, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 365, + 128 + ], + "type": "text", + "content": "2.1 Data-Driven Motion Modelling Methods" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 137, + 481, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 137, + 481, + 232 + ], + "spans": [ + { + "bbox": [ + 130, + 137, + 481, + 232 + ], + "type": "text", + "content": "Contemporary research in motion modelling predominantly explore learned data-driven methods, using neural networks to capture, model, and establish correlations between the motion features of various skeletal configurations [29, 35, 41, 51, 55]. Motion features alone have demonstrated considerable efficacy in cross-skeleton imitation using a variety of strategies, including task-driven objectives [19, 45], diffusion-based generation [25, 49, 52], and latent feature consistency [1-3, 7, 46]. Latent consistency in particular learns inter-skeleton correlations without the need for paired datasets, which reduces its barrier of entry." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 232, + 482, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 232, + 482, + 351 + ], + "spans": [ + { + "bbox": [ + 130, + 232, + 482, + 351 + ], + "type": "text", + "content": "Likewise, motion interpolation strategies have evolved through the use of learning based approaches. The inherent numerical precision of motion data has traditionally posed a challenge for the adoption of the learning approaches, due to their approximate nature [21,27]. This challenge is further amplified by the temporally sparse distribution of keyframes. Recently, a number of recurrent neural network based methods [17, 18, 43, 53] and transformer-based networks [12, 34, 37] have demonstrated encouraging interpolation performance for the transition between keyframes. Additionally, various other motion modelling methods are capable for keyframe based interpolation, albeit with limited intermediate frame representation capabilities [19, 25, 45]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 352, + 482, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 352, + 482, + 519 + ], + "spans": [ + { + "bbox": [ + 130, + 352, + 482, + 519 + ], + "type": "text", + "content": "A major constraint inherent in all data-driven motion modelling methods is their dependency on datasets featuring specific skeletal configurations. To address this, skeleton-free training strategies have been explored for elementary tasks involving 3D characters. These include vertex-based pose transfer [7,8,31,47], and point cloud-based human shape reconstruction [23,48]. Notably, in these methods, 3D mesh and point cloud coordinates serve as general 3D data representations. Concurrently, other research has significantly minimized the required volume of training data, bringing it down to single example sequences. This reduction has been achieved through the use of patch matching techniques [28] and imitation learning within physics-based simulations [30,39]. Reinforcement learning has facilitated the generation of goal-driven motion without pre-existing datasets [32,38,50]. However, these simulation-based methods have typically been either exceedingly complex to implement in practice, or yield results that are too inflexible and/or prone to errors for animation workflows." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 537, + 371, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 537, + 371, + 550 + ], + "spans": [ + { + "bbox": [ + 132, + 537, + 371, + 550 + ], + "type": "text", + "content": "2.2 Motion Re-Targeting without Supervision" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 558, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 481, + 665 + ], + "type": "text", + "content": "Conventional approaches in motion re-targeting have been a cornerstone of 3D animation for decades, primarily dedicated to converting motion capture data into 3D skeletal motions [4, 15]. Typically, raw motion capture data is optically recorded, yielding primarily spatial information. In contrast, motion data is largely rotational in nature [5], often represented in " + }, + { + "bbox": [ + 130, + 558, + 481, + 665 + ], + "type": "inline_equation", + "content": "SO(3)" + }, + { + "bbox": [ + 130, + 558, + 481, + 665 + ], + "type": "text", + "content": " space. Therefore, a principal objective of motion re-targeting solutions has been to effectively bridge these two distinct types of representations. In [14], extensive manually adjustable constraints were introduced for flexible re-targeting between skeletons of identical topology (i.e., same hierarchy, different bone lengths and rest poses). Key" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "C. Mo et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 137, + 121, + 473, + 222 + ], + "blocks": [ + { + "bbox": [ + 137, + 121, + 473, + 222 + ], + "lines": [ + { + "bbox": [ + 137, + 121, + 473, + 222 + ], + "spans": [ + { + "bbox": [ + 137, + 121, + 473, + 222 + ], + "type": "image", + "image_path": "b2e7a16acdf66b05482e71dd89d4c86ffdc781ecfb2d07b2bd8b9030fd37b84b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 238, + 482, + 304 + ], + "lines": [ + { + "bbox": [ + 130, + 238, + 482, + 304 + ], + "spans": [ + { + "bbox": [ + 130, + 238, + 482, + 304 + ], + "type": "text", + "content": "Fig. 3: Illustration of point cloud-based motion representation learning (PC-MRL) pipeline for keyframe interpolation to any human skeleton. Our unsupervised point cloud-to-pose decoding approach (I) first adapts motion features from existing datasets to new skeletons. We then employ the adapted motion data to supervise motion interpolation models (II), using FOQ and rest pose augmentations to address remaining discrepancies between the adapted and expected target skeleton motion patterns." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 328, + 479, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 328, + 479, + 388 + ], + "spans": [ + { + "bbox": [ + 130, + 328, + 479, + 388 + ], + "type": "text", + "content": "practices in this approach, including end-effector matching, joint contacts, and inter-skeleton point correlations, are still relevant to current animation pipelines. Point correlations using Inverse Kinematics (IK) have also been proposed as an intermediate representation to facilitate the adaptation between topologically distinct skeletons [9, 36], albeit with hierarchical pairing limitations." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 388, + 480, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 388, + 480, + 472 + ], + "spans": [ + { + "bbox": [ + 130, + 388, + 480, + 472 + ], + "type": "text", + "content": "In response, morphologically-independent motion control systems were proposed using IK joint chains and joint-wise constraints [20,26]. Such systems often expect highly standardised motion skeleton features, and are generally impractical for re-targeting motions produced for different control schemes. Additionally, the reliance on Euler angles for extensive classical re-targeting constraints poses a variety of issues regarding rotational freedom [11], as well as compatibility with animation workflows [33], which predominantly use quaternion-based systems." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 472, + 481, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 472, + 481, + 520 + ], + "spans": [ + { + "bbox": [ + 130, + 472, + 481, + 520 + ], + "type": "text", + "content": "Despite these advancements, the challenge of arbitrary skeleton re-targeting from motion datasets still remains largely unresolved, limiting any widespread application of data-driven pipelines. Our paper aims to effectively address this pervasive limitation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 540, + 233, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 540, + 233, + 553 + ], + "spans": [ + { + "bbox": [ + 131, + 540, + 233, + 553 + ], + "type": "text", + "content": "3 Methodology" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 566, + 479, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 566, + 479, + 590 + ], + "spans": [ + { + "bbox": [ + 130, + 566, + 479, + 590 + ], + "type": "text", + "content": "As illustrated in Fig. 3, our proposed method consists of two main components: point cloud-to-motion reconstruction and motion keyframe interpolation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 609, + 373, + 620 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 609, + 373, + 620 + ], + "spans": [ + { + "bbox": [ + 131, + 609, + 373, + 620 + ], + "type": "text", + "content": "3.1 Quaternion-based Motion Representations" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "content": "We first declare a set of notations for representing skeletons. A skeleton " + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "content": " is defined by a tree of bones " + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "content": ". A bone " + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_n \\in \\mathcal{B}" + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "content": " is defined by its bone parent " + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_m \\in \\mathcal{B}" + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "content": ", and a head position " + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{H}_n \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "content": " which denotes the bone's base offset from" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 215, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 215, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 215, + 91, + 447, + 102 + ], + "type": "text", + "content": "Motion Keyframe Interpolation for Any Human Skeleton" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": "its parent head " + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "inline_equation", + "content": "\\mathbf{H}_m" + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": ". In addition, a tail position " + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_n \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": " helps to define the end point of the bone's line segment representation, starting from its head position, of " + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_n" + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": ". The root bone " + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_0" + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": " has no parent, and thus no head position " + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "inline_equation", + "content": "\\mathbf{H}_0" + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": ". The set of " + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "inline_equation", + "content": "\\mathbf{H}_n" + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_n" + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": " for each bone " + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_n \\in \\mathcal{B}" + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": " make up the rest position of the skeleton." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 164, + 480, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 164, + 480, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 164, + 480, + 224 + ], + "type": "text", + "content": "Next, we define the properties of motion data. At any given frame " + }, + { + "bbox": [ + 130, + 164, + 480, + 224 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 164, + 480, + 224 + ], + "type": "text", + "content": ", a skeleton's pose position can be represented as global 3D positions " + }, + { + "bbox": [ + 130, + 164, + 480, + 224 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_{n,t} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 130, + 164, + 480, + 224 + ], + "type": "text", + "content": " and global quaternion rotations " + }, + { + "bbox": [ + 130, + 164, + 480, + 224 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_{n,t} \\in \\mathbb{R}^4" + }, + { + "bbox": [ + 130, + 164, + 480, + 224 + ], + "type": "text", + "content": " for each bone " + }, + { + "bbox": [ + 130, + 164, + 480, + 224 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_n \\in \\mathcal{B}" + }, + { + "bbox": [ + 130, + 164, + 480, + 224 + ], + "type": "text", + "content": ". For a given motion sequence " + }, + { + "bbox": [ + 130, + 164, + 480, + 224 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 164, + 480, + 224 + ], + "type": "text", + "content": ", only the root bone is given a 3D position in pose position, i.e. " + }, + { + "bbox": [ + 130, + 164, + 480, + 224 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_{0,t}" + }, + { + "bbox": [ + 130, + 164, + 480, + 224 + ], + "type": "text", + "content": " as all other pose positions are derived using Forwards Kinematics (FK [11,24]):" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 244, + 232, + 480, + 243 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 244, + 232, + 480, + 243 + ], + "spans": [ + { + "bbox": [ + 244, + 232, + 480, + 243 + ], + "type": "interline_equation", + "content": "\\mathbf {p} _ {n, t} = \\mathbf {p} _ {m, t} + Q R \\left(\\mathbf {H} _ {n}, \\mathbf {q} _ {m, t}\\right), \\tag {1}", + "image_path": "0411daee976950c200e00ed7a01435345d7c6c3e346558d5d0917ea91c532109.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 249, + 479, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 249, + 479, + 274 + ], + "spans": [ + { + "bbox": [ + 130, + 249, + 479, + 274 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 249, + 479, + 274 + ], + "type": "inline_equation", + "content": "QR" + }, + { + "bbox": [ + 130, + 249, + 479, + 274 + ], + "type": "text", + "content": " is the quaternion rotation function, and " + }, + { + "bbox": [ + 130, + 249, + 479, + 274 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_m" + }, + { + "bbox": [ + 130, + 249, + 479, + 274 + ], + "type": "text", + "content": " is the parent joint. In summary, a motion sequence " + }, + { + "bbox": [ + 130, + 249, + 479, + 274 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 249, + 479, + 274 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 130, + 249, + 479, + 274 + ], + "type": "inline_equation", + "content": "|M|" + }, + { + "bbox": [ + 130, + 249, + 479, + 274 + ], + "type": "text", + "content": " frames is defined by:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 213, + 281, + 480, + 293 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 213, + 281, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 213, + 281, + 480, + 293 + ], + "type": "interline_equation", + "content": "M = \\left\\{\\mathbf {p} _ {0, t} \\cup \\left\\{\\mathbf {q} _ {n, t} \\mid \\mathbf {v} _ {n} \\in \\mathcal {B} \\right\\} \\mid 0 \\leq t < | M | \\right\\} \\tag {2}", + "image_path": "d33c503fab10aab846cff0516e3b445f1f641aef96908472a34e8c33877f7015.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 308, + 433, + 319 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 308, + 433, + 319 + ], + "spans": [ + { + "bbox": [ + 131, + 308, + 433, + 319 + ], + "type": "text", + "content": "3.2 Point Cloud Obfuscation and Skeleton Reconstruction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 325, + 480, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 325, + 480, + 469 + ], + "spans": [ + { + "bbox": [ + 130, + 325, + 480, + 469 + ], + "type": "text", + "content": "Our approach is based on the hypothesis that 3D motion sequences can be visually represented by a more universal format: 3D point cloud sequences. By geometrically sampling the volume around a skeleton, point cloud data becomes a non-hierarchical representation of its pose(s), obfuscating the original skeletal configuration. Due to this decoupling between the skeleton configuration and motion representation, a successful motion data reconstruction from point cloud data would be the first step to enabling learnable cross-skeleton motion features. Formally, for a given source skeleton " + }, + { + "bbox": [ + 130, + 325, + 480, + 469 + ], + "type": "inline_equation", + "content": "S_A" + }, + { + "bbox": [ + 130, + 325, + 480, + 469 + ], + "type": "text", + "content": " with an associated motion sequence " + }, + { + "bbox": [ + 130, + 325, + 480, + 469 + ], + "type": "inline_equation", + "content": "M_A" + }, + { + "bbox": [ + 130, + 325, + 480, + 469 + ], + "type": "text", + "content": ", and any humanoid skeletal configuration " + }, + { + "bbox": [ + 130, + 325, + 480, + 469 + ], + "type": "inline_equation", + "content": "S_B" + }, + { + "bbox": [ + 130, + 325, + 480, + 469 + ], + "type": "text", + "content": ", the point cloud obfuscation and reconstruction module is to learn a function that can adequately perform " + }, + { + "bbox": [ + 130, + 325, + 480, + 469 + ], + "type": "inline_equation", + "content": "F_{S_A \\to S_B}(M_A) = \\hat{M}_B" + }, + { + "bbox": [ + 130, + 325, + 480, + 469 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 325, + 480, + 469 + ], + "type": "inline_equation", + "content": "\\hat{M}_B" + }, + { + "bbox": [ + 130, + 325, + 480, + 469 + ], + "type": "text", + "content": " is a visually similar motion adaptation of " + }, + { + "bbox": [ + 130, + 325, + 480, + 469 + ], + "type": "inline_equation", + "content": "M_A" + }, + { + "bbox": [ + 130, + 325, + 480, + 469 + ], + "type": "text", + "content": " on skeleton " + }, + { + "bbox": [ + 130, + 325, + 480, + 469 + ], + "type": "inline_equation", + "content": "S_B" + }, + { + "bbox": [ + 130, + 325, + 480, + 469 + ], + "type": "text", + "content": ", which is an estimation of " + }, + { + "bbox": [ + 130, + 325, + 480, + 469 + ], + "type": "inline_equation", + "content": "M_B" + }, + { + "bbox": [ + 130, + 325, + 480, + 469 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 484, + 480, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 484, + 480, + 579 + ], + "spans": [ + { + "bbox": [ + 130, + 484, + 480, + 579 + ], + "type": "text", + "content": "Motion Data Obfuscation with Point Clouds To sample the point cloud " + }, + { + "bbox": [ + 130, + 484, + 480, + 579 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 130, + 484, + 480, + 579 + ], + "type": "text", + "content": ", we first generate a set of points " + }, + { + "bbox": [ + 130, + 484, + 480, + 579 + ], + "type": "inline_equation", + "content": "u\\in \\mathcal{X}" + }, + { + "bbox": [ + 130, + 484, + 480, + 579 + ], + "type": "text", + "content": " along the line segments defined by the global head and tail positions for each bone " + }, + { + "bbox": [ + 130, + 484, + 480, + 579 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_n" + }, + { + "bbox": [ + 130, + 484, + 480, + 579 + ], + "type": "text", + "content": ". In the bone-local space to " + }, + { + "bbox": [ + 130, + 484, + 480, + 579 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_n" + }, + { + "bbox": [ + 130, + 484, + 480, + 579 + ], + "type": "text", + "content": ", this would mean sampling " + }, + { + "bbox": [ + 130, + 484, + 480, + 579 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 130, + 484, + 480, + 579 + ], + "type": "text", + "content": " by a uniformly distributed factor " + }, + { + "bbox": [ + 130, + 484, + 480, + 579 + ], + "type": "inline_equation", + "content": "\\alpha \\sim \\mathcal{U}_{[0,1]}" + }, + { + "bbox": [ + 130, + 484, + 480, + 579 + ], + "type": "text", + "content": ". Specifically, " + }, + { + "bbox": [ + 130, + 484, + 480, + 579 + ], + "type": "inline_equation", + "content": "\\alpha = 0" + }, + { + "bbox": [ + 130, + 484, + 480, + 579 + ], + "type": "text", + "content": " aligns with the head position of the bone, and " + }, + { + "bbox": [ + 130, + 484, + 480, + 579 + ], + "type": "inline_equation", + "content": "\\alpha = 1" + }, + { + "bbox": [ + 130, + 484, + 480, + 579 + ], + "type": "text", + "content": " indicates the tail position. By further sampling " + }, + { + "bbox": [ + 130, + 484, + 480, + 579 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 130, + 484, + 480, + 579 + ], + "type": "text", + "content": " via a normal distribution, our sampling strategy can capture the surrounding volume around the skeleton. Formally, to sample " + }, + { + "bbox": [ + 130, + 484, + 480, + 579 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 130, + 484, + 480, + 579 + ], + "type": "text", + "content": " from its associated bone " + }, + { + "bbox": [ + 130, + 484, + 480, + 579 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_n" + }, + { + "bbox": [ + 130, + 484, + 480, + 579 + ], + "type": "text", + "content": " within a standard deviation of " + }, + { + "bbox": [ + 130, + 484, + 480, + 579 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 130, + 484, + 480, + 579 + ], + "type": "text", + "content": ", we have:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 250, + 597, + 328, + 609 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 250, + 597, + 328, + 609 + ], + "spans": [ + { + "bbox": [ + 250, + 597, + 328, + 609 + ], + "type": "interline_equation", + "content": "\\mathbf {H} _ {x} \\sim \\mathcal {N} (\\alpha \\mathbf {T} _ {n}, \\sigma),", + "image_path": "98d372866df518fdf08783175be7e7541bd89ce3a2800dbc702d5bb7b75bd7ad.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 247, + 605, + 480, + 624 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 247, + 605, + 480, + 624 + ], + "spans": [ + { + "bbox": [ + 247, + 605, + 480, + 624 + ], + "type": "interline_equation", + "content": "\\mathbf {p} _ {u, t} = \\mathbf {p} _ {n, t} + Q R (\\mathbf {H} _ {x}, \\mathbf {q} _ {n, t}) \\tag {3}", + "image_path": "6d1f11ba8dd30daa6ffebb250980b369ed413eb1abe85697d2c8b5ea63f93a4c.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 629, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 479, + 665 + ], + "type": "text", + "content": "Notably, this sampling process as in Eq. 3 is associated with the skeletal configuration only, by which the sampled points are temporally consistent for further processing with the motion sequence. Specifically, each point maintains a" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "text", + "content": "C. Mo et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 151 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 151 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 151 + ], + "type": "text", + "content": "constant position from its associated bone in the bone-local space. The lateral distance of each point from their parent bone segment allows point clouds to react to the bone's rotational roll axis." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 480, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 480, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 480, + 248 + ], + "type": "text", + "content": "In a pure point cloud representation, it is possible to remove the relationships between bones. However, to clearly distinguish symmetrical body features in different skeletons, we introduce and embed body group associations for the points during our reconstruction process. Specifically, every bone is categorised by one of these body groups: [spine, left arm, right arm, left leg, right leg]. Each sampled point is assigned q group attribute in line with the associated bone. The attribute is represented by a one-hot feature vector " + }, + { + "bbox": [ + 130, + 152, + 480, + 248 + ], + "type": "inline_equation", + "content": "\\mathbf{g}_u" + }, + { + "bbox": [ + 130, + 152, + 480, + 248 + ], + "type": "text", + "content": ". To this end, we characterize " + }, + { + "bbox": [ + 130, + 152, + 480, + 248 + ], + "type": "inline_equation", + "content": "u = \\{\\mathbf{p}_{u,t}, \\mathbf{g}_u\\} \\in \\mathcal{X}" + }, + { + "bbox": [ + 130, + 152, + 480, + 248 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 268, + 479, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 268, + 479, + 351 + ], + "spans": [ + { + "bbox": [ + 130, + 268, + 479, + 351 + ], + "type": "text", + "content": "Skeleton Reconstruction Given a sampled point cloud representation " + }, + { + "bbox": [ + 130, + 268, + 479, + 351 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_A" + }, + { + "bbox": [ + 130, + 268, + 479, + 351 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 130, + 268, + 479, + 351 + ], + "type": "inline_equation", + "content": "S_A" + }, + { + "bbox": [ + 130, + 268, + 479, + 351 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 268, + 479, + 351 + ], + "type": "inline_equation", + "content": "M_A" + }, + { + "bbox": [ + 130, + 268, + 479, + 351 + ], + "type": "text", + "content": " as input, we aim to train a motion adaptation function " + }, + { + "bbox": [ + 130, + 268, + 479, + 351 + ], + "type": "inline_equation", + "content": "F_{S_A\\to S_B}" + }, + { + "bbox": [ + 130, + 268, + 479, + 351 + ], + "type": "text", + "content": " which is tasked with producing a visually similar skeletal motion " + }, + { + "bbox": [ + 130, + 268, + 479, + 351 + ], + "type": "inline_equation", + "content": "\\hat{M}_B" + }, + { + "bbox": [ + 130, + 268, + 479, + 351 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 130, + 268, + 479, + 351 + ], + "type": "inline_equation", + "content": "S_B" + }, + { + "bbox": [ + 130, + 268, + 479, + 351 + ], + "type": "text", + "content": " using a temporal and set-based neural network. As shown in Fig. 3, to construct " + }, + { + "bbox": [ + 130, + 268, + 479, + 351 + ], + "type": "inline_equation", + "content": "F_{S_A\\rightarrow S_B}" + }, + { + "bbox": [ + 130, + 268, + 479, + 351 + ], + "type": "text", + "content": ", we employ a Point Transformer [54] to learn embeddings for the unordered point cloud sets frame-wisely, and a standard temporal transformer to process the sequence and decode " + }, + { + "bbox": [ + 130, + 268, + 479, + 351 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_A" + }, + { + "bbox": [ + 130, + 268, + 479, + 351 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 130, + 268, + 479, + 351 + ], + "type": "inline_equation", + "content": "\\hat{M}_B" + }, + { + "bbox": [ + 130, + 268, + 479, + 351 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 353, + 480, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 353, + 480, + 472 + ], + "spans": [ + { + "bbox": [ + 130, + 353, + 480, + 472 + ], + "type": "text", + "content": "To optimize " + }, + { + "bbox": [ + 130, + 353, + 480, + 472 + ], + "type": "inline_equation", + "content": "F_{S_A \\to S_B}" + }, + { + "bbox": [ + 130, + 353, + 480, + 472 + ], + "type": "text", + "content": ", our objective function maximises the similarity between the sampled point clouds of both " + }, + { + "bbox": [ + 130, + 353, + 480, + 472 + ], + "type": "inline_equation", + "content": "M_A" + }, + { + "bbox": [ + 130, + 353, + 480, + 472 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 353, + 480, + 472 + ], + "type": "inline_equation", + "content": "\\hat{M}_B" + }, + { + "bbox": [ + 130, + 353, + 480, + 472 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 130, + 353, + 480, + 472 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_A" + }, + { + "bbox": [ + 130, + 353, + 480, + 472 + ], + "type": "text", + "content": " can be viewed as the ground truth to " + }, + { + "bbox": [ + 130, + 353, + 480, + 472 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_B" + }, + { + "bbox": [ + 130, + 353, + 480, + 472 + ], + "type": "text", + "content": ", which enables an unsupervised strategy for " + }, + { + "bbox": [ + 130, + 353, + 480, + 472 + ], + "type": "inline_equation", + "content": "F_{S_A \\to S_B}" + }, + { + "bbox": [ + 130, + 353, + 480, + 472 + ], + "type": "text", + "content": ". In detail, we first sample a point cloud " + }, + { + "bbox": [ + 130, + 353, + 480, + 472 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_B" + }, + { + "bbox": [ + 130, + 353, + 480, + 472 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 130, + 353, + 480, + 472 + ], + "type": "inline_equation", + "content": "S_B" + }, + { + "bbox": [ + 130, + 353, + 480, + 472 + ], + "type": "text", + "content": " based on " + }, + { + "bbox": [ + 130, + 353, + 480, + 472 + ], + "type": "inline_equation", + "content": "\\hat{M}_B" + }, + { + "bbox": [ + 130, + 353, + 480, + 472 + ], + "type": "text", + "content": " derived by " + }, + { + "bbox": [ + 130, + 353, + 480, + 472 + ], + "type": "inline_equation", + "content": "F_{S_A \\to S_B}" + }, + { + "bbox": [ + 130, + 353, + 480, + 472 + ], + "type": "text", + "content": ". Next, a temporally consistent KNN-Loss objective is devised to minimise the " + }, + { + "bbox": [ + 130, + 353, + 480, + 472 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 130, + 353, + 480, + 472 + ], + "type": "text", + "content": " distance between any given point of " + }, + { + "bbox": [ + 130, + 353, + 480, + 472 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_B" + }, + { + "bbox": [ + 130, + 353, + 480, + 472 + ], + "type": "text", + "content": " from its " + }, + { + "bbox": [ + 130, + 353, + 480, + 472 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 353, + 480, + 472 + ], + "type": "text", + "content": " nearest neighbouring points in " + }, + { + "bbox": [ + 130, + 353, + 480, + 472 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_A" + }, + { + "bbox": [ + 130, + 353, + 480, + 472 + ], + "type": "text", + "content": ". The " + }, + { + "bbox": [ + 130, + 353, + 480, + 472 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 353, + 480, + 472 + ], + "type": "text", + "content": " nearest neighbours are determined based on inter-point distance throughout the entire motion sequence, instead of a frame-wise setting, as shown in Fig. 4. The points with differing body groups are excluded from the neighbour search. Mathematically, we have:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 194, + 482, + 480, + 513 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 482, + 480, + 513 + ], + "spans": [ + { + "bbox": [ + 194, + 482, + 480, + 513 + ], + "type": "interline_equation", + "content": "\\delta \\left(u _ {A}, u _ {B}\\right) = \\left\\{ \\begin{array}{l l} \\sum_ {t} \\left\\| \\mathbf {p} _ {u _ {A}, t} - \\mathbf {p} _ {u _ {B}, t} \\right\\| _ {2}, & \\text {i f} \\mathbf {g} _ {u _ {A}} = \\mathbf {g} _ {u _ {B}}, \\\\ \\inf , & \\text {o t h e r w i s e}, \\end{array} \\right. \\tag {4}", + "image_path": "24bfa3b059cef4713ab312cec62d6348346f6644a58b6f8cc12a3beaebe66064.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 194, + 516, + 480, + 543 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 516, + 480, + 543 + ], + "spans": [ + { + "bbox": [ + 194, + 516, + 480, + 543 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {K N N}} \\left(\\mathcal {X} _ {A}, \\mathcal {X} _ {B}\\right) = \\sum_ {u _ {A} \\in \\mathcal {X} _ {A}, u _ {B} \\in N _ {k} \\left(u _ {A}, \\mathcal {X} _ {B}\\right)} \\delta \\left(u _ {A}, u _ {B}\\right), \\tag {5}", + "image_path": "1db072b1319e520a7803fbd651bca2a94f9b0d65110136a8023ac295ed293790.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 552, + 479, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 552, + 479, + 576 + ], + "spans": [ + { + "bbox": [ + 130, + 552, + 479, + 576 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 552, + 479, + 576 + ], + "type": "inline_equation", + "content": "N_{k}(u_{A},\\mathcal{X}_{B})" + }, + { + "bbox": [ + 130, + 552, + 479, + 576 + ], + "type": "text", + "content": " is the set containing points that are the k-nearest neighbors of " + }, + { + "bbox": [ + 130, + 552, + 479, + 576 + ], + "type": "inline_equation", + "content": "u_{A}" + }, + { + "bbox": [ + 130, + 552, + 479, + 576 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 130, + 552, + 479, + 576 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_B" + }, + { + "bbox": [ + 130, + 552, + 479, + 576 + ], + "type": "text", + "content": ", based on " + }, + { + "bbox": [ + 130, + 552, + 479, + 576 + ], + "type": "inline_equation", + "content": "\\delta (u_A,u_B)" + }, + { + "bbox": [ + 130, + 552, + 479, + 576 + ], + "type": "text", + "content": " distance." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 577, + 479, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 577, + 479, + 613 + ], + "spans": [ + { + "bbox": [ + 130, + 577, + 479, + 613 + ], + "type": "text", + "content": "In addition to the KNN-Loss, we introduce an optional end-effector loss in the skeleton space for the hand, foot, and head bones, which provides direct positional guidance for matched end-effector pairs " + }, + { + "bbox": [ + 130, + 577, + 479, + 613 + ], + "type": "inline_equation", + "content": "E_{AB}" + }, + { + "bbox": [ + 130, + 577, + 479, + 613 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 205, + 622, + 480, + 651 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 622, + 480, + 651 + ], + "spans": [ + { + "bbox": [ + 205, + 622, + 480, + 651 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {e n d}} = \\sum_ {(\\mathbf {v} _ {A}, \\mathbf {v} _ {B}) \\in E _ {A B}} \\frac {1}{| M _ {A} |} \\sum_ {t} \\left\\| \\mathbf {p} _ {\\mathbf {v} _ {A}, t} - \\mathbf {p} _ {\\mathbf {v} _ {B}, t} \\right\\| _ {2}. \\tag {6}", + "image_path": "5566065ac85f4f06bdca9615f0eb84c1db66ae27dd224591b39f8c862f657c89.jpg" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 215, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 215, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 215, + 91, + 447, + 102 + ], + "type": "text", + "content": "Motion Keyframe Interpolation for Any Human Skeleton" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 138, + 116, + 475, + 222 + ], + "blocks": [ + { + "bbox": [ + 138, + 116, + 475, + 222 + ], + "lines": [ + { + "bbox": [ + 138, + 116, + 475, + 222 + ], + "spans": [ + { + "bbox": [ + 138, + 116, + 475, + 222 + ], + "type": "image", + "image_path": "3a14d93f1d2874a63ea21ea43eb030712e6913a3b9a1c82718ed631ffec26b07.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 231, + 482, + 276 + ], + "lines": [ + { + "bbox": [ + 131, + 231, + 482, + 276 + ], + "spans": [ + { + "bbox": [ + 131, + 231, + 482, + 276 + ], + "type": "text", + "content": "Fig. 4: Visualisation of our temporally consistent KNN objective, processed independently per body group. For each point from " + }, + { + "bbox": [ + 131, + 231, + 482, + 276 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_A" + }, + { + "bbox": [ + 131, + 231, + 482, + 276 + ], + "type": "text", + "content": " (a), the total " + }, + { + "bbox": [ + 131, + 231, + 482, + 276 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 131, + 231, + 482, + 276 + ], + "type": "text", + "content": " distance throughout the sequence is measured between each point (b) from " + }, + { + "bbox": [ + 131, + 231, + 482, + 276 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_B" + }, + { + "bbox": [ + 131, + 231, + 482, + 276 + ], + "type": "text", + "content": ". The " + }, + { + "bbox": [ + 131, + 231, + 482, + 276 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 131, + 231, + 482, + 276 + ], + "type": "text", + "content": " distances from the K-nearest points are summed to produce KNN-Loss (c)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 300, + 319, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 300, + 319, + 312 + ], + "spans": [ + { + "bbox": [ + 132, + 300, + 319, + 312 + ], + "type": "text", + "content": "3.3 First-Frame Offset Quaternions" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 321, + 482, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 321, + 482, + 417 + ], + "spans": [ + { + "bbox": [ + 130, + 321, + 482, + 417 + ], + "type": "text", + "content": "Point clouds lack absolute roll axis information for each bone. As such, we introduce a First-frame Offset Quaternion (FOQ) transform to incorporate the available relative roll data for motion modelling. We formulate all quaternions of a motion sequence relative to their values from the initial frame. In detail, for any quaternion sequence " + }, + { + "bbox": [ + 130, + 321, + 482, + 417 + ], + "type": "inline_equation", + "content": "Q = \\{\\mathbf{q}_0, \\mathbf{q}_1, \\mathbf{q}_2, \\ldots\\}" + }, + { + "bbox": [ + 130, + 321, + 482, + 417 + ], + "type": "text", + "content": ", we can simply obtain its FOQ transform as " + }, + { + "bbox": [ + 130, + 321, + 482, + 417 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{q}_0 \\times \\mathbf{q}_0^{-1}, \\mathbf{q}_1 \\times \\mathbf{q}_0^{-1}, \\mathbf{q}_2 \\times \\mathbf{q}_0^{-1}, \\ldots\\}" + }, + { + "bbox": [ + 130, + 321, + 482, + 417 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 321, + 482, + 417 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_0^{-1}" + }, + { + "bbox": [ + 130, + 321, + 482, + 417 + ], + "type": "text", + "content": " is the quaternion conjugate of " + }, + { + "bbox": [ + 130, + 321, + 482, + 417 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_0" + }, + { + "bbox": [ + 130, + 321, + 482, + 417 + ], + "type": "text", + "content": ". An FOQ sequence can easily be converted back to a quaternion sequence as long as any " + }, + { + "bbox": [ + 130, + 321, + 482, + 417 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_t \\in Q" + }, + { + "bbox": [ + 130, + 321, + 482, + 417 + ], + "type": "text", + "content": " is known." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 418, + 482, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 418, + 482, + 513 + ], + "spans": [ + { + "bbox": [ + 130, + 418, + 482, + 513 + ], + "type": "text", + "content": "Crucially, FOQ is a roll-invariant rotation representation, that is, FOQ remains identical regardless of a bone's absolute roll position in the original data. Trivially, to adjust the rotational roll of a bone " + }, + { + "bbox": [ + 130, + 418, + 482, + 513 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_n" + }, + { + "bbox": [ + 130, + 418, + 482, + 513 + ], + "type": "text", + "content": ", we perform the quaternion multiplication for each frame: " + }, + { + "bbox": [ + 130, + 418, + 482, + 513 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_{n,t} \\times \\mathbf{r}" + }, + { + "bbox": [ + 130, + 418, + 482, + 513 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 418, + 482, + 513 + ], + "type": "inline_equation", + "content": "\\mathbf{r}" + }, + { + "bbox": [ + 130, + 418, + 482, + 513 + ], + "type": "text", + "content": " is a roll quaternion. As a quaternion on the roll axis of " + }, + { + "bbox": [ + 130, + 418, + 482, + 513 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_n" + }, + { + "bbox": [ + 130, + 418, + 482, + 513 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 418, + 482, + 513 + ], + "type": "inline_equation", + "content": "\\mathbf{r}" + }, + { + "bbox": [ + 130, + 418, + 482, + 513 + ], + "type": "text", + "content": " is constrained within " + }, + { + "bbox": [ + 130, + 418, + 482, + 513 + ], + "type": "inline_equation", + "content": "\\mathbf{r} = \\alpha + \\beta (x\\mathbf{i} + y\\mathbf{j} + z\\mathbf{k})" + }, + { + "bbox": [ + 130, + 418, + 482, + 513 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 418, + 482, + 513 + ], + "type": "inline_equation", + "content": "x,y,z" + }, + { + "bbox": [ + 130, + 418, + 482, + 513 + ], + "type": "text", + "content": " are the 3D XYZ values of " + }, + { + "bbox": [ + 130, + 418, + 482, + 513 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_n" + }, + { + "bbox": [ + 130, + 418, + 482, + 513 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 418, + 482, + 513 + ], + "type": "inline_equation", + "content": "\\alpha, \\beta" + }, + { + "bbox": [ + 130, + 418, + 482, + 513 + ], + "type": "text", + "content": " are scalars that control the roll magnitude. Note that " + }, + { + "bbox": [ + 130, + 418, + 482, + 513 + ], + "type": "inline_equation", + "content": "\\mathbf{r}" + }, + { + "bbox": [ + 130, + 418, + 482, + 513 + ], + "type": "text", + "content": " is constant throughout the sequence. Therefore, we can deduce the roll-invariant nature of any constant " + }, + { + "bbox": [ + 130, + 418, + 482, + 513 + ], + "type": "inline_equation", + "content": "\\mathbf{r}" + }, + { + "bbox": [ + 130, + 418, + 482, + 513 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 200, + 521, + 411, + 554 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 521, + 411, + 554 + ], + "spans": [ + { + "bbox": [ + 200, + 521, + 411, + 554 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\left(\\mathbf {q} _ {n, t} \\times \\mathbf {r}\\right) \\times \\left(\\mathbf {q} _ {0, t} \\times \\mathbf {r}\\right) ^ {- 1} = \\mathbf {q} _ {n, t} \\times \\mathbf {r} \\times \\mathbf {r} ^ {- 1} \\times \\mathbf {q} _ {0, t} ^ {- 1} \\\\ = \\mathbf {q} _ {n, t} \\mathbf {q} _ {0, t} ^ {- 1} \\\\ \\end{array}", + "image_path": "416ee2de67439f7ad9e7a84fafa475d09ef28cf2a13fdfca55c3fbfc2443b44d.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 571, + 481, + 585 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 571, + 481, + 585 + ], + "spans": [ + { + "bbox": [ + 131, + 571, + 481, + 585 + ], + "type": "text", + "content": "3.4 Motion Interpolation Transformer and Rest Pose Augmentation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "type": "text", + "content": "To leverage our point cloud-based representation learning for cross-skeleton interpolation, a transformer model - CITL [34] is adapted, allowing for an efficient interpolation pipeline with existing datasets to alternative skeletons. We introduce two key modifications to CITL and its training strategy, focusing on enhancing motion quality. First, quaternion predictions are substituted with FOQ predictions. This is inspired by the RNN-based method TGcomplete [18], which predicts" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "C. Mo et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "code", + "bbox": [ + 134, + 130, + 482, + 217 + ], + "blocks": [ + { + "bbox": [ + 133, + 115, + 476, + 129 + ], + "lines": [ + { + "bbox": [ + 133, + 115, + 476, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 115, + 476, + 129 + ], + "type": "text", + "content": "Algorithm 1 IK target alignment with maximum real quaternion component" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 134, + 130, + 482, + 217 + ], + "lines": [ + { + "bbox": [ + 134, + 130, + 482, + 217 + ], + "spans": [ + { + "bbox": [ + 134, + 130, + 482, + 217 + ], + "type": "text", + "content": "1: Input: " + }, + { + "bbox": [ + 134, + 130, + 482, + 217 + ], + "type": "inline_equation", + "content": "\\mathbf{p_{IK}} = 3D" + }, + { + "bbox": [ + 134, + 130, + 482, + 217 + ], + "type": "text", + "content": " IK target unit vector \n2: " + }, + { + "bbox": [ + 134, + 130, + 482, + 217 + ], + "type": "inline_equation", + "content": "\\mathbf{T} =" + }, + { + "bbox": [ + 134, + 130, + 482, + 217 + ], + "type": "text", + "content": " resting tail position of associated bone \n3: Output: " + }, + { + "bbox": [ + 134, + 130, + 482, + 217 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_{\\mathrm{max}} =" + }, + { + "bbox": [ + 134, + 130, + 482, + 217 + ], + "type": "text", + "content": " output quaternion, to be applied on original bone quaternion \n4: " + }, + { + "bbox": [ + 134, + 130, + 482, + 217 + ], + "type": "inline_equation", + "content": "\\eta \\leftarrow \\frac{\\mathbf{T}}{|\\mathbf{T}|}" + }, + { + "bbox": [ + 134, + 130, + 482, + 217 + ], + "type": "inline_equation", + "content": "\\triangleright" + }, + { + "bbox": [ + 134, + 130, + 482, + 217 + ], + "type": "text", + "content": " 3D unit coordinates " + }, + { + "bbox": [ + 134, + 130, + 482, + 217 + ], + "type": "inline_equation", + "content": "[\\eta_x,\\eta_y,\\eta_z]" + }, + { + "bbox": [ + 134, + 130, + 482, + 217 + ], + "type": "text", + "content": " \n5: " + }, + { + "bbox": [ + 134, + 130, + 482, + 217 + ], + "type": "inline_equation", + "content": "\\omega \\leftarrow \\frac{\\mathbf{p_{IK}}}{|\\mathbf{p_{IK}}|}" + }, + { + "bbox": [ + 134, + 130, + 482, + 217 + ], + "type": "inline_equation", + "content": "\\triangleright" + }, + { + "bbox": [ + 134, + 130, + 482, + 217 + ], + "type": "text", + "content": " 3D unit coordinates " + }, + { + "bbox": [ + 134, + 130, + 482, + 217 + ], + "type": "inline_equation", + "content": "[\\omega_x,\\omega_y,\\omega_z]" + }, + { + "bbox": [ + 134, + 130, + 482, + 217 + ], + "type": "text", + "content": " \n6: " + }, + { + "bbox": [ + 134, + 130, + 482, + 217 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_{\\mathrm{min}} \\gets \\frac{0 + (\\eta_x + \\omega_x)\\mathbf{i} + (\\eta_y + \\omega_y)\\mathbf{j} + (\\eta_z + \\omega_z)\\mathbf{k}}{2}" + }, + { + "bbox": [ + 134, + 130, + 482, + 217 + ], + "type": "text", + "content": " \n7: " + }, + { + "bbox": [ + 134, + 130, + 482, + 217 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_{\\mathrm{max}} \\gets \\mathbf{q}_{\\mathrm{min}} \\times (0 + \\eta_x\\mathbf{i} + \\eta_y\\mathbf{j} + \\eta_z\\mathbf{k})" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "code_body" + } + ], + "index": 3, + "sub_type": "algorithm" + }, + { + "bbox": [ + 130, + 239, + 479, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 239, + 479, + 263 + ], + "spans": [ + { + "bbox": [ + 130, + 239, + 479, + 263 + ], + "type": "text", + "content": "quaternion offsets from the preceding frame rather than raw quaternions, thereby achieving greater generalisability and superior quality in interpolation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 263, + 480, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 263, + 480, + 370 + ], + "spans": [ + { + "bbox": [ + 130, + 263, + 480, + 370 + ], + "type": "text", + "content": "For our model's CITL-based architecture, we substitute the original sinusoidal positional encoding scheme with learned relative positions [44]. While the original implementation relies on the continual nature of sinusoidal functions, we observe that relative positional embeddings with zero initialisation converges upon similarly continuous behaviour, and significantly lowers the complexity of positional attention. The lessened focus on positional relations allows the model to synthesise deeper behaviours within the pose space, improving overall generalisability. As a side effect, the model also accepts arbitrary length inputs beyond the training data length." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 371, + 481, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 371, + 481, + 515 + ], + "spans": [ + { + "bbox": [ + 130, + 371, + 481, + 515 + ], + "type": "text", + "content": "Notably, the intended rest pose of a skeletal configuration is often geometrically sub-optimal for point cloud matching, and can manifest in multiple forms. To address this, we devise a rest position augmentation strategy, leveraging a per-bone Inverse Kinematics (IK) system. RPA broadens the range of intended rest poses that the learned interpolation model can accommodate. In detail, during each phase of the Forward Kinematics (FK) process, we augment the bone tails with a random additional offset. We re-align the augmented global tail position as closely as possible to its original location using a quaternion rotation with the maximum possible real component. Algorithm 1 describes the method by which this quaternion can be derived. Since the real component describes, in cosine, the amount of rotation around an axis, this minimizes numerical disturbance from RPA while altering the visual representation to the desired extent." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 544, + 282, + 558 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 544, + 282, + 558 + ], + "spans": [ + { + "bbox": [ + 132, + 544, + 282, + 558 + ], + "type": "text", + "content": "4 Experimental Results" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 567, + 204, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 204, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 204, + 578 + ], + "type": "text", + "content": "4.1 Datasets" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 586, + 479, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 586, + 479, + 611 + ], + "spans": [ + { + "bbox": [ + 130, + 586, + 479, + 611 + ], + "type": "text", + "content": "To thoroughly evaluate the effectiveness of the proposed method, we conducted extensive experiments on the following widely used motion capture datasets:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 137, + 617, + 485, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 617, + 485, + 665 + ], + "spans": [ + { + "bbox": [ + 137, + 617, + 485, + 665 + ], + "type": "text", + "content": "- LaFAN1 [18] contains long, high quality motions in controllable video game character styles. The dataset was recorded on a large motion capture stage, and manually cleaned to production standards. The native skeleton of LaFAN1 contains 6 spinal bones, and 4 bones for each limb." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 215, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 215, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 215, + 91, + 447, + 102 + ], + "type": "text", + "content": "Motion Keyframe Interpolation for Any Human Skeleton" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 137, + 115, + 481, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 115, + 481, + 163 + ], + "spans": [ + { + "bbox": [ + 137, + 115, + 481, + 163 + ], + "type": "text", + "content": "- Human3.6M [6,22] comprises of 3.6 million motion frames. The motions captured in Human3.6M tend to be less dynamic, as they were recorded on a relatively compact " + }, + { + "bbox": [ + 137, + 115, + 481, + 163 + ], + "type": "inline_equation", + "content": "4 \\times 3 \\mathrm{~m}" + }, + { + "bbox": [ + 137, + 115, + 481, + 163 + ], + "type": "text", + "content": " stage. Its native skeleton includes 5 spinal bones, 4 bones for each limb, and an additional finger bone for each hand." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 163, + 481, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 163, + 481, + 212 + ], + "spans": [ + { + "bbox": [ + 138, + 163, + 481, + 212 + ], + "type": "text", + "content": "- CMU Mocap [10] is a diverse motion dataset encompassing 113 categories. It features a mix of static and dynamic motions, captured on a " + }, + { + "bbox": [ + 138, + 163, + 481, + 212 + ], + "type": "inline_equation", + "content": "3 \\times 8 \\mathrm{~m}" + }, + { + "bbox": [ + 138, + 163, + 481, + 212 + ], + "type": "text", + "content": " stage. The skeleton is identical to that of Human3.6M, with the exception of the pelvis, which is divided into two hip bones and a lower back bone." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 215, + 481, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 215, + 481, + 252 + ], + "spans": [ + { + "bbox": [ + 132, + 215, + 481, + 252 + ], + "type": "text", + "content": "For PC-MRL, we designated the CMU MoCap skeleton as " + }, + { + "bbox": [ + 132, + 215, + 481, + 252 + ], + "type": "inline_equation", + "content": "S_B" + }, + { + "bbox": [ + 132, + 215, + 481, + 252 + ], + "type": "text", + "content": ". Its native dataset is exclusively utilized for testing purposes, while LaFAN1 and Human3.6M are used simultaneously for training." + }, + { + "bbox": [ + 132, + 215, + 481, + 252 + ], + "type": "inline_equation", + "content": "^3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 266, + 279, + 278 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 266, + 279, + 278 + ], + "spans": [ + { + "bbox": [ + 132, + 266, + 279, + 278 + ], + "type": "text", + "content": "4.2 Implementation Details" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 283, + 482, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 283, + 482, + 475 + ], + "spans": [ + { + "bbox": [ + 130, + 283, + 482, + 475 + ], + "type": "text", + "content": "In our experiments, point cloud obfuscation adopted a 256-point setting for the sampling process with " + }, + { + "bbox": [ + 130, + 283, + 482, + 475 + ], + "type": "inline_equation", + "content": "\\sigma = 0.05" + }, + { + "bbox": [ + 130, + 283, + 482, + 475 + ], + "type": "text", + "content": " and each point was characterized by a 64-element vector. To maintain consistency, we ensure that all bone offsets and motion positions are defined in metre units. For skeletal reconstruction, 4 Point Transformer [54] layers was utilized to construct the neural network. Within each layer, the feature size was doubled, and the size of the point set was downsampled by a factor of 4 through a farthest-first traversal approach, eventually resulting in a singular feature vector. In the final step, two linear layers with ReLU activations were utilized to transform the resulted feature vector in a frame-wise manner, producing the output skeletal representation. During training, we set " + }, + { + "bbox": [ + 130, + 283, + 482, + 475 + ], + "type": "inline_equation", + "content": "K = 8" + }, + { + "bbox": [ + 130, + 283, + 482, + 475 + ], + "type": "text", + "content": " for the KNN loss to optimize the network. Additionally, the use of unit quaternion values in practice relies on modulus division for constraint, a process that can potentially lead to exploding gradients. To address this, we introduced an objective function " + }, + { + "bbox": [ + 130, + 283, + 482, + 475 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_q" + }, + { + "bbox": [ + 130, + 283, + 482, + 475 + ], + "type": "text", + "content": " that measures the " + }, + { + "bbox": [ + 130, + 283, + 482, + 475 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 130, + 283, + 482, + 475 + ], + "type": "text", + "content": " norms' difference between the raw quaternion output and the expected unit quaternion. All experiments were trained and evaluated on a single NVIDIA GeForce RTX 3090 GPU." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 489, + 477, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 489, + 477, + 502 + ], + "spans": [ + { + "bbox": [ + 132, + 489, + 477, + 502 + ], + "type": "text", + "content": "4.3 Motion Interpolation Comparison against Supervised Methods" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 506, + 482, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 506, + 482, + 626 + ], + "spans": [ + { + "bbox": [ + 130, + 506, + 482, + 626 + ], + "type": "text", + "content": "To demonstrate the efficacy of our method, we produce and compare our method against the conventional LERP method, as well as three state-of-the-art interpolation models trained on the original CMU MoCap dataset. Specifically, we measure the performance of the RNN-based " + }, + { + "bbox": [ + 130, + 506, + 482, + 626 + ], + "type": "inline_equation", + "content": "\\mathrm{TG}_{\\mathrm{complete}}" + }, + { + "bbox": [ + 130, + 506, + 482, + 626 + ], + "type": "text", + "content": " model [18], the BERT-based motion interpolation adaptation [12], and a transformer-based encoder-decoder approach [34]. We additionally provide results on a variant of our method trained without our RPA strategy. The experimental configuration for each state-of-the-art model is identical to their original implementations. To measure interpolation accuracy, we employ the standard " + }, + { + "bbox": [ + 130, + 506, + 482, + 626 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 130, + 506, + 482, + 626 + ], + "type": "text", + "content": " positional distance (L2P), " + }, + { + "bbox": [ + 130, + 506, + 482, + 626 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 130, + 506, + 482, + 626 + ], + "type": "text", + "content": " quaternion difference (L2Q), and NPSS [16] for visual similarity evaluation [18]." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "C. Mo et al." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 631, + 481, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 631, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 631, + 481, + 665 + ], + "type": "text", + "content": "3 The authors Clinton Mo, Kun Hu and Zhiyong Wang are signatories of the dataset licenses and produced all experimental results of the paper. Meta Inc. was not granted access to the datasets." + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 137, + 114, + 479, + 308 + ], + "blocks": [ + { + "bbox": [ + 137, + 114, + 479, + 308 + ], + "lines": [ + { + "bbox": [ + 137, + 114, + 479, + 308 + ], + "spans": [ + { + "bbox": [ + 137, + 114, + 479, + 308 + ], + "type": "table", + "html": "
Motion categoryMethodL2P↓L2Q↓NPSS↓
515305153051530
BasketballLERP0.07870.40090.80330.15880.58340.96110.13150.61201.8202
TGcomplete [18]0.10500.44190.71040.18590.64390.93030.16000.65791.2136
BERT [12]0.08890.37370.68650.15850.55600.89920.16890.61771.4587
CITL [34]0.06250.29730.57080.14100.49320.80110.10470.40981.0959
PC-MRL w/o RPA0.07670.35340.67150.15050.54310.90490.13640.55961.3471
PC-MRL (Ours)0.07520.34430.66620.15610.54950.90880.13970.55721.3015
GolfLERP0.01890.12290.31870.04550.18700.39870.05870.39791.0312
TGcomplete [18]0.04730.26160.45460.06930.31910.52970.09540.50621.1534
BERT [12]0.02760.11590.29670.06120.20820.42580.10990.39060.9255
CITL [34]0.02010.10430.25890.04760.17660.35520.06360.25810.7799
PC-MRL w/out RPA0.02740.13040.36220.05550.20650.49230.09110.37591.2083
PC-MRL (Ours)0.02730.11300.27060.05980.20120.38440.10490.37160.8150
SwimmingLERP0.07310.36800.73740.13260.53270.94760.21470.92221.6972
TGcomplete [18]0.12920.57960.91410.18450.74201.11990.26600.97161.7698
BERT [12]0.10880.38960.75160.18190.55180.96220.28770.90741.6682
CITL [34]0.09050.38950.74750.14840.55350.95330.20840.83171.6122
PC-MRL w/o RPA0.07220.37540.74590.13450.54420.95940.19440.87311.6608
PC-MRL (Ours)0.07340.35950.71390.13550.53310.92480.19560.85911.6370
Walking & RunningLERP0.04550.25050.55270.11200.39970.62330.06550.26820.9071
TGcomplete [18]0.05400.17860.31030.11470.28660.38660.08950.26160.4265
BERT [12]0.05520.19980.35440.11110.31010.47620.09010.27600.6458
CITL [34]0.03280.10340.24270.09420.21020.33840.05260.15680.4080
PC-MRL w/o RPA0.03970.16570.34170.10580.28790.47220.06400.22880.6021
PC-MRL (Ours)0.03770.13920.28480.10640.27780.42810.06580.21290.5098
", + "image_path": "d6f4353838fb737cf53e19a26d670c71b8f1f903ef28e0ecfc73dc2930b5e69b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 308, + 482, + 352 + ], + "lines": [ + { + "bbox": [ + 132, + 308, + 482, + 352 + ], + "spans": [ + { + "bbox": [ + 132, + 308, + 482, + 352 + ], + "type": "text", + "content": "Table 1: Performance comparisons for various motion categories, measured in L2P, L2Q, and NPSS. All methods except PC-MRL are trained on the original CMU MoCap dataset. Each motion contains 128 frames with keyframes placed every 5, 15, or 30 frames. The top 2 results for each test are highlighted in purple and blue respectively." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 388, + 482, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 388, + 482, + 519 + ], + "spans": [ + { + "bbox": [ + 130, + 388, + 482, + 519 + ], + "type": "text", + "content": "The results listed in Table 1 clearly demonstrate our method's ability to approach the accuracy levels exhibited by directly supervised state-of-the-art methods. PC-MRL consistently outperforms both the LERP standard and RNN-based " + }, + { + "bbox": [ + 130, + 388, + 482, + 519 + ], + "type": "inline_equation", + "content": "\\mathrm{TG}_{\\mathrm{complete}}" + }, + { + "bbox": [ + 130, + 388, + 482, + 519 + ], + "type": "text", + "content": " model, particularly in longer keyframe interval scenarios where the quality and depth of learned motion features is most critical and impactful. Likewise, the inclusion of RPA also tends to improve the long interval performance of PC-MRL, particularly when precise movements (i.e. golf) or deeper motion features (i.e. locomotion) are expected. Given this, we can observe that RPA definitively improves the overall consistency of the PC-MRL method in scenarios that may be unseen or less effectively represented by our point cloud-to-motion system." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": "Native supervision, i.e. with CITL, unsurprisingly produces the most accurate interpolation model, with a notable exception of swimming motions, where our PC-MRL supervision produces even stronger results. We believe the abundance of low crawl motions, a similar motion to swimming in LaFAN1, provides more meaningful supervision over the original CMU dataset, which conversely has few similar motions once swimming motions are filtered out. On the other end of the spectrum, i.e. for high data scenarios such as walking and running locomotion from Fig. 5, all models including our PC-MRL approach observe the strongest improvements for learned methods over the naive LERP method. Both cases strongly support our method's capability to supervise highly intricate motion patterns, despite the incompleteness of the point cloud representation and reliance on relative rotations and geometric optimisation assumptions." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 215, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 215, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 215, + 91, + 447, + 102 + ], + "type": "text", + "content": "Motion Keyframe Interpolation for Any Human Skeleton" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 149, + 123, + 465, + 326 + ], + "blocks": [ + { + "bbox": [ + 149, + 123, + 465, + 326 + ], + "lines": [ + { + "bbox": [ + 149, + 123, + 465, + 326 + ], + "spans": [ + { + "bbox": [ + 149, + 123, + 465, + 326 + ], + "type": "image", + "image_path": "5bf89b54c8e79c1caf8b33c94c9efab4976a017e1a19a022bbde137256c2059a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 341, + 480, + 364 + ], + "lines": [ + { + "bbox": [ + 131, + 341, + 480, + 364 + ], + "spans": [ + { + "bbox": [ + 131, + 341, + 480, + 364 + ], + "type": "text", + "content": "Fig. 5: Walking keyframe interpolation examples produced by methods from Table 1. Close-up layered comparisons against the original motion are provided on the right." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 387, + 411, + 400 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 387, + 411, + 400 + ], + "spans": [ + { + "bbox": [ + 131, + 387, + 411, + 400 + ], + "type": "text", + "content": "4.4 Cross-Skeleton Motion Re-Targeting Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 407, + 482, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 407, + 482, + 503 + ], + "spans": [ + { + "bbox": [ + 130, + 407, + 482, + 503 + ], + "type": "text", + "content": "We further conduct an experiment on motion re-targeting to directly benchmark our method against the sole existing state-of-the-art method for unpaired motion re-targeting with unrestricted skeletons, i.e. primal skeletons (PS) [1]. For direct motion re-targeting evaluation, we utilized the KNN-Loss and end-effector loss as metrics, owing to their effective measurement of visual similarity and non-sensitivity towards absolute roll axis correctness. Though absolute rolls are generally necessary for a complete motion re-targeting solution, this aspect is out of scope for our project as our main goal is motion interpolation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 504, + 482, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 504, + 482, + 541 + ], + "spans": [ + { + "bbox": [ + 130, + 504, + 482, + 541 + ], + "type": "text", + "content": "Table 2 indicates the superior performance of our method over the state-of-the-art PS method in terms of visual similarity. At low " + }, + { + "bbox": [ + 130, + 504, + 482, + 541 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 130, + 504, + 482, + 541 + ], + "type": "text", + "content": " values, the higher relative " + }, + { + "bbox": [ + 130, + 504, + 482, + 541 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{KNN}}" + }, + { + "bbox": [ + 130, + 504, + 482, + 541 + ], + "type": "text", + "content": " of the primal skeleton method underscores its geometric deviations" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 164, + 560, + 452, + 609 + ], + "blocks": [ + { + "bbox": [ + 164, + 560, + 452, + 609 + ], + "lines": [ + { + "bbox": [ + 164, + 560, + 452, + 609 + ], + "spans": [ + { + "bbox": [ + 164, + 560, + 452, + 609 + ], + "type": "table", + "html": "
Model\\( L_{\\text{KNN}}(\\mathcal{X}_A,\\mathcal{X}_B) \\downarrow \\)\\( L_{\\text{end}} \\downarrow \\)
\\( \\sigma = 0 \\)\\( \\sigma = 0.01 \\)\\( \\sigma = 0.05 \\)\\( \\sigma = 0.1 \\)\\( \\sigma = 0.2 \\)
Original motion1.00992.02345.73999.663616.92230
Primal skeleton [1]8.11098.03539.420612.829120.63329.6990
PC-MRL (Ours)3.87144.12316.727710.307817.52615.6258
", + "image_path": "1b9559e395566ecda0d7e8684a2e77835c2ba857c8b6aa16239b6973f57981d4.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 608, + 482, + 653 + ], + "lines": [ + { + "bbox": [ + 130, + 608, + 482, + 653 + ], + "spans": [ + { + "bbox": [ + 130, + 608, + 482, + 653 + ], + "type": "text", + "content": "Table 2: Average unpaired LaFAN1 to CMU motion re-targeting performance of the primal skeleton method and our point cloud method. " + }, + { + "bbox": [ + 130, + 608, + 482, + 653 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{KNN}}" + }, + { + "bbox": [ + 130, + 608, + 482, + 653 + ], + "type": "text", + "content": " is measured using 1024-point clouds at " + }, + { + "bbox": [ + 130, + 608, + 482, + 653 + ], + "type": "inline_equation", + "content": "K = 8" + }, + { + "bbox": [ + 130, + 608, + 482, + 653 + ], + "type": "text", + "content": ". Independently sampled point clouds from the original LaFAN1 skeleton are provided as optimal " + }, + { + "bbox": [ + 130, + 608, + 482, + 653 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{KNN}}" + }, + { + "bbox": [ + 130, + 608, + 482, + 653 + ], + "type": "text", + "content": " reference. All values are scaled by " + }, + { + "bbox": [ + 130, + 608, + 482, + 653 + ], + "type": "inline_equation", + "content": "\\times 100" + }, + { + "bbox": [ + 130, + 608, + 482, + 653 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "C. Mo et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 143, + 116, + 470, + 220 + ], + "blocks": [ + { + "bbox": [ + 143, + 116, + 470, + 220 + ], + "lines": [ + { + "bbox": [ + 143, + 116, + 470, + 220 + ], + "spans": [ + { + "bbox": [ + 143, + 116, + 470, + 220 + ], + "type": "image", + "image_path": "4d71934c2946edcc88669cce7457ddcefb2a49921e2cbd04202c403bdc4878ed.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 231, + 482, + 265 + ], + "lines": [ + { + "bbox": [ + 130, + 231, + 482, + 265 + ], + "spans": [ + { + "bbox": [ + 130, + 231, + 482, + 265 + ], + "type": "text", + "content": "Fig. 6: Jumping motion re-targeting examples produced on the CMU skeleton by the primal skeleton method (middle row) and our method (bottom row). Sampled point cloud representations of each motion sequence are provided on the right." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 297, + 482, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 297, + 482, + 406 + ], + "spans": [ + { + "bbox": [ + 130, + 297, + 482, + 406 + ], + "type": "text", + "content": "from the original motion. In contrast, our method demonstrates significantly better visual adherence to the original geometry. As " + }, + { + "bbox": [ + 130, + 297, + 482, + 406 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 130, + 297, + 482, + 406 + ], + "type": "text", + "content": " increases, our method approaches near-perfect alignment with the original motion. Figure 6 visually indicates this, showing our method follows the original motion closely, while the latent consistency-based primal skeleton method struggles to generalise and decode the skeleton's internal structures without directly supervised objectives, such as the end-effector positions provided during training. In addition, Fig 7 demonstrates that, like existing re-targeting approaches, our method is able to produce adequate results on disproportionate skeletons." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 409, + 482, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 409, + 482, + 458 + ], + "spans": [ + { + "bbox": [ + 130, + 409, + 482, + 458 + ], + "type": "text", + "content": "Since our approach is largely focused on enabling non-native motion interpolation supervision, we additionally compare our PC-MRL method against a CITL model supervised by PS in Table 3. Due to the suboptimal re-targeting performance of PS, such interpolation models exhibit expectedly poor accuracy." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 178, + 488, + 440, + 604 + ], + "blocks": [ + { + "bbox": [ + 178, + 488, + 440, + 604 + ], + "lines": [ + { + "bbox": [ + 178, + 488, + 440, + 604 + ], + "spans": [ + { + "bbox": [ + 178, + 488, + 440, + 604 + ], + "type": "image", + "image_path": "22c90d7f153bf66827f9df7e59a0f616eb3eed7b990ed10dbb5963987623cddf.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 612, + 482, + 635 + ], + "lines": [ + { + "bbox": [ + 130, + 612, + 482, + 635 + ], + "spans": [ + { + "bbox": [ + 130, + 612, + 482, + 635 + ], + "type": "text", + "content": "Fig. 7: Re-targeted falling motion from LaFAN1 (top) on the CMU skeleton (middle) and a short custom skeleton (bottom), performed by our point cloud-to-motion model." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 215, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 215, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 215, + 91, + 447, + 102 + ], + "type": "text", + "content": "Motion Keyframe Interpolation for Any Human Skeleton" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 137, + 114, + 479, + 164 + ], + "blocks": [ + { + "bbox": [ + 137, + 114, + 479, + 164 + ], + "lines": [ + { + "bbox": [ + 137, + 114, + 479, + 164 + ], + "spans": [ + { + "bbox": [ + 137, + 114, + 479, + 164 + ], + "type": "table", + "html": "
MethodL2P↓L2Q↓NPSS↓
515305153051530
CITL - PS0.20430.63321.15240.23180.74281.14420.11220.42761.4622
CITL - Native0.04950.23060.48980.09790.36040.66090.11490.46841.1591
PC-MRL0.04940.23370.49600.10300.37710.68130.13700.54331.1760
", + "image_path": "78c7a253c6c77c9a25a2144dec57a7944e36688644f4412cc77be868fa8bb19c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 132, + 164, + 481, + 185 + ], + "lines": [ + { + "bbox": [ + 132, + 164, + 481, + 185 + ], + "spans": [ + { + "bbox": [ + 132, + 164, + 481, + 185 + ], + "type": "text", + "content": "Table 3: Average L2P, L2Q, and NPSS performance of CITL trained on native and PS-generated datasets, compared to our method." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 217, + 222, + 229 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 217, + 222, + 229 + ], + "spans": [ + { + "bbox": [ + 132, + 217, + 222, + 229 + ], + "type": "text", + "content": "5 Limitations" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 241, + 482, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 241, + 482, + 361 + ], + "spans": [ + { + "bbox": [ + 130, + 241, + 482, + 361 + ], + "type": "text", + "content": "Due to the inherent limitations of the point cloud representation, certain constraints on the applicability of our point cloud-to-motion method are inevitable. Primarily, the reliance of our point cloud re-targeting method on the relative nature of First-frame Offset Quaternions (FOQ) for most motion learning tasks necessitates the presence of at least one known pose. In the absence of a known pose, relative quaternions are incapable of reconstructing absolute rotational values, which are essential for motion applications. Secondly, due to our method of sampling point clouds with a consistent deviation from each bone segment, the resultant representation inevitably obscures finer details. This includes elements like fingers and facial controls, to an extent that is beyond rectification." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 361, + 482, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 361, + 482, + 445 + ], + "spans": [ + { + "bbox": [ + 130, + 361, + 482, + 445 + ], + "type": "text", + "content": "Our proposed objective function relies entirely on geometric segments, necessitating that all bones possess a non-zero length. This stipulation is deemed a reasonable prerequisite for generic human skeletons, as each bone is typically responsible for manipulating a discernible segment of the human body. Due to technical considerations, the skeletal configurations in each of our datasets included at least one bone of zero length, which we ignored in our experiments in favour of global quaternion rotations of all bones with non-zero length." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 445, + 481, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 445, + 481, + 469 + ], + "spans": [ + { + "bbox": [ + 130, + 445, + 481, + 469 + ], + "type": "text", + "content": "As a result of these limitations, we refrained from claiming state-of-the-art performance for general motion re-targeting tasks." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 486, + 220, + 498 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 486, + 220, + 498 + ], + "spans": [ + { + "bbox": [ + 132, + 486, + 220, + 498 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": "This paper introduces a novel learning-based motion interpolation method designed to enable cross-skeleton compatibility with existing motion datasets. Our proposed method PC-MRL employs a process of point cloud obfuscation and skeleton reconstruction. The point cloud space represents human motions in a non-hierarchical and skeleton-agnostic manner. It enables a KNN-based objective to be optimised without dataset supervision, guiding a neural network to generate high-quality motion features with any target skeleton. We address the rotational information loss of our point cloud format by presenting an offset quaternion strategy compatible with concurrent transformer-based models. Through extensive experiments, we have demonstrated the efficacy of PC-MRL in performing motion interpolation without relying on native motion data. Moreover, PC-MRL has achieved superior visual similarity metrics in the domain of motion re-targeting. We concluded our work by discussing the limitations of PC-MRL." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "C. Mo et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 140, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 138, + 140, + 480, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 140, + 480, + 161 + ], + "spans": [ + { + "bbox": [ + 138, + 140, + 480, + 161 + ], + "type": "text", + "content": "1. Aberman, K., Li, P., Lischinski, D., Sorkine-Hornung, O., Cohen-Or, D., Chen, B.: Skeleton-aware networks for deep motion retargeting 39(4), 62–1 (2020) 4, 12" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 162, + 480, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 162, + 480, + 183 + ], + "spans": [ + { + "bbox": [ + 138, + 162, + 480, + 183 + ], + "type": "text", + "content": "2. Aberman, K., Wu, R., Lischinski, D., Chen, B., Cohen-Or, D.: Learning character-agnostic motion for motion retargeting in 2d 38(4), 1-14 (jul 2019) 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 184, + 480, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 184, + 480, + 215 + ], + "spans": [ + { + "bbox": [ + 138, + 184, + 480, + 215 + ], + "type": "text", + "content": "3. Annabi, L., Ma, Z., et al.: Unsupervised motion retargeting for human-robot imitation. In: Companion of the 2024 ACM/IEEE International Conference on Human-Robot Interactio. ACM (2024) 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 216, + 480, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 216, + 480, + 248 + ], + "spans": [ + { + "bbox": [ + 138, + 216, + 480, + 248 + ], + "type": "text", + "content": "4. Bodenheimer, B., Rose, C., Rosenthal, S., Pella, J.: The process of motion capture: Dealing with the data. In: Computer Animation and Simulation. pp. 3-18. Springer (1997) 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 249, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 249, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 138, + 249, + 480, + 270 + ], + "type": "text", + "content": "5. Burtnyk, N., Wein, M.: Computer-generated key-frame animation. Journal of the SMPTE 80(3), 149-153 (1971) 1, 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 271, + 480, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 271, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 138, + 271, + 480, + 293 + ], + "type": "text", + "content": "6. Catalin Ionescu, Fuxin Li, C.S.: Latent structured models for human pose estimation. In: International Conference on Computer Vision (2011) 10" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 293, + 480, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 293, + 480, + 325 + ], + "spans": [ + { + "bbox": [ + 138, + 293, + 480, + 325 + ], + "type": "text", + "content": "7. Chen, H., Tang, H., Timofte, R., Gool, L.V., Zhao, G.: Lart: Neural correspondence learning with latent regularization transformer for 3d motion transfer. Advances in Neural Information Processing Systems 36 (2024) 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 326, + 480, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 326, + 480, + 347 + ], + "spans": [ + { + "bbox": [ + 138, + 326, + 480, + 347 + ], + "type": "text", + "content": "8. Chen, J., Li, C., Lee, G.H.: Weakly-supervised 3D pose transfer with keypoints. In: ICCV. pp. 15156-15165 (2023) 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 348, + 480, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 348, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 138, + 348, + 480, + 369 + ], + "type": "text", + "content": "9. Choi, K.J., Ko, H.S.: Online motion retargetting. The Journal of Visualization and Computer Animation 11(5), 223-235 (2000) 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 370, + 480, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 370, + 480, + 391 + ], + "spans": [ + { + "bbox": [ + 138, + 370, + 480, + 391 + ], + "type": "text", + "content": "0. CMU: Carnegie-mellon university motion capture database (2003), http://mocap.cs.cmu.edu/ 10" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 392, + 480, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 392, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 138, + 392, + 480, + 413 + ], + "type": "text", + "content": "1. Dam, E.B., Koch, M., Lillholm, M.: Quaternions, interpolation and animation, vol. 2. CiteSeer (1998) 5, 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 414, + 480, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 414, + 480, + 435 + ], + "spans": [ + { + "bbox": [ + 138, + 414, + 480, + 435 + ], + "type": "text", + "content": "2. Duan, Y., Lin, Y., Zou, Z., Yuan, Y., Qian, Z., Zhang, B.: A unified framework for real time motion completion. In: AAAI. vol. 36, pp. 4459-4467 (2022) 2, 4, 10, 11" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 436, + 480, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 436, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 138, + 436, + 480, + 456 + ], + "type": "text", + "content": "3. Frank, T., Johnston, O.: Disney Animation: The Illusion of Life. Abbeville Publishing Group, New York (1981) 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 457, + 480, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 457, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 138, + 457, + 480, + 479 + ], + "type": "text", + "content": "4. Gleicher, M.: Retargetting motion to new characters. In: Computer Graphics and Interactive Techniques. pp. 33-42 (1998) 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 479, + 480, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 479, + 480, + 501 + ], + "spans": [ + { + "bbox": [ + 138, + 479, + 480, + 501 + ], + "type": "text", + "content": "5. Gleicher, M.: Animation from observation: Motion capture and motion editing 33(4), 51-54 (1999) 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 502, + 480, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 502, + 480, + 523 + ], + "spans": [ + { + "bbox": [ + 138, + 502, + 480, + 523 + ], + "type": "text", + "content": "6. Gopalakrishnan, A., Mali, A., Kifer, D., Giles, L., Ororbia, A.G.: A neural temporal model for human motion prediction. In: CVPR. pp. 12116-12125 (2019) 10" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 523, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 523, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 138, + 523, + 480, + 544 + ], + "type": "text", + "content": "7. Harvey, F.G., Pal, C.: Recurrent transition networks for character locomotion. In: SIGGRAPH Asia 2018 Technical Briefs, pp. 1-4 (2018) 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 545, + 480, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 545, + 480, + 567 + ], + "spans": [ + { + "bbox": [ + 138, + 545, + 480, + 567 + ], + "type": "text", + "content": "8. Harvey, F.G., Yurick, M., Nowrouzezahrai, D., Pal, C.: Robust motion in betweening 39(4), 60-1 (2020) 2, 4, 8, 9, 10, 11" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 138, + 567, + 480, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 567, + 480, + 588 + ], + "spans": [ + { + "bbox": [ + 138, + 567, + 480, + 588 + ], + "type": "text", + "content": "9. He, C., Saito, J., Zachary, J., Rushmeier, H., Zhou, Y.: Nemf: Neural motion fields for kinematic animation 35, 4244-4256 (2022) 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 138, + 589, + 480, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 589, + 480, + 620 + ], + "spans": [ + { + "bbox": [ + 138, + 589, + 480, + 620 + ], + "type": "text", + "content": "20. Hecker, C., Raabe, B., Enslow, R.W., DeWeese, J., Maynard, J., van Prooijen, K.: Real-time motion retargeting to highly varied user-created morphologies 27(3), 1-11 (2008) 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 138, + 621, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 621, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 138, + 621, + 480, + 643 + ], + "type": "text", + "content": "21. Hernandez, A., Gall, J., Moreno-Noguer, F.: Human motion prediction via spatiotemporal inpainting. In: ICCV. pp. 7134-7143 (2019) 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "text", + "content": "22. Ionescu, C., Papava, D., Olaru, V., Sminchisescu, C.: Human3.6m: Large scale datasets and predictive methods for 3D human sensing in natural environments." + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 215, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 215, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 215, + 91, + 447, + 102 + ], + "type": "text", + "content": "Motion Keyframe Interpolation for Any Human Skeleton" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 149, + 116, + 481, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 116, + 481, + 138 + ], + "spans": [ + { + "bbox": [ + 149, + 116, + 481, + 138 + ], + "type": "text", + "content": "IEEE Transactions on Pattern Analysis and Machine Intelligence 36(7), 1325-1339 (jul 2014) 10" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 139, + 480, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 139, + 480, + 159 + ], + "spans": [ + { + "bbox": [ + 133, + 139, + 480, + 159 + ], + "type": "text", + "content": "23. Jiang, H., Cai, J., Zheng, J.: Skeleton-aware 3D human shape reconstruction from point clouds. In: ICCV. pp. 5431-5441 (2019) 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 160, + 480, + 180 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 160, + 480, + 180 + ], + "spans": [ + { + "bbox": [ + 133, + 160, + 480, + 180 + ], + "type": "text", + "content": "24. Kantor, I.L., Solodovnikov, A.S., Shenitzer, A.: Hypercomplex numbers: an elementary introduction to algebras, vol. 302. Springer (1989) 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 181, + 480, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 181, + 480, + 212 + ], + "spans": [ + { + "bbox": [ + 132, + 181, + 480, + 212 + ], + "type": "text", + "content": "25. Karunratanakul, K., Preechakul, K., Suwajanakorn, S., Tang, S.: Guided motion diffusion for controllable human motion synthesis. In: ICCV. pp. 2151-2162 (2023) 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 213, + 480, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 213, + 480, + 235 + ], + "spans": [ + { + "bbox": [ + 132, + 213, + 480, + 235 + ], + "type": "text", + "content": "26. Kulpa, R., Multon, F., Arnaldi, B.: Morphology-independent representation of motions for interactive human-like animation. In: Eurographics (2005) 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 235, + 480, + 256 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 235, + 480, + 256 + ], + "spans": [ + { + "bbox": [ + 132, + 235, + 480, + 256 + ], + "type": "text", + "content": "27. Lehrmann, A.M., Gehler, P.V., Nowozin, S.: Efficient nonlinear markov models for human motion. In: CVPR. pp. 1314-1321 (2014) 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 257, + 480, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 257, + 480, + 277 + ], + "spans": [ + { + "bbox": [ + 132, + 257, + 480, + 277 + ], + "type": "text", + "content": "28. Li, P., Aberman, K., Zhang, Z., Hanocka, R., Sorkine-Hornung, O.: Ganimator: Neural motion synthesis from a single sequence 41(4), 1-12 (2022) 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 278, + 480, + 310 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 278, + 480, + 310 + ], + "spans": [ + { + "bbox": [ + 132, + 278, + 480, + 310 + ], + "type": "text", + "content": "29. Li, W., Wang, Z., Mai, R., Ren, P., Zhang, Q., Zhou, Y., Xu, N., Zhuang, J., Xin, B., Gao, L., et al.: Modular design automation of the morphologies, controllers, and vision systems for intelligent robots: a survey. Visual Intelligence 1(1), 2 (2023) 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 311, + 480, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 311, + 480, + 342 + ], + "spans": [ + { + "bbox": [ + 132, + 311, + 480, + 342 + ], + "type": "text", + "content": "30. Li, Z., Peng, X.B., Abbeel, P., Levine, S., Berseth, G., Sreenath, K.: Reinforcement learning for versatile, dynamic, and robust bipedal locomotion control. arXiv preprint arXiv:2401.16889 (2024) 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 343, + 480, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 343, + 480, + 363 + ], + "spans": [ + { + "bbox": [ + 132, + 343, + 480, + 363 + ], + "type": "text", + "content": "31. Liao, Z., Yang, J., Saito, J., Pons-Moll, G., Zhou, Y.: Skeleton-free pose transfer for stylized 3D characters. In: ECCV. pp. 640-656. Springer (2022) 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 364, + 480, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 364, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 364, + 480, + 407 + ], + "type": "text", + "content": "32. Makoviychuk, V., Wawrzyniak, L., Guo, Y., Lu, M., Storey, K., Macklin, M., Hoeller, D., Rudin, N., Allshire, A., Handa, A., et al.: Isaac gym: High performancegpu-based physics simulation for robot learning. arXiv preprint arXiv:2108.10470 (2021) 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 407, + 480, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 407, + 480, + 439 + ], + "spans": [ + { + "bbox": [ + 132, + 407, + 480, + 439 + ], + "type": "text", + "content": "33. Merry, B., Marais, P., Gain, J.: Animation space: A truly linear framework for character animation. ACM Transactions on Graphics (TOG) 25(4), 1400-1423 (2006) 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 440, + 480, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 440, + 480, + 471 + ], + "spans": [ + { + "bbox": [ + 132, + 440, + 480, + 471 + ], + "type": "text", + "content": "34. Mo, C.A., Hu, K., Long, C., Wang, Z.: Continuous intermediate token learning with implicit motion manifold for keyframe based motion interpolation. In: CVPR. pp. 13894-13903 (2023) 2, 4, 8, 10, 11" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 472, + 480, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 472, + 480, + 504 + ], + "spans": [ + { + "bbox": [ + 132, + 472, + 480, + 504 + ], + "type": "text", + "content": "35. Mo, C.A., Hu, K., Mei, S., Chen, Z., Wang, Z.: Keyframe extraction from motion capture sequences with graph based deep reinforcement learning. In: ACM International Conference on Multimedia. pp. 5194-5202 (2021) 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 505, + 480, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 505, + 480, + 536 + ], + "spans": [ + { + "bbox": [ + 132, + 505, + 480, + 536 + ], + "type": "text", + "content": "36. Monzani, J.S., Baerlocher, P., Boulic, R., Thalmann, D.: Using an intermediate skeleton and inverse kinematics for motion retargeting. In: Computer Graphics Forum. vol. 19, pp. 11-19. Wiley Online Library (2000) 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 537, + 480, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 537, + 480, + 557 + ], + "spans": [ + { + "bbox": [ + 132, + 537, + 480, + 557 + ], + "type": "text", + "content": "37. Oreshkin, B.N., Valkanas, A., Harvey, F.G., Menard, L.S., Bocquelet, F., Coates, M.J.: Motion inbetweening via deep " + }, + { + "bbox": [ + 132, + 537, + 480, + 557 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 132, + 537, + 480, + 557 + ], + "type": "text", + "content": "-interpolator (2022) 2, 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 558, + 480, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 558, + 480, + 590 + ], + "spans": [ + { + "bbox": [ + 132, + 558, + 480, + 590 + ], + "type": "text", + "content": "38. Peng, X.B., Abbeel, P., Levine, S., van de Panne, M.: Deepmimic: Example-guided deep reinforcement learning of physics-based character skills 37(4), 143:1-143:14 (Jul 2018) 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 590, + 480, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 590, + 480, + 611 + ], + "spans": [ + { + "bbox": [ + 132, + 590, + 480, + 611 + ], + "type": "text", + "content": "39. Peng, X.B., Ma, Z., Abbeel, P., Levine, S., Kanazawa, A.: Amp: Adversarial motion priors for stylized physics-based character control 40(4) (Jul 2021) 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 612, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 612, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 612, + 480, + 632 + ], + "type": "text", + "content": "40. Qin, J., Zheng, Y., Zhou, K.: Motion in-between via two-stage transformers 41(6), 1-16 (2022) 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "type": "text", + "content": "41. Reda, D., Won, J., Ye, Y., van de Panne, M., Winkler, A.: Physics-based motion retargeting from sparse inputs. ACM Computer Graphics and Interactive Techniques 6(3), 1-19 (2023) 4" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "C. Mo et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 500 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 117, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 117, + 480, + 138 + ], + "type": "text", + "content": "42. Reeves, W.T.: Inbetweening for computer animation utilizing moving point constraints 15(3), 263-269 (1981) 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 139, + 480, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 139, + 480, + 171 + ], + "spans": [ + { + "bbox": [ + 132, + 139, + 480, + 171 + ], + "type": "text", + "content": "43. Ren, T., Yu, J., Guo, S., Ma, Y., Ouyang, Y., Zeng, Z., Zhang, Y., Qin, Y.: Diverse motion in-betweening from sparse keyframes with dual posture stitching. IEEE Transactions on Visualization and Computer Graphics (2024) 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 172, + 480, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 172, + 480, + 193 + ], + "spans": [ + { + "bbox": [ + 133, + 172, + 480, + 193 + ], + "type": "text", + "content": "44. Shaw, P., Uszkoreit, J., Vaswani, A.: Self-attention with relative position representations. In: NAACL. Association for Computational Linguistics (2018) 9" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 194, + 480, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 194, + 480, + 226 + ], + "spans": [ + { + "bbox": [ + 133, + 194, + 480, + 226 + ], + "type": "text", + "content": "45. Tiwari, G., Antic, D., Lenssen, J.E., Sarafianos, N., Tung, T., Pons-Moll, G.: Posendf: Modeling human pose manifolds with neural distance fields. In: ECCV. pp. 572-589. Springer (2022) 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 227, + 480, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 480, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 480, + 248 + ], + "type": "text", + "content": "46. Villegas, R., Yang, J., Ceylan, D., Lee, H.: Neural kinematic networks for unsupervised motion retargeting. In: CVPR. pp. 8639-8648 (2018) 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 249, + 480, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 480, + 280 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 480, + 280 + ], + "type": "text", + "content": "47. Wang, J., Li, X., Liu, S., De Mello, S., Gallo, O., Wang, X., Kautz, J.: Zero-shot pose transfer for unrigged stylized 3D characters. In: CVPR. pp. 8704-8714 (2023) 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 281, + 480, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 281, + 480, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 281, + 480, + 303 + ], + "type": "text", + "content": "48. Wang, K., Xie, J., Zhang, G., Liu, L., Yang, J.: Sequential 3D human pose and shape estimation from point clouds. In: CVPR. pp. 7275-7284 (2020) 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 304, + 480, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 480, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 480, + 325 + ], + "type": "text", + "content": "49. Yuan, Y., Song, J., Iqbal, U., Vahdat, A., Kautz, J.: Physdiff: Physics-guided human motion diffusion model. In: ICCV. pp. 16010-16021 (2023) 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 325, + 480, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 325, + 480, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 325, + 480, + 346 + ], + "type": "text", + "content": "50. Yuan, Y., Wei, S.E., Simon, T., Kitani, K., Saragih, J.: Simpoe: Simulated character control for 3D human pose estimation. In: CVPR. pp. 7159-7169 (2021) 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 347, + 480, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 480, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 480, + 380 + ], + "type": "text", + "content": "51. Zhang, J., Weng, J., Kang, D., Zhao, F., Huang, S., Zhe, X., Bao, L., Shan, Y., Wang, J., Tu, Z.: Skinned motion retargeting with residual perception of motion semantics & geometry. In: CVPR. pp. 13864-13872 (2023) 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 380, + 480, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 480, + 413 + ], + "type": "text", + "content": "52. Zhang, M., Li, H., Cai, Z., Ren, J., Yang, L., Liu, Z.: Finemogen: Fine-grained spatio-temporal motion generation and editing. Advances in Neural Information Processing Systems 36 (2024) 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 413, + 480, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 413, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 413, + 480, + 445 + ], + "type": "text", + "content": "53. Zhang, X., van de Panne, M.: Data-driven autocompletion for keyframe animation. In: ACM SIGGRAPH Conference on Motion, Interaction and Games. pp. 1-11 (2018) 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 446, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 446, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 446, + 480, + 468 + ], + "type": "text", + "content": "54. Zhao, H., Jiang, L., Jia, J., Torr, P.H., Koltun, V.: Point transformer. In: ICCV. pp. 16259-16268 (2021) 7, 10" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 468, + 480, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 480, + 500 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 480, + 500 + ], + "type": "text", + "content": "55. Zhu, W., Yang, Z., Di, Z., Wu, W., Wang, Y., Loy, C.C.: Mocanet: Motion retargeting in-the-wild via canonicalization networks. In: AAAI. vol. 36, pp. 3617-3625 (2022) 4" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 215, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 215, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 215, + 91, + 447, + 102 + ], + "type": "text", + "content": "Motion Keyframe Interpolation for Any Human Skeleton" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Motion Mamba_ Efficient and Long Sequence Motion Generation/1b586349-9221-4bad-a56d-e274ec79c852_content_list.json b/2024/Motion Mamba_ Efficient and Long Sequence Motion Generation/1b586349-9221-4bad-a56d-e274ec79c852_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..dfbcbad03cb09fc81f74d030a2228af394802b40 --- /dev/null +++ b/2024/Motion Mamba_ Efficient and Long Sequence Motion Generation/1b586349-9221-4bad-a56d-e274ec79c852_content_list.json @@ -0,0 +1,1595 @@ +[ + { + "type": "text", + "text": "Motion Mamba: Efficient and Long Sequence Motion Generation", + "text_level": 1, + "bbox": [ + 243, + 140, + 759, + 184 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zeyu Zhang $^{12*}$ , Akide Liu $^{1*}$ , Ian Reid $^{3}$ , Richard Hartley $^{2}$ , Bohan Zhuang $^{1}$ , and Hao Tang $^{4\\boxtimes}$", + "bbox": [ + 238, + 212, + 764, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Monash University", + "bbox": [ + 431, + 253, + 573, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 The Australian National University", + "bbox": [ + 375, + 268, + 627, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 Mohamed bin Zayed University of Artificial Intelligence", + "bbox": [ + 308, + 282, + 692, + 297 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{4}$ National Key Laboratory for Multimedia Information Processing, School of Computer Science, Peking University", + "bbox": [ + 276, + 296, + 725, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://steve-zeyu-zhang.github.io/MotionMamba", + "bbox": [ + 334, + 335, + 666, + 348 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9d50515d2181f0c7d935e2bbb103cb5293cf413d46914332a1dd381edc6ed7a8.jpg", + "image_caption": [ + "Fig.1: Motion Mamba has achieved significantly superior performance on long sequence modeling and motion generation efficiency compared with other well-designed state-of-the-art methods such as MLD [6], MotionDiffuse [54], and MDM [49]." + ], + "image_footnote": [], + "bbox": [ + 222, + 362, + 419, + 513 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/48cf11f3798a6aedad48113acb76dca9df21fbaad9e72b652f24ec0f0b7e17e3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 421, + 361, + 614, + 513 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/210741e7ade2cfcbcf250c0c7992f4b4a689a716569803d67cdd28922bb7d8d0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 627, + 361, + 787, + 513 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Human motion generation stands as a significant pursuit in generative computer vision, while achieving long-sequence and efficient motion generation remains challenging. Recent advancements in state space models (SSMs), notably Mamba, have showcased considerable promise in long sequence modeling with an efficient hardware-aware design, which appears to be a significant direction upon building motion generation model. Nevertheless, adapting SSMs to motion generation faces hurdles since the lack of a specialized design architecture to model motion sequence. To address these challenges, we propose Motion Mamba, a simple yet efficient approach that presents the pioneering motion generation model utilized SSMs. Specifically, we design a Hierarchical Temporal Mamba (HTM) block to process temporal data by ensemble varying numbers of isolated SSM modules across a symmetric U-Net architecture aimed at preserving motion consistency between frames. We also design a Bidirectional Spatial Mamba (BSM) block", + "bbox": [ + 259, + 583, + 743, + 792 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution.", + "bbox": [ + 232, + 804, + 354, + 816 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "†Work done while being a research assistant at Monash University.", + "bbox": [ + 233, + 816, + 617, + 827 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "Corresponding author: bjdxtanghao@gmail.com", + "bbox": [ + 233, + 827, + 517, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "to bidirectionally process latent poses, to enhance accurate motion generation within a temporal frame. Our proposed method achieves up to $50\\%$ FID improvement and up to 4 times faster on the HumanML3D and KIT-ML datasets compared to the previous best diffusion-based method, which demonstrates strong capabilities of high-quality long sequence motion modeling and real-time human motion generation.", + "bbox": [ + 259, + 147, + 743, + 231 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Keywords: Human Motion Generation $\\cdot$ Selective State Space Models $\\cdot$ Latent Diffusion Models", + "bbox": [ + 261, + 244, + 743, + 272 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 303, + 375, + 319 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Human motion generation involves creating high-quality 3D human motion based on given conditions, such as a text prompt, which generates Euler angles for each joint in a human skeleton. To emulate human motion effectively, virtual characters must respond to the conditional context, exhibit natural movement, and perform motion accurately. Recent motion generation models are categorized into four main approaches: autoencoder-based [1,10,17,36,48,56], utilizing transformers for latent space compression and motion synthesis; GAN-based [4,20,30], using discriminators to enhance the realism of generated motions; autoregressive models [24], treating motion sequences as languages with specialized codebooks; and diffusion-based [6,49,54], employing denoising steps for motion generation. Challenges vary across methods, with autoencoder models struggling to generate accurate motions from detailed descriptions due to textual information compression, GAN-based models facing training difficulties, especially in conditional tasks, and diffusion-based models relying on complex transformer-based architectures results in inefficient of motion prediction.", + "bbox": [ + 212, + 339, + 787, + 566 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Although diffusion-based models excel at generating motion with robust performance and often exhibit superior diversity, they encounter two limitations. 1) Convolutional or transformer-based diffusion methods exhibit limitations in generating long-range motion sequences. Previous transformer-based methodologies [6,49,54] have focused on modeling long-range dependencies and acquiring comprehensive global context information. Despite these advances, they are frequently associated with a substantial increase in computational requirements. Furthermore, transformer architectures are not intrinsically designed for temporal sequential modeling, which poses an inherent limitation. 2) The efficiency of inference in transformer-based diffusion methods is constrained. Although prior research has attempted to leverage the Variational Autoencoder for denoising operations in the latent space [6], the inference speed remains adversely affected by the attention mechanism's quadratic scaling, leading to inefficient motion generation. Consequently, exploring a new architectural paradigm that accommodates long-range dependencies and maintains a linear computational complexity is crucial for sustaining motion generation tasks.", + "bbox": [ + 212, + 566, + 787, + 808 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent advances have sparked renewed interest in state space models (SSMs) [14, 15], a field that originated from the foundational classic state space model", + "bbox": [ + 215, + 809, + 785, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Zeyu and Akide et al.", + "bbox": [ + 271, + 114, + 416, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "[25]. Modern versions of SSMs stand out due to their ability to effectively capture long-range dependencies, a capability greatly improved by the introduction of parallel training techniques. This evolution has led to various methodologies based on SSM, notably the linear state space layers (LSSL) [15], the structured state-space sequence model (S4) [14], the diagonal state space (DSS) [19], and S4D [13]. These methods have been carefully designed to handle sequential data across various tasks and modalities, paying special attention to modeling long-range dependencies. Their efficacy in managing long sequences is attributed to the implementation of convolutional computations [15] and near-linear computational strategies, such as mamba [12], marking a significant stride in sequentially oriented tasks, including large language model decoding and motion sequence generation.", + "bbox": [ + 212, + 145, + 787, + 327 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Adapting selective state space modules for motion generation tasks presents notable challenges, primarily due to the lack of specialized design in SSMs for capturing the sensitive motion details required for temporal representation and the complexities involved in aggregating latent space. In response to these challenges, we have meticulously developed a motion generation architecture, specifically tailored to address the intricacies of long-term sequence generation, while optimizing for computational efficiency with near-linear-time complexity. This innovation is embodied in the Motion Mamba model, a simple yet potent approach to motion generation. The Motion Mamba framework pioneers a diffusion-based generative system, incorporating two key components oriented toward SSM as shown in Figure. 2: (1) a Hierarchical Temporal Mamba (HTM) block: This component is ingeniously crafted to arrange motion frames in sequential order, using hierarchically adjusted scanning. It is adept at identifying temporal dependencies at various depths, thereby facilitating a thorough comprehension of the dynamics inherent in motion sequences. (2) a Bidirectional Spatial Mamba (BSM) block: This block is designed to unravel the structured latent skeleton by evaluating data from both forward and reverse directions. Its primary goal is to safeguard the continuity of information flow, significantly bolstering the model's capacity for precise motion generation through the retention of dense informational exchange.", + "bbox": [ + 212, + 330, + 787, + 633 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The Motion Mamba introduces a new approach to motion generation that strikes an exceptional trade-off between accuracy and efficiency, shown in Fig. 1. Our experimental results underscore the significant improvements brought about by Motion Mamba, showcasing a remarkable improvement in the Fréchet Inception Distance (FID), with a reduction of up to $50\\%$ from the prior state-of-the-art metric of 0.473 to an impressive 0.281 on the HumanML3D dataset [17]. Furthermore, we emphasize our framework's unparalleled inference speed, which is four times faster than previous methods, achieving an average inference time of only 0.058 seconds per sequence compared to the 0.217 seconds required by the MLD [6] method per sequence. These outcomes unequivocally establish Motion Mamba's state-of-the-art performance, concurrently ensuring fast inference speeds for conditional human motion generation tasks.", + "bbox": [ + 212, + 638, + 787, + 820 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our contributions to the field of motion generation can be summarized as:", + "bbox": [ + 238, + 824, + 774, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Motion Mamba", + "bbox": [ + 625, + 114, + 730, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. We introduce a simple yet effective framework, named Motion Mamba, which is a pioneering method integrates a selective scanning mechanism into motion generation tasks.", + "2. Motion Mamba is comprised of two modules: Hierarchical Temporal Mamba (HTM) and Bidirectional Spatial Mamba (BSM), which are designed for temporal and spatial modeling, respectively. HTM blocks are tasked with processing temporal motion data, aiming to enhance motion consistency across frames. BSM blocks are engineered to bidirectionally capture the channel-wise flow of hidden information within the latent pose representations.", + "3. Motion Mamba framework demonstrated exceptional performance on text-to-motion generation task, through experimental validation on the HumanML3D [17] and KIT-ML [38] datasets. Our methodology achieved state-of-the-art generation quality and significantly improved long-squence modeling, meanwhile optimizing inference speed." + ], + "bbox": [ + 215, + 146, + 785, + 357 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2 Related Works", + "text_level": 1, + "bbox": [ + 215, + 383, + 395, + 398 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Human Motion Generation. Generating human motion is a significant application of computer vision, essential for various applications like 3D modelling and robot manipulation. Recently, the predominant method of achieving human motion generation, known as the Text-to-Motion task, involves learning a common latent space for both language and motion.", + "bbox": [ + 212, + 416, + 785, + 491 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "DVGAN [30] creates the GAN [11] discriminator by densely validating at each time-scale and perturbing the discriminator input for translation invariance, enabling motion generation and completion. ERD-QV [20] enhances latent representations through two additive modifiers: a time-to-arrival embedding applied universally and an additive scheduled target noise vector used during extended transitions. It further improves transition quality by incorporating a GAN framework with two discriminators operating at different timescales. HP-GAN [4], trained with a modified version of the improved WGAN-GP [16], utilizes a custom loss function designed for human motion prediction. It learns a probability density function of future human poses conditioned on previous poses.", + "bbox": [ + 212, + 492, + 785, + 643 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Autoencoders [27, 43] are notable generative models known for their ability to represent data robustly by compressing high-dimensional data into a latent space, which is widely adopted in human motion generation. JL2P [1] uses RNN-based autoencoders [23] to learn a combined representation of language and pose. It restricts the direct mapping from text to motion to a one-to-one relationship. MotionCLIP [48] uses Transformer-based Autoencoders [51] to reconstruct motion while ensuring alignment with the corresponding text label in the CLIP [40] space. This alignment effectively integrates the semantic knowledge from CLIP into the human motion manifold. TEMOS [36] and T2M [17] combine a Transformer-based VAE [26] with a text encoder to generate distribution parameters that work within the VAE latent space. AttT2M [56] and TM2D [10] incorporate a body-part spatio-temporal encoder into VQ-VAE [50] for enhanced learning of a discrete latent space with increased expressiveness.", + "bbox": [ + 212, + 643, + 785, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Zeyu and Akide et al.", + "bbox": [ + 271, + 114, + 416, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Diffusion models [7, 22, 41, 47] have recently surpassed GANs and VAEs in generating 2D images. Developing a motion generation model based on diffusion models is obviously an attractive direction. MotionDiffuse [54] introduces the inaugural framework for text-driven motion generation based on diffusion models. It showcases several desirable properties, including probabilistic mapping, realistic synthesis, and multi-level manipulation. MDM [49] utilizes a classifier-free Transformer-based diffusion model for the human motion domain to predict sample rather than noise in each diffusion step. MLD [6] performs a diffusion process in latent motion space, rather than using a diffusion model to establish connections between raw motion sequences and conditional inputs.", + "bbox": [ + 218, + 145, + 785, + 296 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "State Space Models. Recently, state space sequence models (SSMs) [14, 15], drawing inspiration from classical state-space models [25], have emerged as a promising architecture for sequence modeling. Mamba [12] introduces a selective SSM architecture, integrating time-varying parameters into the SSM framework, and proposes a hardware-aware algorithm to facilitate highly efficient training and inference processes. Some research works leverage SSM in computer vision to process 2D data. The 2D SSM [3] introduces an SSM block at the beginning of each transformer block [8, 51]. This approach aims to achieve efficient parameterization, accelerated computation, and a suitable normalization scheme. SGConvNeXt [29] presents a structured global convolution method inspired by ConvNeXt [32], incorporating multi-scale sub-kernels to achieve both parameterization efficiency and effective long sequence modeling. ConvSSM [46] integrates the tensor modeling principles of ConvLSTM [45] with SSMs, elucidating the utilization of parallel scans in convolutional recurrences. This approach enables subquadratic parallelization and rapid autoregressive generation. Vim [57] introduces a bidirectional SSM block [52] for efficient and versatile visual representation learning, achieving performance comparable to established ViT [8] methods. VMamba [31] introduces a Cross-Scan Module (CSM) designed to traverse the spatial domain and transform any non-causal visual image into ordered patch sequences. This approach achieves linear complexity while preserving global receptive fields. There have also been attempts to utilize SSMs to handle higher-dimensional data. Mamba-ND [28] explores various combinations of SSM and different scan directions within the SSM block to adapt Mamba [12] to higher-dimensional tasks. Recent efforts have sought to replace the traditional transformer-based U-Net within the diffusion denoiser with the SSM block, with the aim of enhancing image generation efficiency. DiffuSSM [53] adeptly manages higher resolutions without relying on global compression, thus maintaining detailed image representation throughout the diffusion process.", + "bbox": [ + 217, + 297, + 785, + 719 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 The Proposed Method", + "text_level": 1, + "bbox": [ + 217, + 744, + 470, + 762 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we delineate the architecture and operational principles of the Motion Mamba framework, designed for generating human motion over long ranges efficiently from textual descriptions. Initially, we discuss the foundational concepts underpinning our approach, including the Mamba Model [12] and the", + "bbox": [ + 217, + 779, + 785, + 839 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Motion Mamba", + "bbox": [ + 625, + 114, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "latent diffusion model [6]. Following this, we detail our uniquely crafted architecture that leverages the Mamba model to enhance motion generation efficiency. This architecture comprises two principal components: the Hierarchical Temporal Mamba (HTM) block, which addresses temporal aspects, and the Bidirectional Spatial Mamba (BSM) block, focusing on spatial dynamics.", + "bbox": [ + 212, + 146, + 787, + 223 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.1 Preliminaries", + "text_level": 1, + "bbox": [ + 215, + 244, + 372, + 258 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Selective Structured State Space Sequence Model. SSMs particularly through the contributions of structured state space sequence models (S4) and Mamba, have demonstrated exceptional proficiency in handling long sequences. These models operationalize the mapping of a 1-D function or sequence, $x(t) \\in \\mathbb{R} \\mapsto y(t) \\in \\mathbb{R}$ , through a hidden state $h(t) \\in \\mathbb{R}^N$ , employing $\\mathbf{A} \\in \\mathbb{R}^{N \\times N}$ as the evolution parameters, $\\mathbf{B} \\in \\mathbb{R}^{N \\times 1}$ and $\\mathbf{C} \\in \\mathbb{R}^{1 \\times N}$ as the projection parameters, respectively.", + "bbox": [ + 212, + 270, + 787, + 375 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The discretized system can then be expressed as follows, incorporating a step size $\\Delta$ :", + "bbox": [ + 212, + 376, + 787, + 404 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{c} h _ {t} = \\overline {{\\mathbf {A}}} h _ {t - 1} + \\overline {{\\mathbf {B}}} x _ {t}, \\\\ \\mathbf {C} ^ {\\prime} \\end{array} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 433, + 402, + 784, + 426 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\ny _ {t} = \\mathbf {C} h _ {t}.\n$$\n", + "text_format": "latex", + "bbox": [ + 434, + 422, + 504, + 436 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "This adaptation facilitates the computation of output through global convolution, leveraging a structured convolutional kernel $\\overline{\\mathbf{K}}$ , which encompasses the entire length $M$ of the input sequence $\\mathbf{x}$ :", + "bbox": [ + 212, + 444, + 787, + 491 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\overline {{\\mathbf {K}}} = \\left(\\underset {-} {\\mathbf {C}} \\overline {{\\mathbf {B}}}, \\mathbf {C} \\overline {{\\mathbf {A}}} \\mathbf {B}, \\dots , \\mathbf {C} \\overline {{\\mathbf {A}}} ^ {M - 1} \\overline {{\\mathbf {B}}}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 382, + 501, + 784, + 527 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {y} = \\mathbf {x} * \\overline {{\\mathbf {K}}}.\n$$\n", + "text_format": "latex", + "bbox": [ + 387, + 525, + 464, + 541 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Selective models like Mamba introduce time-varying parameters, deviating from the linear time invariance (LTI) assumption and complicating parallel computation. However, hardware-aware optimizations, such as associative scans, have been developed to address these computational challenges, highlighting the ongoing evolution and application of SSMs in modeling complex temporal dynamics.", + "bbox": [ + 212, + 553, + 787, + 628 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Latent Motion Diffusion Model. Diffusion probabilistic models offer a significant advancement in motion generation by gradually reducing noise from a Gaussian distribution to a target data distribution $p(x)$ through a T-length learned Markov process [7,22,41,44,47,49,54], giving $\\{\\pmb{x}_t\\}_{t=1}^T$ . In the motion generation, we define our trainable diffusion models with a denoiser $\\epsilon_{\\theta}(x_t,t)$ which anneal the random noise to motion sequence $\\{\\hat{x}_t^{1:N}\\}_{t=1}^T$ iteratively. To address the inefficiencies of applying diffusion models directly to raw motion sequences, we employ a low-dimensional motion latent space for the diffusion process. Given an input condition $c$ , such as a descriptive sentence $\\pmb{w}^{1:N} = \\{w^i\\}_{i=1}^N$ , an action label $a$ from a predefined set $\\mathcal{A}$ , or an empty condition $c = \\emptyset$ , and the motion representation that combines 3D joint rotations, positions, velocities, and foot contact as proposed in [17]. The frozen CLIP [40] text encoder $\\tau_{\\theta}^w$ has been employed to obtain projected text embedding $\\tau_{\\theta}^w(w^{1:N}) \\in \\mathbb{R}^{1 \\times d}$ , thereby conditional denoiser comprised in term of $\\epsilon_{\\theta}(z_t,t,\\tau_{\\theta}(c))$ . The latent diffusion model $\\epsilon_{\\theta}(x_t,t)$ aimed", + "bbox": [ + 212, + 628, + 787, + 843 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Zeyu and Akide et al.", + "bbox": [ + 271, + 114, + 418, + 128 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c47e35af41944d8524c810593fb0260c0e75549c04159ef1f62bdcc5c0d563b4.jpg", + "image_caption": [ + "Fig. 2: This figure illustrates the architecture of the proposed Motion Mamba model. Each of encoder and decoder blocks consists of a Hierarchical Temporal Mamba block (HTM) and a Bidirectional Spatial Mamba (BSM) block, which possess hierarchical scan and bidirectional scan within SSM layers respectively. This symmetric distribution of scans ensure a balanced and coherence framework across the encoder-decoder architecture." + ], + "image_footnote": [], + "bbox": [ + 212, + 143, + 787, + 359 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "to generate the human motion sequence in terms of $\\hat{x}^{1:L} = \\{\\hat{x}^i\\}_{i=1}^L$ , where L denotes the sequence length or number of frames [34,35,37,55]. After that we reused the motion Variational AutoEncoder (VAE) $\\mathcal{V} = \\{\\mathcal{E},\\mathcal{D}\\}$ proposed in MLD [6] to manipulate the motion sequence in latent space $z = \\mathcal{E}(x^{1:L})$ , and decompress the intermediate representation to motion sequence by $\\hat{x}^{1:L} = \\mathcal{D}(z) = \\mathcal{D}\\mathcal{E}(x^{1:L})$ [26, 42, 51]. Finally, our latent diffusion model is trained with an objective focusing on minimization of MSE between true and predicted noise in the latent space, facilitating efficient and high-quality motion generation [2, 22].", + "bbox": [ + 212, + 468, + 787, + 592 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.2 Motion Mamba", + "text_level": 1, + "bbox": [ + 215, + 611, + 392, + 625 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The architecture of the proposed Motion Mamba framework is illustrated in Figure. 2. At its core, Motion Mamba utilizes a denoising U-Net architecture, which is distinguished for its effectiveness in modeling the continuous, temporal sequences of motion frames. This effectiveness is attributed to the inherent long-sequence modeling capacity of the Mamba model. The denoiser, denoted by $\\epsilon_{\\theta}$ , comprises $N$ blocks including encoder $E_{1..N}$ and decoder $D_{1..N}$ . Additionally, the architecture is enhanced with a transformer-based attention mixer block $M$ , designed to augment the model's ability to capture complex temporal dynamics.", + "bbox": [ + 212, + 636, + 787, + 758 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\epsilon_ {\\theta} (x) \\equiv \\{E _ {1 \\dots N}, M, D _ {1 \\dots N} \\}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 401, + 771, + 785, + 789 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The encoder blocks are represented as $E_{1..N}$ , arranged sequentially, and the decoder blocks as $D_{1..N}$ , configured in reverse order to facilitate effective bottom-up and top-down information flow. Given that selective operations have signif", + "bbox": [ + 212, + 794, + 787, + 842 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Motion Mamba", + "bbox": [ + 625, + 114, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 Hierarchical Temporal Mamba (HTM) Block." + ], + "code_body": "Require: compressed latent representations $z:(T,B,C)$ \nEnsure: transformed representations $z_{\\mathrm{HTM}}:(T,B,E)$ \n1:/\\* linear projection layer \\*/ \n2: $x,z:(T,B,E)\\gets \\mathrm{Linear}(z)$ \n3:/\\* set of scans and memory matrices \\*/ \n4: $K = \\{S_{2N_n - 1},S_{2N_{n - 1} - 1},\\ldots ,S_1\\}$ \n5: Memory matrices: $\\{A_1,\\dots ,A_k\\}$ \n6: for each scan $S_{i}$ in $K$ do \n7: $x_o^{\\prime}:(T,B,E)\\gets \\mathrm{Conv1D}(x)$ \n8: $B_{o},C_{o},\\varDelta_{o}\\gets \\mathrm{Linear}(x_{o}^{\\prime})$ \n9: Transform $\\overline{A}_o$ and $\\overline{B}_o$ using $\\varDelta_{o}$ \n10: $O_{i}\\gets \\mathrm{SSMA}_{i,x}(x_{o}^{\\prime})$ \n11: end for \n12:/\\* aggregation of outputs \\*/ \n13: $z_{\\mathrm{HTM}}:(T,B,E)\\gets \\mathrm{Linear(Aggregate(\\{O_1,\\ldots,O_k\\}))}$ \n14: Return: $z_{\\mathrm{HTM}}$", + "bbox": [ + 215, + 166, + 602, + 391 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "icantly lower computational complexity compared to attention-based methods, we have increased the number of scans to achieve higher quality generations. Concurrently, it is imperative to maintain a balance between the model's parameters and its efficiency. Thereby, a novel aspect of our model is the introduction of a hierarchical scan strategy, characterized by a sequence of scan numbers as,", + "bbox": [ + 212, + 421, + 787, + 497 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nK = \\left\\{S _ {2 N - 1}, S _ {2 (N - 1) - 1}, \\dots , S _ {1} \\right\\}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 379, + 507, + 785, + 525 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This sequence specifies the number of scans allocated to each layer, in descending order of complexity. For instance, the uppermost encoder layer, $E_{1}$ , and the lowermost decoder layer, $D_{N}$ , are allocated $S_{2N - 1}$ scans, indicating the highest scanning complexity. Conversely, the lowest encoder layer, $E_{N}$ , and the uppermost decoder layer, $D_{1}$ , are assigned $S_{1}$ scans, reflecting the lowest level of scanning complexity.", + "bbox": [ + 212, + 534, + 787, + 626 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nE _ {i} (S) = \\left\\{ \\begin{array}{l l} S _ {2 N - 1} & \\text {f o r} i = 1 \\\\ S _ {2 (N - i) - 1} & \\text {f o r} i = 2, \\dots , N - 1 \\\\ S _ {1} & \\text {f o r} i = N \\end{array} \\right. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 344, + 635, + 785, + 691 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nD _ {j} (S) = \\left\\{ \\begin{array}{l l} S _ {2 N - 1} & \\text {f o r} j = N \\\\ S _ {2 (N - j) - 1} & \\text {f o r} j = N - 1, \\dots , 2 \\\\ S _ {1} & \\text {f o r} j = 1 \\end{array} \\right. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 341, + 703, + 785, + 760 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This hierarchical scanning approach ensures that processing capabilities are evenly distributed throughout the encoder-decoder architecture., facilitating a detailed and nuanced analysis of temporal sequences. Within this structured framework, each denoiser is equipped with a specialized Hierarchical Temporal Mamba (HTM) block, which serves to augment the model's ability to process", + "bbox": [ + 212, + 763, + 787, + 842 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Zeyu and Akide et al.", + "bbox": [ + 271, + 114, + 418, + 128 + ], + "page_idx": 7 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 2 Bidirectional Spatial Mamba (BSM) Block." + ], + "code_body": "Require: compressed latent representations $z:(T,B,C)$ \nEnsure: transformed representations $z_{\\mathrm{BSM}}:(C,B,E)$ \n1:/\\* dimension rearrangement \\*/ \n2: $z^{\\prime}:(C,B,T)\\gets$ Rearrange(z) \n3:/\\* linear projection after normalization \\*/ \n4: $z^{\\prime}:(C,B,T)\\gets$ Norm(z) \n5:x,z:(C,B,E) $\\leftarrow$ Linear(z') \n6: for o in {forward, backward} do \n7: $x_o':(C,B,E)\\gets$ Conv1D(x) \n8: $B_{o},C_{o},\\varDelta_{o}\\gets$ Linear $(x_o^{\\prime})$ \n9: Transform $\\overline{A}_o$ and $\\overline{B}_o$ using $\\varDelta_{o}$ \n10: $y_{o}\\gets$ SSM(Ao,Bo,Co) \n11: end for \n12:/\\* gating and summing outputs \\*/ \n13: $z_{\\mathrm{BSM}}:(C,B,T)\\gets$ GateAndSum(yforward,ybackward,z) \n14: $z_{\\mathrm{BSM}}^{\\prime}:(T,B,C)\\gets$ Rearrange(z) \n15: Return: $z_{\\mathrm{BSM}}^{\\prime}$", + "bbox": [ + 215, + 166, + 619, + 404 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "temporal information effectively. Additionally, the proposed Motion Mamba incorporates an attention-based mixer block denoted as $M$ , strategically integrated to enhance conditional fusion.", + "bbox": [ + 212, + 436, + 784, + 479 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Hierarchical Temporal Mamba (HTM) block processes compressed latent representations, denoted as $z$ , with the dimensions $(T,B,C)$ , of which procedure shown in Algorithm 1. Here, $T$ signifies the temporal dimension, as specified in the Variational AutoEncoder (VAE) framework. Initially, the input $z$ is subjected to a linear projection layer, producing transformed representations $x$ and $z$ with dimension $E$ . Our analysis revealed an increased density of motion within the lower-level feature spaces. Consequently, we developed a hierarchical scanning methodology that is executed at various depths of the network. This approach not only accommodates the diverse motion densities, but also significantly reduces computational overhead. This step utilizes a hierarchically structured set of scans, $K = \\{S_{2N_n - 1}, S_{2N_{n - 1} - 1}, \\ldots, S_1\\}$ , in conjunction with a corresponding series of memory matrices $\\{A_1, \\ldots, A_k\\}$ . Each sub-SSM scan first applies a 1-D convolution to $x$ , resulting in $x_o'$ . $x_o'$ is then linearly projected to derive $B_o$ , $C_o$ , and $\\Delta_o$ . These projections $B_o$ , $C_o$ use $\\Delta_o$ to effect transformations in $\\overline{A}_o$ and $\\overline{B}_o$ , respectively. After executing a sequence of SSM scans $\\{SSM_{A_1,x}, SSM_{A_2,x}, \\ldots, SSM_{A_k,x}\\}$ , a set of outputs $\\{O_1, \\ldots, O_k\\}$ is compiled. This collection is subsequently aggregated via a linear projection to obtain the final output of the HTM block.", + "bbox": [ + 212, + 482, + 787, + 753 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Bidirectional Spatial Mamba (BSM) block focuses on enhancing latent representation learning through a novel approach of dimension rearrangement and bidirectional scanning, of which the process is shown in Algorithm 2. Initially, it alters the original input dimensions from $(T,B,C)$ to $(C,B,T)$ , effectively swap", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Motion Mamba", + "bbox": [ + 625, + 114, + 730, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "ping the temporal and channel dimensions. After this rearrangement, the input, now denoted $z'$ , undergoes a linear projection after normalization, resulting in dimensions $x$ and $z$ of size $E$ . The process involves bidirectional scanning of the latent channel dimension, where $\\mathbf{x}$ is subjected to a 1-D convolution, yielding $\\mathbf{x}_o'$ for both forward and backward directions. Each $\\mathbf{x}_o'$ is then linearly projected to obtain $B_o$ , $C_o$ , and $\\Delta_o$ , which are utilized to transform $\\overline{A}_o$ and $\\overline{B}_o$ , respectively. The final output token sequence, $\\mathbf{z}_1$ , is computed by gating and summing the forward $y_{\\mathrm{forward}}$ and backward $y_{\\mathrm{backward}}$ output with $\\mathbf{z}$ . This component is engineered to decode the structured latent skeleton by analyzing data from both forward and reverse viewpoints. Its main objective is to ensure the seamless continuity of information flow, thereby substantially enhancing the model's ability to generate accurate motion. This is achieved through the maintenance of a dense informational exchange, which is critical for the model's performance.", + "bbox": [ + 212, + 146, + 787, + 343 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 362, + 375, + 380 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.1 Datasets", + "text_level": 1, + "bbox": [ + 215, + 390, + 334, + 402 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We evaluate our proposed Motion Mamba on two prominent Text-to-Motion synthesis benchmarks as follows:", + "bbox": [ + 212, + 412, + 785, + 440 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "HumanML3D. The HumanML3D [17] dataset aggregates 14,616 motions sourced from the AMASS [33] and HumanAct12 [18] datasets, with each motion accompanied by three textual descriptions, culminating in 44,970 scripts. This dataset spans a wide range of actions such as exercising, dancing, and acrobatics, presenting a rich motion-language corpus.", + "bbox": [ + 212, + 441, + 797, + 516 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "KIT-ML. The KIT-ML dataset [38] is comprised of 3,911 motions paired with 6,278 textual descriptions, serving as a compact yet effective benchmark for evaluation. For both datasets, the pose representation adopted is derived from T2M [17], ensuring consistency in motion representation across evaluations.", + "bbox": [ + 212, + 517, + 785, + 579 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 Evaluation Metrics", + "text_level": 1, + "bbox": [ + 215, + 595, + 419, + 611 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We adapt the standard evaluation metrics on following aspects throughout our experiments, including: Generation Quality. We implement a Fréchet inception distance (FID) [21] to quantify the realism and diversity of motion generated by models. Moreover, we use multi-modal distance (MM Dist) to measure the distance between motions and texts and assess motion-text alignment. Diversity. We use the diversity metric to measure motion diversity, which calculates variance in features extracted from the motions. Additionally, we employ multimodality (MModality) to assess diversity within generated motions sharing the same text description.", + "bbox": [ + 212, + 618, + 787, + 753 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.3 Comparative Studies", + "text_level": 1, + "bbox": [ + 215, + 772, + 434, + 787 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We evaluate our method against the state-of-the-art methods on the HumanML3D [17] and KIT-ML [38] datasets. We train our Motion Mamba with HTM arrangement strategy MM ( $\\{S_{2N_n - 1},\\ldots ,S_1\\}$ ), BSM bidirectional block strategy on the", + "bbox": [ + 212, + 794, + 794, + 842 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Zeyu and Akide et al.", + "bbox": [ + 271, + 114, + 416, + 128 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/f22668e60db39236e4993bff3b2a318b4fa3c2d281126a45986137457738618f.jpg", + "table_caption": [ + "Table 1: Comparison of text-conditional motion synthesis on HumanML3D [17]. These metrics are evaluated by the motion encoder from [17]. Empty MModality indicates the non-diverse generation methods. We employ real motion as a reference and sort all methods by descending FIDs. The right arrow $\\rightarrow$ means that the closer to the real motion, the better. Bold and underline indicate the best and second best result." + ], + "table_footnote": [], + "table_body": "
MethodR Precision ↑FID↓MM Dist↓Diversity→MModality↑
Top 1Top 2Top 3
Real0.511±.0030.703±.0030.797±.0020.002±.0002.974±.0089.503±.065-
Seq2Seq [39]0.180±.0020.300±.0020.396±.00211.75±.0355.529±.0076.223±.061-
LJ2P [1]0.246±.0010.387±.0020.486±.00211.02±.0465.296±.0087.676±.058-
T2G [5]0.165±.0010.267±.0020.345±.0027.664±.0306.030±.0086.409±.071-
Hier [9]0.301±.0020.425±.0020.552±.0046.532±.0245.012±.0188.332±.042-
TEMOS [37]0.424±.0020.612±.0020.722±.0023.734±.0283.703±.0088.973±.0710.368±.018
T2M [17]0.457±.0020.639±.0030.740±.0031.067±.0023.340±.0089.188±.0022.090±.083
MDM [49]0.320±.0050.498±.0040.611±.0070.544±.0445.566±.0279.559±.0862.799±.072
MotionDiffuse [54]0.491±.0010.681±.0010.782±.0010.630±.0013.113±.0019.410±.0491.553±.042
MLD [6]0.481±.0030.673±.0030.772±.0020.473±.0133.196±.0109.724±.0822.413±.079
Motion Mamba (Ours)0.502±.0030.693±.0020.792±.0020.281±.0093.060±.0589.871±.0842.294±.058
", + "bbox": [ + 215, + 213, + 785, + 380 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/d4a6902799c96d2278dd1a098881e5ed26a9fd074019736e497d0d9b158b0591.jpg", + "table_caption": [ + "Table 2: We involve KIT-ML [38] dataset and evaluate the SOTA methods on the text-to-motion task." + ], + "table_footnote": [], + "table_body": "
MethodR Precision ↑FID↓MM Dist↓Diversity→MModality↑
Top 1Top 2Top 3
Real0.424±.0050.649±.0060.779±.0060.031±.0042.788±.01211.08±.097-
Seq2Seq [39]0.103±.0030.178±.0050.241±.00624.86±.3487.960±.0316.744±.106-
T2G [5]0.156±.0040.255±.0040.338±.00512.12±.1836.964±.0299.334±.079-
LJ2P [1]0.221±.0050.373±.0040.483±.0056.545±.0725.147±.0309.073±.100-
Hier [9]0.255±.0060.432±.0070.531±.0075.203±.1074.986±.0279.563±.0722.090±.083
TEMOS [37]0.353±.0060.561±.0070.687±.0053.717±.0513.417±.01910.84±.1000.532±.034
T2M [17]0.370±.0050.569±.0070.693±.0072.770±.1093.401±.00810.91±.1191.482±.065
MDM [49]0.164±.0040.291±.0040.396±.0040.497±.0219.191±.02210.85±.1091.907±.214
MotionDiffuse [54]0.417±.0040.621±.0040.739±.0041.954±.0622.958±.00511.10±.1430.730±.013
MLD [6]0.390±.0080.609±.0080.734±.0070.404±.0273.204±.02710.80±.1172.192±.071
Motion Mamba (Ours)0.419±.0060.645±.0050.765±.0060.307±.0413.021±.02511.02±.0981.678±.064
", + "bbox": [ + 215, + 414, + 785, + 583 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "latent dimension $= 2$ with 11 layers. We evaluate our model and previous works with suggested metrics in HumanML3D [17] and calculate $95\\%$ confidence interval by repeat evaluation 20 times. The results for the HumanML3D dataset are presented in Table 1. Our model outperforms other methods significantly across various evaluation metrics, including FID, R precision, multi-modal distance, and diversity. For instance, our Motion Mamba outperforms previous best diffusion based motion generation MLD by $40.5\\%$ in terms of FID, and up to $10\\%$ improvement on R Precision, we aslo obatined best MModality by 3.060. The results for the KIT-ML dataset are presented in Table 2. We have also outperformed other well-established methods in FID and multi-modal distance.", + "bbox": [ + 212, + 595, + 787, + 748 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.4 Ablation Studies", + "text_level": 1, + "bbox": [ + 215, + 768, + 401, + 782 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We concluded the ablation studies including long sequence evaluation, hierarchical design with HTM, bidirectional design in the BSM, number of latent dimensions, and number of layers of our proposed motion mamba in Table 4.", + "bbox": [ + 212, + 792, + 785, + 838 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Motion Mamba", + "bbox": [ + 625, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/f10dee507e46e0ced2ba07214fcf83530aab51e42cf950b0ed7dd0baa6212667.jpg", + "table_caption": [ + "Table 3: In order to evaluate the models' capability in long sequence motion generation, we compared our method with an existing approach on the recently introduced HumanML3D-LS dataset. This dataset comprises motion sequences longer than 190 frames from the original evaluation set. Our model demonstrates superior performance compared to other methods." + ], + "table_footnote": [], + "table_body": "
MethodR Precision ↑FID↓MM Dist↓Diversity→MModality↑
Top 1Top 2Top 3
Real0.437±.0030.622±.0040.721±.0040.004±.0003.343±.0158.423±.090-
MDM [49]0.368±.0050.553±.0060.672±.0050.802±.0443.860±.0258.817±.068-
MotionDiffuse [54]0.367±.0040.521±.0040.623±.0042.460±.0623.789±.0058.707±.1431.602±.013
MLD [6]0.403±.0050.584±.0050.690±.0050.952±.0203.580±.0169.050±.0852.711±.104
Motion Mamba (Ours)0.417±.0030.606±.0030.713±.0040.668±.0193.435±.0159.021±.0702.373±.084
", + "bbox": [ + 215, + 212, + 785, + 316 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Long Sequence Motion Generation. The HumanML3D [17] dataset exhibits a long-tailed and right-skewed distribution with a significant proportion of long-sequence human motions, as shown in Figure 3. We suggest previous study overlooked the challenges in the long-sequence generation problem. Thus, we introduce a new dataset variant, HumanML3D-LS, comprising motion sequences longer than 190 frames extracted from the original test set. This addition allows us to showcase our capability in generating long-sequence motions. Subsequently, we eval", + "bbox": [ + 212, + 335, + 544, + 532 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/aae2391766f3828f2e11babdcf89aea649a1a4ec6a3b1b4196a8a1944d3acb76.jpg", + "image_caption": [ + "Fig. 3: The figure shows a long tail distribution of the HumanML3D [17], which has a significant proportion of long-sequence human motions." + ], + "image_footnote": [], + "bbox": [ + 555, + 334, + 787, + 443 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "uate the performance of our method on HumanML3D-LS and compare it with other diffusion-based motion generation approaches. The comparative results are presented in Table 3. Motion Mamba by leverage the benefits on long-range dependency modeling make it well suitable for long sequence motion generation.", + "bbox": [ + 212, + 532, + 787, + 594 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Hierarchical Design with HTM. In our ablation studies, we observed a slight improvement upon reversing the scan order from a lower to a higher level, specifically transitioning from MM $\\{S_1,\\ldots ,S_N\\}$ to MM $\\{S_N,\\dots ,S_1\\}$ . This enhancement suggests a correlation with the increase in temporal motion density within the lower-level feature spaces. Furthermore, to achieve the optimal result, we introduce the hierarchical design to arrange the scanning frequency, resulting in the sequence MM $\\{S_{2N_n - 1},\\ldots ,S_1\\}$ . This expansion in the number of scans led to a performance increase. We attribute this enhancement to the observation that individual selective scan operations significantly reduce the parameter count, especially when compared to the parameter-intensive constructs of self-attention and feedforward network blocks prevalent in transformer architectures.", + "bbox": [ + 212, + 595, + 787, + 762 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Bidirectional Design in BSM. We developed three distinct variations of latent scanning mechanisms, differentiated by their scanning directions. In the context of motion generation tasks, we posit that the flow of hidden information within the structured latent skeleton holds significance, an aspect previously underexplored. Our ablation study reveals that a single scan across the la", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Zeyu and Akide et al.", + "bbox": [ + 271, + 114, + 416, + 128 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/91cc7db9915584cb2721c037bc8a8073a2e1c0bf79422ad9a14143816aef9ff8.jpg", + "table_caption": [ + "Table 4: Evaluation of text-based motion synthesis on HumanML3D [17]: we use metrics in Table 1 and provides real reference, we evaluate the various HTM and BSM design choices, the dimension of the latent input, the different number of layer of Motion Mamba model." + ], + "table_footnote": [], + "table_body": "
ModelsR Precision Top 3↑FID↓MM Dist.↓Diversity→MModality↑
Real0.797±.0020.002±.0002.974±.0089.503±.065-
MM ( {S1, ..., SN})0.673±.0031.278±.0123.802±.0418.678±.0963.127±.024
MM ( {SN, ..., S1})0.738±.0020.962±.0113.433±.0039.180±.0712.723±.033
MM ( {S1, ..., S2Nn-1})0.698±.0020.856±.0083.624±.0379.229±.0672.826±.017
MM ( {S2Nn-1, ..., S1})0.792±.0020.281±.0093.060±.0589.871±.0842.294±.058
MM (SingleScan)0.736±.0031.063±.0103.443±.0269.180±.0672.676±.041
MM (BiScan, layer)0.735±.0040.789±.0073.408±.0349.374±.0592.591±.046
MM (BiScan, block)0.792±.0020.281±.0093.060±.0589.871±.0842.294±.058
MM (Dim, 1)0.706±.0030.652±.0113.541±.0729.141±.0822.612±.055
MM (Dim, 2)0.792±.0020.281±.0093.060±.0589.871±.0842.294±.058
MM (Dim, 5)0.741±.0080.728±.0093.307±.0279.427±.0992.314±.062
MM (Dim, 7)0.738±.0040.599±.0073.359±.0689.166±.0752.488±.037
MM (Dim, 10)0.715±.0030.628±.0083.548±.0439.200±.0752.884±.096
MM (9 layers)0.755±.0021.080±.0123.309±.0579.721±.0812.974±.039
MM (11 layers)0.792±.0020.281±.0093.060±.0589.871±.0842.294±.058
MM (27 layers)0.750±.0030.975±.0083.336±.0969.249±.0712.821±.063
MM (37 layers)0.754±.0050.809±.0103.338±.0619.355±.0622.741±.077
", + "bbox": [ + 267, + 199, + 728, + 444 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "tent dimension yields minimal improvement. Subsequently, we investigated both layer-based and block-based bidirectional scans. Our findings indicate that the block-based bidirectional scan achieves optimal performance. This suggests that spatial information flows are encoded within the latent spaces and that bidirectional scanning facilitates the exchange of this information, thereby enhancing the efficacy of motion generation tasks.", + "bbox": [ + 212, + 460, + 787, + 551 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Architecture Design for Motion Mamba. The proposed Motion Mamba which is grounded in a standardized motion latent diffusion system. We delved into the interplay between dimensional aspects and the module's capacity (measured by the number of layers) to ascertain their impact on system performance. Experimental results demonstrate that the Motion Mamba achieves superior performance at a latent dimension of 2, diverging from prior works where the optimal dimension was identified as 1. We attribute this discrepancy to our HTM, which necessitates multiple scans correlating with the sequence length, thus implicating dimensionality as a pivotal factor. A marginal increase in dimensionality enabled us to attain peak performance, simultaneously enhancing efficiency compared to models with a dimensionality of 10. Furthermore, we conducted experiments to determine the optimal layer count for Motion Mamba, inspired by the design of its selective scanning mechanism. Notably, a single Mamba layer achieves a parameter reduction of approximately $75\\%$ compared to a conventional transformer encoder block. By increasing the number of layers, we aim to uncover the relationship between model capacity and its performance. Our findings reveal that, through the integration of our specially designed HTM and BSM (Bidirectional Scanning Module) blocks, the Motion Mamba reaches its optimal performance with 11 layers. This represents a slight increase over the MLD [6] baseline. How-", + "bbox": [ + 212, + 551, + 787, + 843 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Motion Mamba", + "bbox": [ + 625, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "ever, due to the reduced parameter count in each layer, Motion Mamba exhibits significantly greater efficiency than previous methodologies.", + "bbox": [ + 212, + 146, + 784, + 176 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "4.5 Inference Time", + "text_level": 1, + "bbox": [ + 215, + 199, + 387, + 213 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Inference time remains a significant challenge for diffusion-based methods. To address this, we enhance the inference speed by incorporating the efficient Mamba block within a lightweight architecture. Compared to the previous strong baseline, such as the MLD model cited in [6], which reports an average inference time of 0.217 seconds, our Motion Mamba model achieves a notable reduction in computational overhead, as shown in Figure 4. Specifically, it requires four times less computational effort, thereby facilitating faster and real-time inference speeds.", + "bbox": [ + 212, + 224, + 517, + 436 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/3f16a76fbbc0786783ac633696339787bd4ab38a49ca1a6d7338b047d5fb89d2.jpg", + "image_caption": [ + "Fig.4: The figure shows the average inference time per sentence (AIT) vs FID, our proposed motion mamba obtained 0.058s AIT and 0.281 FID over all outperform previous methods. We evaluate all methods on a single V100 GPU." + ], + "image_footnote": [], + "bbox": [ + 537, + 220, + 764, + 353 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Discussion and Conclusion", + "text_level": 1, + "bbox": [ + 215, + 458, + 509, + 474 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this study, we introduced Motion Mamba, a novel framework designed for efficient and extended sequence motion generation. Our approach represents the inaugural integration of the Mamba model within the domain of motion generation, featuring significant advancements including the implementation of Hierarchical Temporal Mamba (HTM) blocks. These blocks are specifically engineered to enhance temporal alignment through hierarchically organized selective scanning. Furthermore, Bidirectional Spatial Mamba (BSM) blocks have been developed to amplify the exchange of information flow within latent spaces, thereby augmenting the model's ability to bidirectionally capture skeleton-level density features with greater precision. Compared to previous diffusion-based motion generation methodologies that predominantly utilize transformer blocks, our Motion Mamba framework achieves SOTA performance, evidencing an improvement of up to $50\\%$ in FID scores and a quadrupled improvement in inference speed. Through comprehensive experimentation across a variety of human motion generation tasks, the effectiveness and efficiency of our proposed Motion Mamba model have been robustly demonstrated, marking a significant leap forward in the field of human motion generation.", + "bbox": [ + 212, + 489, + 787, + 750 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 215, + 771, + 403, + 787 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "This work is partially supported by the Fundamental Research Funds for the Central Universities, Peking University.", + "bbox": [ + 212, + 803, + 785, + 834 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Zeyu and Akide et al.", + "bbox": [ + 271, + 114, + 416, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Ahuja, C., Morency, L.P.: Language2pose: Natural language grounded pose forecasting. In: 2019 International Conference on 3D Vision (3DV). pp. 719-728. IEEE (2019)", + "2. Bao, F., Li, C., Cao, Y., Zhu, J.: All are worth words: a vit backbone for score-based diffusion models. arXiv preprint arXiv:2209.12152 (2022)", + "3. Baron, E., Zimerman, I., Wolf, L.: 2-d ssm: A general spatial layer for visual transformers. arXiv preprint arXiv:2306.06635 (2023)", + "4. Barsoum, E., Kender, J., Liu, Z.: Hp-gan: Probabilistic 3d human motion prediction via gan. In: Proceedings of the IEEE conference on computer vision and pattern recognition workshops. pp. 1418-1427 (2018)", + "5. Bhattacharya, U., Rewkowski, N., Banerjee, A., Guhan, P., Bera, A., Manocha, D.: Text2gestures: A transformer-based network for generating emotive body gestures for virtual agents. In: 2021 IEEE virtual reality and 3D user interfaces (VR). pp. 1-10. IEEE (2021)", + "6. Chen, X., Jiang, B., Liu, W., Huang, Z., Fu, B., Chen, T., Yu, G.: Executing your commands via motion diffusion in latent space. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18000-18010 (2023)", + "7. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems 34, 8780-8794 (2021)", + "8. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. In: International Conference on Learning Representations (2020)", + "9. Ghosh, A., Cheema, N., Oguz, C., Theobalt, C., Slusallek, P.: Synthesis of compositional animations from textual descriptions. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 1396-1406 (2021)", + "0. Gong, K., Lian, D., Chang, H., Guo, C., Jiang, Z., Zuo, X., Mi, M.B., Wang, X.: Tm2d: Bimodality driven 3d dance generation via music-text integration. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9942-9952 (2023)", + "1. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial nets. Advances in neural information processing systems 27 (2014)", + "2. Gu, A., Dao, T.: Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752 (2023)", + "3. Gu, A., Goel, K., Gupta, A., Ré, C.: On the parameterization and initialization of diagonal state space models. Advances in Neural Information Processing Systems 35, 35971-35983 (2022)", + "4. Gu, A., Goel, K., Ré, C.: Efficiently modeling long sequences with structured state spaces. arXiv preprint arXiv:2111.00396 (2021)", + "5. Gu, A., Johnson, I., Goel, K., Saab, K., Dao, T., Rudra, A., Ré, C.: Combining recurrent, convolutional, and continuous-time models with linear state space layers. Advances in neural information processing systems 34, 572-585 (2021)", + "6. Gulrajani, I., Ahmed, F., Arjovsky, M., Dumoulin, V., Courville, A.C.: Improved training of wasserstein gans. Advances in neural information processing systems 30 (2017)", + "7. Guo, C., Zou, S., Zuo, X., Wang, S., Ji, W., Li, X., Cheng, L.: Generating diverse and natural 3d human motions from text. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5152-5161 (2022)" + ], + "bbox": [ + 225, + 172, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Motion Mamba", + "bbox": [ + 625, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "18. Guo, C., Zuo, X., Wang, S., Zou, S., Sun, Q., Deng, A., Gong, M., Cheng, L.: Action2motion: Conditioned generation of 3d human motions. In: Proceedings of the 28th ACM International Conference on Multimedia. pp. 2021-2029 (2020)", + "19. Gupta, A., Gu, A., Berant, J.: Diagonal state spaces are as effective as structured state spaces. Advances in Neural Information Processing Systems 35, 22982-22994 (2022)", + "20. Harvey, F.G., Yurick, M., Nowrouzezahrai, D., Pal, C.: Robust motion in-between. ACM Transactions on Graphics (TOG) 39(4), 60-1 (2020)", + "21. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems 30 (2017)", + "22. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems 33, 6840-6851 (2020)", + "23. Hopfield, J.J.: Neural networks and physical systems with emergent collective computational abilities. Proceedings of the national academy of sciences 79(8), 2554-2558 (1982)", + "24. Jiang, B., Chen, X., Liu, W., Yu, J., Yu, G., Chen, T.: Motiongpt: Human motion as a foreign language. Advances in Neural Information Processing Systems 36 (2024)", + "25. Kalman, R.E.: A New Approach to Linear Filtering and Prediction Problems. Journal of Basic Engineering 82(1), 35-45 (1960)", + "26. Kingma, D.P., Welling, M.: Auto-encoding variational bayes. stat 1050, 1 (2014)", + "27. Kramer, M.A.: Nonlinear principal component analysis using autoassociative neural networks. AIChE journal 37(2), 233-243 (1991)", + "28. Li, S., Singh, H., Grover, A.: Mamba-nd: Selective state space modeling for multidimensional data. arXiv preprint arXiv:2402.05892 (2024)", + "29. Li, Y., Cai, T., Zhang, Y., Chen, D., Dey, D.: What makes convolutional models great on long sequence modeling? In: The Eleventh International Conference on Learning Representations (2022)", + "30. Lin, X., Amer, M.R.: Human motion modeling using dvgans. arXiv preprint arXiv:1804.10652 (2018)", + "31. Liu, Y., Tian, Y., Zhao, Y., Yu, H., Xie, L., Wang, Y., Ye, Q., Liu, Y.: Vmamba: Visual state space model. arXiv preprint arXiv:2401.10166 (2024)", + "32. Liu, Z., Mao, H., Wu, C.Y., Feichtenhofer, C., Darrell, T., Xie, S.: A convnet for the 2020s. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11976-11986 (2022)", + "33. Mahmood, N., Ghorbani, N., Troje, N.F., Pons-Moll, G., Black, M.J.: Amass: Archive of motion capture as surface shapes. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5442-5451 (2019)", + "34. Pavlakos, G., Choutas, V., Ghorbani, N., Bolkart, T., Osman, A.A.A., Tzionas, D., Black, M.J.: Expressive body capture: 3d hands, face, and body from a single image. In: Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR) (2019)", + "35. Petrovich, M., Black, M.J., Varol, G.: Action-conditioned 3D human motion synthesis with transformer VAE. In: International Conference on Computer Vision (ICCV) (2021)", + "36. Petrovich, M., Black, M.J., Varol, G.: Temos: Generating diverse human motions from textual descriptions. In: European Conference on Computer Vision. pp. 480-497. Springer (2022)" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Zeyu and Akide et al.", + "bbox": [ + 271, + 114, + 416, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "37. Petrovich, M., Black, M.J., Varol, G.: TEMOS: Generating diverse human motions from textual descriptions. In: European Conference on Computer Vision (ECCV) (2022)", + "38. Plappert, M., Mandery, C., Asfour, T.: The kit motion-language dataset. Big data 4(4), 236-252 (2016)", + "39. Plappert, M., Mandery, C., Asfour, T.: Learning a bidirectional mapping between human whole-body motion and natural language using deep recurrent neural networks. Robotics and Autonomous Systems 109, 13-26 (2018)", + "40. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)", + "41. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022)", + "42. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: International Conference on Medical image computing and computer-assisted intervention. pp. 234-241. Springer (2015)", + "43. Rumelhart, D.E., Hinton, G.E., Williams, R.J.: Learning internal representations by error propagation. Parallel Distributed Processing pp. 318-362 (1986)", + "44. Sahara, C., Ho, J., Chan, W., Salimans, T., Fleet, D.J., Norouzi, M.: Image superresolution via iterative refinement. IEEE Transactions on Pattern Analysis and Machine Intelligence (2022)", + "45. Shi, X., Chen, Z., Wang, H., Yeung, D.Y., Wong, W.K., Woo, W.c.: Convolutional LSTM network: A machine learning approach for precipitation nowcasting. Advances in neural information processing systems 28 (2015)", + "46. Smith, J., De Mello, S., Kautz, J., Linderman, S., Byeon, W.: Convolutional state space models for long-range spatiotemporal modeling. Advances in Neural Information Processing Systems 36 (2024)", + "47. Sohl-Dickstein, J., Weiss, E., Maheswaranathan, N., Ganguli, S.: Deep unsupervised learning using nonequilibrium thermodynamics. In: International conference on machine learning. pp. 2256-2265. PMLR (2015)", + "48. Tevet, G., Gordon, B., Hertz, A., Bermano, A.H., Cohen-Or, D.: Motionclip: Exposing human motion generation to clip space. In: European Conference on Computer Vision. pp. 358-374. Springer (2022)", + "49. Tevet, G., Raab, S., Gordon, B., Shafir, Y., Cohen-or, D., Bermano, A.H.: Human motion diffusion model. In: The Eleventh International Conference on Learning Representations (2022)", + "50. Van Den Oord, A., Vinyals, O., et al.: Neural discrete representation learning. Advances in neural information processing systems 30 (2017)", + "51. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017)", + "52. Wang, J., Yan, J.N., Gu, A., Rush, A.M.: Pretraining without attention. arXiv preprint arXiv:2212.10544 (2022)", + "53. Yan, J.N., Gu, J., Rush, A.M.: Diffusion models without attention. arXiv preprint arXiv:2311.18257 (2023)", + "54. Zhang, M., Cai, Z., Pan, L., Hong, F., Guo, X., Yang, L., Liu, Z.: Motiondiffuse: Text-driven human motion generation with diffusion model. IEEE Transactions on Pattern Analysis and Machine Intelligence (2024)" + ], + "bbox": [ + 212, + 146, + 787, + 840 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Motion Mamba", + "bbox": [ + 625, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "55. Zhang, Y., Black, M.J., Tang, S.: We are more than our joints: Predicting how 3d bodies move. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3372-3382 (2021)", + "56. Zhong, C., Hu, L., Zhang, Z., Xia, S.: Attt2m: Text-driven human motion generation with multi-perspective attention mechanism. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 509-519 (2023)", + "57. Zhu, L., Liao, B., Zhang, Q., Wang, X., Liu, W., Wang, X.: Vision mamba: Efficient visual representation learning with bidirectional state space model. arXiv preprint arXiv:2401.09417 (2024)" + ], + "bbox": [ + 215, + 146, + 784, + 272 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Zeyu and Akide et al.", + "bbox": [ + 271, + 114, + 416, + 128 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/Motion Mamba_ Efficient and Long Sequence Motion Generation/1b586349-9221-4bad-a56d-e274ec79c852_model.json b/2024/Motion Mamba_ Efficient and Long Sequence Motion Generation/1b586349-9221-4bad-a56d-e274ec79c852_model.json new file mode 100644 index 0000000000000000000000000000000000000000..e708ba6b957d84a79785441c3ddcfaa989d2aea0 --- /dev/null +++ b/2024/Motion Mamba_ Efficient and Long Sequence Motion Generation/1b586349-9221-4bad-a56d-e274ec79c852_model.json @@ -0,0 +1,2271 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.245, + 0.141, + 0.761, + 0.185 + ], + "angle": 0, + "content": "Motion Mamba: Efficient and Long Sequence Motion Generation" + }, + { + "type": "text", + "bbox": [ + 0.239, + 0.213, + 0.765, + 0.246 + ], + "angle": 0, + "content": "Zeyu Zhang\\(^{12*}\\), Akide Liu\\(^{1*}\\), Ian Reid\\(^{3}\\), Richard Hartley\\(^{2}\\), Bohan Zhuang\\(^{1}\\), and Hao Tang\\(^{4\\boxtimes}\\)" + }, + { + "type": "text", + "bbox": [ + 0.432, + 0.255, + 0.574, + 0.27 + ], + "angle": 0, + "content": "1 Monash University" + }, + { + "type": "text", + "bbox": [ + 0.376, + 0.27, + 0.628, + 0.284 + ], + "angle": 0, + "content": "2 The Australian National University" + }, + { + "type": "text", + "bbox": [ + 0.309, + 0.283, + 0.694, + 0.298 + ], + "angle": 0, + "content": "3 Mohamed bin Zayed University of Artificial Intelligence" + }, + { + "type": "text", + "bbox": [ + 0.277, + 0.297, + 0.726, + 0.327 + ], + "angle": 0, + "content": "\\(^{4}\\) National Key Laboratory for Multimedia Information Processing, School of Computer Science, Peking University" + }, + { + "type": "text", + "bbox": [ + 0.336, + 0.336, + 0.668, + 0.349 + ], + "angle": 0, + "content": "https://steve-zeyu-zhang.github.io/MotionMamba" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.363, + 0.42, + 0.514 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.423, + 0.362, + 0.615, + 0.514 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.629, + 0.362, + 0.789, + 0.514 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.525, + 0.789, + 0.57 + ], + "angle": 0, + "content": "Fig.1: Motion Mamba has achieved significantly superior performance on long sequence modeling and motion generation efficiency compared with other well-designed state-of-the-art methods such as MLD [6], MotionDiffuse [54], and MDM [49]." + }, + { + "type": "text", + "bbox": [ + 0.26, + 0.584, + 0.744, + 0.794 + ], + "angle": 0, + "content": "Abstract. Human motion generation stands as a significant pursuit in generative computer vision, while achieving long-sequence and efficient motion generation remains challenging. Recent advancements in state space models (SSMs), notably Mamba, have showcased considerable promise in long sequence modeling with an efficient hardware-aware design, which appears to be a significant direction upon building motion generation model. Nevertheless, adapting SSMs to motion generation faces hurdles since the lack of a specialized design architecture to model motion sequence. To address these challenges, we propose Motion Mamba, a simple yet efficient approach that presents the pioneering motion generation model utilized SSMs. Specifically, we design a Hierarchical Temporal Mamba (HTM) block to process temporal data by ensemble varying numbers of isolated SSM modules across a symmetric U-Net architecture aimed at preserving motion consistency between frames. We also design a Bidirectional Spatial Mamba (BSM) block" + }, + { + "type": "page_footnote", + "bbox": [ + 0.233, + 0.805, + 0.356, + 0.817 + ], + "angle": 0, + "content": "*Equal contribution." + }, + { + "type": "page_footnote", + "bbox": [ + 0.235, + 0.817, + 0.618, + 0.828 + ], + "angle": 0, + "content": "†Work done while being a research assistant at Monash University." + }, + { + "type": "page_footnote", + "bbox": [ + 0.235, + 0.828, + 0.519, + 0.84 + ], + "angle": 0, + "content": "Corresponding author: bjdxtanghao@gmail.com" + }, + { + "type": "list", + "bbox": [ + 0.233, + 0.805, + 0.618, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.418, + 0.129 + ], + "angle": 0, + "content": "Zeyu and Akide et al." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.148, + 0.744, + 0.232 + ], + "angle": 0, + "content": "to bidirectionally process latent poses, to enhance accurate motion generation within a temporal frame. Our proposed method achieves up to \\(50\\%\\) FID improvement and up to 4 times faster on the HumanML3D and KIT-ML datasets compared to the previous best diffusion-based method, which demonstrates strong capabilities of high-quality long sequence motion modeling and real-time human motion generation." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.246, + 0.744, + 0.273 + ], + "angle": 0, + "content": "Keywords: Human Motion Generation \\(\\cdot\\) Selective State Space Models \\(\\cdot\\) Latent Diffusion Models" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.304, + 0.376, + 0.32 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.34, + 0.788, + 0.567 + ], + "angle": 0, + "content": "Human motion generation involves creating high-quality 3D human motion based on given conditions, such as a text prompt, which generates Euler angles for each joint in a human skeleton. To emulate human motion effectively, virtual characters must respond to the conditional context, exhibit natural movement, and perform motion accurately. Recent motion generation models are categorized into four main approaches: autoencoder-based [1,10,17,36,48,56], utilizing transformers for latent space compression and motion synthesis; GAN-based [4,20,30], using discriminators to enhance the realism of generated motions; autoregressive models [24], treating motion sequences as languages with specialized codebooks; and diffusion-based [6,49,54], employing denoising steps for motion generation. Challenges vary across methods, with autoencoder models struggling to generate accurate motions from detailed descriptions due to textual information compression, GAN-based models facing training difficulties, especially in conditional tasks, and diffusion-based models relying on complex transformer-based architectures results in inefficient of motion prediction." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.568, + 0.788, + 0.809 + ], + "angle": 0, + "content": "Although diffusion-based models excel at generating motion with robust performance and often exhibit superior diversity, they encounter two limitations. 1) Convolutional or transformer-based diffusion methods exhibit limitations in generating long-range motion sequences. Previous transformer-based methodologies [6,49,54] have focused on modeling long-range dependencies and acquiring comprehensive global context information. Despite these advances, they are frequently associated with a substantial increase in computational requirements. Furthermore, transformer architectures are not intrinsically designed for temporal sequential modeling, which poses an inherent limitation. 2) The efficiency of inference in transformer-based diffusion methods is constrained. Although prior research has attempted to leverage the Variational Autoencoder for denoising operations in the latent space [6], the inference speed remains adversely affected by the attention mechanism's quadratic scaling, leading to inefficient motion generation. Consequently, exploring a new architectural paradigm that accommodates long-range dependencies and maintains a linear computational complexity is crucial for sustaining motion generation tasks." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Recent advances have sparked renewed interest in state space models (SSMs) [14, 15], a field that originated from the foundational classic state space model" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.627, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Motion Mamba" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.788, + 0.328 + ], + "angle": 0, + "content": "[25]. Modern versions of SSMs stand out due to their ability to effectively capture long-range dependencies, a capability greatly improved by the introduction of parallel training techniques. This evolution has led to various methodologies based on SSM, notably the linear state space layers (LSSL) [15], the structured state-space sequence model (S4) [14], the diagonal state space (DSS) [19], and S4D [13]. These methods have been carefully designed to handle sequential data across various tasks and modalities, paying special attention to modeling long-range dependencies. Their efficacy in managing long sequences is attributed to the implementation of convolutional computations [15] and near-linear computational strategies, such as mamba [12], marking a significant stride in sequentially oriented tasks, including large language model decoding and motion sequence generation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.332, + 0.788, + 0.635 + ], + "angle": 0, + "content": "Adapting selective state space modules for motion generation tasks presents notable challenges, primarily due to the lack of specialized design in SSMs for capturing the sensitive motion details required for temporal representation and the complexities involved in aggregating latent space. In response to these challenges, we have meticulously developed a motion generation architecture, specifically tailored to address the intricacies of long-term sequence generation, while optimizing for computational efficiency with near-linear-time complexity. This innovation is embodied in the Motion Mamba model, a simple yet potent approach to motion generation. The Motion Mamba framework pioneers a diffusion-based generative system, incorporating two key components oriented toward SSM as shown in Figure. 2: (1) a Hierarchical Temporal Mamba (HTM) block: This component is ingeniously crafted to arrange motion frames in sequential order, using hierarchically adjusted scanning. It is adept at identifying temporal dependencies at various depths, thereby facilitating a thorough comprehension of the dynamics inherent in motion sequences. (2) a Bidirectional Spatial Mamba (BSM) block: This block is designed to unravel the structured latent skeleton by evaluating data from both forward and reverse directions. Its primary goal is to safeguard the continuity of information flow, significantly bolstering the model's capacity for precise motion generation through the retention of dense informational exchange." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.64, + 0.788, + 0.821 + ], + "angle": 0, + "content": "The Motion Mamba introduces a new approach to motion generation that strikes an exceptional trade-off between accuracy and efficiency, shown in Fig. 1. Our experimental results underscore the significant improvements brought about by Motion Mamba, showcasing a remarkable improvement in the Fréchet Inception Distance (FID), with a reduction of up to \\(50\\%\\) from the prior state-of-the-art metric of 0.473 to an impressive 0.281 on the HumanML3D dataset [17]. Furthermore, we emphasize our framework's unparalleled inference speed, which is four times faster than previous methods, achieving an average inference time of only 0.058 seconds per sequence compared to the 0.217 seconds required by the MLD [6] method per sequence. These outcomes unequivocally establish Motion Mamba's state-of-the-art performance, concurrently ensuring fast inference speeds for conditional human motion generation tasks." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.825, + 0.776, + 0.84 + ], + "angle": 0, + "content": "Our contributions to the field of motion generation can be summarized as:" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.418, + 0.129 + ], + "angle": 0, + "content": "Zeyu and Akide et al." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.786, + 0.191 + ], + "angle": 0, + "content": "1. We introduce a simple yet effective framework, named Motion Mamba, which is a pioneering method integrates a selective scanning mechanism into motion generation tasks." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.192, + 0.787, + 0.284 + ], + "angle": 0, + "content": "2. Motion Mamba is comprised of two modules: Hierarchical Temporal Mamba (HTM) and Bidirectional Spatial Mamba (BSM), which are designed for temporal and spatial modeling, respectively. HTM blocks are tasked with processing temporal motion data, aiming to enhance motion consistency across frames. BSM blocks are engineered to bidirectionally capture the channel-wise flow of hidden information within the latent pose representations." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.284, + 0.787, + 0.358 + ], + "angle": 0, + "content": "3. Motion Mamba framework demonstrated exceptional performance on text-to-motion generation task, through experimental validation on the HumanML3D [17] and KIT-ML [38] datasets. Our methodology achieved state-of-the-art generation quality and significantly improved long-squence modeling, meanwhile optimizing inference speed." + }, + { + "type": "list", + "bbox": [ + 0.216, + 0.147, + 0.787, + 0.358 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.384, + 0.396, + 0.4 + ], + "angle": 0, + "content": "2 Related Works" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.417, + 0.786, + 0.492 + ], + "angle": 0, + "content": "Human Motion Generation. Generating human motion is a significant application of computer vision, essential for various applications like 3D modelling and robot manipulation. Recently, the predominant method of achieving human motion generation, known as the Text-to-Motion task, involves learning a common latent space for both language and motion." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.493, + 0.787, + 0.644 + ], + "angle": 0, + "content": "DVGAN [30] creates the GAN [11] discriminator by densely validating at each time-scale and perturbing the discriminator input for translation invariance, enabling motion generation and completion. ERD-QV [20] enhances latent representations through two additive modifiers: a time-to-arrival embedding applied universally and an additive scheduled target noise vector used during extended transitions. It further improves transition quality by incorporating a GAN framework with two discriminators operating at different timescales. HP-GAN [4], trained with a modified version of the improved WGAN-GP [16], utilizes a custom loss function designed for human motion prediction. It learns a probability density function of future human poses conditioned on previous poses." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.645, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Autoencoders [27, 43] are notable generative models known for their ability to represent data robustly by compressing high-dimensional data into a latent space, which is widely adopted in human motion generation. JL2P [1] uses RNN-based autoencoders [23] to learn a combined representation of language and pose. It restricts the direct mapping from text to motion to a one-to-one relationship. MotionCLIP [48] uses Transformer-based Autoencoders [51] to reconstruct motion while ensuring alignment with the corresponding text label in the CLIP [40] space. This alignment effectively integrates the semantic knowledge from CLIP into the human motion manifold. TEMOS [36] and T2M [17] combine a Transformer-based VAE [26] with a text encoder to generate distribution parameters that work within the VAE latent space. AttT2M [56] and TM2D [10] incorporate a body-part spatio-temporal encoder into VQ-VAE [50] for enhanced learning of a discrete latent space with increased expressiveness." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.627, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Motion Mamba" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.219, + 0.146, + 0.786, + 0.297 + ], + "angle": 0, + "content": "Diffusion models [7, 22, 41, 47] have recently surpassed GANs and VAEs in generating 2D images. Developing a motion generation model based on diffusion models is obviously an attractive direction. MotionDiffuse [54] introduces the inaugural framework for text-driven motion generation based on diffusion models. It showcases several desirable properties, including probabilistic mapping, realistic synthesis, and multi-level manipulation. MDM [49] utilizes a classifier-free Transformer-based diffusion model for the human motion domain to predict sample rather than noise in each diffusion step. MLD [6] performs a diffusion process in latent motion space, rather than using a diffusion model to establish connections between raw motion sequences and conditional inputs." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.298, + 0.786, + 0.72 + ], + "angle": 0, + "content": "State Space Models. Recently, state space sequence models (SSMs) [14, 15], drawing inspiration from classical state-space models [25], have emerged as a promising architecture for sequence modeling. Mamba [12] introduces a selective SSM architecture, integrating time-varying parameters into the SSM framework, and proposes a hardware-aware algorithm to facilitate highly efficient training and inference processes. Some research works leverage SSM in computer vision to process 2D data. The 2D SSM [3] introduces an SSM block at the beginning of each transformer block [8, 51]. This approach aims to achieve efficient parameterization, accelerated computation, and a suitable normalization scheme. SGConvNeXt [29] presents a structured global convolution method inspired by ConvNeXt [32], incorporating multi-scale sub-kernels to achieve both parameterization efficiency and effective long sequence modeling. ConvSSM [46] integrates the tensor modeling principles of ConvLSTM [45] with SSMs, elucidating the utilization of parallel scans in convolutional recurrences. This approach enables subquadratic parallelization and rapid autoregressive generation. Vim [57] introduces a bidirectional SSM block [52] for efficient and versatile visual representation learning, achieving performance comparable to established ViT [8] methods. VMamba [31] introduces a Cross-Scan Module (CSM) designed to traverse the spatial domain and transform any non-causal visual image into ordered patch sequences. This approach achieves linear complexity while preserving global receptive fields. There have also been attempts to utilize SSMs to handle higher-dimensional data. Mamba-ND [28] explores various combinations of SSM and different scan directions within the SSM block to adapt Mamba [12] to higher-dimensional tasks. Recent efforts have sought to replace the traditional transformer-based U-Net within the diffusion denoiser with the SSM block, with the aim of enhancing image generation efficiency. DiffuSSM [53] adeptly manages higher resolutions without relying on global compression, thus maintaining detailed image representation throughout the diffusion process." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.746, + 0.471, + 0.763 + ], + "angle": 0, + "content": "3 The Proposed Method" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.78, + 0.786, + 0.84 + ], + "angle": 0, + "content": "In this section, we delineate the architecture and operational principles of the Motion Mamba framework, designed for generating human motion over long ranges efficiently from textual descriptions. Initially, we discuss the foundational concepts underpinning our approach, including the Mamba Model [12] and the" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.419, + 0.129 + ], + "angle": 0, + "content": "Zeyu and Akide et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.224 + ], + "angle": 0, + "content": "latent diffusion model [6]. Following this, we detail our uniquely crafted architecture that leverages the Mamba model to enhance motion generation efficiency. This architecture comprises two principal components: the Hierarchical Temporal Mamba (HTM) block, which addresses temporal aspects, and the Bidirectional Spatial Mamba (BSM) block, focusing on spatial dynamics." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.246, + 0.373, + 0.26 + ], + "angle": 0, + "content": "3.1 Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.271, + 0.788, + 0.375 + ], + "angle": 0, + "content": "Selective Structured State Space Sequence Model. SSMs particularly through the contributions of structured state space sequence models (S4) and Mamba, have demonstrated exceptional proficiency in handling long sequences. These models operationalize the mapping of a 1-D function or sequence, \\( x(t) \\in \\mathbb{R} \\mapsto y(t) \\in \\mathbb{R} \\), through a hidden state \\( h(t) \\in \\mathbb{R}^N \\), employing \\( \\mathbf{A} \\in \\mathbb{R}^{N \\times N} \\) as the evolution parameters, \\( \\mathbf{B} \\in \\mathbb{R}^{N \\times 1} \\) and \\( \\mathbf{C} \\in \\mathbb{R}^{1 \\times N} \\) as the projection parameters, respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.377, + 0.788, + 0.405 + ], + "angle": 0, + "content": "The discretized system can then be expressed as follows, incorporating a step size \\(\\Delta\\):" + }, + { + "type": "equation", + "bbox": [ + 0.434, + 0.404, + 0.785, + 0.428 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{c} h _ {t} = \\overline {{\\mathbf {A}}} h _ {t - 1} + \\overline {{\\mathbf {B}}} x _ {t}, \\\\ \\mathbf {C} ^ {\\prime} \\end{array} \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.435, + 0.424, + 0.505, + 0.438 + ], + "angle": 0, + "content": "\\[\ny _ {t} = \\mathbf {C} h _ {t}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.445, + 0.788, + 0.492 + ], + "angle": 0, + "content": "This adaptation facilitates the computation of output through global convolution, leveraging a structured convolutional kernel \\(\\overline{\\mathbf{K}}\\), which encompasses the entire length \\(M\\) of the input sequence \\(\\mathbf{x}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.383, + 0.502, + 0.785, + 0.529 + ], + "angle": 0, + "content": "\\[\n\\overline {{\\mathbf {K}}} = \\left(\\underset {-} {\\mathbf {C}} \\overline {{\\mathbf {B}}}, \\mathbf {C} \\overline {{\\mathbf {A}}} \\mathbf {B}, \\dots , \\mathbf {C} \\overline {{\\mathbf {A}}} ^ {M - 1} \\overline {{\\mathbf {B}}}\\right), \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.388, + 0.526, + 0.465, + 0.542 + ], + "angle": 0, + "content": "\\[\n\\mathbf {y} = \\mathbf {x} * \\overline {{\\mathbf {K}}}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.554, + 0.788, + 0.63 + ], + "angle": 0, + "content": "Selective models like Mamba introduce time-varying parameters, deviating from the linear time invariance (LTI) assumption and complicating parallel computation. However, hardware-aware optimizations, such as associative scans, have been developed to address these computational challenges, highlighting the ongoing evolution and application of SSMs in modeling complex temporal dynamics." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.63, + 0.789, + 0.844 + ], + "angle": 0, + "content": "Latent Motion Diffusion Model. Diffusion probabilistic models offer a significant advancement in motion generation by gradually reducing noise from a Gaussian distribution to a target data distribution \\( p(x) \\) through a T-length learned Markov process [7,22,41,44,47,49,54], giving \\( \\{\\pmb{x}_t\\}_{t=1}^T \\). In the motion generation, we define our trainable diffusion models with a denoiser \\( \\epsilon_{\\theta}(x_t,t) \\) which anneal the random noise to motion sequence \\( \\{\\hat{x}_t^{1:N}\\}_{t=1}^T \\) iteratively. To address the inefficiencies of applying diffusion models directly to raw motion sequences, we employ a low-dimensional motion latent space for the diffusion process. Given an input condition \\( c \\), such as a descriptive sentence \\( \\pmb{w}^{1:N} = \\{w^i\\}_{i=1}^N \\), an action label \\( a \\) from a predefined set \\( \\mathcal{A} \\), or an empty condition \\( c = \\emptyset \\), and the motion representation that combines 3D joint rotations, positions, velocities, and foot contact as proposed in [17]. The frozen CLIP [40] text encoder \\( \\tau_{\\theta}^w \\) has been employed to obtain projected text embedding \\( \\tau_{\\theta}^w(w^{1:N}) \\in \\mathbb{R}^{1 \\times d} \\), thereby conditional denoiser comprised in term of \\( \\epsilon_{\\theta}(z_t,t,\\tau_{\\theta}(c)) \\). The latent diffusion model \\( \\epsilon_{\\theta}(x_t,t) \\) aimed" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.626, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Motion Mamba" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.214, + 0.144, + 0.788, + 0.361 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.371, + 0.788, + 0.453 + ], + "angle": 0, + "content": "Fig. 2: This figure illustrates the architecture of the proposed Motion Mamba model. Each of encoder and decoder blocks consists of a Hierarchical Temporal Mamba block (HTM) and a Bidirectional Spatial Mamba (BSM) block, which possess hierarchical scan and bidirectional scan within SSM layers respectively. This symmetric distribution of scans ensure a balanced and coherence framework across the encoder-decoder architecture." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.469, + 0.788, + 0.593 + ], + "angle": 0, + "content": "to generate the human motion sequence in terms of \\(\\hat{x}^{1:L} = \\{\\hat{x}^i\\}_{i=1}^L\\), where L denotes the sequence length or number of frames [34,35,37,55]. After that we reused the motion Variational AutoEncoder (VAE) \\(\\mathcal{V} = \\{\\mathcal{E},\\mathcal{D}\\}\\) proposed in MLD [6] to manipulate the motion sequence in latent space \\(z = \\mathcal{E}(x^{1:L})\\), and decompress the intermediate representation to motion sequence by \\(\\hat{x}^{1:L} = \\mathcal{D}(z) = \\mathcal{D}\\mathcal{E}(x^{1:L})\\) [26, 42, 51]. Finally, our latent diffusion model is trained with an objective focusing on minimization of MSE between true and predicted noise in the latent space, facilitating efficient and high-quality motion generation [2, 22]." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.612, + 0.393, + 0.626 + ], + "angle": 0, + "content": "3.2 Motion Mamba" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.637, + 0.788, + 0.759 + ], + "angle": 0, + "content": "The architecture of the proposed Motion Mamba framework is illustrated in Figure. 2. At its core, Motion Mamba utilizes a denoising U-Net architecture, which is distinguished for its effectiveness in modeling the continuous, temporal sequences of motion frames. This effectiveness is attributed to the inherent long-sequence modeling capacity of the Mamba model. The denoiser, denoted by \\(\\epsilon_{\\theta}\\), comprises \\(N\\) blocks including encoder \\(E_{1..N}\\) and decoder \\(D_{1..N}\\). Additionally, the architecture is enhanced with a transformer-based attention mixer block \\(M\\), designed to augment the model's ability to capture complex temporal dynamics." + }, + { + "type": "equation", + "bbox": [ + 0.402, + 0.772, + 0.786, + 0.79 + ], + "angle": 0, + "content": "\\[\n\\epsilon_ {\\theta} (x) \\equiv \\{E _ {1 \\dots N}, M, D _ {1 \\dots N} \\}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.843 + ], + "angle": 0, + "content": "The encoder blocks are represented as \\( E_{1..N} \\), arranged sequentially, and the decoder blocks as \\( D_{1..N} \\), configured in reverse order to facilitate effective bottom-up and top-down information flow. Given that selective operations have signif" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.419, + 0.129 + ], + "angle": 0, + "content": "Zeyu and Akide et al." + }, + { + "type": "code_caption", + "bbox": [ + 0.218, + 0.147, + 0.651, + 0.163 + ], + "angle": 0, + "content": "Algorithm 1 Hierarchical Temporal Mamba (HTM) Block." + }, + { + "type": "algorithm", + "bbox": [ + 0.217, + 0.167, + 0.603, + 0.392 + ], + "angle": 0, + "content": "Require: compressed latent representations \\(z:(T,B,C)\\) \nEnsure: transformed representations \\(z_{\\mathrm{HTM}}:(T,B,E)\\) \n1:/\\* linear projection layer \\*/ \n2: \\(x,z:(T,B,E)\\gets \\mathrm{Linear}(z)\\) \n3:/\\* set of scans and memory matrices \\*/ \n4: \\(K = \\{S_{2N_n - 1},S_{2N_{n - 1} - 1},\\ldots ,S_1\\}\\) \n5: Memory matrices: \\(\\{A_1,\\dots ,A_k\\}\\) \n6: for each scan \\(S_{i}\\) in \\(K\\) do \n7: \\(x_o^{\\prime}:(T,B,E)\\gets \\mathrm{Conv1D}(x)\\) \n8: \\(B_{o},C_{o},\\varDelta_{o}\\gets \\mathrm{Linear}(x_{o}^{\\prime})\\) \n9: Transform \\(\\overline{A}_o\\) and \\(\\overline{B}_o\\) using \\(\\varDelta_{o}\\) \n10: \\(O_{i}\\gets \\mathrm{SSMA}_{i,x}(x_{o}^{\\prime})\\) \n11: end for \n12:/\\* aggregation of outputs \\*/ \n13: \\(z_{\\mathrm{HTM}}:(T,B,E)\\gets \\mathrm{Linear(Aggregate(\\{O_1,\\ldots,O_k\\}))}\\) \n14: Return: \\(z_{\\mathrm{HTM}}\\)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.422, + 0.788, + 0.498 + ], + "angle": 0, + "content": "icantly lower computational complexity compared to attention-based methods, we have increased the number of scans to achieve higher quality generations. Concurrently, it is imperative to maintain a balance between the model's parameters and its efficiency. Thereby, a novel aspect of our model is the introduction of a hierarchical scan strategy, characterized by a sequence of scan numbers as," + }, + { + "type": "equation", + "bbox": [ + 0.38, + 0.508, + 0.787, + 0.526 + ], + "angle": 0, + "content": "\\[\nK = \\left\\{S _ {2 N - 1}, S _ {2 (N - 1) - 1}, \\dots , S _ {1} \\right\\}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.535, + 0.788, + 0.627 + ], + "angle": 0, + "content": "This sequence specifies the number of scans allocated to each layer, in descending order of complexity. For instance, the uppermost encoder layer, \\( E_{1} \\), and the lowermost decoder layer, \\( D_{N} \\), are allocated \\( S_{2N - 1} \\) scans, indicating the highest scanning complexity. Conversely, the lowest encoder layer, \\( E_{N} \\), and the uppermost decoder layer, \\( D_{1} \\), are assigned \\( S_{1} \\) scans, reflecting the lowest level of scanning complexity." + }, + { + "type": "equation", + "bbox": [ + 0.346, + 0.636, + 0.786, + 0.693 + ], + "angle": 0, + "content": "\\[\nE _ {i} (S) = \\left\\{ \\begin{array}{l l} S _ {2 N - 1} & \\text {f o r} i = 1 \\\\ S _ {2 (N - i) - 1} & \\text {f o r} i = 2, \\dots , N - 1 \\\\ S _ {1} & \\text {f o r} i = N \\end{array} \\right. \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.343, + 0.704, + 0.787, + 0.761 + ], + "angle": 0, + "content": "\\[\nD _ {j} (S) = \\left\\{ \\begin{array}{l l} S _ {2 N - 1} & \\text {f o r} j = N \\\\ S _ {2 (N - j) - 1} & \\text {f o r} j = N - 1, \\dots , 2 \\\\ S _ {1} & \\text {f o r} j = 1 \\end{array} \\right. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.843 + ], + "angle": 0, + "content": "This hierarchical scanning approach ensures that processing capabilities are evenly distributed throughout the encoder-decoder architecture., facilitating a detailed and nuanced analysis of temporal sequences. Within this structured framework, each denoiser is equipped with a specialized Hierarchical Temporal Mamba (HTM) block, which serves to augment the model's ability to process" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.626, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Motion Mamba" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "code_caption", + "bbox": [ + 0.218, + 0.147, + 0.634, + 0.163 + ], + "angle": 0, + "content": "Algorithm 2 Bidirectional Spatial Mamba (BSM) Block." + }, + { + "type": "algorithm", + "bbox": [ + 0.217, + 0.167, + 0.62, + 0.405 + ], + "angle": 0, + "content": "Require: compressed latent representations \\(z:(T,B,C)\\) \nEnsure: transformed representations \\(z_{\\mathrm{BSM}}:(C,B,E)\\) \n1:/\\* dimension rearrangement \\*/ \n2: \\(z^{\\prime}:(C,B,T)\\gets\\) Rearrange(z) \n3:/\\* linear projection after normalization \\*/ \n4: \\(z^{\\prime}:(C,B,T)\\gets\\) Norm(z) \n5:x,z:(C,B,E) \\(\\leftarrow\\) Linear(z') \n6: for o in {forward, backward} do \n7: \\(x_o':(C,B,E)\\gets\\) Conv1D(x) \n8: \\(B_{o},C_{o},\\varDelta_{o}\\gets\\) Linear \\((x_o^{\\prime})\\) \n9: Transform \\(\\overline{A}_o\\) and \\(\\overline{B}_o\\) using \\(\\varDelta_{o}\\) \n10: \\(y_{o}\\gets\\) SSM(Ao,Bo,Co) \n11: end for \n12:/\\* gating and summing outputs \\*/ \n13: \\(z_{\\mathrm{BSM}}:(C,B,T)\\gets\\) GateAndSum(yforward,ybackward,z) \n14: \\(z_{\\mathrm{BSM}}^{\\prime}:(T,B,C)\\gets\\) Rearrange(z) \n15: Return: \\(z_{\\mathrm{BSM}}^{\\prime}\\)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.438, + 0.785, + 0.481 + ], + "angle": 0, + "content": "temporal information effectively. Additionally, the proposed Motion Mamba incorporates an attention-based mixer block denoted as \\(M\\), strategically integrated to enhance conditional fusion." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.483, + 0.788, + 0.755 + ], + "angle": 0, + "content": "Hierarchical Temporal Mamba (HTM) block processes compressed latent representations, denoted as \\( z \\), with the dimensions \\( (T,B,C) \\), of which procedure shown in Algorithm 1. Here, \\( T \\) signifies the temporal dimension, as specified in the Variational AutoEncoder (VAE) framework. Initially, the input \\( z \\) is subjected to a linear projection layer, producing transformed representations \\( x \\) and \\( z \\) with dimension \\( E \\). Our analysis revealed an increased density of motion within the lower-level feature spaces. Consequently, we developed a hierarchical scanning methodology that is executed at various depths of the network. This approach not only accommodates the diverse motion densities, but also significantly reduces computational overhead. This step utilizes a hierarchically structured set of scans, \\( K = \\{S_{2N_n - 1}, S_{2N_{n - 1} - 1}, \\ldots, S_1\\} \\), in conjunction with a corresponding series of memory matrices \\( \\{A_1, \\ldots, A_k\\} \\). Each sub-SSM scan first applies a 1-D convolution to \\( x \\), resulting in \\( x_o' \\). \\( x_o' \\) is then linearly projected to derive \\( B_o \\), \\( C_o \\), and \\( \\Delta_o \\). These projections \\( B_o \\), \\( C_o \\) use \\( \\Delta_o \\) to effect transformations in \\( \\overline{A}_o \\) and \\( \\overline{B}_o \\), respectively. After executing a sequence of SSM scans \\( \\{SSM_{A_1,x}, SSM_{A_2,x}, \\ldots, SSM_{A_k,x}\\} \\), a set of outputs \\( \\{O_1, \\ldots, O_k\\} \\) is compiled. This collection is subsequently aggregated via a linear projection to obtain the final output of the HTM block." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Bidirectional Spatial Mamba (BSM) block focuses on enhancing latent representation learning through a novel approach of dimension rearrangement and bidirectional scanning, of which the process is shown in Algorithm 2. Initially, it alters the original input dimensions from \\((T,B,C)\\) to \\((C,B,T)\\), effectively swap" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.418, + 0.129 + ], + "angle": 0, + "content": "Zeyu and Akide et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.344 + ], + "angle": 0, + "content": "ping the temporal and channel dimensions. After this rearrangement, the input, now denoted \\( z' \\), undergoes a linear projection after normalization, resulting in dimensions \\( x \\) and \\( z \\) of size \\( E \\). The process involves bidirectional scanning of the latent channel dimension, where \\( \\mathbf{x} \\) is subjected to a 1-D convolution, yielding \\( \\mathbf{x}_o' \\) for both forward and backward directions. Each \\( \\mathbf{x}_o' \\) is then linearly projected to obtain \\( B_o \\), \\( C_o \\), and \\( \\Delta_o \\), which are utilized to transform \\( \\overline{A}_o \\) and \\( \\overline{B}_o \\), respectively. The final output token sequence, \\( \\mathbf{z}_1 \\), is computed by gating and summing the forward \\( y_{\\mathrm{forward}} \\) and backward \\( y_{\\mathrm{backward}} \\) output with \\( \\mathbf{z} \\). This component is engineered to decode the structured latent skeleton by analyzing data from both forward and reverse viewpoints. Its main objective is to ensure the seamless continuity of information flow, thereby substantially enhancing the model's ability to generate accurate motion. This is achieved through the maintenance of a dense informational exchange, which is critical for the model's performance." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.363, + 0.377, + 0.381 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.391, + 0.335, + 0.404 + ], + "angle": 0, + "content": "4.1 Datasets" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.413, + 0.786, + 0.441 + ], + "angle": 0, + "content": "We evaluate our proposed Motion Mamba on two prominent Text-to-Motion synthesis benchmarks as follows:" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.442, + 0.799, + 0.517 + ], + "angle": 0, + "content": "HumanML3D. The HumanML3D [17] dataset aggregates 14,616 motions sourced from the AMASS [33] and HumanAct12 [18] datasets, with each motion accompanied by three textual descriptions, culminating in 44,970 scripts. This dataset spans a wide range of actions such as exercising, dancing, and acrobatics, presenting a rich motion-language corpus." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.518, + 0.787, + 0.58 + ], + "angle": 0, + "content": "KIT-ML. The KIT-ML dataset [38] is comprised of 3,911 motions paired with 6,278 textual descriptions, serving as a compact yet effective benchmark for evaluation. For both datasets, the pose representation adopted is derived from T2M [17], ensuring consistency in motion representation across evaluations." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.597, + 0.421, + 0.612 + ], + "angle": 0, + "content": "4.2 Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.619, + 0.788, + 0.755 + ], + "angle": 0, + "content": "We adapt the standard evaluation metrics on following aspects throughout our experiments, including: Generation Quality. We implement a Fréchet inception distance (FID) [21] to quantify the realism and diversity of motion generated by models. Moreover, we use multi-modal distance (MM Dist) to measure the distance between motions and texts and assess motion-text alignment. Diversity. We use the diversity metric to measure motion diversity, which calculates variance in features extracted from the motions. Additionally, we employ multimodality (MModality) to assess diversity within generated motions sharing the same text description." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.773, + 0.436, + 0.789 + ], + "angle": 0, + "content": "4.3 Comparative Studies" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.795, + 0.843 + ], + "angle": 0, + "content": "We evaluate our method against the state-of-the-art methods on the HumanML3D [17] and KIT-ML [38] datasets. We train our Motion Mamba with HTM arrangement strategy MM (\\( \\{S_{2N_n - 1},\\ldots ,S_1\\} \\)), BSM bidirectional block strategy on the" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.626, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Motion Mamba" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.214 + ], + "angle": 0, + "content": "Table 1: Comparison of text-conditional motion synthesis on HumanML3D [17]. These metrics are evaluated by the motion encoder from [17]. Empty MModality indicates the non-diverse generation methods. We employ real motion as a reference and sort all methods by descending FIDs. The right arrow \\(\\rightarrow\\) means that the closer to the real motion, the better. Bold and underline indicate the best and second best result." + }, + { + "type": "table", + "bbox": [ + 0.216, + 0.214, + 0.787, + 0.381 + ], + "angle": 0, + "content": "
MethodR Precision ↑FID↓MM Dist↓Diversity→MModality↑
Top 1Top 2Top 3
Real0.511±.0030.703±.0030.797±.0020.002±.0002.974±.0089.503±.065-
Seq2Seq [39]0.180±.0020.300±.0020.396±.00211.75±.0355.529±.0076.223±.061-
LJ2P [1]0.246±.0010.387±.0020.486±.00211.02±.0465.296±.0087.676±.058-
T2G [5]0.165±.0010.267±.0020.345±.0027.664±.0306.030±.0086.409±.071-
Hier [9]0.301±.0020.425±.0020.552±.0046.532±.0245.012±.0188.332±.042-
TEMOS [37]0.424±.0020.612±.0020.722±.0023.734±.0283.703±.0088.973±.0710.368±.018
T2M [17]0.457±.0020.639±.0030.740±.0031.067±.0023.340±.0089.188±.0022.090±.083
MDM [49]0.320±.0050.498±.0040.611±.0070.544±.0445.566±.0279.559±.0862.799±.072
MotionDiffuse [54]0.491±.0010.681±.0010.782±.0010.630±.0013.113±.0019.410±.0491.553±.042
MLD [6]0.481±.0030.673±.0030.772±.0020.473±.0133.196±.0109.724±.0822.413±.079
Motion Mamba (Ours)0.502±.0030.693±.0020.792±.0020.281±.0093.060±.0589.871±.0842.294±.058
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.388, + 0.787, + 0.414 + ], + "angle": 0, + "content": "Table 2: We involve KIT-ML [38] dataset and evaluate the SOTA methods on the text-to-motion task." + }, + { + "type": "table", + "bbox": [ + 0.216, + 0.415, + 0.787, + 0.584 + ], + "angle": 0, + "content": "
MethodR Precision ↑FID↓MM Dist↓Diversity→MModality↑
Top 1Top 2Top 3
Real0.424±.0050.649±.0060.779±.0060.031±.0042.788±.01211.08±.097-
Seq2Seq [39]0.103±.0030.178±.0050.241±.00624.86±.3487.960±.0316.744±.106-
T2G [5]0.156±.0040.255±.0040.338±.00512.12±.1836.964±.0299.334±.079-
LJ2P [1]0.221±.0050.373±.0040.483±.0056.545±.0725.147±.0309.073±.100-
Hier [9]0.255±.0060.432±.0070.531±.0075.203±.1074.986±.0279.563±.0722.090±.083
TEMOS [37]0.353±.0060.561±.0070.687±.0053.717±.0513.417±.01910.84±.1000.532±.034
T2M [17]0.370±.0050.569±.0070.693±.0072.770±.1093.401±.00810.91±.1191.482±.065
MDM [49]0.164±.0040.291±.0040.396±.0040.497±.0219.191±.02210.85±.1091.907±.214
MotionDiffuse [54]0.417±.0040.621±.0040.739±.0041.954±.0622.958±.00511.10±.1430.730±.013
MLD [6]0.390±.0080.609±.0080.734±.0070.404±.0273.204±.02710.80±.1172.192±.071
Motion Mamba (Ours)0.419±.0060.645±.0050.765±.0060.307±.0413.021±.02511.02±.0981.678±.064
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.597, + 0.788, + 0.749 + ], + "angle": 0, + "content": "latent dimension \\(= 2\\) with 11 layers. We evaluate our model and previous works with suggested metrics in HumanML3D [17] and calculate \\(95\\%\\) confidence interval by repeat evaluation 20 times. The results for the HumanML3D dataset are presented in Table 1. Our model outperforms other methods significantly across various evaluation metrics, including FID, R precision, multi-modal distance, and diversity. For instance, our Motion Mamba outperforms previous best diffusion based motion generation MLD by \\(40.5\\%\\) in terms of FID, and up to \\(10\\%\\) improvement on R Precision, we aslo obatined best MModality by 3.060. The results for the KIT-ML dataset are presented in Table 2. We have also outperformed other well-established methods in FID and multi-modal distance." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.769, + 0.403, + 0.784 + ], + "angle": 0, + "content": "4.4 Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.793, + 0.787, + 0.839 + ], + "angle": 0, + "content": "We concluded the ablation studies including long sequence evaluation, hierarchical design with HTM, bidirectional design in the BSM, number of latent dimensions, and number of layers of our proposed motion mamba in Table 4." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.418, + 0.129 + ], + "angle": 0, + "content": "Zeyu and Akide et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.214 + ], + "angle": 0, + "content": "Table 3: In order to evaluate the models' capability in long sequence motion generation, we compared our method with an existing approach on the recently introduced HumanML3D-LS dataset. This dataset comprises motion sequences longer than 190 frames from the original evaluation set. Our model demonstrates superior performance compared to other methods." + }, + { + "type": "table", + "bbox": [ + 0.216, + 0.213, + 0.787, + 0.317 + ], + "angle": 0, + "content": "
MethodR Precision ↑FID↓MM Dist↓Diversity→MModality↑
Top 1Top 2Top 3
Real0.437±.0030.622±.0040.721±.0040.004±.0003.343±.0158.423±.090-
MDM [49]0.368±.0050.553±.0060.672±.0050.802±.0443.860±.0258.817±.068-
MotionDiffuse [54]0.367±.0040.521±.0040.623±.0042.460±.0623.789±.0058.707±.1431.602±.013
MLD [6]0.403±.0050.584±.0050.690±.0050.952±.0203.580±.0169.050±.0852.711±.104
Motion Mamba (Ours)0.417±.0030.606±.0030.713±.0040.668±.0193.435±.0159.021±.0702.373±.084
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.337, + 0.545, + 0.533 + ], + "angle": 0, + "content": "Long Sequence Motion Generation. The HumanML3D [17] dataset exhibits a long-tailed and right-skewed distribution with a significant proportion of long-sequence human motions, as shown in Figure 3. We suggest previous study overlooked the challenges in the long-sequence generation problem. Thus, we introduce a new dataset variant, HumanML3D-LS, comprising motion sequences longer than 190 frames extracted from the original test set. This addition allows us to showcase our capability in generating long-sequence motions. Subsequently, we eval" + }, + { + "type": "image", + "bbox": [ + 0.556, + 0.335, + 0.789, + 0.444 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.554, + 0.453, + 0.788, + 0.522 + ], + "angle": 0, + "content": "Fig. 3: The figure shows a long tail distribution of the HumanML3D [17], which has a significant proportion of long-sequence human motions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.534, + 0.788, + 0.595 + ], + "angle": 0, + "content": "uate the performance of our method on HumanML3D-LS and compare it with other diffusion-based motion generation approaches. The comparative results are presented in Table 3. Motion Mamba by leverage the benefits on long-range dependency modeling make it well suitable for long sequence motion generation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.596, + 0.788, + 0.763 + ], + "angle": 0, + "content": "Hierarchical Design with HTM. In our ablation studies, we observed a slight improvement upon reversing the scan order from a lower to a higher level, specifically transitioning from MM \\(\\{S_1,\\ldots ,S_N\\}\\) to MM \\(\\{S_N,\\dots ,S_1\\}\\). This enhancement suggests a correlation with the increase in temporal motion density within the lower-level feature spaces. Furthermore, to achieve the optimal result, we introduce the hierarchical design to arrange the scanning frequency, resulting in the sequence MM \\(\\{S_{2N_n - 1},\\ldots ,S_1\\}\\). This expansion in the number of scans led to a performance increase. We attribute this enhancement to the observation that individual selective scan operations significantly reduce the parameter count, especially when compared to the parameter-intensive constructs of self-attention and feedforward network blocks prevalent in transformer architectures." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Bidirectional Design in BSM. We developed three distinct variations of latent scanning mechanisms, differentiated by their scanning directions. In the context of motion generation tasks, we posit that the flow of hidden information within the structured latent skeleton holds significance, an aspect previously underexplored. Our ablation study reveals that a single scan across the la" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.626, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Motion Mamba" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.145, + 0.788, + 0.199 + ], + "angle": 0, + "content": "Table 4: Evaluation of text-based motion synthesis on HumanML3D [17]: we use metrics in Table 1 and provides real reference, we evaluate the various HTM and BSM design choices, the dimension of the latent input, the different number of layer of Motion Mamba model." + }, + { + "type": "table", + "bbox": [ + 0.269, + 0.2, + 0.73, + 0.445 + ], + "angle": 0, + "content": "
ModelsR Precision Top 3↑FID↓MM Dist.↓Diversity→MModality↑
Real0.797±.0020.002±.0002.974±.0089.503±.065-
MM ( {S1, ..., SN})0.673±.0031.278±.0123.802±.0418.678±.0963.127±.024
MM ( {SN, ..., S1})0.738±.0020.962±.0113.433±.0039.180±.0712.723±.033
MM ( {S1, ..., S2Nn-1})0.698±.0020.856±.0083.624±.0379.229±.0672.826±.017
MM ( {S2Nn-1, ..., S1})0.792±.0020.281±.0093.060±.0589.871±.0842.294±.058
MM (SingleScan)0.736±.0031.063±.0103.443±.0269.180±.0672.676±.041
MM (BiScan, layer)0.735±.0040.789±.0073.408±.0349.374±.0592.591±.046
MM (BiScan, block)0.792±.0020.281±.0093.060±.0589.871±.0842.294±.058
MM (Dim, 1)0.706±.0030.652±.0113.541±.0729.141±.0822.612±.055
MM (Dim, 2)0.792±.0020.281±.0093.060±.0589.871±.0842.294±.058
MM (Dim, 5)0.741±.0080.728±.0093.307±.0279.427±.0992.314±.062
MM (Dim, 7)0.738±.0040.599±.0073.359±.0689.166±.0752.488±.037
MM (Dim, 10)0.715±.0030.628±.0083.548±.0439.200±.0752.884±.096
MM (9 layers)0.755±.0021.080±.0123.309±.0579.721±.0812.974±.039
MM (11 layers)0.792±.0020.281±.0093.060±.0589.871±.0842.294±.058
MM (27 layers)0.750±.0030.975±.0083.336±.0969.249±.0712.821±.063
MM (37 layers)0.754±.0050.809±.0103.338±.0619.355±.0622.741±.077
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.461, + 0.788, + 0.552 + ], + "angle": 0, + "content": "tent dimension yields minimal improvement. Subsequently, we investigated both layer-based and block-based bidirectional scans. Our findings indicate that the block-based bidirectional scan achieves optimal performance. This suggests that spatial information flows are encoded within the latent spaces and that bidirectional scanning facilitates the exchange of this information, thereby enhancing the efficacy of motion generation tasks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.553, + 0.789, + 0.844 + ], + "angle": 0, + "content": "Architecture Design for Motion Mamba. The proposed Motion Mamba which is grounded in a standardized motion latent diffusion system. We delved into the interplay between dimensional aspects and the module's capacity (measured by the number of layers) to ascertain their impact on system performance. Experimental results demonstrate that the Motion Mamba achieves superior performance at a latent dimension of 2, diverging from prior works where the optimal dimension was identified as 1. We attribute this discrepancy to our HTM, which necessitates multiple scans correlating with the sequence length, thus implicating dimensionality as a pivotal factor. A marginal increase in dimensionality enabled us to attain peak performance, simultaneously enhancing efficiency compared to models with a dimensionality of 10. Furthermore, we conducted experiments to determine the optimal layer count for Motion Mamba, inspired by the design of its selective scanning mechanism. Notably, a single Mamba layer achieves a parameter reduction of approximately \\(75\\%\\) compared to a conventional transformer encoder block. By increasing the number of layers, we aim to uncover the relationship between model capacity and its performance. Our findings reveal that, through the integration of our specially designed HTM and BSM (Bidirectional Scanning Module) blocks, the Motion Mamba reaches its optimal performance with 11 layers. This represents a slight increase over the MLD [6] baseline. How-" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.418, + 0.129 + ], + "angle": 0, + "content": "Zeyu and Akide et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.785, + 0.178 + ], + "angle": 0, + "content": "ever, due to the reduced parameter count in each layer, Motion Mamba exhibits significantly greater efficiency than previous methodologies." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.2, + 0.388, + 0.214 + ], + "angle": 0, + "content": "4.5 Inference Time" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.225, + 0.518, + 0.438 + ], + "angle": 0, + "content": "Inference time remains a significant challenge for diffusion-based methods. To address this, we enhance the inference speed by incorporating the efficient Mamba block within a lightweight architecture. Compared to the previous strong baseline, such as the MLD model cited in [6], which reports an average inference time of 0.217 seconds, our Motion Mamba model achieves a notable reduction in computational overhead, as shown in Figure 4. Specifically, it requires four times less computational effort, thereby facilitating faster and real-time inference speeds." + }, + { + "type": "image", + "bbox": [ + 0.539, + 0.221, + 0.765, + 0.354 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.526, + 0.366, + 0.789, + 0.463 + ], + "angle": 0, + "content": "Fig.4: The figure shows the average inference time per sentence (AIT) vs FID, our proposed motion mamba obtained 0.058s AIT and 0.281 FID over all outperform previous methods. We evaluate all methods on a single V100 GPU." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.459, + 0.51, + 0.475 + ], + "angle": 0, + "content": "5 Discussion and Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.491, + 0.789, + 0.75 + ], + "angle": 0, + "content": "In this study, we introduced Motion Mamba, a novel framework designed for efficient and extended sequence motion generation. Our approach represents the inaugural integration of the Mamba model within the domain of motion generation, featuring significant advancements including the implementation of Hierarchical Temporal Mamba (HTM) blocks. These blocks are specifically engineered to enhance temporal alignment through hierarchically organized selective scanning. Furthermore, Bidirectional Spatial Mamba (BSM) blocks have been developed to amplify the exchange of information flow within latent spaces, thereby augmenting the model's ability to bidirectionally capture skeleton-level density features with greater precision. Compared to previous diffusion-based motion generation methodologies that predominantly utilize transformer blocks, our Motion Mamba framework achieves SOTA performance, evidencing an improvement of up to \\(50\\%\\) in FID scores and a quadrupled improvement in inference speed. Through comprehensive experimentation across a variety of human motion generation tasks, the effectiveness and efficiency of our proposed Motion Mamba model have been robustly demonstrated, marking a significant leap forward in the field of human motion generation." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.772, + 0.405, + 0.789 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.804, + 0.787, + 0.835 + ], + "angle": 0, + "content": "This work is partially supported by the Fundamental Research Funds for the Central Universities, Peking University." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.626, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Motion Mamba" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.173, + 0.786, + 0.214 + ], + "angle": 0, + "content": "1. Ahuja, C., Morency, L.P.: Language2pose: Natural language grounded pose forecasting. In: 2019 International Conference on 3D Vision (3DV). pp. 719-728. IEEE (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.215, + 0.786, + 0.242 + ], + "angle": 0, + "content": "2. Bao, F., Li, C., Cao, Y., Zhu, J.: All are worth words: a vit backbone for score-based diffusion models. arXiv preprint arXiv:2209.12152 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.242, + 0.786, + 0.269 + ], + "angle": 0, + "content": "3. Baron, E., Zimerman, I., Wolf, L.: 2-d ssm: A general spatial layer for visual transformers. arXiv preprint arXiv:2306.06635 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.269, + 0.786, + 0.31 + ], + "angle": 0, + "content": "4. Barsoum, E., Kender, J., Liu, Z.: Hp-gan: Probabilistic 3d human motion prediction via gan. In: Proceedings of the IEEE conference on computer vision and pattern recognition workshops. pp. 1418-1427 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.31, + 0.786, + 0.364 + ], + "angle": 0, + "content": "5. Bhattacharya, U., Rewkowski, N., Banerjee, A., Guhan, P., Bera, A., Manocha, D.: Text2gestures: A transformer-based network for generating emotive body gestures for virtual agents. In: 2021 IEEE virtual reality and 3D user interfaces (VR). pp. 1-10. IEEE (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.365, + 0.786, + 0.405 + ], + "angle": 0, + "content": "6. Chen, X., Jiang, B., Liu, W., Huang, Z., Fu, B., Chen, T., Yu, G.: Executing your commands via motion diffusion in latent space. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18000-18010 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.405, + 0.786, + 0.433 + ], + "angle": 0, + "content": "7. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems 34, 8780-8794 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.433, + 0.786, + 0.487 + ], + "angle": 0, + "content": "8. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. In: International Conference on Learning Representations (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.487, + 0.786, + 0.528 + ], + "angle": 0, + "content": "9. Ghosh, A., Cheema, N., Oguz, C., Theobalt, C., Slusallek, P.: Synthesis of compositional animations from textual descriptions. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 1396-1406 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.528, + 0.786, + 0.581 + ], + "angle": 0, + "content": "0. Gong, K., Lian, D., Chang, H., Guo, C., Jiang, Z., Zuo, X., Mi, M.B., Wang, X.: Tm2d: Bimodality driven 3d dance generation via music-text integration. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9942-9952 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.582, + 0.786, + 0.624 + ], + "angle": 0, + "content": "1. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial nets. Advances in neural information processing systems 27 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.624, + 0.786, + 0.651 + ], + "angle": 0, + "content": "2. Gu, A., Dao, T.: Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.651, + 0.786, + 0.69 + ], + "angle": 0, + "content": "3. Gu, A., Goel, K., Gupta, A., Ré, C.: On the parameterization and initialization of diagonal state space models. Advances in Neural Information Processing Systems 35, 35971-35983 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.691, + 0.786, + 0.718 + ], + "angle": 0, + "content": "4. Gu, A., Goel, K., Ré, C.: Efficiently modeling long sequences with structured state spaces. arXiv preprint arXiv:2111.00396 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.718, + 0.786, + 0.759 + ], + "angle": 0, + "content": "5. Gu, A., Johnson, I., Goel, K., Saab, K., Dao, T., Rudra, A., Ré, C.: Combining recurrent, convolutional, and continuous-time models with linear state space layers. Advances in neural information processing systems 34, 572-585 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.759, + 0.786, + 0.8 + ], + "angle": 0, + "content": "6. Gulrajani, I., Ahmed, F., Arjovsky, M., Dumoulin, V., Courville, A.C.: Improved training of wasserstein gans. Advances in neural information processing systems 30 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.8, + 0.786, + 0.84 + ], + "angle": 0, + "content": "7. Guo, C., Zou, S., Zuo, X., Wang, S., Ji, W., Li, X., Cheng, L.: Generating diverse and natural 3d human motions from text. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5152-5161 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.173, + 0.786, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.418, + 0.128 + ], + "angle": 0, + "content": "Zeyu and Akide et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "18. Guo, C., Zuo, X., Wang, S., Zou, S., Sun, Q., Deng, A., Gong, M., Cheng, L.: Action2motion: Conditioned generation of 3d human motions. In: Proceedings of the 28th ACM International Conference on Multimedia. pp. 2021-2029 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.787, + 0.232 + ], + "angle": 0, + "content": "19. Gupta, A., Gu, A., Berant, J.: Diagonal state spaces are as effective as structured state spaces. Advances in Neural Information Processing Systems 35, 22982-22994 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.233, + 0.786, + 0.26 + ], + "angle": 0, + "content": "20. Harvey, F.G., Yurick, M., Nowrouzezahrai, D., Pal, C.: Robust motion in-between. ACM Transactions on Graphics (TOG) 39(4), 60-1 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.261, + 0.787, + 0.302 + ], + "angle": 0, + "content": "21. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems 30 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.303, + 0.787, + 0.331 + ], + "angle": 0, + "content": "22. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems 33, 6840-6851 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.332, + 0.787, + 0.374 + ], + "angle": 0, + "content": "23. Hopfield, J.J.: Neural networks and physical systems with emergent collective computational abilities. Proceedings of the national academy of sciences 79(8), 2554-2558 (1982)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.375, + 0.787, + 0.416 + ], + "angle": 0, + "content": "24. Jiang, B., Chen, X., Liu, W., Yu, J., Yu, G., Chen, T.: Motiongpt: Human motion as a foreign language. Advances in Neural Information Processing Systems 36 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.417, + 0.787, + 0.444 + ], + "angle": 0, + "content": "25. Kalman, R.E.: A New Approach to Linear Filtering and Prediction Problems. Journal of Basic Engineering 82(1), 35-45 (1960)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.445, + 0.787, + 0.459 + ], + "angle": 0, + "content": "26. Kingma, D.P., Welling, M.: Auto-encoding variational bayes. stat 1050, 1 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.46, + 0.787, + 0.487 + ], + "angle": 0, + "content": "27. Kramer, M.A.: Nonlinear principal component analysis using autoassociative neural networks. AIChE journal 37(2), 233-243 (1991)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.488, + 0.787, + 0.516 + ], + "angle": 0, + "content": "28. Li, S., Singh, H., Grover, A.: Mamba-nd: Selective state space modeling for multidimensional data. arXiv preprint arXiv:2402.05892 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.517, + 0.787, + 0.558 + ], + "angle": 0, + "content": "29. Li, Y., Cai, T., Zhang, Y., Chen, D., Dey, D.: What makes convolutional models great on long sequence modeling? In: The Eleventh International Conference on Learning Representations (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.559, + 0.787, + 0.586 + ], + "angle": 0, + "content": "30. Lin, X., Amer, M.R.: Human motion modeling using dvgans. arXiv preprint arXiv:1804.10652 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.587, + 0.787, + 0.615 + ], + "angle": 0, + "content": "31. Liu, Y., Tian, Y., Zhao, Y., Yu, H., Xie, L., Wang, Y., Ye, Q., Liu, Y.: Vmamba: Visual state space model. arXiv preprint arXiv:2401.10166 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.616, + 0.787, + 0.657 + ], + "angle": 0, + "content": "32. Liu, Z., Mao, H., Wu, C.Y., Feichtenhofer, C., Darrell, T., Xie, S.: A convnet for the 2020s. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11976-11986 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.658, + 0.787, + 0.699 + ], + "angle": 0, + "content": "33. Mahmood, N., Ghorbani, N., Troje, N.F., Pons-Moll, G., Black, M.J.: Amass: Archive of motion capture as surface shapes. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5442-5451 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.7, + 0.787, + 0.755 + ], + "angle": 0, + "content": "34. Pavlakos, G., Choutas, V., Ghorbani, N., Bolkart, T., Osman, A.A.A., Tzionas, D., Black, M.J.: Expressive body capture: 3d hands, face, and body from a single image. In: Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR) (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.757, + 0.787, + 0.798 + ], + "angle": 0, + "content": "35. Petrovich, M., Black, M.J., Varol, G.: Action-conditioned 3D human motion synthesis with transformer VAE. In: International Conference on Computer Vision (ICCV) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.787, + 0.84 + ], + "angle": 0, + "content": "36. Petrovich, M., Black, M.J., Varol, G.: Temos: Generating diverse human motions from textual descriptions. In: European Conference on Computer Vision. pp. 480-497. Springer (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.627, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Motion Mamba" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "37. Petrovich, M., Black, M.J., Varol, G.: TEMOS: Generating diverse human motions from textual descriptions. In: European Conference on Computer Vision (ECCV) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.787, + 0.218 + ], + "angle": 0, + "content": "38. Plappert, M., Mandery, C., Asfour, T.: The kit motion-language dataset. Big data 4(4), 236-252 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.218, + 0.787, + 0.259 + ], + "angle": 0, + "content": "39. Plappert, M., Mandery, C., Asfour, T.: Learning a bidirectional mapping between human whole-body motion and natural language using deep recurrent neural networks. Robotics and Autonomous Systems 109, 13-26 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.259, + 0.787, + 0.313 + ], + "angle": 0, + "content": "40. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.787, + 0.357 + ], + "angle": 0, + "content": "41. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.357, + 0.787, + 0.398 + ], + "angle": 0, + "content": "42. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: International Conference on Medical image computing and computer-assisted intervention. pp. 234-241. Springer (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.787, + 0.425 + ], + "angle": 0, + "content": "43. Rumelhart, D.E., Hinton, G.E., Williams, R.J.: Learning internal representations by error propagation. Parallel Distributed Processing pp. 318-362 (1986)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.425, + 0.787, + 0.467 + ], + "angle": 0, + "content": "44. Sahara, C., Ho, J., Chan, W., Salimans, T., Fleet, D.J., Norouzi, M.: Image superresolution via iterative refinement. IEEE Transactions on Pattern Analysis and Machine Intelligence (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.467, + 0.787, + 0.509 + ], + "angle": 0, + "content": "45. Shi, X., Chen, Z., Wang, H., Yeung, D.Y., Wong, W.K., Woo, W.c.: Convolutional LSTM network: A machine learning approach for precipitation nowcasting. Advances in neural information processing systems 28 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.509, + 0.787, + 0.55 + ], + "angle": 0, + "content": "46. Smith, J., De Mello, S., Kautz, J., Linderman, S., Byeon, W.: Convolutional state space models for long-range spatiotemporal modeling. Advances in Neural Information Processing Systems 36 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.55, + 0.787, + 0.592 + ], + "angle": 0, + "content": "47. Sohl-Dickstein, J., Weiss, E., Maheswaranathan, N., Ganguli, S.: Deep unsupervised learning using nonequilibrium thermodynamics. In: International conference on machine learning. pp. 2256-2265. PMLR (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.592, + 0.787, + 0.633 + ], + "angle": 0, + "content": "48. Tevet, G., Gordon, B., Hertz, A., Bermano, A.H., Cohen-Or, D.: Motionclip: Exposing human motion generation to clip space. In: European Conference on Computer Vision. pp. 358-374. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.633, + 0.787, + 0.675 + ], + "angle": 0, + "content": "49. Tevet, G., Raab, S., Gordon, B., Shafir, Y., Cohen-or, D., Bermano, A.H.: Human motion diffusion model. In: The Eleventh International Conference on Learning Representations (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.675, + 0.787, + 0.702 + ], + "angle": 0, + "content": "50. Van Den Oord, A., Vinyals, O., et al.: Neural discrete representation learning. Advances in neural information processing systems 30 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.702, + 0.787, + 0.743 + ], + "angle": 0, + "content": "51. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.743, + 0.787, + 0.772 + ], + "angle": 0, + "content": "52. Wang, J., Yan, J.N., Gu, A., Rush, A.M.: Pretraining without attention. arXiv preprint arXiv:2212.10544 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.772, + 0.787, + 0.799 + ], + "angle": 0, + "content": "53. Yan, J.N., Gu, J., Rush, A.M.: Diffusion models without attention. arXiv preprint arXiv:2311.18257 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.787, + 0.841 + ], + "angle": 0, + "content": "54. Zhang, M., Cai, Z., Pan, L., Hong, F., Guo, X., Yang, L., Liu, Z.: Motiondiffuse: Text-driven human motion generation with diffusion model. IEEE Transactions on Pattern Analysis and Machine Intelligence (2024)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.418, + 0.129 + ], + "angle": 0, + "content": "Zeyu and Akide et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "55. Zhang, Y., Black, M.J., Tang, S.: We are more than our joints: Predicting how 3d bodies move. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3372-3382 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.785, + 0.232 + ], + "angle": 0, + "content": "56. Zhong, C., Hu, L., Zhang, Z., Xia, S.: Attt2m: Text-driven human motion generation with multi-perspective attention mechanism. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 509-519 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.785, + 0.273 + ], + "angle": 0, + "content": "57. Zhu, L., Liao, B., Zhang, Q., Wang, X., Liu, W., Wang, X.: Vision mamba: Efficient visual representation learning with bidirectional state space model. arXiv preprint arXiv:2401.09417 (2024)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.273 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Motion Mamba_ Efficient and Long Sequence Motion Generation/1b586349-9221-4bad-a56d-e274ec79c852_origin.pdf b/2024/Motion Mamba_ Efficient and Long Sequence Motion Generation/1b586349-9221-4bad-a56d-e274ec79c852_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..3e72778a1a067dc74b1426b38990b90891438098 --- /dev/null +++ b/2024/Motion Mamba_ Efficient and Long Sequence Motion Generation/1b586349-9221-4bad-a56d-e274ec79c852_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c0a72d2370bc1dd6ec9376c87822020048d65fa2bde9842164183544c9e0a152 +size 3543768 diff --git a/2024/Motion Mamba_ Efficient and Long Sequence Motion Generation/full.md b/2024/Motion Mamba_ Efficient and Long Sequence Motion Generation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..9dc8c8eb44b0241d3ba441af65349657d01a6a67 --- /dev/null +++ b/2024/Motion Mamba_ Efficient and Long Sequence Motion Generation/full.md @@ -0,0 +1,305 @@ +# Motion Mamba: Efficient and Long Sequence Motion Generation + +Zeyu Zhang $^{12*}$ , Akide Liu $^{1*}$ , Ian Reid $^{3}$ , Richard Hartley $^{2}$ , Bohan Zhuang $^{1}$ , and Hao Tang $^{4\boxtimes}$ + +1 Monash University + +2 The Australian National University + +3 Mohamed bin Zayed University of Artificial Intelligence + +$^{4}$ National Key Laboratory for Multimedia Information Processing, School of Computer Science, Peking University + +https://steve-zeyu-zhang.github.io/MotionMamba + +![](images/9d50515d2181f0c7d935e2bbb103cb5293cf413d46914332a1dd381edc6ed7a8.jpg) +Fig.1: Motion Mamba has achieved significantly superior performance on long sequence modeling and motion generation efficiency compared with other well-designed state-of-the-art methods such as MLD [6], MotionDiffuse [54], and MDM [49]. + +![](images/48cf11f3798a6aedad48113acb76dca9df21fbaad9e72b652f24ec0f0b7e17e3.jpg) + +![](images/210741e7ade2cfcbcf250c0c7992f4b4a689a716569803d67cdd28922bb7d8d0.jpg) + +Abstract. Human motion generation stands as a significant pursuit in generative computer vision, while achieving long-sequence and efficient motion generation remains challenging. Recent advancements in state space models (SSMs), notably Mamba, have showcased considerable promise in long sequence modeling with an efficient hardware-aware design, which appears to be a significant direction upon building motion generation model. Nevertheless, adapting SSMs to motion generation faces hurdles since the lack of a specialized design architecture to model motion sequence. To address these challenges, we propose Motion Mamba, a simple yet efficient approach that presents the pioneering motion generation model utilized SSMs. Specifically, we design a Hierarchical Temporal Mamba (HTM) block to process temporal data by ensemble varying numbers of isolated SSM modules across a symmetric U-Net architecture aimed at preserving motion consistency between frames. We also design a Bidirectional Spatial Mamba (BSM) block + +to bidirectionally process latent poses, to enhance accurate motion generation within a temporal frame. Our proposed method achieves up to $50\%$ FID improvement and up to 4 times faster on the HumanML3D and KIT-ML datasets compared to the previous best diffusion-based method, which demonstrates strong capabilities of high-quality long sequence motion modeling and real-time human motion generation. + +Keywords: Human Motion Generation $\cdot$ Selective State Space Models $\cdot$ Latent Diffusion Models + +# 1 Introduction + +Human motion generation involves creating high-quality 3D human motion based on given conditions, such as a text prompt, which generates Euler angles for each joint in a human skeleton. To emulate human motion effectively, virtual characters must respond to the conditional context, exhibit natural movement, and perform motion accurately. Recent motion generation models are categorized into four main approaches: autoencoder-based [1,10,17,36,48,56], utilizing transformers for latent space compression and motion synthesis; GAN-based [4,20,30], using discriminators to enhance the realism of generated motions; autoregressive models [24], treating motion sequences as languages with specialized codebooks; and diffusion-based [6,49,54], employing denoising steps for motion generation. Challenges vary across methods, with autoencoder models struggling to generate accurate motions from detailed descriptions due to textual information compression, GAN-based models facing training difficulties, especially in conditional tasks, and diffusion-based models relying on complex transformer-based architectures results in inefficient of motion prediction. + +Although diffusion-based models excel at generating motion with robust performance and often exhibit superior diversity, they encounter two limitations. 1) Convolutional or transformer-based diffusion methods exhibit limitations in generating long-range motion sequences. Previous transformer-based methodologies [6,49,54] have focused on modeling long-range dependencies and acquiring comprehensive global context information. Despite these advances, they are frequently associated with a substantial increase in computational requirements. Furthermore, transformer architectures are not intrinsically designed for temporal sequential modeling, which poses an inherent limitation. 2) The efficiency of inference in transformer-based diffusion methods is constrained. Although prior research has attempted to leverage the Variational Autoencoder for denoising operations in the latent space [6], the inference speed remains adversely affected by the attention mechanism's quadratic scaling, leading to inefficient motion generation. Consequently, exploring a new architectural paradigm that accommodates long-range dependencies and maintains a linear computational complexity is crucial for sustaining motion generation tasks. + +Recent advances have sparked renewed interest in state space models (SSMs) [14, 15], a field that originated from the foundational classic state space model + +[25]. Modern versions of SSMs stand out due to their ability to effectively capture long-range dependencies, a capability greatly improved by the introduction of parallel training techniques. This evolution has led to various methodologies based on SSM, notably the linear state space layers (LSSL) [15], the structured state-space sequence model (S4) [14], the diagonal state space (DSS) [19], and S4D [13]. These methods have been carefully designed to handle sequential data across various tasks and modalities, paying special attention to modeling long-range dependencies. Their efficacy in managing long sequences is attributed to the implementation of convolutional computations [15] and near-linear computational strategies, such as mamba [12], marking a significant stride in sequentially oriented tasks, including large language model decoding and motion sequence generation. + +Adapting selective state space modules for motion generation tasks presents notable challenges, primarily due to the lack of specialized design in SSMs for capturing the sensitive motion details required for temporal representation and the complexities involved in aggregating latent space. In response to these challenges, we have meticulously developed a motion generation architecture, specifically tailored to address the intricacies of long-term sequence generation, while optimizing for computational efficiency with near-linear-time complexity. This innovation is embodied in the Motion Mamba model, a simple yet potent approach to motion generation. The Motion Mamba framework pioneers a diffusion-based generative system, incorporating two key components oriented toward SSM as shown in Figure. 2: (1) a Hierarchical Temporal Mamba (HTM) block: This component is ingeniously crafted to arrange motion frames in sequential order, using hierarchically adjusted scanning. It is adept at identifying temporal dependencies at various depths, thereby facilitating a thorough comprehension of the dynamics inherent in motion sequences. (2) a Bidirectional Spatial Mamba (BSM) block: This block is designed to unravel the structured latent skeleton by evaluating data from both forward and reverse directions. Its primary goal is to safeguard the continuity of information flow, significantly bolstering the model's capacity for precise motion generation through the retention of dense informational exchange. + +The Motion Mamba introduces a new approach to motion generation that strikes an exceptional trade-off between accuracy and efficiency, shown in Fig. 1. Our experimental results underscore the significant improvements brought about by Motion Mamba, showcasing a remarkable improvement in the Fréchet Inception Distance (FID), with a reduction of up to $50\%$ from the prior state-of-the-art metric of 0.473 to an impressive 0.281 on the HumanML3D dataset [17]. Furthermore, we emphasize our framework's unparalleled inference speed, which is four times faster than previous methods, achieving an average inference time of only 0.058 seconds per sequence compared to the 0.217 seconds required by the MLD [6] method per sequence. These outcomes unequivocally establish Motion Mamba's state-of-the-art performance, concurrently ensuring fast inference speeds for conditional human motion generation tasks. + +Our contributions to the field of motion generation can be summarized as: + +1. We introduce a simple yet effective framework, named Motion Mamba, which is a pioneering method integrates a selective scanning mechanism into motion generation tasks. +2. Motion Mamba is comprised of two modules: Hierarchical Temporal Mamba (HTM) and Bidirectional Spatial Mamba (BSM), which are designed for temporal and spatial modeling, respectively. HTM blocks are tasked with processing temporal motion data, aiming to enhance motion consistency across frames. BSM blocks are engineered to bidirectionally capture the channel-wise flow of hidden information within the latent pose representations. +3. Motion Mamba framework demonstrated exceptional performance on text-to-motion generation task, through experimental validation on the HumanML3D [17] and KIT-ML [38] datasets. Our methodology achieved state-of-the-art generation quality and significantly improved long-squence modeling, meanwhile optimizing inference speed. + +# 2 Related Works + +Human Motion Generation. Generating human motion is a significant application of computer vision, essential for various applications like 3D modelling and robot manipulation. Recently, the predominant method of achieving human motion generation, known as the Text-to-Motion task, involves learning a common latent space for both language and motion. + +DVGAN [30] creates the GAN [11] discriminator by densely validating at each time-scale and perturbing the discriminator input for translation invariance, enabling motion generation and completion. ERD-QV [20] enhances latent representations through two additive modifiers: a time-to-arrival embedding applied universally and an additive scheduled target noise vector used during extended transitions. It further improves transition quality by incorporating a GAN framework with two discriminators operating at different timescales. HP-GAN [4], trained with a modified version of the improved WGAN-GP [16], utilizes a custom loss function designed for human motion prediction. It learns a probability density function of future human poses conditioned on previous poses. + +Autoencoders [27, 43] are notable generative models known for their ability to represent data robustly by compressing high-dimensional data into a latent space, which is widely adopted in human motion generation. JL2P [1] uses RNN-based autoencoders [23] to learn a combined representation of language and pose. It restricts the direct mapping from text to motion to a one-to-one relationship. MotionCLIP [48] uses Transformer-based Autoencoders [51] to reconstruct motion while ensuring alignment with the corresponding text label in the CLIP [40] space. This alignment effectively integrates the semantic knowledge from CLIP into the human motion manifold. TEMOS [36] and T2M [17] combine a Transformer-based VAE [26] with a text encoder to generate distribution parameters that work within the VAE latent space. AttT2M [56] and TM2D [10] incorporate a body-part spatio-temporal encoder into VQ-VAE [50] for enhanced learning of a discrete latent space with increased expressiveness. + +Diffusion models [7, 22, 41, 47] have recently surpassed GANs and VAEs in generating 2D images. Developing a motion generation model based on diffusion models is obviously an attractive direction. MotionDiffuse [54] introduces the inaugural framework for text-driven motion generation based on diffusion models. It showcases several desirable properties, including probabilistic mapping, realistic synthesis, and multi-level manipulation. MDM [49] utilizes a classifier-free Transformer-based diffusion model for the human motion domain to predict sample rather than noise in each diffusion step. MLD [6] performs a diffusion process in latent motion space, rather than using a diffusion model to establish connections between raw motion sequences and conditional inputs. + +State Space Models. Recently, state space sequence models (SSMs) [14, 15], drawing inspiration from classical state-space models [25], have emerged as a promising architecture for sequence modeling. Mamba [12] introduces a selective SSM architecture, integrating time-varying parameters into the SSM framework, and proposes a hardware-aware algorithm to facilitate highly efficient training and inference processes. Some research works leverage SSM in computer vision to process 2D data. The 2D SSM [3] introduces an SSM block at the beginning of each transformer block [8, 51]. This approach aims to achieve efficient parameterization, accelerated computation, and a suitable normalization scheme. SGConvNeXt [29] presents a structured global convolution method inspired by ConvNeXt [32], incorporating multi-scale sub-kernels to achieve both parameterization efficiency and effective long sequence modeling. ConvSSM [46] integrates the tensor modeling principles of ConvLSTM [45] with SSMs, elucidating the utilization of parallel scans in convolutional recurrences. This approach enables subquadratic parallelization and rapid autoregressive generation. Vim [57] introduces a bidirectional SSM block [52] for efficient and versatile visual representation learning, achieving performance comparable to established ViT [8] methods. VMamba [31] introduces a Cross-Scan Module (CSM) designed to traverse the spatial domain and transform any non-causal visual image into ordered patch sequences. This approach achieves linear complexity while preserving global receptive fields. There have also been attempts to utilize SSMs to handle higher-dimensional data. Mamba-ND [28] explores various combinations of SSM and different scan directions within the SSM block to adapt Mamba [12] to higher-dimensional tasks. Recent efforts have sought to replace the traditional transformer-based U-Net within the diffusion denoiser with the SSM block, with the aim of enhancing image generation efficiency. DiffuSSM [53] adeptly manages higher resolutions without relying on global compression, thus maintaining detailed image representation throughout the diffusion process. + +# 3 The Proposed Method + +In this section, we delineate the architecture and operational principles of the Motion Mamba framework, designed for generating human motion over long ranges efficiently from textual descriptions. Initially, we discuss the foundational concepts underpinning our approach, including the Mamba Model [12] and the + +latent diffusion model [6]. Following this, we detail our uniquely crafted architecture that leverages the Mamba model to enhance motion generation efficiency. This architecture comprises two principal components: the Hierarchical Temporal Mamba (HTM) block, which addresses temporal aspects, and the Bidirectional Spatial Mamba (BSM) block, focusing on spatial dynamics. + +# 3.1 Preliminaries + +Selective Structured State Space Sequence Model. SSMs particularly through the contributions of structured state space sequence models (S4) and Mamba, have demonstrated exceptional proficiency in handling long sequences. These models operationalize the mapping of a 1-D function or sequence, $x(t) \in \mathbb{R} \mapsto y(t) \in \mathbb{R}$ , through a hidden state $h(t) \in \mathbb{R}^N$ , employing $\mathbf{A} \in \mathbb{R}^{N \times N}$ as the evolution parameters, $\mathbf{B} \in \mathbb{R}^{N \times 1}$ and $\mathbf{C} \in \mathbb{R}^{1 \times N}$ as the projection parameters, respectively. + +The discretized system can then be expressed as follows, incorporating a step size $\Delta$ : + +$$ +\begin{array}{c} h _ {t} = \overline {{\mathbf {A}}} h _ {t - 1} + \overline {{\mathbf {B}}} x _ {t}, \\ \mathbf {C} ^ {\prime} \end{array} \tag {1} +$$ + +$$ +y _ {t} = \mathbf {C} h _ {t}. +$$ + +This adaptation facilitates the computation of output through global convolution, leveraging a structured convolutional kernel $\overline{\mathbf{K}}$ , which encompasses the entire length $M$ of the input sequence $\mathbf{x}$ : + +$$ +\overline {{\mathbf {K}}} = \left(\underset {-} {\mathbf {C}} \overline {{\mathbf {B}}}, \mathbf {C} \overline {{\mathbf {A}}} \mathbf {B}, \dots , \mathbf {C} \overline {{\mathbf {A}}} ^ {M - 1} \overline {{\mathbf {B}}}\right), \tag {2} +$$ + +$$ +\mathbf {y} = \mathbf {x} * \overline {{\mathbf {K}}}. +$$ + +Selective models like Mamba introduce time-varying parameters, deviating from the linear time invariance (LTI) assumption and complicating parallel computation. However, hardware-aware optimizations, such as associative scans, have been developed to address these computational challenges, highlighting the ongoing evolution and application of SSMs in modeling complex temporal dynamics. + +Latent Motion Diffusion Model. Diffusion probabilistic models offer a significant advancement in motion generation by gradually reducing noise from a Gaussian distribution to a target data distribution $p(x)$ through a T-length learned Markov process [7,22,41,44,47,49,54], giving $\{\pmb{x}_t\}_{t=1}^T$ . In the motion generation, we define our trainable diffusion models with a denoiser $\epsilon_{\theta}(x_t,t)$ which anneal the random noise to motion sequence $\{\hat{x}_t^{1:N}\}_{t=1}^T$ iteratively. To address the inefficiencies of applying diffusion models directly to raw motion sequences, we employ a low-dimensional motion latent space for the diffusion process. Given an input condition $c$ , such as a descriptive sentence $\pmb{w}^{1:N} = \{w^i\}_{i=1}^N$ , an action label $a$ from a predefined set $\mathcal{A}$ , or an empty condition $c = \emptyset$ , and the motion representation that combines 3D joint rotations, positions, velocities, and foot contact as proposed in [17]. The frozen CLIP [40] text encoder $\tau_{\theta}^w$ has been employed to obtain projected text embedding $\tau_{\theta}^w(w^{1:N}) \in \mathbb{R}^{1 \times d}$ , thereby conditional denoiser comprised in term of $\epsilon_{\theta}(z_t,t,\tau_{\theta}(c))$ . The latent diffusion model $\epsilon_{\theta}(x_t,t)$ aimed + +![](images/c47e35af41944d8524c810593fb0260c0e75549c04159ef1f62bdcc5c0d563b4.jpg) +Fig. 2: This figure illustrates the architecture of the proposed Motion Mamba model. Each of encoder and decoder blocks consists of a Hierarchical Temporal Mamba block (HTM) and a Bidirectional Spatial Mamba (BSM) block, which possess hierarchical scan and bidirectional scan within SSM layers respectively. This symmetric distribution of scans ensure a balanced and coherence framework across the encoder-decoder architecture. + +to generate the human motion sequence in terms of $\hat{x}^{1:L} = \{\hat{x}^i\}_{i=1}^L$ , where L denotes the sequence length or number of frames [34,35,37,55]. After that we reused the motion Variational AutoEncoder (VAE) $\mathcal{V} = \{\mathcal{E},\mathcal{D}\}$ proposed in MLD [6] to manipulate the motion sequence in latent space $z = \mathcal{E}(x^{1:L})$ , and decompress the intermediate representation to motion sequence by $\hat{x}^{1:L} = \mathcal{D}(z) = \mathcal{D}\mathcal{E}(x^{1:L})$ [26, 42, 51]. Finally, our latent diffusion model is trained with an objective focusing on minimization of MSE between true and predicted noise in the latent space, facilitating efficient and high-quality motion generation [2, 22]. + +# 3.2 Motion Mamba + +The architecture of the proposed Motion Mamba framework is illustrated in Figure. 2. At its core, Motion Mamba utilizes a denoising U-Net architecture, which is distinguished for its effectiveness in modeling the continuous, temporal sequences of motion frames. This effectiveness is attributed to the inherent long-sequence modeling capacity of the Mamba model. The denoiser, denoted by $\epsilon_{\theta}$ , comprises $N$ blocks including encoder $E_{1..N}$ and decoder $D_{1..N}$ . Additionally, the architecture is enhanced with a transformer-based attention mixer block $M$ , designed to augment the model's ability to capture complex temporal dynamics. + +$$ +\epsilon_ {\theta} (x) \equiv \{E _ {1 \dots N}, M, D _ {1 \dots N} \}. \tag {3} +$$ + +The encoder blocks are represented as $E_{1..N}$ , arranged sequentially, and the decoder blocks as $D_{1..N}$ , configured in reverse order to facilitate effective bottom-up and top-down information flow. Given that selective operations have signif + +Algorithm 1 Hierarchical Temporal Mamba (HTM) Block. +Require: compressed latent representations $z:(T,B,C)$ +Ensure: transformed representations $z_{\mathrm{HTM}}:(T,B,E)$ +1:/\* linear projection layer \*/ +2: $x,z:(T,B,E)\gets \mathrm{Linear}(z)$ +3:/\* set of scans and memory matrices \*/ +4: $K = \{S_{2N_n - 1},S_{2N_{n - 1} - 1},\ldots ,S_1\}$ +5: Memory matrices: $\{A_1,\dots ,A_k\}$ +6: for each scan $S_{i}$ in $K$ do +7: $x_o^{\prime}:(T,B,E)\gets \mathrm{Conv1D}(x)$ +8: $B_{o},C_{o},\varDelta_{o}\gets \mathrm{Linear}(x_{o}^{\prime})$ +9: Transform $\overline{A}_o$ and $\overline{B}_o$ using $\varDelta_{o}$ +10: $O_{i}\gets \mathrm{SSMA}_{i,x}(x_{o}^{\prime})$ +11: end for +12:/\* aggregation of outputs \*/ +13: $z_{\mathrm{HTM}}:(T,B,E)\gets \mathrm{Linear(Aggregate(\{O_1,\ldots,O_k\}))}$ +14: Return: $z_{\mathrm{HTM}}$ + +icantly lower computational complexity compared to attention-based methods, we have increased the number of scans to achieve higher quality generations. Concurrently, it is imperative to maintain a balance between the model's parameters and its efficiency. Thereby, a novel aspect of our model is the introduction of a hierarchical scan strategy, characterized by a sequence of scan numbers as, + +$$ +K = \left\{S _ {2 N - 1}, S _ {2 (N - 1) - 1}, \dots , S _ {1} \right\}. \tag {4} +$$ + +This sequence specifies the number of scans allocated to each layer, in descending order of complexity. For instance, the uppermost encoder layer, $E_{1}$ , and the lowermost decoder layer, $D_{N}$ , are allocated $S_{2N - 1}$ scans, indicating the highest scanning complexity. Conversely, the lowest encoder layer, $E_{N}$ , and the uppermost decoder layer, $D_{1}$ , are assigned $S_{1}$ scans, reflecting the lowest level of scanning complexity. + +$$ +E _ {i} (S) = \left\{ \begin{array}{l l} S _ {2 N - 1} & \text {f o r} i = 1 \\ S _ {2 (N - i) - 1} & \text {f o r} i = 2, \dots , N - 1 \\ S _ {1} & \text {f o r} i = N \end{array} \right. \tag {5} +$$ + +$$ +D _ {j} (S) = \left\{ \begin{array}{l l} S _ {2 N - 1} & \text {f o r} j = N \\ S _ {2 (N - j) - 1} & \text {f o r} j = N - 1, \dots , 2 \\ S _ {1} & \text {f o r} j = 1 \end{array} \right. \tag {6} +$$ + +This hierarchical scanning approach ensures that processing capabilities are evenly distributed throughout the encoder-decoder architecture., facilitating a detailed and nuanced analysis of temporal sequences. Within this structured framework, each denoiser is equipped with a specialized Hierarchical Temporal Mamba (HTM) block, which serves to augment the model's ability to process + +Algorithm 2 Bidirectional Spatial Mamba (BSM) Block. +Require: compressed latent representations $z:(T,B,C)$ +Ensure: transformed representations $z_{\mathrm{BSM}}:(C,B,E)$ +1:/\* dimension rearrangement \*/ +2: $z^{\prime}:(C,B,T)\gets$ Rearrange(z) +3:/\* linear projection after normalization \*/ +4: $z^{\prime}:(C,B,T)\gets$ Norm(z) +5:x,z:(C,B,E) $\leftarrow$ Linear(z') +6: for o in {forward, backward} do +7: $x_o':(C,B,E)\gets$ Conv1D(x) +8: $B_{o},C_{o},\varDelta_{o}\gets$ Linear $(x_o^{\prime})$ +9: Transform $\overline{A}_o$ and $\overline{B}_o$ using $\varDelta_{o}$ +10: $y_{o}\gets$ SSM(Ao,Bo,Co) +11: end for +12:/\* gating and summing outputs \*/ +13: $z_{\mathrm{BSM}}:(C,B,T)\gets$ GateAndSum(yforward,ybackward,z) +14: $z_{\mathrm{BSM}}^{\prime}:(T,B,C)\gets$ Rearrange(z) +15: Return: $z_{\mathrm{BSM}}^{\prime}$ + +temporal information effectively. Additionally, the proposed Motion Mamba incorporates an attention-based mixer block denoted as $M$ , strategically integrated to enhance conditional fusion. + +Hierarchical Temporal Mamba (HTM) block processes compressed latent representations, denoted as $z$ , with the dimensions $(T,B,C)$ , of which procedure shown in Algorithm 1. Here, $T$ signifies the temporal dimension, as specified in the Variational AutoEncoder (VAE) framework. Initially, the input $z$ is subjected to a linear projection layer, producing transformed representations $x$ and $z$ with dimension $E$ . Our analysis revealed an increased density of motion within the lower-level feature spaces. Consequently, we developed a hierarchical scanning methodology that is executed at various depths of the network. This approach not only accommodates the diverse motion densities, but also significantly reduces computational overhead. This step utilizes a hierarchically structured set of scans, $K = \{S_{2N_n - 1}, S_{2N_{n - 1} - 1}, \ldots, S_1\}$ , in conjunction with a corresponding series of memory matrices $\{A_1, \ldots, A_k\}$ . Each sub-SSM scan first applies a 1-D convolution to $x$ , resulting in $x_o'$ . $x_o'$ is then linearly projected to derive $B_o$ , $C_o$ , and $\Delta_o$ . These projections $B_o$ , $C_o$ use $\Delta_o$ to effect transformations in $\overline{A}_o$ and $\overline{B}_o$ , respectively. After executing a sequence of SSM scans $\{SSM_{A_1,x}, SSM_{A_2,x}, \ldots, SSM_{A_k,x}\}$ , a set of outputs $\{O_1, \ldots, O_k\}$ is compiled. This collection is subsequently aggregated via a linear projection to obtain the final output of the HTM block. + +Bidirectional Spatial Mamba (BSM) block focuses on enhancing latent representation learning through a novel approach of dimension rearrangement and bidirectional scanning, of which the process is shown in Algorithm 2. Initially, it alters the original input dimensions from $(T,B,C)$ to $(C,B,T)$ , effectively swap + +ping the temporal and channel dimensions. After this rearrangement, the input, now denoted $z'$ , undergoes a linear projection after normalization, resulting in dimensions $x$ and $z$ of size $E$ . The process involves bidirectional scanning of the latent channel dimension, where $\mathbf{x}$ is subjected to a 1-D convolution, yielding $\mathbf{x}_o'$ for both forward and backward directions. Each $\mathbf{x}_o'$ is then linearly projected to obtain $B_o$ , $C_o$ , and $\Delta_o$ , which are utilized to transform $\overline{A}_o$ and $\overline{B}_o$ , respectively. The final output token sequence, $\mathbf{z}_1$ , is computed by gating and summing the forward $y_{\mathrm{forward}}$ and backward $y_{\mathrm{backward}}$ output with $\mathbf{z}$ . This component is engineered to decode the structured latent skeleton by analyzing data from both forward and reverse viewpoints. Its main objective is to ensure the seamless continuity of information flow, thereby substantially enhancing the model's ability to generate accurate motion. This is achieved through the maintenance of a dense informational exchange, which is critical for the model's performance. + +# 4 Experiments + +# 4.1 Datasets + +We evaluate our proposed Motion Mamba on two prominent Text-to-Motion synthesis benchmarks as follows: + +HumanML3D. The HumanML3D [17] dataset aggregates 14,616 motions sourced from the AMASS [33] and HumanAct12 [18] datasets, with each motion accompanied by three textual descriptions, culminating in 44,970 scripts. This dataset spans a wide range of actions such as exercising, dancing, and acrobatics, presenting a rich motion-language corpus. + +KIT-ML. The KIT-ML dataset [38] is comprised of 3,911 motions paired with 6,278 textual descriptions, serving as a compact yet effective benchmark for evaluation. For both datasets, the pose representation adopted is derived from T2M [17], ensuring consistency in motion representation across evaluations. + +# 4.2 Evaluation Metrics + +We adapt the standard evaluation metrics on following aspects throughout our experiments, including: Generation Quality. We implement a Fréchet inception distance (FID) [21] to quantify the realism and diversity of motion generated by models. Moreover, we use multi-modal distance (MM Dist) to measure the distance between motions and texts and assess motion-text alignment. Diversity. We use the diversity metric to measure motion diversity, which calculates variance in features extracted from the motions. Additionally, we employ multimodality (MModality) to assess diversity within generated motions sharing the same text description. + +# 4.3 Comparative Studies + +We evaluate our method against the state-of-the-art methods on the HumanML3D [17] and KIT-ML [38] datasets. We train our Motion Mamba with HTM arrangement strategy MM ( $\{S_{2N_n - 1},\ldots ,S_1\}$ ), BSM bidirectional block strategy on the + +Table 1: Comparison of text-conditional motion synthesis on HumanML3D [17]. These metrics are evaluated by the motion encoder from [17]. Empty MModality indicates the non-diverse generation methods. We employ real motion as a reference and sort all methods by descending FIDs. The right arrow $\rightarrow$ means that the closer to the real motion, the better. Bold and underline indicate the best and second best result. + +
MethodR Precision ↑FID↓MM Dist↓Diversity→MModality↑
Top 1Top 2Top 3
Real0.511±.0030.703±.0030.797±.0020.002±.0002.974±.0089.503±.065-
Seq2Seq [39]0.180±.0020.300±.0020.396±.00211.75±.0355.529±.0076.223±.061-
LJ2P [1]0.246±.0010.387±.0020.486±.00211.02±.0465.296±.0087.676±.058-
T2G [5]0.165±.0010.267±.0020.345±.0027.664±.0306.030±.0086.409±.071-
Hier [9]0.301±.0020.425±.0020.552±.0046.532±.0245.012±.0188.332±.042-
TEMOS [37]0.424±.0020.612±.0020.722±.0023.734±.0283.703±.0088.973±.0710.368±.018
T2M [17]0.457±.0020.639±.0030.740±.0031.067±.0023.340±.0089.188±.0022.090±.083
MDM [49]0.320±.0050.498±.0040.611±.0070.544±.0445.566±.0279.559±.0862.799±.072
MotionDiffuse [54]0.491±.0010.681±.0010.782±.0010.630±.0013.113±.0019.410±.0491.553±.042
MLD [6]0.481±.0030.673±.0030.772±.0020.473±.0133.196±.0109.724±.0822.413±.079
Motion Mamba (Ours)0.502±.0030.693±.0020.792±.0020.281±.0093.060±.0589.871±.0842.294±.058
+ +Table 2: We involve KIT-ML [38] dataset and evaluate the SOTA methods on the text-to-motion task. + +
MethodR Precision ↑FID↓MM Dist↓Diversity→MModality↑
Top 1Top 2Top 3
Real0.424±.0050.649±.0060.779±.0060.031±.0042.788±.01211.08±.097-
Seq2Seq [39]0.103±.0030.178±.0050.241±.00624.86±.3487.960±.0316.744±.106-
T2G [5]0.156±.0040.255±.0040.338±.00512.12±.1836.964±.0299.334±.079-
LJ2P [1]0.221±.0050.373±.0040.483±.0056.545±.0725.147±.0309.073±.100-
Hier [9]0.255±.0060.432±.0070.531±.0075.203±.1074.986±.0279.563±.0722.090±.083
TEMOS [37]0.353±.0060.561±.0070.687±.0053.717±.0513.417±.01910.84±.1000.532±.034
T2M [17]0.370±.0050.569±.0070.693±.0072.770±.1093.401±.00810.91±.1191.482±.065
MDM [49]0.164±.0040.291±.0040.396±.0040.497±.0219.191±.02210.85±.1091.907±.214
MotionDiffuse [54]0.417±.0040.621±.0040.739±.0041.954±.0622.958±.00511.10±.1430.730±.013
MLD [6]0.390±.0080.609±.0080.734±.0070.404±.0273.204±.02710.80±.1172.192±.071
Motion Mamba (Ours)0.419±.0060.645±.0050.765±.0060.307±.0413.021±.02511.02±.0981.678±.064
+ +latent dimension $= 2$ with 11 layers. We evaluate our model and previous works with suggested metrics in HumanML3D [17] and calculate $95\%$ confidence interval by repeat evaluation 20 times. The results for the HumanML3D dataset are presented in Table 1. Our model outperforms other methods significantly across various evaluation metrics, including FID, R precision, multi-modal distance, and diversity. For instance, our Motion Mamba outperforms previous best diffusion based motion generation MLD by $40.5\%$ in terms of FID, and up to $10\%$ improvement on R Precision, we aslo obatined best MModality by 3.060. The results for the KIT-ML dataset are presented in Table 2. We have also outperformed other well-established methods in FID and multi-modal distance. + +# 4.4 Ablation Studies + +We concluded the ablation studies including long sequence evaluation, hierarchical design with HTM, bidirectional design in the BSM, number of latent dimensions, and number of layers of our proposed motion mamba in Table 4. + +Table 3: In order to evaluate the models' capability in long sequence motion generation, we compared our method with an existing approach on the recently introduced HumanML3D-LS dataset. This dataset comprises motion sequences longer than 190 frames from the original evaluation set. Our model demonstrates superior performance compared to other methods. + +
MethodR Precision ↑FID↓MM Dist↓Diversity→MModality↑
Top 1Top 2Top 3
Real0.437±.0030.622±.0040.721±.0040.004±.0003.343±.0158.423±.090-
MDM [49]0.368±.0050.553±.0060.672±.0050.802±.0443.860±.0258.817±.068-
MotionDiffuse [54]0.367±.0040.521±.0040.623±.0042.460±.0623.789±.0058.707±.1431.602±.013
MLD [6]0.403±.0050.584±.0050.690±.0050.952±.0203.580±.0169.050±.0852.711±.104
Motion Mamba (Ours)0.417±.0030.606±.0030.713±.0040.668±.0193.435±.0159.021±.0702.373±.084
+ +Long Sequence Motion Generation. The HumanML3D [17] dataset exhibits a long-tailed and right-skewed distribution with a significant proportion of long-sequence human motions, as shown in Figure 3. We suggest previous study overlooked the challenges in the long-sequence generation problem. Thus, we introduce a new dataset variant, HumanML3D-LS, comprising motion sequences longer than 190 frames extracted from the original test set. This addition allows us to showcase our capability in generating long-sequence motions. Subsequently, we eval + +![](images/aae2391766f3828f2e11babdcf89aea649a1a4ec6a3b1b4196a8a1944d3acb76.jpg) +Fig. 3: The figure shows a long tail distribution of the HumanML3D [17], which has a significant proportion of long-sequence human motions. + +uate the performance of our method on HumanML3D-LS and compare it with other diffusion-based motion generation approaches. The comparative results are presented in Table 3. Motion Mamba by leverage the benefits on long-range dependency modeling make it well suitable for long sequence motion generation. + +Hierarchical Design with HTM. In our ablation studies, we observed a slight improvement upon reversing the scan order from a lower to a higher level, specifically transitioning from MM $\{S_1,\ldots ,S_N\}$ to MM $\{S_N,\dots ,S_1\}$ . This enhancement suggests a correlation with the increase in temporal motion density within the lower-level feature spaces. Furthermore, to achieve the optimal result, we introduce the hierarchical design to arrange the scanning frequency, resulting in the sequence MM $\{S_{2N_n - 1},\ldots ,S_1\}$ . This expansion in the number of scans led to a performance increase. We attribute this enhancement to the observation that individual selective scan operations significantly reduce the parameter count, especially when compared to the parameter-intensive constructs of self-attention and feedforward network blocks prevalent in transformer architectures. + +Bidirectional Design in BSM. We developed three distinct variations of latent scanning mechanisms, differentiated by their scanning directions. In the context of motion generation tasks, we posit that the flow of hidden information within the structured latent skeleton holds significance, an aspect previously underexplored. Our ablation study reveals that a single scan across the la + +Table 4: Evaluation of text-based motion synthesis on HumanML3D [17]: we use metrics in Table 1 and provides real reference, we evaluate the various HTM and BSM design choices, the dimension of the latent input, the different number of layer of Motion Mamba model. + +
ModelsR Precision Top 3↑FID↓MM Dist.↓Diversity→MModality↑
Real0.797±.0020.002±.0002.974±.0089.503±.065-
MM ( {S1, ..., SN})0.673±.0031.278±.0123.802±.0418.678±.0963.127±.024
MM ( {SN, ..., S1})0.738±.0020.962±.0113.433±.0039.180±.0712.723±.033
MM ( {S1, ..., S2Nn-1})0.698±.0020.856±.0083.624±.0379.229±.0672.826±.017
MM ( {S2Nn-1, ..., S1})0.792±.0020.281±.0093.060±.0589.871±.0842.294±.058
MM (SingleScan)0.736±.0031.063±.0103.443±.0269.180±.0672.676±.041
MM (BiScan, layer)0.735±.0040.789±.0073.408±.0349.374±.0592.591±.046
MM (BiScan, block)0.792±.0020.281±.0093.060±.0589.871±.0842.294±.058
MM (Dim, 1)0.706±.0030.652±.0113.541±.0729.141±.0822.612±.055
MM (Dim, 2)0.792±.0020.281±.0093.060±.0589.871±.0842.294±.058
MM (Dim, 5)0.741±.0080.728±.0093.307±.0279.427±.0992.314±.062
MM (Dim, 7)0.738±.0040.599±.0073.359±.0689.166±.0752.488±.037
MM (Dim, 10)0.715±.0030.628±.0083.548±.0439.200±.0752.884±.096
MM (9 layers)0.755±.0021.080±.0123.309±.0579.721±.0812.974±.039
MM (11 layers)0.792±.0020.281±.0093.060±.0589.871±.0842.294±.058
MM (27 layers)0.750±.0030.975±.0083.336±.0969.249±.0712.821±.063
MM (37 layers)0.754±.0050.809±.0103.338±.0619.355±.0622.741±.077
+ +tent dimension yields minimal improvement. Subsequently, we investigated both layer-based and block-based bidirectional scans. Our findings indicate that the block-based bidirectional scan achieves optimal performance. This suggests that spatial information flows are encoded within the latent spaces and that bidirectional scanning facilitates the exchange of this information, thereby enhancing the efficacy of motion generation tasks. + +Architecture Design for Motion Mamba. The proposed Motion Mamba which is grounded in a standardized motion latent diffusion system. We delved into the interplay between dimensional aspects and the module's capacity (measured by the number of layers) to ascertain their impact on system performance. Experimental results demonstrate that the Motion Mamba achieves superior performance at a latent dimension of 2, diverging from prior works where the optimal dimension was identified as 1. We attribute this discrepancy to our HTM, which necessitates multiple scans correlating with the sequence length, thus implicating dimensionality as a pivotal factor. A marginal increase in dimensionality enabled us to attain peak performance, simultaneously enhancing efficiency compared to models with a dimensionality of 10. Furthermore, we conducted experiments to determine the optimal layer count for Motion Mamba, inspired by the design of its selective scanning mechanism. Notably, a single Mamba layer achieves a parameter reduction of approximately $75\%$ compared to a conventional transformer encoder block. By increasing the number of layers, we aim to uncover the relationship between model capacity and its performance. Our findings reveal that, through the integration of our specially designed HTM and BSM (Bidirectional Scanning Module) blocks, the Motion Mamba reaches its optimal performance with 11 layers. This represents a slight increase over the MLD [6] baseline. How- + +ever, due to the reduced parameter count in each layer, Motion Mamba exhibits significantly greater efficiency than previous methodologies. + +# 4.5 Inference Time + +Inference time remains a significant challenge for diffusion-based methods. To address this, we enhance the inference speed by incorporating the efficient Mamba block within a lightweight architecture. Compared to the previous strong baseline, such as the MLD model cited in [6], which reports an average inference time of 0.217 seconds, our Motion Mamba model achieves a notable reduction in computational overhead, as shown in Figure 4. Specifically, it requires four times less computational effort, thereby facilitating faster and real-time inference speeds. + +![](images/3f16a76fbbc0786783ac633696339787bd4ab38a49ca1a6d7338b047d5fb89d2.jpg) +Fig.4: The figure shows the average inference time per sentence (AIT) vs FID, our proposed motion mamba obtained 0.058s AIT and 0.281 FID over all outperform previous methods. We evaluate all methods on a single V100 GPU. + +# 5 Discussion and Conclusion + +In this study, we introduced Motion Mamba, a novel framework designed for efficient and extended sequence motion generation. Our approach represents the inaugural integration of the Mamba model within the domain of motion generation, featuring significant advancements including the implementation of Hierarchical Temporal Mamba (HTM) blocks. These blocks are specifically engineered to enhance temporal alignment through hierarchically organized selective scanning. Furthermore, Bidirectional Spatial Mamba (BSM) blocks have been developed to amplify the exchange of information flow within latent spaces, thereby augmenting the model's ability to bidirectionally capture skeleton-level density features with greater precision. Compared to previous diffusion-based motion generation methodologies that predominantly utilize transformer blocks, our Motion Mamba framework achieves SOTA performance, evidencing an improvement of up to $50\%$ in FID scores and a quadrupled improvement in inference speed. Through comprehensive experimentation across a variety of human motion generation tasks, the effectiveness and efficiency of our proposed Motion Mamba model have been robustly demonstrated, marking a significant leap forward in the field of human motion generation. + +# Acknowledgements + +This work is partially supported by the Fundamental Research Funds for the Central Universities, Peking University. + +# References + +1. Ahuja, C., Morency, L.P.: Language2pose: Natural language grounded pose forecasting. In: 2019 International Conference on 3D Vision (3DV). pp. 719-728. IEEE (2019) +2. Bao, F., Li, C., Cao, Y., Zhu, J.: All are worth words: a vit backbone for score-based diffusion models. arXiv preprint arXiv:2209.12152 (2022) +3. Baron, E., Zimerman, I., Wolf, L.: 2-d ssm: A general spatial layer for visual transformers. arXiv preprint arXiv:2306.06635 (2023) +4. Barsoum, E., Kender, J., Liu, Z.: Hp-gan: Probabilistic 3d human motion prediction via gan. In: Proceedings of the IEEE conference on computer vision and pattern recognition workshops. pp. 1418-1427 (2018) +5. Bhattacharya, U., Rewkowski, N., Banerjee, A., Guhan, P., Bera, A., Manocha, D.: Text2gestures: A transformer-based network for generating emotive body gestures for virtual agents. In: 2021 IEEE virtual reality and 3D user interfaces (VR). pp. 1-10. IEEE (2021) +6. Chen, X., Jiang, B., Liu, W., Huang, Z., Fu, B., Chen, T., Yu, G.: Executing your commands via motion diffusion in latent space. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18000-18010 (2023) +7. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems 34, 8780-8794 (2021) +8. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. In: International Conference on Learning Representations (2020) +9. Ghosh, A., Cheema, N., Oguz, C., Theobalt, C., Slusallek, P.: Synthesis of compositional animations from textual descriptions. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 1396-1406 (2021) +0. Gong, K., Lian, D., Chang, H., Guo, C., Jiang, Z., Zuo, X., Mi, M.B., Wang, X.: Tm2d: Bimodality driven 3d dance generation via music-text integration. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9942-9952 (2023) +1. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial nets. Advances in neural information processing systems 27 (2014) +2. Gu, A., Dao, T.: Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752 (2023) +3. Gu, A., Goel, K., Gupta, A., Ré, C.: On the parameterization and initialization of diagonal state space models. Advances in Neural Information Processing Systems 35, 35971-35983 (2022) +4. Gu, A., Goel, K., Ré, C.: Efficiently modeling long sequences with structured state spaces. arXiv preprint arXiv:2111.00396 (2021) +5. Gu, A., Johnson, I., Goel, K., Saab, K., Dao, T., Rudra, A., Ré, C.: Combining recurrent, convolutional, and continuous-time models with linear state space layers. Advances in neural information processing systems 34, 572-585 (2021) +6. Gulrajani, I., Ahmed, F., Arjovsky, M., Dumoulin, V., Courville, A.C.: Improved training of wasserstein gans. Advances in neural information processing systems 30 (2017) +7. Guo, C., Zou, S., Zuo, X., Wang, S., Ji, W., Li, X., Cheng, L.: Generating diverse and natural 3d human motions from text. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5152-5161 (2022) + +18. Guo, C., Zuo, X., Wang, S., Zou, S., Sun, Q., Deng, A., Gong, M., Cheng, L.: Action2motion: Conditioned generation of 3d human motions. In: Proceedings of the 28th ACM International Conference on Multimedia. pp. 2021-2029 (2020) +19. Gupta, A., Gu, A., Berant, J.: Diagonal state spaces are as effective as structured state spaces. Advances in Neural Information Processing Systems 35, 22982-22994 (2022) +20. Harvey, F.G., Yurick, M., Nowrouzezahrai, D., Pal, C.: Robust motion in-between. ACM Transactions on Graphics (TOG) 39(4), 60-1 (2020) +21. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems 30 (2017) +22. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems 33, 6840-6851 (2020) +23. Hopfield, J.J.: Neural networks and physical systems with emergent collective computational abilities. Proceedings of the national academy of sciences 79(8), 2554-2558 (1982) +24. Jiang, B., Chen, X., Liu, W., Yu, J., Yu, G., Chen, T.: Motiongpt: Human motion as a foreign language. Advances in Neural Information Processing Systems 36 (2024) +25. Kalman, R.E.: A New Approach to Linear Filtering and Prediction Problems. Journal of Basic Engineering 82(1), 35-45 (1960) +26. Kingma, D.P., Welling, M.: Auto-encoding variational bayes. stat 1050, 1 (2014) +27. Kramer, M.A.: Nonlinear principal component analysis using autoassociative neural networks. AIChE journal 37(2), 233-243 (1991) +28. Li, S., Singh, H., Grover, A.: Mamba-nd: Selective state space modeling for multidimensional data. arXiv preprint arXiv:2402.05892 (2024) +29. Li, Y., Cai, T., Zhang, Y., Chen, D., Dey, D.: What makes convolutional models great on long sequence modeling? In: The Eleventh International Conference on Learning Representations (2022) +30. Lin, X., Amer, M.R.: Human motion modeling using dvgans. arXiv preprint arXiv:1804.10652 (2018) +31. Liu, Y., Tian, Y., Zhao, Y., Yu, H., Xie, L., Wang, Y., Ye, Q., Liu, Y.: Vmamba: Visual state space model. arXiv preprint arXiv:2401.10166 (2024) +32. Liu, Z., Mao, H., Wu, C.Y., Feichtenhofer, C., Darrell, T., Xie, S.: A convnet for the 2020s. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11976-11986 (2022) +33. Mahmood, N., Ghorbani, N., Troje, N.F., Pons-Moll, G., Black, M.J.: Amass: Archive of motion capture as surface shapes. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5442-5451 (2019) +34. Pavlakos, G., Choutas, V., Ghorbani, N., Bolkart, T., Osman, A.A.A., Tzionas, D., Black, M.J.: Expressive body capture: 3d hands, face, and body from a single image. In: Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR) (2019) +35. Petrovich, M., Black, M.J., Varol, G.: Action-conditioned 3D human motion synthesis with transformer VAE. In: International Conference on Computer Vision (ICCV) (2021) +36. Petrovich, M., Black, M.J., Varol, G.: Temos: Generating diverse human motions from textual descriptions. In: European Conference on Computer Vision. pp. 480-497. Springer (2022) + +37. Petrovich, M., Black, M.J., Varol, G.: TEMOS: Generating diverse human motions from textual descriptions. In: European Conference on Computer Vision (ECCV) (2022) +38. Plappert, M., Mandery, C., Asfour, T.: The kit motion-language dataset. Big data 4(4), 236-252 (2016) +39. Plappert, M., Mandery, C., Asfour, T.: Learning a bidirectional mapping between human whole-body motion and natural language using deep recurrent neural networks. Robotics and Autonomous Systems 109, 13-26 (2018) +40. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021) +41. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022) +42. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: International Conference on Medical image computing and computer-assisted intervention. pp. 234-241. Springer (2015) +43. Rumelhart, D.E., Hinton, G.E., Williams, R.J.: Learning internal representations by error propagation. Parallel Distributed Processing pp. 318-362 (1986) +44. Sahara, C., Ho, J., Chan, W., Salimans, T., Fleet, D.J., Norouzi, M.: Image superresolution via iterative refinement. IEEE Transactions on Pattern Analysis and Machine Intelligence (2022) +45. Shi, X., Chen, Z., Wang, H., Yeung, D.Y., Wong, W.K., Woo, W.c.: Convolutional LSTM network: A machine learning approach for precipitation nowcasting. Advances in neural information processing systems 28 (2015) +46. Smith, J., De Mello, S., Kautz, J., Linderman, S., Byeon, W.: Convolutional state space models for long-range spatiotemporal modeling. Advances in Neural Information Processing Systems 36 (2024) +47. Sohl-Dickstein, J., Weiss, E., Maheswaranathan, N., Ganguli, S.: Deep unsupervised learning using nonequilibrium thermodynamics. In: International conference on machine learning. pp. 2256-2265. PMLR (2015) +48. Tevet, G., Gordon, B., Hertz, A., Bermano, A.H., Cohen-Or, D.: Motionclip: Exposing human motion generation to clip space. In: European Conference on Computer Vision. pp. 358-374. Springer (2022) +49. Tevet, G., Raab, S., Gordon, B., Shafir, Y., Cohen-or, D., Bermano, A.H.: Human motion diffusion model. In: The Eleventh International Conference on Learning Representations (2022) +50. Van Den Oord, A., Vinyals, O., et al.: Neural discrete representation learning. Advances in neural information processing systems 30 (2017) +51. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017) +52. Wang, J., Yan, J.N., Gu, A., Rush, A.M.: Pretraining without attention. arXiv preprint arXiv:2212.10544 (2022) +53. Yan, J.N., Gu, J., Rush, A.M.: Diffusion models without attention. arXiv preprint arXiv:2311.18257 (2023) +54. Zhang, M., Cai, Z., Pan, L., Hong, F., Guo, X., Yang, L., Liu, Z.: Motiondiffuse: Text-driven human motion generation with diffusion model. IEEE Transactions on Pattern Analysis and Machine Intelligence (2024) + +55. Zhang, Y., Black, M.J., Tang, S.: We are more than our joints: Predicting how 3d bodies move. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3372-3382 (2021) +56. Zhong, C., Hu, L., Zhang, Z., Xia, S.: Attt2m: Text-driven human motion generation with multi-perspective attention mechanism. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 509-519 (2023) +57. Zhu, L., Liao, B., Zhang, Q., Wang, X., Liu, W., Wang, X.: Vision mamba: Efficient visual representation learning with bidirectional state space model. arXiv preprint arXiv:2401.09417 (2024) \ No newline at end of file diff --git a/2024/Motion Mamba_ Efficient and Long Sequence Motion Generation/images.zip b/2024/Motion Mamba_ Efficient and Long Sequence Motion Generation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..5a6505773c683c9785b55c5ce90f9e19712b58b2 --- /dev/null +++ b/2024/Motion Mamba_ Efficient and Long Sequence Motion Generation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1197c102592c1467b060b73dd3f39de20deb581b7121d862525565437c2688b7 +size 447387 diff --git a/2024/Motion Mamba_ Efficient and Long Sequence Motion Generation/layout.json b/2024/Motion Mamba_ Efficient and Long Sequence Motion Generation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..896dc0666cb73d37c7b217418b8a0b90fdccd0ad --- /dev/null +++ b/2024/Motion Mamba_ Efficient and Long Sequence Motion Generation/layout.json @@ -0,0 +1,9533 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 149, + 111, + 465, + 146 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 111, + 465, + 146 + ], + "spans": [ + { + "bbox": [ + 149, + 111, + 465, + 146 + ], + "type": "text", + "content": "Motion Mamba: Efficient and Long Sequence Motion Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 168, + 468, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 168, + 468, + 194 + ], + "spans": [ + { + "bbox": [ + 146, + 168, + 468, + 194 + ], + "type": "text", + "content": "Zeyu Zhang" + }, + { + "bbox": [ + 146, + 168, + 468, + 194 + ], + "type": "inline_equation", + "content": "^{12*}" + }, + { + "bbox": [ + 146, + 168, + 468, + 194 + ], + "type": "text", + "content": ", Akide Liu" + }, + { + "bbox": [ + 146, + 168, + 468, + 194 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 146, + 168, + 468, + 194 + ], + "type": "text", + "content": ", Ian Reid" + }, + { + "bbox": [ + 146, + 168, + 468, + 194 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 146, + 168, + 468, + 194 + ], + "type": "text", + "content": ", Richard Hartley" + }, + { + "bbox": [ + 146, + 168, + 468, + 194 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 146, + 168, + 468, + 194 + ], + "type": "text", + "content": ", Bohan Zhuang" + }, + { + "bbox": [ + 146, + 168, + 468, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 146, + 168, + 468, + 194 + ], + "type": "text", + "content": ", and Hao Tang" + }, + { + "bbox": [ + 146, + 168, + 468, + 194 + ], + "type": "inline_equation", + "content": "^{4\\boxtimes}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 264, + 201, + 351, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 264, + 201, + 351, + 213 + ], + "spans": [ + { + "bbox": [ + 264, + 201, + 351, + 213 + ], + "type": "text", + "content": "1 Monash University" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 230, + 213, + 384, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 230, + 213, + 384, + 224 + ], + "spans": [ + { + "bbox": [ + 230, + 213, + 384, + 224 + ], + "type": "text", + "content": "2 The Australian National University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 189, + 224, + 424, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 189, + 224, + 424, + 236 + ], + "spans": [ + { + "bbox": [ + 189, + 224, + 424, + 236 + ], + "type": "text", + "content": "3 Mohamed bin Zayed University of Artificial Intelligence" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 169, + 235, + 444, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 169, + 235, + 444, + 258 + ], + "spans": [ + { + "bbox": [ + 169, + 235, + 444, + 258 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 169, + 235, + 444, + 258 + ], + "type": "text", + "content": " National Key Laboratory for Multimedia Information Processing, School of Computer Science, Peking University" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 205, + 266, + 408, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 266, + 408, + 276 + ], + "spans": [ + { + "bbox": [ + 205, + 266, + 408, + 276 + ], + "type": "text", + "content": "https://steve-zeyu-zhang.github.io/MotionMamba" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 136, + 287, + 257, + 407 + ], + "blocks": [ + { + "bbox": [ + 136, + 287, + 257, + 407 + ], + "lines": [ + { + "bbox": [ + 136, + 287, + 257, + 407 + ], + "spans": [ + { + "bbox": [ + 136, + 287, + 257, + 407 + ], + "type": "image", + "image_path": "9d50515d2181f0c7d935e2bbb103cb5293cf413d46914332a1dd381edc6ed7a8.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 415, + 482, + 451 + ], + "lines": [ + { + "bbox": [ + 130, + 415, + 482, + 451 + ], + "spans": [ + { + "bbox": [ + 130, + 415, + 482, + 451 + ], + "type": "text", + "content": "Fig.1: Motion Mamba has achieved significantly superior performance on long sequence modeling and motion generation efficiency compared with other well-designed state-of-the-art methods such as MLD [6], MotionDiffuse [54], and MDM [49]." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 258, + 286, + 376, + 407 + ], + "blocks": [ + { + "bbox": [ + 258, + 286, + 376, + 407 + ], + "lines": [ + { + "bbox": [ + 258, + 286, + 376, + 407 + ], + "spans": [ + { + "bbox": [ + 258, + 286, + 376, + 407 + ], + "type": "image", + "image_path": "48cf11f3798a6aedad48113acb76dca9df21fbaad9e72b652f24ec0f0b7e17e3.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 384, + 286, + 482, + 407 + ], + "blocks": [ + { + "bbox": [ + 384, + 286, + 482, + 407 + ], + "lines": [ + { + "bbox": [ + 384, + 286, + 482, + 407 + ], + "spans": [ + { + "bbox": [ + 384, + 286, + 482, + 407 + ], + "type": "image", + "image_path": "210741e7ade2cfcbcf250c0c7992f4b4a689a716569803d67cdd28922bb7d8d0.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 159, + 462, + 455, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 462, + 455, + 628 + ], + "spans": [ + { + "bbox": [ + 159, + 462, + 455, + 628 + ], + "type": "text", + "content": "Abstract. Human motion generation stands as a significant pursuit in generative computer vision, while achieving long-sequence and efficient motion generation remains challenging. Recent advancements in state space models (SSMs), notably Mamba, have showcased considerable promise in long sequence modeling with an efficient hardware-aware design, which appears to be a significant direction upon building motion generation model. Nevertheless, adapting SSMs to motion generation faces hurdles since the lack of a specialized design architecture to model motion sequence. To address these challenges, we propose Motion Mamba, a simple yet efficient approach that presents the pioneering motion generation model utilized SSMs. Specifically, we design a Hierarchical Temporal Mamba (HTM) block to process temporal data by ensemble varying numbers of isolated SSM modules across a symmetric U-Net architecture aimed at preserving motion consistency between frames. We also design a Bidirectional Spatial Mamba (BSM) block" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 142, + 637, + 217, + 647 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 637, + 217, + 647 + ], + "spans": [ + { + "bbox": [ + 142, + 637, + 217, + 647 + ], + "type": "text", + "content": "*Equal contribution." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 143, + 647, + 378, + 655 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 647, + 378, + 655 + ], + "spans": [ + { + "bbox": [ + 143, + 647, + 378, + 655 + ], + "type": "text", + "content": "†Work done while being a research assistant at Monash University." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 143, + 655, + 317, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 655, + 317, + 665 + ], + "spans": [ + { + "bbox": [ + 143, + 655, + 317, + 665 + ], + "type": "text", + "content": "Corresponding author: bjdxtanghao@gmail.com" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 159, + 117, + 455, + 183 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 117, + 455, + 183 + ], + "spans": [ + { + "bbox": [ + 159, + 117, + 455, + 183 + ], + "type": "text", + "content": "to bidirectionally process latent poses, to enhance accurate motion generation within a temporal frame. Our proposed method achieves up to " + }, + { + "bbox": [ + 159, + 117, + 455, + 183 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 159, + 117, + 455, + 183 + ], + "type": "text", + "content": " FID improvement and up to 4 times faster on the HumanML3D and KIT-ML datasets compared to the previous best diffusion-based method, which demonstrates strong capabilities of high-quality long sequence motion modeling and real-time human motion generation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 194, + 455, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 194, + 455, + 216 + ], + "spans": [ + { + "bbox": [ + 160, + 194, + 455, + 216 + ], + "type": "text", + "content": "Keywords: Human Motion Generation " + }, + { + "bbox": [ + 160, + 194, + 455, + 216 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 194, + 455, + 216 + ], + "type": "text", + "content": " Selective State Space Models " + }, + { + "bbox": [ + 160, + 194, + 455, + 216 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 194, + 455, + 216 + ], + "type": "text", + "content": " Latent Diffusion Models" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 240, + 230, + 253 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 240, + 230, + 253 + ], + "spans": [ + { + "bbox": [ + 132, + 240, + 230, + 253 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 269, + 482, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 269, + 482, + 449 + ], + "spans": [ + { + "bbox": [ + 130, + 269, + 482, + 449 + ], + "type": "text", + "content": "Human motion generation involves creating high-quality 3D human motion based on given conditions, such as a text prompt, which generates Euler angles for each joint in a human skeleton. To emulate human motion effectively, virtual characters must respond to the conditional context, exhibit natural movement, and perform motion accurately. Recent motion generation models are categorized into four main approaches: autoencoder-based [1,10,17,36,48,56], utilizing transformers for latent space compression and motion synthesis; GAN-based [4,20,30], using discriminators to enhance the realism of generated motions; autoregressive models [24], treating motion sequences as languages with specialized codebooks; and diffusion-based [6,49,54], employing denoising steps for motion generation. Challenges vary across methods, with autoencoder models struggling to generate accurate motions from detailed descriptions due to textual information compression, GAN-based models facing training difficulties, especially in conditional tasks, and diffusion-based models relying on complex transformer-based architectures results in inefficient of motion prediction." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 449, + 482, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 482, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 482, + 640 + ], + "type": "text", + "content": "Although diffusion-based models excel at generating motion with robust performance and often exhibit superior diversity, they encounter two limitations. 1) Convolutional or transformer-based diffusion methods exhibit limitations in generating long-range motion sequences. Previous transformer-based methodologies [6,49,54] have focused on modeling long-range dependencies and acquiring comprehensive global context information. Despite these advances, they are frequently associated with a substantial increase in computational requirements. Furthermore, transformer architectures are not intrinsically designed for temporal sequential modeling, which poses an inherent limitation. 2) The efficiency of inference in transformer-based diffusion methods is constrained. Although prior research has attempted to leverage the Variational Autoencoder for denoising operations in the latent space [6], the inference speed remains adversely affected by the attention mechanism's quadratic scaling, leading to inefficient motion generation. Consequently, exploring a new architectural paradigm that accommodates long-range dependencies and maintains a linear computational complexity is crucial for sustaining motion generation tasks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 641, + 481, + 666 + ], + "type": "text", + "content": "Recent advances have sparked renewed interest in state space models (SSMs) [14, 15], a field that originated from the foundational classic state space model" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 255, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 255, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 255, + 102 + ], + "type": "text", + "content": "Zeyu and Akide et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 482, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 482, + 259 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 482, + 259 + ], + "type": "text", + "content": "[25]. Modern versions of SSMs stand out due to their ability to effectively capture long-range dependencies, a capability greatly improved by the introduction of parallel training techniques. This evolution has led to various methodologies based on SSM, notably the linear state space layers (LSSL) [15], the structured state-space sequence model (S4) [14], the diagonal state space (DSS) [19], and S4D [13]. These methods have been carefully designed to handle sequential data across various tasks and modalities, paying special attention to modeling long-range dependencies. Their efficacy in managing long sequences is attributed to the implementation of convolutional computations [15] and near-linear computational strategies, such as mamba [12], marking a significant stride in sequentially oriented tasks, including large language model decoding and motion sequence generation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 262, + 482, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 262, + 482, + 502 + ], + "spans": [ + { + "bbox": [ + 130, + 262, + 482, + 502 + ], + "type": "text", + "content": "Adapting selective state space modules for motion generation tasks presents notable challenges, primarily due to the lack of specialized design in SSMs for capturing the sensitive motion details required for temporal representation and the complexities involved in aggregating latent space. In response to these challenges, we have meticulously developed a motion generation architecture, specifically tailored to address the intricacies of long-term sequence generation, while optimizing for computational efficiency with near-linear-time complexity. This innovation is embodied in the Motion Mamba model, a simple yet potent approach to motion generation. The Motion Mamba framework pioneers a diffusion-based generative system, incorporating two key components oriented toward SSM as shown in Figure. 2: (1) a Hierarchical Temporal Mamba (HTM) block: This component is ingeniously crafted to arrange motion frames in sequential order, using hierarchically adjusted scanning. It is adept at identifying temporal dependencies at various depths, thereby facilitating a thorough comprehension of the dynamics inherent in motion sequences. (2) a Bidirectional Spatial Mamba (BSM) block: This block is designed to unravel the structured latent skeleton by evaluating data from both forward and reverse directions. Its primary goal is to safeguard the continuity of information flow, significantly bolstering the model's capacity for precise motion generation through the retention of dense informational exchange." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 506, + 482, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 506, + 482, + 650 + ], + "spans": [ + { + "bbox": [ + 130, + 506, + 482, + 650 + ], + "type": "text", + "content": "The Motion Mamba introduces a new approach to motion generation that strikes an exceptional trade-off between accuracy and efficiency, shown in Fig. 1. Our experimental results underscore the significant improvements brought about by Motion Mamba, showcasing a remarkable improvement in the Fréchet Inception Distance (FID), with a reduction of up to " + }, + { + "bbox": [ + 130, + 506, + 482, + 650 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 130, + 506, + 482, + 650 + ], + "type": "text", + "content": " from the prior state-of-the-art metric of 0.473 to an impressive 0.281 on the HumanML3D dataset [17]. Furthermore, we emphasize our framework's unparalleled inference speed, which is four times faster than previous methods, achieving an average inference time of only 0.058 seconds per sequence compared to the 0.217 seconds required by the MLD [6] method per sequence. These outcomes unequivocally establish Motion Mamba's state-of-the-art performance, concurrently ensuring fast inference speeds for conditional human motion generation tasks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 146, + 653, + 474, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 653, + 474, + 665 + ], + "spans": [ + { + "bbox": [ + 146, + 653, + 474, + 665 + ], + "type": "text", + "content": "Our contributions to the field of motion generation can be summarized as:" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 383, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 383, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 383, + 91, + 447, + 100 + ], + "type": "text", + "content": "Motion Mamba" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 283 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 151 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 481, + 151 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 481, + 151 + ], + "type": "text", + "content": "1. We introduce a simple yet effective framework, named Motion Mamba, which is a pioneering method integrates a selective scanning mechanism into motion generation tasks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 152, + 481, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 152, + 481, + 224 + ], + "spans": [ + { + "bbox": [ + 132, + 152, + 481, + 224 + ], + "type": "text", + "content": "2. Motion Mamba is comprised of two modules: Hierarchical Temporal Mamba (HTM) and Bidirectional Spatial Mamba (BSM), which are designed for temporal and spatial modeling, respectively. HTM blocks are tasked with processing temporal motion data, aiming to enhance motion consistency across frames. BSM blocks are engineered to bidirectionally capture the channel-wise flow of hidden information within the latent pose representations." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 224, + 481, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 224, + 481, + 283 + ], + "spans": [ + { + "bbox": [ + 132, + 224, + 481, + 283 + ], + "type": "text", + "content": "3. Motion Mamba framework demonstrated exceptional performance on text-to-motion generation task, through experimental validation on the HumanML3D [17] and KIT-ML [38] datasets. Our methodology achieved state-of-the-art generation quality and significantly improved long-squence modeling, meanwhile optimizing inference speed." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 304, + 242, + 316 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 242, + 316 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 242, + 316 + ], + "type": "text", + "content": "2 Related Works" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 330, + 481, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 330, + 481, + 389 + ], + "spans": [ + { + "bbox": [ + 130, + 330, + 481, + 389 + ], + "type": "text", + "content": "Human Motion Generation. Generating human motion is a significant application of computer vision, essential for various applications like 3D modelling and robot manipulation. Recently, the predominant method of achieving human motion generation, known as the Text-to-Motion task, involves learning a common latent space for both language and motion." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 390, + 481, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 390, + 481, + 510 + ], + "spans": [ + { + "bbox": [ + 130, + 390, + 481, + 510 + ], + "type": "text", + "content": "DVGAN [30] creates the GAN [11] discriminator by densely validating at each time-scale and perturbing the discriminator input for translation invariance, enabling motion generation and completion. ERD-QV [20] enhances latent representations through two additive modifiers: a time-to-arrival embedding applied universally and an additive scheduled target noise vector used during extended transitions. It further improves transition quality by incorporating a GAN framework with two discriminators operating at different timescales. HP-GAN [4], trained with a modified version of the improved WGAN-GP [16], utilizes a custom loss function designed for human motion prediction. It learns a probability density function of future human poses conditioned on previous poses." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 510, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 481, + 665 + ], + "type": "text", + "content": "Autoencoders [27, 43] are notable generative models known for their ability to represent data robustly by compressing high-dimensional data into a latent space, which is widely adopted in human motion generation. JL2P [1] uses RNN-based autoencoders [23] to learn a combined representation of language and pose. It restricts the direct mapping from text to motion to a one-to-one relationship. MotionCLIP [48] uses Transformer-based Autoencoders [51] to reconstruct motion while ensuring alignment with the corresponding text label in the CLIP [40] space. This alignment effectively integrates the semantic knowledge from CLIP into the human motion manifold. TEMOS [36] and T2M [17] combine a Transformer-based VAE [26] with a text encoder to generate distribution parameters that work within the VAE latent space. AttT2M [56] and TM2D [10] incorporate a body-part spatio-temporal encoder into VQ-VAE [50] for enhanced learning of a discrete latent space with increased expressiveness." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 255, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 255, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 255, + 102 + ], + "type": "text", + "content": "Zeyu and Akide et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 134, + 115, + 481, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 115, + 481, + 235 + ], + "spans": [ + { + "bbox": [ + 134, + 115, + 481, + 235 + ], + "type": "text", + "content": "Diffusion models [7, 22, 41, 47] have recently surpassed GANs and VAEs in generating 2D images. Developing a motion generation model based on diffusion models is obviously an attractive direction. MotionDiffuse [54] introduces the inaugural framework for text-driven motion generation based on diffusion models. It showcases several desirable properties, including probabilistic mapping, realistic synthesis, and multi-level manipulation. MDM [49] utilizes a classifier-free Transformer-based diffusion model for the human motion domain to predict sample rather than noise in each diffusion step. MLD [6] performs a diffusion process in latent motion space, rather than using a diffusion model to establish connections between raw motion sequences and conditional inputs." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 236, + 481, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 236, + 481, + 570 + ], + "spans": [ + { + "bbox": [ + 133, + 236, + 481, + 570 + ], + "type": "text", + "content": "State Space Models. Recently, state space sequence models (SSMs) [14, 15], drawing inspiration from classical state-space models [25], have emerged as a promising architecture for sequence modeling. Mamba [12] introduces a selective SSM architecture, integrating time-varying parameters into the SSM framework, and proposes a hardware-aware algorithm to facilitate highly efficient training and inference processes. Some research works leverage SSM in computer vision to process 2D data. The 2D SSM [3] introduces an SSM block at the beginning of each transformer block [8, 51]. This approach aims to achieve efficient parameterization, accelerated computation, and a suitable normalization scheme. SGConvNeXt [29] presents a structured global convolution method inspired by ConvNeXt [32], incorporating multi-scale sub-kernels to achieve both parameterization efficiency and effective long sequence modeling. ConvSSM [46] integrates the tensor modeling principles of ConvLSTM [45] with SSMs, elucidating the utilization of parallel scans in convolutional recurrences. This approach enables subquadratic parallelization and rapid autoregressive generation. Vim [57] introduces a bidirectional SSM block [52] for efficient and versatile visual representation learning, achieving performance comparable to established ViT [8] methods. VMamba [31] introduces a Cross-Scan Module (CSM) designed to traverse the spatial domain and transform any non-causal visual image into ordered patch sequences. This approach achieves linear complexity while preserving global receptive fields. There have also been attempts to utilize SSMs to handle higher-dimensional data. Mamba-ND [28] explores various combinations of SSM and different scan directions within the SSM block to adapt Mamba [12] to higher-dimensional tasks. Recent efforts have sought to replace the traditional transformer-based U-Net within the diffusion denoiser with the SSM block, with the aim of enhancing image generation efficiency. DiffuSSM [53] adeptly manages higher resolutions without relying on global compression, thus maintaining detailed image representation throughout the diffusion process." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 590, + 288, + 604 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 590, + 288, + 604 + ], + "spans": [ + { + "bbox": [ + 133, + 590, + 288, + 604 + ], + "type": "text", + "content": "3 The Proposed Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 617, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 617, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 617, + 481, + 665 + ], + "type": "text", + "content": "In this section, we delineate the architecture and operational principles of the Motion Mamba framework, designed for generating human motion over long ranges efficiently from textual descriptions. Initially, we discuss the foundational concepts underpinning our approach, including the Mamba Model [12] and the" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 383, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 383, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 383, + 91, + 447, + 100 + ], + "type": "text", + "content": "Motion Mamba" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": "latent diffusion model [6]. Following this, we detail our uniquely crafted architecture that leverages the Mamba model to enhance motion generation efficiency. This architecture comprises two principal components: the Hierarchical Temporal Mamba (HTM) block, which addresses temporal aspects, and the Bidirectional Spatial Mamba (BSM) block, focusing on spatial dynamics." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 194, + 228, + 205 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 228, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 228, + 205 + ], + "type": "text", + "content": "3.1 Preliminaries" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 214, + 482, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 214, + 482, + 297 + ], + "spans": [ + { + "bbox": [ + 130, + 214, + 482, + 297 + ], + "type": "text", + "content": "Selective Structured State Space Sequence Model. SSMs particularly through the contributions of structured state space sequence models (S4) and Mamba, have demonstrated exceptional proficiency in handling long sequences. These models operationalize the mapping of a 1-D function or sequence, " + }, + { + "bbox": [ + 130, + 214, + 482, + 297 + ], + "type": "inline_equation", + "content": "x(t) \\in \\mathbb{R} \\mapsto y(t) \\in \\mathbb{R}" + }, + { + "bbox": [ + 130, + 214, + 482, + 297 + ], + "type": "text", + "content": ", through a hidden state " + }, + { + "bbox": [ + 130, + 214, + 482, + 297 + ], + "type": "inline_equation", + "content": "h(t) \\in \\mathbb{R}^N" + }, + { + "bbox": [ + 130, + 214, + 482, + 297 + ], + "type": "text", + "content": ", employing " + }, + { + "bbox": [ + 130, + 214, + 482, + 297 + ], + "type": "inline_equation", + "content": "\\mathbf{A} \\in \\mathbb{R}^{N \\times N}" + }, + { + "bbox": [ + 130, + 214, + 482, + 297 + ], + "type": "text", + "content": " as the evolution parameters, " + }, + { + "bbox": [ + 130, + 214, + 482, + 297 + ], + "type": "inline_equation", + "content": "\\mathbf{B} \\in \\mathbb{R}^{N \\times 1}" + }, + { + "bbox": [ + 130, + 214, + 482, + 297 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 214, + 482, + 297 + ], + "type": "inline_equation", + "content": "\\mathbf{C} \\in \\mathbb{R}^{1 \\times N}" + }, + { + "bbox": [ + 130, + 214, + 482, + 297 + ], + "type": "text", + "content": " as the projection parameters, respectively." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 298, + 482, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 298, + 482, + 320 + ], + "spans": [ + { + "bbox": [ + 130, + 298, + 482, + 320 + ], + "type": "text", + "content": "The discretized system can then be expressed as follows, incorporating a step size " + }, + { + "bbox": [ + 130, + 298, + 482, + 320 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 130, + 298, + 482, + 320 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 265, + 319, + 480, + 338 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 265, + 319, + 480, + 338 + ], + "spans": [ + { + "bbox": [ + 265, + 319, + 480, + 338 + ], + "type": "interline_equation", + "content": "\\begin{array}{c} h _ {t} = \\overline {{\\mathbf {A}}} h _ {t - 1} + \\overline {{\\mathbf {B}}} x _ {t}, \\\\ \\mathbf {C} ^ {\\prime} \\end{array} \\tag {1}", + "image_path": "49ba42053833407b648d31486360c6acef990b3e7e1c1648ee355ea4da260adb.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 266, + 335, + 309, + 346 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 266, + 335, + 309, + 346 + ], + "spans": [ + { + "bbox": [ + 266, + 335, + 309, + 346 + ], + "type": "interline_equation", + "content": "y _ {t} = \\mathbf {C} h _ {t}.", + "image_path": "679a6f973aef3efcba6932bcc334020fc1abceabd33af63e9db30db4b5902474.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 352, + 482, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 352, + 482, + 389 + ], + "spans": [ + { + "bbox": [ + 130, + 352, + 482, + 389 + ], + "type": "text", + "content": "This adaptation facilitates the computation of output through global convolution, leveraging a structured convolutional kernel " + }, + { + "bbox": [ + 130, + 352, + 482, + 389 + ], + "type": "inline_equation", + "content": "\\overline{\\mathbf{K}}" + }, + { + "bbox": [ + 130, + 352, + 482, + 389 + ], + "type": "text", + "content": ", which encompasses the entire length " + }, + { + "bbox": [ + 130, + 352, + 482, + 389 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 352, + 482, + 389 + ], + "type": "text", + "content": " of the input sequence " + }, + { + "bbox": [ + 130, + 352, + 482, + 389 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 130, + 352, + 482, + 389 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 234, + 397, + 480, + 418 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 234, + 397, + 480, + 418 + ], + "spans": [ + { + "bbox": [ + 234, + 397, + 480, + 418 + ], + "type": "interline_equation", + "content": "\\overline {{\\mathbf {K}}} = \\left(\\underset {-} {\\mathbf {C}} \\overline {{\\mathbf {B}}}, \\mathbf {C} \\overline {{\\mathbf {A}}} \\mathbf {B}, \\dots , \\mathbf {C} \\overline {{\\mathbf {A}}} ^ {M - 1} \\overline {{\\mathbf {B}}}\\right), \\tag {2}", + "image_path": "a65e82bc1a02d667a8c41fc7ebe3cdde18e9618144b6297505545f1b754ed012.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 237, + 416, + 284, + 429 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 237, + 416, + 284, + 429 + ], + "spans": [ + { + "bbox": [ + 237, + 416, + 284, + 429 + ], + "type": "interline_equation", + "content": "\\mathbf {y} = \\mathbf {x} * \\overline {{\\mathbf {K}}}.", + "image_path": "7e2f6308a7b4ca78c2da997faeed26f9156b934054c464aaf970b41fd817039c.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 438, + 482, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 438, + 482, + 498 + ], + "spans": [ + { + "bbox": [ + 130, + 438, + 482, + 498 + ], + "type": "text", + "content": "Selective models like Mamba introduce time-varying parameters, deviating from the linear time invariance (LTI) assumption and complicating parallel computation. However, hardware-aware optimizations, such as associative scans, have been developed to address these computational challenges, highlighting the ongoing evolution and application of SSMs in modeling complex temporal dynamics." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 498, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 668 + ], + "type": "text", + "content": "Latent Motion Diffusion Model. Diffusion probabilistic models offer a significant advancement in motion generation by gradually reducing noise from a Gaussian distribution to a target data distribution " + }, + { + "bbox": [ + 130, + 498, + 482, + 668 + ], + "type": "inline_equation", + "content": "p(x)" + }, + { + "bbox": [ + 130, + 498, + 482, + 668 + ], + "type": "text", + "content": " through a T-length learned Markov process [7,22,41,44,47,49,54], giving " + }, + { + "bbox": [ + 130, + 498, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\{\\pmb{x}_t\\}_{t=1}^T" + }, + { + "bbox": [ + 130, + 498, + 482, + 668 + ], + "type": "text", + "content": ". In the motion generation, we define our trainable diffusion models with a denoiser " + }, + { + "bbox": [ + 130, + 498, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(x_t,t)" + }, + { + "bbox": [ + 130, + 498, + 482, + 668 + ], + "type": "text", + "content": " which anneal the random noise to motion sequence " + }, + { + "bbox": [ + 130, + 498, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\{\\hat{x}_t^{1:N}\\}_{t=1}^T" + }, + { + "bbox": [ + 130, + 498, + 482, + 668 + ], + "type": "text", + "content": " iteratively. To address the inefficiencies of applying diffusion models directly to raw motion sequences, we employ a low-dimensional motion latent space for the diffusion process. Given an input condition " + }, + { + "bbox": [ + 130, + 498, + 482, + 668 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 498, + 482, + 668 + ], + "type": "text", + "content": ", such as a descriptive sentence " + }, + { + "bbox": [ + 130, + 498, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\pmb{w}^{1:N} = \\{w^i\\}_{i=1}^N" + }, + { + "bbox": [ + 130, + 498, + 482, + 668 + ], + "type": "text", + "content": ", an action label " + }, + { + "bbox": [ + 130, + 498, + 482, + 668 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 130, + 498, + 482, + 668 + ], + "type": "text", + "content": " from a predefined set " + }, + { + "bbox": [ + 130, + 498, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 130, + 498, + 482, + 668 + ], + "type": "text", + "content": ", or an empty condition " + }, + { + "bbox": [ + 130, + 498, + 482, + 668 + ], + "type": "inline_equation", + "content": "c = \\emptyset" + }, + { + "bbox": [ + 130, + 498, + 482, + 668 + ], + "type": "text", + "content": ", and the motion representation that combines 3D joint rotations, positions, velocities, and foot contact as proposed in [17]. The frozen CLIP [40] text encoder " + }, + { + "bbox": [ + 130, + 498, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\tau_{\\theta}^w" + }, + { + "bbox": [ + 130, + 498, + 482, + 668 + ], + "type": "text", + "content": " has been employed to obtain projected text embedding " + }, + { + "bbox": [ + 130, + 498, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\tau_{\\theta}^w(w^{1:N}) \\in \\mathbb{R}^{1 \\times d}" + }, + { + "bbox": [ + 130, + 498, + 482, + 668 + ], + "type": "text", + "content": ", thereby conditional denoiser comprised in term of " + }, + { + "bbox": [ + 130, + 498, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(z_t,t,\\tau_{\\theta}(c))" + }, + { + "bbox": [ + 130, + 498, + 482, + 668 + ], + "type": "text", + "content": ". The latent diffusion model " + }, + { + "bbox": [ + 130, + 498, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(x_t,t)" + }, + { + "bbox": [ + 130, + 498, + 482, + 668 + ], + "type": "text", + "content": " aimed" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 256, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 256, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 256, + 102 + ], + "type": "text", + "content": "Zeyu and Akide et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 130, + 114, + 482, + 285 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 285 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 285 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 285 + ], + "type": "image", + "image_path": "c47e35af41944d8524c810593fb0260c0e75549c04159ef1f62bdcc5c0d563b4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 293, + 482, + 358 + ], + "lines": [ + { + "bbox": [ + 130, + 293, + 482, + 358 + ], + "spans": [ + { + "bbox": [ + 130, + 293, + 482, + 358 + ], + "type": "text", + "content": "Fig. 2: This figure illustrates the architecture of the proposed Motion Mamba model. Each of encoder and decoder blocks consists of a Hierarchical Temporal Mamba block (HTM) and a Bidirectional Spatial Mamba (BSM) block, which possess hierarchical scan and bidirectional scan within SSM layers respectively. This symmetric distribution of scans ensure a balanced and coherence framework across the encoder-decoder architecture." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 371, + 482, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 371, + 482, + 469 + ], + "spans": [ + { + "bbox": [ + 130, + 371, + 482, + 469 + ], + "type": "text", + "content": "to generate the human motion sequence in terms of " + }, + { + "bbox": [ + 130, + 371, + 482, + 469 + ], + "type": "inline_equation", + "content": "\\hat{x}^{1:L} = \\{\\hat{x}^i\\}_{i=1}^L" + }, + { + "bbox": [ + 130, + 371, + 482, + 469 + ], + "type": "text", + "content": ", where L denotes the sequence length or number of frames [34,35,37,55]. After that we reused the motion Variational AutoEncoder (VAE) " + }, + { + "bbox": [ + 130, + 371, + 482, + 469 + ], + "type": "inline_equation", + "content": "\\mathcal{V} = \\{\\mathcal{E},\\mathcal{D}\\}" + }, + { + "bbox": [ + 130, + 371, + 482, + 469 + ], + "type": "text", + "content": " proposed in MLD [6] to manipulate the motion sequence in latent space " + }, + { + "bbox": [ + 130, + 371, + 482, + 469 + ], + "type": "inline_equation", + "content": "z = \\mathcal{E}(x^{1:L})" + }, + { + "bbox": [ + 130, + 371, + 482, + 469 + ], + "type": "text", + "content": ", and decompress the intermediate representation to motion sequence by " + }, + { + "bbox": [ + 130, + 371, + 482, + 469 + ], + "type": "inline_equation", + "content": "\\hat{x}^{1:L} = \\mathcal{D}(z) = \\mathcal{D}\\mathcal{E}(x^{1:L})" + }, + { + "bbox": [ + 130, + 371, + 482, + 469 + ], + "type": "text", + "content": " [26, 42, 51]. Finally, our latent diffusion model is trained with an objective focusing on minimization of MSE between true and predicted noise in the latent space, facilitating efficient and high-quality motion generation [2, 22]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 484, + 240, + 495 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 484, + 240, + 495 + ], + "spans": [ + { + "bbox": [ + 132, + 484, + 240, + 495 + ], + "type": "text", + "content": "3.2 Motion Mamba" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 504, + 482, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 504, + 482, + 601 + ], + "spans": [ + { + "bbox": [ + 130, + 504, + 482, + 601 + ], + "type": "text", + "content": "The architecture of the proposed Motion Mamba framework is illustrated in Figure. 2. At its core, Motion Mamba utilizes a denoising U-Net architecture, which is distinguished for its effectiveness in modeling the continuous, temporal sequences of motion frames. This effectiveness is attributed to the inherent long-sequence modeling capacity of the Mamba model. The denoiser, denoted by " + }, + { + "bbox": [ + 130, + 504, + 482, + 601 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 130, + 504, + 482, + 601 + ], + "type": "text", + "content": ", comprises " + }, + { + "bbox": [ + 130, + 504, + 482, + 601 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 504, + 482, + 601 + ], + "type": "text", + "content": " blocks including encoder " + }, + { + "bbox": [ + 130, + 504, + 482, + 601 + ], + "type": "inline_equation", + "content": "E_{1..N}" + }, + { + "bbox": [ + 130, + 504, + 482, + 601 + ], + "type": "text", + "content": " and decoder " + }, + { + "bbox": [ + 130, + 504, + 482, + 601 + ], + "type": "inline_equation", + "content": "D_{1..N}" + }, + { + "bbox": [ + 130, + 504, + 482, + 601 + ], + "type": "text", + "content": ". Additionally, the architecture is enhanced with a transformer-based attention mixer block " + }, + { + "bbox": [ + 130, + 504, + 482, + 601 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 504, + 482, + 601 + ], + "type": "text", + "content": ", designed to augment the model's ability to capture complex temporal dynamics." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 246, + 611, + 481, + 625 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 246, + 611, + 481, + 625 + ], + "spans": [ + { + "bbox": [ + 246, + 611, + 481, + 625 + ], + "type": "interline_equation", + "content": "\\epsilon_ {\\theta} (x) \\equiv \\{E _ {1 \\dots N}, M, D _ {1 \\dots N} \\}. \\tag {3}", + "image_path": "eb31375dac6581f139e5988d06c2c215982d7170a22a675d544b01a678c66e3b.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": "The encoder blocks are represented as " + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "inline_equation", + "content": "E_{1..N}" + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": ", arranged sequentially, and the decoder blocks as " + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "inline_equation", + "content": "D_{1..N}" + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": ", configured in reverse order to facilitate effective bottom-up and top-down information flow. Given that selective operations have signif" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 383, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 383, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 383, + 91, + 447, + 100 + ], + "type": "text", + "content": "Motion Mamba" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "code", + "bbox": [ + 132, + 132, + 369, + 310 + ], + "blocks": [ + { + "bbox": [ + 133, + 116, + 398, + 129 + ], + "lines": [ + { + "bbox": [ + 133, + 116, + 398, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 398, + 129 + ], + "type": "text", + "content": "Algorithm 1 Hierarchical Temporal Mamba (HTM) Block." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 132, + 132, + 369, + 310 + ], + "lines": [ + { + "bbox": [ + 132, + 132, + 369, + 310 + ], + "spans": [ + { + "bbox": [ + 132, + 132, + 369, + 310 + ], + "type": "text", + "content": "Require: compressed latent representations " + }, + { + "bbox": [ + 132, + 132, + 369, + 310 + ], + "type": "inline_equation", + "content": "z:(T,B,C)" + }, + { + "bbox": [ + 132, + 132, + 369, + 310 + ], + "type": "text", + "content": " \nEnsure: transformed representations " + }, + { + "bbox": [ + 132, + 132, + 369, + 310 + ], + "type": "inline_equation", + "content": "z_{\\mathrm{HTM}}:(T,B,E)" + }, + { + "bbox": [ + 132, + 132, + 369, + 310 + ], + "type": "text", + "content": " \n1:/\\* linear projection layer \\*/ \n2: " + }, + { + "bbox": [ + 132, + 132, + 369, + 310 + ], + "type": "inline_equation", + "content": "x,z:(T,B,E)\\gets \\mathrm{Linear}(z)" + }, + { + "bbox": [ + 132, + 132, + 369, + 310 + ], + "type": "text", + "content": " \n3:/\\* set of scans and memory matrices \\*/ \n4: " + }, + { + "bbox": [ + 132, + 132, + 369, + 310 + ], + "type": "inline_equation", + "content": "K = \\{S_{2N_n - 1},S_{2N_{n - 1} - 1},\\ldots ,S_1\\}" + }, + { + "bbox": [ + 132, + 132, + 369, + 310 + ], + "type": "text", + "content": " \n5: Memory matrices: " + }, + { + "bbox": [ + 132, + 132, + 369, + 310 + ], + "type": "inline_equation", + "content": "\\{A_1,\\dots ,A_k\\}" + }, + { + "bbox": [ + 132, + 132, + 369, + 310 + ], + "type": "text", + "content": " \n6: for each scan " + }, + { + "bbox": [ + 132, + 132, + 369, + 310 + ], + "type": "inline_equation", + "content": "S_{i}" + }, + { + "bbox": [ + 132, + 132, + 369, + 310 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 132, + 132, + 369, + 310 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 132, + 132, + 369, + 310 + ], + "type": "text", + "content": " do \n7: " + }, + { + "bbox": [ + 132, + 132, + 369, + 310 + ], + "type": "inline_equation", + "content": "x_o^{\\prime}:(T,B,E)\\gets \\mathrm{Conv1D}(x)" + }, + { + "bbox": [ + 132, + 132, + 369, + 310 + ], + "type": "text", + "content": " \n8: " + }, + { + "bbox": [ + 132, + 132, + 369, + 310 + ], + "type": "inline_equation", + "content": "B_{o},C_{o},\\varDelta_{o}\\gets \\mathrm{Linear}(x_{o}^{\\prime})" + }, + { + "bbox": [ + 132, + 132, + 369, + 310 + ], + "type": "text", + "content": " \n9: Transform " + }, + { + "bbox": [ + 132, + 132, + 369, + 310 + ], + "type": "inline_equation", + "content": "\\overline{A}_o" + }, + { + "bbox": [ + 132, + 132, + 369, + 310 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 132, + 369, + 310 + ], + "type": "inline_equation", + "content": "\\overline{B}_o" + }, + { + "bbox": [ + 132, + 132, + 369, + 310 + ], + "type": "text", + "content": " using " + }, + { + "bbox": [ + 132, + 132, + 369, + 310 + ], + "type": "inline_equation", + "content": "\\varDelta_{o}" + }, + { + "bbox": [ + 132, + 132, + 369, + 310 + ], + "type": "text", + "content": " \n10: " + }, + { + "bbox": [ + 132, + 132, + 369, + 310 + ], + "type": "inline_equation", + "content": "O_{i}\\gets \\mathrm{SSMA}_{i,x}(x_{o}^{\\prime})" + }, + { + "bbox": [ + 132, + 132, + 369, + 310 + ], + "type": "text", + "content": " \n11: end for \n12:/\\* aggregation of outputs \\*/ \n13: " + }, + { + "bbox": [ + 132, + 132, + 369, + 310 + ], + "type": "inline_equation", + "content": "z_{\\mathrm{HTM}}:(T,B,E)\\gets \\mathrm{Linear(Aggregate(\\{O_1,\\ldots,O_k\\}))}" + }, + { + "bbox": [ + 132, + 132, + 369, + 310 + ], + "type": "text", + "content": " \n14: Return: " + }, + { + "bbox": [ + 132, + 132, + 369, + 310 + ], + "type": "inline_equation", + "content": "z_{\\mathrm{HTM}}" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "code_body" + } + ], + "index": 3, + "sub_type": "algorithm" + }, + { + "bbox": [ + 130, + 334, + 482, + 394 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 334, + 482, + 394 + ], + "spans": [ + { + "bbox": [ + 130, + 334, + 482, + 394 + ], + "type": "text", + "content": "icantly lower computational complexity compared to attention-based methods, we have increased the number of scans to achieve higher quality generations. Concurrently, it is imperative to maintain a balance between the model's parameters and its efficiency. Thereby, a novel aspect of our model is the introduction of a hierarchical scan strategy, characterized by a sequence of scan numbers as," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 232, + 402, + 481, + 416 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 232, + 402, + 481, + 416 + ], + "spans": [ + { + "bbox": [ + 232, + 402, + 481, + 416 + ], + "type": "interline_equation", + "content": "K = \\left\\{S _ {2 N - 1}, S _ {2 (N - 1) - 1}, \\dots , S _ {1} \\right\\}. \\tag {4}", + "image_path": "1ea531ac217f083de4bb88129c9c96f8af1a1ad378a0dc5852d0b6cd37c22441.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 423, + 482, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 423, + 482, + 496 + ], + "spans": [ + { + "bbox": [ + 130, + 423, + 482, + 496 + ], + "type": "text", + "content": "This sequence specifies the number of scans allocated to each layer, in descending order of complexity. For instance, the uppermost encoder layer, " + }, + { + "bbox": [ + 130, + 423, + 482, + 496 + ], + "type": "inline_equation", + "content": "E_{1}" + }, + { + "bbox": [ + 130, + 423, + 482, + 496 + ], + "type": "text", + "content": ", and the lowermost decoder layer, " + }, + { + "bbox": [ + 130, + 423, + 482, + 496 + ], + "type": "inline_equation", + "content": "D_{N}" + }, + { + "bbox": [ + 130, + 423, + 482, + 496 + ], + "type": "text", + "content": ", are allocated " + }, + { + "bbox": [ + 130, + 423, + 482, + 496 + ], + "type": "inline_equation", + "content": "S_{2N - 1}" + }, + { + "bbox": [ + 130, + 423, + 482, + 496 + ], + "type": "text", + "content": " scans, indicating the highest scanning complexity. Conversely, the lowest encoder layer, " + }, + { + "bbox": [ + 130, + 423, + 482, + 496 + ], + "type": "inline_equation", + "content": "E_{N}" + }, + { + "bbox": [ + 130, + 423, + 482, + 496 + ], + "type": "text", + "content": ", and the uppermost decoder layer, " + }, + { + "bbox": [ + 130, + 423, + 482, + 496 + ], + "type": "inline_equation", + "content": "D_{1}" + }, + { + "bbox": [ + 130, + 423, + 482, + 496 + ], + "type": "text", + "content": ", are assigned " + }, + { + "bbox": [ + 130, + 423, + 482, + 496 + ], + "type": "inline_equation", + "content": "S_{1}" + }, + { + "bbox": [ + 130, + 423, + 482, + 496 + ], + "type": "text", + "content": " scans, reflecting the lowest level of scanning complexity." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 211, + 503, + 481, + 548 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 211, + 503, + 481, + 548 + ], + "spans": [ + { + "bbox": [ + 211, + 503, + 481, + 548 + ], + "type": "interline_equation", + "content": "E _ {i} (S) = \\left\\{ \\begin{array}{l l} S _ {2 N - 1} & \\text {f o r} i = 1 \\\\ S _ {2 (N - i) - 1} & \\text {f o r} i = 2, \\dots , N - 1 \\\\ S _ {1} & \\text {f o r} i = N \\end{array} \\right. \\tag {5}", + "image_path": "8e9eec8be5aa54dffc70d5b397c802766706a4df148c78ffbf8a58017f64888d.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 209, + 557, + 481, + 602 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 209, + 557, + 481, + 602 + ], + "spans": [ + { + "bbox": [ + 209, + 557, + 481, + 602 + ], + "type": "interline_equation", + "content": "D _ {j} (S) = \\left\\{ \\begin{array}{l l} S _ {2 N - 1} & \\text {f o r} j = N \\\\ S _ {2 (N - j) - 1} & \\text {f o r} j = N - 1, \\dots , 2 \\\\ S _ {1} & \\text {f o r} j = 1 \\end{array} \\right. \\tag {6}", + "image_path": "dd4a158d8c567da07c34678c729562fb569220ee3f08631c3700df443d8a15ad.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": "This hierarchical scanning approach ensures that processing capabilities are evenly distributed throughout the encoder-decoder architecture., facilitating a detailed and nuanced analysis of temporal sequences. Within this structured framework, each denoiser is equipped with a specialized Hierarchical Temporal Mamba (HTM) block, which serves to augment the model's ability to process" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 256, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 256, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 256, + 102 + ], + "type": "text", + "content": "Zeyu and Akide et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "code", + "bbox": [ + 132, + 132, + 379, + 320 + ], + "blocks": [ + { + "bbox": [ + 133, + 116, + 388, + 129 + ], + "lines": [ + { + "bbox": [ + 133, + 116, + 388, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 388, + 129 + ], + "type": "text", + "content": "Algorithm 2 Bidirectional Spatial Mamba (BSM) Block." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 132, + 132, + 379, + 320 + ], + "lines": [ + { + "bbox": [ + 132, + 132, + 379, + 320 + ], + "spans": [ + { + "bbox": [ + 132, + 132, + 379, + 320 + ], + "type": "text", + "content": "Require: compressed latent representations " + }, + { + "bbox": [ + 132, + 132, + 379, + 320 + ], + "type": "inline_equation", + "content": "z:(T,B,C)" + }, + { + "bbox": [ + 132, + 132, + 379, + 320 + ], + "type": "text", + "content": " \nEnsure: transformed representations " + }, + { + "bbox": [ + 132, + 132, + 379, + 320 + ], + "type": "inline_equation", + "content": "z_{\\mathrm{BSM}}:(C,B,E)" + }, + { + "bbox": [ + 132, + 132, + 379, + 320 + ], + "type": "text", + "content": " \n1:/\\* dimension rearrangement \\*/ \n2: " + }, + { + "bbox": [ + 132, + 132, + 379, + 320 + ], + "type": "inline_equation", + "content": "z^{\\prime}:(C,B,T)\\gets" + }, + { + "bbox": [ + 132, + 132, + 379, + 320 + ], + "type": "text", + "content": " Rearrange(z) \n3:/\\* linear projection after normalization \\*/ \n4: " + }, + { + "bbox": [ + 132, + 132, + 379, + 320 + ], + "type": "inline_equation", + "content": "z^{\\prime}:(C,B,T)\\gets" + }, + { + "bbox": [ + 132, + 132, + 379, + 320 + ], + "type": "text", + "content": " Norm(z) \n5:x,z:(C,B,E) " + }, + { + "bbox": [ + 132, + 132, + 379, + 320 + ], + "type": "inline_equation", + "content": "\\leftarrow" + }, + { + "bbox": [ + 132, + 132, + 379, + 320 + ], + "type": "text", + "content": " Linear(z') \n6: for o in {forward, backward} do \n7: " + }, + { + "bbox": [ + 132, + 132, + 379, + 320 + ], + "type": "inline_equation", + "content": "x_o':(C,B,E)\\gets" + }, + { + "bbox": [ + 132, + 132, + 379, + 320 + ], + "type": "text", + "content": " Conv1D(x) \n8: " + }, + { + "bbox": [ + 132, + 132, + 379, + 320 + ], + "type": "inline_equation", + "content": "B_{o},C_{o},\\varDelta_{o}\\gets" + }, + { + "bbox": [ + 132, + 132, + 379, + 320 + ], + "type": "text", + "content": " Linear " + }, + { + "bbox": [ + 132, + 132, + 379, + 320 + ], + "type": "inline_equation", + "content": "(x_o^{\\prime})" + }, + { + "bbox": [ + 132, + 132, + 379, + 320 + ], + "type": "text", + "content": " \n9: Transform " + }, + { + "bbox": [ + 132, + 132, + 379, + 320 + ], + "type": "inline_equation", + "content": "\\overline{A}_o" + }, + { + "bbox": [ + 132, + 132, + 379, + 320 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 132, + 379, + 320 + ], + "type": "inline_equation", + "content": "\\overline{B}_o" + }, + { + "bbox": [ + 132, + 132, + 379, + 320 + ], + "type": "text", + "content": " using " + }, + { + "bbox": [ + 132, + 132, + 379, + 320 + ], + "type": "inline_equation", + "content": "\\varDelta_{o}" + }, + { + "bbox": [ + 132, + 132, + 379, + 320 + ], + "type": "text", + "content": " \n10: " + }, + { + "bbox": [ + 132, + 132, + 379, + 320 + ], + "type": "inline_equation", + "content": "y_{o}\\gets" + }, + { + "bbox": [ + 132, + 132, + 379, + 320 + ], + "type": "text", + "content": " SSM(Ao,Bo,Co) \n11: end for \n12:/\\* gating and summing outputs \\*/ \n13: " + }, + { + "bbox": [ + 132, + 132, + 379, + 320 + ], + "type": "inline_equation", + "content": "z_{\\mathrm{BSM}}:(C,B,T)\\gets" + }, + { + "bbox": [ + 132, + 132, + 379, + 320 + ], + "type": "text", + "content": " GateAndSum(yforward,ybackward,z) \n14: " + }, + { + "bbox": [ + 132, + 132, + 379, + 320 + ], + "type": "inline_equation", + "content": "z_{\\mathrm{BSM}}^{\\prime}:(T,B,C)\\gets" + }, + { + "bbox": [ + 132, + 132, + 379, + 320 + ], + "type": "text", + "content": " Rearrange(z) \n15: Return: " + }, + { + "bbox": [ + 132, + 132, + 379, + 320 + ], + "type": "inline_equation", + "content": "z_{\\mathrm{BSM}}^{\\prime}" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "code_body" + } + ], + "index": 3, + "sub_type": "algorithm" + }, + { + "bbox": [ + 130, + 346, + 480, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 346, + 480, + 380 + ], + "spans": [ + { + "bbox": [ + 130, + 346, + 480, + 380 + ], + "type": "text", + "content": "temporal information effectively. Additionally, the proposed Motion Mamba incorporates an attention-based mixer block denoted as " + }, + { + "bbox": [ + 130, + 346, + 480, + 380 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 346, + 480, + 380 + ], + "type": "text", + "content": ", strategically integrated to enhance conditional fusion." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 382, + 482, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 382, + 482, + 597 + ], + "spans": [ + { + "bbox": [ + 130, + 382, + 482, + 597 + ], + "type": "text", + "content": "Hierarchical Temporal Mamba (HTM) block processes compressed latent representations, denoted as " + }, + { + "bbox": [ + 130, + 382, + 482, + 597 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 130, + 382, + 482, + 597 + ], + "type": "text", + "content": ", with the dimensions " + }, + { + "bbox": [ + 130, + 382, + 482, + 597 + ], + "type": "inline_equation", + "content": "(T,B,C)" + }, + { + "bbox": [ + 130, + 382, + 482, + 597 + ], + "type": "text", + "content": ", of which procedure shown in Algorithm 1. Here, " + }, + { + "bbox": [ + 130, + 382, + 482, + 597 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 382, + 482, + 597 + ], + "type": "text", + "content": " signifies the temporal dimension, as specified in the Variational AutoEncoder (VAE) framework. Initially, the input " + }, + { + "bbox": [ + 130, + 382, + 482, + 597 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 130, + 382, + 482, + 597 + ], + "type": "text", + "content": " is subjected to a linear projection layer, producing transformed representations " + }, + { + "bbox": [ + 130, + 382, + 482, + 597 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 382, + 482, + 597 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 382, + 482, + 597 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 130, + 382, + 482, + 597 + ], + "type": "text", + "content": " with dimension " + }, + { + "bbox": [ + 130, + 382, + 482, + 597 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 130, + 382, + 482, + 597 + ], + "type": "text", + "content": ". Our analysis revealed an increased density of motion within the lower-level feature spaces. Consequently, we developed a hierarchical scanning methodology that is executed at various depths of the network. This approach not only accommodates the diverse motion densities, but also significantly reduces computational overhead. This step utilizes a hierarchically structured set of scans, " + }, + { + "bbox": [ + 130, + 382, + 482, + 597 + ], + "type": "inline_equation", + "content": "K = \\{S_{2N_n - 1}, S_{2N_{n - 1} - 1}, \\ldots, S_1\\}" + }, + { + "bbox": [ + 130, + 382, + 482, + 597 + ], + "type": "text", + "content": ", in conjunction with a corresponding series of memory matrices " + }, + { + "bbox": [ + 130, + 382, + 482, + 597 + ], + "type": "inline_equation", + "content": "\\{A_1, \\ldots, A_k\\}" + }, + { + "bbox": [ + 130, + 382, + 482, + 597 + ], + "type": "text", + "content": ". Each sub-SSM scan first applies a 1-D convolution to " + }, + { + "bbox": [ + 130, + 382, + 482, + 597 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 382, + 482, + 597 + ], + "type": "text", + "content": ", resulting in " + }, + { + "bbox": [ + 130, + 382, + 482, + 597 + ], + "type": "inline_equation", + "content": "x_o'" + }, + { + "bbox": [ + 130, + 382, + 482, + 597 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 130, + 382, + 482, + 597 + ], + "type": "inline_equation", + "content": "x_o'" + }, + { + "bbox": [ + 130, + 382, + 482, + 597 + ], + "type": "text", + "content": " is then linearly projected to derive " + }, + { + "bbox": [ + 130, + 382, + 482, + 597 + ], + "type": "inline_equation", + "content": "B_o" + }, + { + "bbox": [ + 130, + 382, + 482, + 597 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 382, + 482, + 597 + ], + "type": "inline_equation", + "content": "C_o" + }, + { + "bbox": [ + 130, + 382, + 482, + 597 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 382, + 482, + 597 + ], + "type": "inline_equation", + "content": "\\Delta_o" + }, + { + "bbox": [ + 130, + 382, + 482, + 597 + ], + "type": "text", + "content": ". These projections " + }, + { + "bbox": [ + 130, + 382, + 482, + 597 + ], + "type": "inline_equation", + "content": "B_o" + }, + { + "bbox": [ + 130, + 382, + 482, + 597 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 382, + 482, + 597 + ], + "type": "inline_equation", + "content": "C_o" + }, + { + "bbox": [ + 130, + 382, + 482, + 597 + ], + "type": "text", + "content": " use " + }, + { + "bbox": [ + 130, + 382, + 482, + 597 + ], + "type": "inline_equation", + "content": "\\Delta_o" + }, + { + "bbox": [ + 130, + 382, + 482, + 597 + ], + "type": "text", + "content": " to effect transformations in " + }, + { + "bbox": [ + 130, + 382, + 482, + 597 + ], + "type": "inline_equation", + "content": "\\overline{A}_o" + }, + { + "bbox": [ + 130, + 382, + 482, + 597 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 382, + 482, + 597 + ], + "type": "inline_equation", + "content": "\\overline{B}_o" + }, + { + "bbox": [ + 130, + 382, + 482, + 597 + ], + "type": "text", + "content": ", respectively. After executing a sequence of SSM scans " + }, + { + "bbox": [ + 130, + 382, + 482, + 597 + ], + "type": "inline_equation", + "content": "\\{SSM_{A_1,x}, SSM_{A_2,x}, \\ldots, SSM_{A_k,x}\\}" + }, + { + "bbox": [ + 130, + 382, + 482, + 597 + ], + "type": "text", + "content": ", a set of outputs " + }, + { + "bbox": [ + 130, + 382, + 482, + 597 + ], + "type": "inline_equation", + "content": "\\{O_1, \\ldots, O_k\\}" + }, + { + "bbox": [ + 130, + 382, + 482, + 597 + ], + "type": "text", + "content": " is compiled. This collection is subsequently aggregated via a linear projection to obtain the final output of the HTM block." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Bidirectional Spatial Mamba (BSM) block focuses on enhancing latent representation learning through a novel approach of dimension rearrangement and bidirectional scanning, of which the process is shown in Algorithm 2. Initially, it alters the original input dimensions from " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "(T,B,C)" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "(C,B,T)" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": ", effectively swap" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 383, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 383, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 383, + 91, + 447, + 100 + ], + "type": "text", + "content": "Motion Mamba" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "text", + "content": "ping the temporal and channel dimensions. After this rearrangement, the input, now denoted " + }, + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "inline_equation", + "content": "z'" + }, + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "text", + "content": ", undergoes a linear projection after normalization, resulting in dimensions " + }, + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "text", + "content": " of size " + }, + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "text", + "content": ". The process involves bidirectional scanning of the latent channel dimension, where " + }, + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "text", + "content": " is subjected to a 1-D convolution, yielding " + }, + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_o'" + }, + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "text", + "content": " for both forward and backward directions. Each " + }, + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_o'" + }, + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "text", + "content": " is then linearly projected to obtain " + }, + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "inline_equation", + "content": "B_o" + }, + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "inline_equation", + "content": "C_o" + }, + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "inline_equation", + "content": "\\Delta_o" + }, + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "text", + "content": ", which are utilized to transform " + }, + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "inline_equation", + "content": "\\overline{A}_o" + }, + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "inline_equation", + "content": "\\overline{B}_o" + }, + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "text", + "content": ", respectively. The final output token sequence, " + }, + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_1" + }, + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "text", + "content": ", is computed by gating and summing the forward " + }, + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "inline_equation", + "content": "y_{\\mathrm{forward}}" + }, + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "text", + "content": " and backward " + }, + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "inline_equation", + "content": "y_{\\mathrm{backward}}" + }, + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "text", + "content": " output with " + }, + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "text", + "content": ". This component is engineered to decode the structured latent skeleton by analyzing data from both forward and reverse viewpoints. Its main objective is to ensure the seamless continuity of information flow, thereby substantially enhancing the model's ability to generate accurate motion. This is achieved through the maintenance of a dense informational exchange, which is critical for the model's performance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 287, + 230, + 301 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 287, + 230, + 301 + ], + "spans": [ + { + "bbox": [ + 132, + 287, + 230, + 301 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 309, + 205, + 319 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 309, + 205, + 319 + ], + "spans": [ + { + "bbox": [ + 132, + 309, + 205, + 319 + ], + "type": "text", + "content": "4.1 Datasets" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 327, + 481, + 349 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 327, + 481, + 349 + ], + "spans": [ + { + "bbox": [ + 130, + 327, + 481, + 349 + ], + "type": "text", + "content": "We evaluate our proposed Motion Mamba on two prominent Text-to-Motion synthesis benchmarks as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 350, + 488, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 350, + 488, + 409 + ], + "spans": [ + { + "bbox": [ + 130, + 350, + 488, + 409 + ], + "type": "text", + "content": "HumanML3D. The HumanML3D [17] dataset aggregates 14,616 motions sourced from the AMASS [33] and HumanAct12 [18] datasets, with each motion accompanied by three textual descriptions, culminating in 44,970 scripts. This dataset spans a wide range of actions such as exercising, dancing, and acrobatics, presenting a rich motion-language corpus." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 410, + 481, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 410, + 481, + 459 + ], + "spans": [ + { + "bbox": [ + 130, + 410, + 481, + 459 + ], + "type": "text", + "content": "KIT-ML. The KIT-ML dataset [38] is comprised of 3,911 motions paired with 6,278 textual descriptions, serving as a compact yet effective benchmark for evaluation. For both datasets, the pose representation adopted is derived from T2M [17], ensuring consistency in motion representation across evaluations." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 472, + 257, + 484 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 472, + 257, + 484 + ], + "spans": [ + { + "bbox": [ + 132, + 472, + 257, + 484 + ], + "type": "text", + "content": "4.2 Evaluation Metrics" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 490, + 482, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 490, + 482, + 597 + ], + "spans": [ + { + "bbox": [ + 130, + 490, + 482, + 597 + ], + "type": "text", + "content": "We adapt the standard evaluation metrics on following aspects throughout our experiments, including: Generation Quality. We implement a Fréchet inception distance (FID) [21] to quantify the realism and diversity of motion generated by models. Moreover, we use multi-modal distance (MM Dist) to measure the distance between motions and texts and assess motion-text alignment. Diversity. We use the diversity metric to measure motion diversity, which calculates variance in features extracted from the motions. Additionally, we employ multimodality (MModality) to assess diversity within generated motions sharing the same text description." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 612, + 266, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 612, + 266, + 624 + ], + "spans": [ + { + "bbox": [ + 132, + 612, + 266, + 624 + ], + "type": "text", + "content": "4.3 Comparative Studies" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 629, + 486, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 486, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 486, + 667 + ], + "type": "text", + "content": "We evaluate our method against the state-of-the-art methods on the HumanML3D [17] and KIT-ML [38] datasets. We train our Motion Mamba with HTM arrangement strategy MM (" + }, + { + "bbox": [ + 130, + 629, + 486, + 667 + ], + "type": "inline_equation", + "content": "\\{S_{2N_n - 1},\\ldots ,S_1\\}" + }, + { + "bbox": [ + 130, + 629, + 486, + 667 + ], + "type": "text", + "content": "), BSM bidirectional block strategy on the" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 255, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 255, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 255, + 102 + ], + "type": "text", + "content": "Zeyu and Akide et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 169, + 481, + 301 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 169 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 169 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 169 + ], + "type": "text", + "content": "Table 1: Comparison of text-conditional motion synthesis on HumanML3D [17]. These metrics are evaluated by the motion encoder from [17]. Empty MModality indicates the non-diverse generation methods. We employ real motion as a reference and sort all methods by descending FIDs. The right arrow " + }, + { + "bbox": [ + 130, + 114, + 482, + 169 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 130, + 114, + 482, + 169 + ], + "type": "text", + "content": " means that the closer to the real motion, the better. Bold and underline indicate the best and second best result." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 169, + 481, + 301 + ], + "lines": [ + { + "bbox": [ + 132, + 169, + 481, + 301 + ], + "spans": [ + { + "bbox": [ + 132, + 169, + 481, + 301 + ], + "type": "table", + "html": "
MethodR Precision ↑FID↓MM Dist↓Diversity→MModality↑
Top 1Top 2Top 3
Real0.511±.0030.703±.0030.797±.0020.002±.0002.974±.0089.503±.065-
Seq2Seq [39]0.180±.0020.300±.0020.396±.00211.75±.0355.529±.0076.223±.061-
LJ2P [1]0.246±.0010.387±.0020.486±.00211.02±.0465.296±.0087.676±.058-
T2G [5]0.165±.0010.267±.0020.345±.0027.664±.0306.030±.0086.409±.071-
Hier [9]0.301±.0020.425±.0020.552±.0046.532±.0245.012±.0188.332±.042-
TEMOS [37]0.424±.0020.612±.0020.722±.0023.734±.0283.703±.0088.973±.0710.368±.018
T2M [17]0.457±.0020.639±.0030.740±.0031.067±.0023.340±.0089.188±.0022.090±.083
MDM [49]0.320±.0050.498±.0040.611±.0070.544±.0445.566±.0279.559±.0862.799±.072
MotionDiffuse [54]0.491±.0010.681±.0010.782±.0010.630±.0013.113±.0019.410±.0491.553±.042
MLD [6]0.481±.0030.673±.0030.772±.0020.473±.0133.196±.0109.724±.0822.413±.079
Motion Mamba (Ours)0.502±.0030.693±.0020.792±.0020.281±.0093.060±.0589.871±.0842.294±.058
", + "image_path": "f22668e60db39236e4993bff3b2a318b4fa3c2d281126a45986137457738618f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 132, + 328, + 481, + 462 + ], + "blocks": [ + { + "bbox": [ + 132, + 307, + 481, + 327 + ], + "lines": [ + { + "bbox": [ + 132, + 307, + 481, + 327 + ], + "spans": [ + { + "bbox": [ + 132, + 307, + 481, + 327 + ], + "type": "text", + "content": "Table 2: We involve KIT-ML [38] dataset and evaluate the SOTA methods on the text-to-motion task." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 328, + 481, + 462 + ], + "lines": [ + { + "bbox": [ + 132, + 328, + 481, + 462 + ], + "spans": [ + { + "bbox": [ + 132, + 328, + 481, + 462 + ], + "type": "table", + "html": "
MethodR Precision ↑FID↓MM Dist↓Diversity→MModality↑
Top 1Top 2Top 3
Real0.424±.0050.649±.0060.779±.0060.031±.0042.788±.01211.08±.097-
Seq2Seq [39]0.103±.0030.178±.0050.241±.00624.86±.3487.960±.0316.744±.106-
T2G [5]0.156±.0040.255±.0040.338±.00512.12±.1836.964±.0299.334±.079-
LJ2P [1]0.221±.0050.373±.0040.483±.0056.545±.0725.147±.0309.073±.100-
Hier [9]0.255±.0060.432±.0070.531±.0075.203±.1074.986±.0279.563±.0722.090±.083
TEMOS [37]0.353±.0060.561±.0070.687±.0053.717±.0513.417±.01910.84±.1000.532±.034
T2M [17]0.370±.0050.569±.0070.693±.0072.770±.1093.401±.00810.91±.1191.482±.065
MDM [49]0.164±.0040.291±.0040.396±.0040.497±.0219.191±.02210.85±.1091.907±.214
MotionDiffuse [54]0.417±.0040.621±.0040.739±.0041.954±.0622.958±.00511.10±.1430.730±.013
MLD [6]0.390±.0080.609±.0080.734±.0070.404±.0273.204±.02710.80±.1172.192±.071
Motion Mamba (Ours)0.419±.0060.645±.0050.765±.0060.307±.0413.021±.02511.02±.0981.678±.064
", + "image_path": "d4a6902799c96d2278dd1a098881e5ed26a9fd074019736e497d0d9b158b0591.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 472, + 482, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 472, + 482, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 472, + 482, + 593 + ], + "type": "text", + "content": "latent dimension " + }, + { + "bbox": [ + 130, + 472, + 482, + 593 + ], + "type": "inline_equation", + "content": "= 2" + }, + { + "bbox": [ + 130, + 472, + 482, + 593 + ], + "type": "text", + "content": " with 11 layers. We evaluate our model and previous works with suggested metrics in HumanML3D [17] and calculate " + }, + { + "bbox": [ + 130, + 472, + 482, + 593 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 130, + 472, + 482, + 593 + ], + "type": "text", + "content": " confidence interval by repeat evaluation 20 times. The results for the HumanML3D dataset are presented in Table 1. Our model outperforms other methods significantly across various evaluation metrics, including FID, R precision, multi-modal distance, and diversity. For instance, our Motion Mamba outperforms previous best diffusion based motion generation MLD by " + }, + { + "bbox": [ + 130, + 472, + 482, + 593 + ], + "type": "inline_equation", + "content": "40.5\\%" + }, + { + "bbox": [ + 130, + 472, + 482, + 593 + ], + "type": "text", + "content": " in terms of FID, and up to " + }, + { + "bbox": [ + 130, + 472, + 482, + 593 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 130, + 472, + 482, + 593 + ], + "type": "text", + "content": " improvement on R Precision, we aslo obatined best MModality by 3.060. The results for the KIT-ML dataset are presented in Table 2. We have also outperformed other well-established methods in FID and multi-modal distance." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 609, + 246, + 620 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 609, + 246, + 620 + ], + "spans": [ + { + "bbox": [ + 132, + 609, + 246, + 620 + ], + "type": "text", + "content": "4.4 Ablation Studies" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 628, + 481, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 628, + 481, + 664 + ], + "spans": [ + { + "bbox": [ + 130, + 628, + 481, + 664 + ], + "type": "text", + "content": "We concluded the ablation studies including long sequence evaluation, hierarchical design with HTM, bidirectional design in the BSM, number of latent dimensions, and number of layers of our proposed motion mamba in Table 4." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 383, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 383, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 383, + 91, + 447, + 100 + ], + "type": "text", + "content": "Motion Mamba" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 168, + 481, + 251 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 169 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 169 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 169 + ], + "type": "text", + "content": "Table 3: In order to evaluate the models' capability in long sequence motion generation, we compared our method with an existing approach on the recently introduced HumanML3D-LS dataset. This dataset comprises motion sequences longer than 190 frames from the original evaluation set. Our model demonstrates superior performance compared to other methods." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 168, + 481, + 251 + ], + "lines": [ + { + "bbox": [ + 132, + 168, + 481, + 251 + ], + "spans": [ + { + "bbox": [ + 132, + 168, + 481, + 251 + ], + "type": "table", + "html": "
MethodR Precision ↑FID↓MM Dist↓Diversity→MModality↑
Top 1Top 2Top 3
Real0.437±.0030.622±.0040.721±.0040.004±.0003.343±.0158.423±.090-
MDM [49]0.368±.0050.553±.0060.672±.0050.802±.0443.860±.0258.817±.068-
MotionDiffuse [54]0.367±.0040.521±.0040.623±.0042.460±.0623.789±.0058.707±.1431.602±.013
MLD [6]0.403±.0050.584±.0050.690±.0050.952±.0203.580±.0169.050±.0852.711±.104
Motion Mamba (Ours)0.417±.0030.606±.0030.713±.0040.668±.0193.435±.0159.021±.0702.373±.084
", + "image_path": "f10dee507e46e0ced2ba07214fcf83530aab51e42cf950b0ed7dd0baa6212667.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 266, + 333, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 266, + 333, + 422 + ], + "spans": [ + { + "bbox": [ + 130, + 266, + 333, + 422 + ], + "type": "text", + "content": "Long Sequence Motion Generation. The HumanML3D [17] dataset exhibits a long-tailed and right-skewed distribution with a significant proportion of long-sequence human motions, as shown in Figure 3. We suggest previous study overlooked the challenges in the long-sequence generation problem. Thus, we introduce a new dataset variant, HumanML3D-LS, comprising motion sequences longer than 190 frames extracted from the original test set. This addition allows us to showcase our capability in generating long-sequence motions. Subsequently, we eval" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 340, + 265, + 482, + 351 + ], + "blocks": [ + { + "bbox": [ + 340, + 265, + 482, + 351 + ], + "lines": [ + { + "bbox": [ + 340, + 265, + 482, + 351 + ], + "spans": [ + { + "bbox": [ + 340, + 265, + 482, + 351 + ], + "type": "image", + "image_path": "aae2391766f3828f2e11babdcf89aea649a1a4ec6a3b1b4196a8a1944d3acb76.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 339, + 358, + 482, + 413 + ], + "lines": [ + { + "bbox": [ + 339, + 358, + 482, + 413 + ], + "spans": [ + { + "bbox": [ + 339, + 358, + 482, + 413 + ], + "type": "text", + "content": "Fig. 3: The figure shows a long tail distribution of the HumanML3D [17], which has a significant proportion of long-sequence human motions." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 422, + 482, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 422, + 482, + 471 + ], + "spans": [ + { + "bbox": [ + 130, + 422, + 482, + 471 + ], + "type": "text", + "content": "uate the performance of our method on HumanML3D-LS and compare it with other diffusion-based motion generation approaches. The comparative results are presented in Table 3. Motion Mamba by leverage the benefits on long-range dependency modeling make it well suitable for long sequence motion generation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 472, + 482, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 472, + 482, + 604 + ], + "spans": [ + { + "bbox": [ + 130, + 472, + 482, + 604 + ], + "type": "text", + "content": "Hierarchical Design with HTM. In our ablation studies, we observed a slight improvement upon reversing the scan order from a lower to a higher level, specifically transitioning from MM " + }, + { + "bbox": [ + 130, + 472, + 482, + 604 + ], + "type": "inline_equation", + "content": "\\{S_1,\\ldots ,S_N\\}" + }, + { + "bbox": [ + 130, + 472, + 482, + 604 + ], + "type": "text", + "content": " to MM " + }, + { + "bbox": [ + 130, + 472, + 482, + 604 + ], + "type": "inline_equation", + "content": "\\{S_N,\\dots ,S_1\\}" + }, + { + "bbox": [ + 130, + 472, + 482, + 604 + ], + "type": "text", + "content": ". This enhancement suggests a correlation with the increase in temporal motion density within the lower-level feature spaces. Furthermore, to achieve the optimal result, we introduce the hierarchical design to arrange the scanning frequency, resulting in the sequence MM " + }, + { + "bbox": [ + 130, + 472, + 482, + 604 + ], + "type": "inline_equation", + "content": "\\{S_{2N_n - 1},\\ldots ,S_1\\}" + }, + { + "bbox": [ + 130, + 472, + 482, + 604 + ], + "type": "text", + "content": ". This expansion in the number of scans led to a performance increase. We attribute this enhancement to the observation that individual selective scan operations significantly reduce the parameter count, especially when compared to the parameter-intensive constructs of self-attention and feedforward network blocks prevalent in transformer architectures." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "Bidirectional Design in BSM. We developed three distinct variations of latent scanning mechanisms, differentiated by their scanning directions. In the context of motion generation tasks, we posit that the flow of hidden information within the structured latent skeleton holds significance, an aspect previously underexplored. Our ablation study reveals that a single scan across the la" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 255, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 255, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 255, + 102 + ], + "type": "text", + "content": "Zeyu and Akide et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 164, + 158, + 446, + 352 + ], + "blocks": [ + { + "bbox": [ + 131, + 114, + 482, + 157 + ], + "lines": [ + { + "bbox": [ + 131, + 114, + 482, + 157 + ], + "spans": [ + { + "bbox": [ + 131, + 114, + 482, + 157 + ], + "type": "text", + "content": "Table 4: Evaluation of text-based motion synthesis on HumanML3D [17]: we use metrics in Table 1 and provides real reference, we evaluate the various HTM and BSM design choices, the dimension of the latent input, the different number of layer of Motion Mamba model." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 164, + 158, + 446, + 352 + ], + "lines": [ + { + "bbox": [ + 164, + 158, + 446, + 352 + ], + "spans": [ + { + "bbox": [ + 164, + 158, + 446, + 352 + ], + "type": "table", + "html": "
ModelsR Precision Top 3↑FID↓MM Dist.↓Diversity→MModality↑
Real0.797±.0020.002±.0002.974±.0089.503±.065-
MM ( {S1, ..., SN})0.673±.0031.278±.0123.802±.0418.678±.0963.127±.024
MM ( {SN, ..., S1})0.738±.0020.962±.0113.433±.0039.180±.0712.723±.033
MM ( {S1, ..., S2Nn-1})0.698±.0020.856±.0083.624±.0379.229±.0672.826±.017
MM ( {S2Nn-1, ..., S1})0.792±.0020.281±.0093.060±.0589.871±.0842.294±.058
MM (SingleScan)0.736±.0031.063±.0103.443±.0269.180±.0672.676±.041
MM (BiScan, layer)0.735±.0040.789±.0073.408±.0349.374±.0592.591±.046
MM (BiScan, block)0.792±.0020.281±.0093.060±.0589.871±.0842.294±.058
MM (Dim, 1)0.706±.0030.652±.0113.541±.0729.141±.0822.612±.055
MM (Dim, 2)0.792±.0020.281±.0093.060±.0589.871±.0842.294±.058
MM (Dim, 5)0.741±.0080.728±.0093.307±.0279.427±.0992.314±.062
MM (Dim, 7)0.738±.0040.599±.0073.359±.0689.166±.0752.488±.037
MM (Dim, 10)0.715±.0030.628±.0083.548±.0439.200±.0752.884±.096
MM (9 layers)0.755±.0021.080±.0123.309±.0579.721±.0812.974±.039
MM (11 layers)0.792±.0020.281±.0093.060±.0589.871±.0842.294±.058
MM (27 layers)0.750±.0030.975±.0083.336±.0969.249±.0712.821±.063
MM (37 layers)0.754±.0050.809±.0103.338±.0619.355±.0622.741±.077
", + "image_path": "91cc7db9915584cb2721c037bc8a8073a2e1c0bf79422ad9a14143816aef9ff8.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 365, + 482, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 365, + 482, + 437 + ], + "spans": [ + { + "bbox": [ + 130, + 365, + 482, + 437 + ], + "type": "text", + "content": "tent dimension yields minimal improvement. Subsequently, we investigated both layer-based and block-based bidirectional scans. Our findings indicate that the block-based bidirectional scan achieves optimal performance. This suggests that spatial information flows are encoded within the latent spaces and that bidirectional scanning facilitates the exchange of this information, thereby enhancing the efficacy of motion generation tasks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 437, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 482, + 668 + ], + "type": "text", + "content": "Architecture Design for Motion Mamba. The proposed Motion Mamba which is grounded in a standardized motion latent diffusion system. We delved into the interplay between dimensional aspects and the module's capacity (measured by the number of layers) to ascertain their impact on system performance. Experimental results demonstrate that the Motion Mamba achieves superior performance at a latent dimension of 2, diverging from prior works where the optimal dimension was identified as 1. We attribute this discrepancy to our HTM, which necessitates multiple scans correlating with the sequence length, thus implicating dimensionality as a pivotal factor. A marginal increase in dimensionality enabled us to attain peak performance, simultaneously enhancing efficiency compared to models with a dimensionality of 10. Furthermore, we conducted experiments to determine the optimal layer count for Motion Mamba, inspired by the design of its selective scanning mechanism. Notably, a single Mamba layer achieves a parameter reduction of approximately " + }, + { + "bbox": [ + 130, + 437, + 482, + 668 + ], + "type": "inline_equation", + "content": "75\\%" + }, + { + "bbox": [ + 130, + 437, + 482, + 668 + ], + "type": "text", + "content": " compared to a conventional transformer encoder block. By increasing the number of layers, we aim to uncover the relationship between model capacity and its performance. Our findings reveal that, through the integration of our specially designed HTM and BSM (Bidirectional Scanning Module) blocks, the Motion Mamba reaches its optimal performance with 11 layers. This represents a slight increase over the MLD [6] baseline. How-" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 383, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 383, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 383, + 91, + 447, + 100 + ], + "type": "text", + "content": "Motion Mamba" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 480, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 480, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 480, + 140 + ], + "type": "text", + "content": "ever, due to the reduced parameter count in each layer, Motion Mamba exhibits significantly greater efficiency than previous methodologies." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 158, + 237, + 169 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 158, + 237, + 169 + ], + "spans": [ + { + "bbox": [ + 132, + 158, + 237, + 169 + ], + "type": "text", + "content": "4.5 Inference Time" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 178, + 317, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 178, + 317, + 346 + ], + "spans": [ + { + "bbox": [ + 130, + 178, + 317, + 346 + ], + "type": "text", + "content": "Inference time remains a significant challenge for diffusion-based methods. To address this, we enhance the inference speed by incorporating the efficient Mamba block within a lightweight architecture. Compared to the previous strong baseline, such as the MLD model cited in [6], which reports an average inference time of 0.217 seconds, our Motion Mamba model achieves a notable reduction in computational overhead, as shown in Figure 4. Specifically, it requires four times less computational effort, thereby facilitating faster and real-time inference speeds." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 329, + 175, + 468, + 280 + ], + "blocks": [ + { + "bbox": [ + 329, + 175, + 468, + 280 + ], + "lines": [ + { + "bbox": [ + 329, + 175, + 468, + 280 + ], + "spans": [ + { + "bbox": [ + 329, + 175, + 468, + 280 + ], + "type": "image", + "image_path": "3f16a76fbbc0786783ac633696339787bd4ab38a49ca1a6d7338b047d5fb89d2.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 321, + 289, + 482, + 366 + ], + "lines": [ + { + "bbox": [ + 321, + 289, + 482, + 366 + ], + "spans": [ + { + "bbox": [ + 321, + 289, + 482, + 366 + ], + "type": "text", + "content": "Fig.4: The figure shows the average inference time per sentence (AIT) vs FID, our proposed motion mamba obtained 0.058s AIT and 0.281 FID over all outperform previous methods. We evaluate all methods on a single V100 GPU." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 363, + 312, + 376 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 363, + 312, + 376 + ], + "spans": [ + { + "bbox": [ + 132, + 363, + 312, + 376 + ], + "type": "text", + "content": "5 Discussion and Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 388, + 482, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 388, + 482, + 594 + ], + "spans": [ + { + "bbox": [ + 130, + 388, + 482, + 594 + ], + "type": "text", + "content": "In this study, we introduced Motion Mamba, a novel framework designed for efficient and extended sequence motion generation. Our approach represents the inaugural integration of the Mamba model within the domain of motion generation, featuring significant advancements including the implementation of Hierarchical Temporal Mamba (HTM) blocks. These blocks are specifically engineered to enhance temporal alignment through hierarchically organized selective scanning. Furthermore, Bidirectional Spatial Mamba (BSM) blocks have been developed to amplify the exchange of information flow within latent spaces, thereby augmenting the model's ability to bidirectionally capture skeleton-level density features with greater precision. Compared to previous diffusion-based motion generation methodologies that predominantly utilize transformer blocks, our Motion Mamba framework achieves SOTA performance, evidencing an improvement of up to " + }, + { + "bbox": [ + 130, + 388, + 482, + 594 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 130, + 388, + 482, + 594 + ], + "type": "text", + "content": " in FID scores and a quadrupled improvement in inference speed. Through comprehensive experimentation across a variety of human motion generation tasks, the effectiveness and efficiency of our proposed Motion Mamba model have been robustly demonstrated, marking a significant leap forward in the field of human motion generation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 611, + 247, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 247, + 624 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 247, + 624 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 636, + 481, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 636, + 481, + 661 + ], + "spans": [ + { + "bbox": [ + 130, + 636, + 481, + 661 + ], + "type": "text", + "content": "This work is partially supported by the Fundamental Research Funds for the Central Universities, Peking University." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 255, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 255, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 255, + 102 + ], + "type": "text", + "content": "Zeyu and Akide et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 137, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 138, + 137, + 481, + 169 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 137, + 481, + 169 + ], + "spans": [ + { + "bbox": [ + 138, + 137, + 481, + 169 + ], + "type": "text", + "content": "1. Ahuja, C., Morency, L.P.: Language2pose: Natural language grounded pose forecasting. In: 2019 International Conference on 3D Vision (3DV). pp. 719-728. IEEE (2019)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 170, + 481, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 170, + 481, + 191 + ], + "spans": [ + { + "bbox": [ + 138, + 170, + 481, + 191 + ], + "type": "text", + "content": "2. Bao, F., Li, C., Cao, Y., Zhu, J.: All are worth words: a vit backbone for score-based diffusion models. arXiv preprint arXiv:2209.12152 (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 191, + 481, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 191, + 481, + 213 + ], + "spans": [ + { + "bbox": [ + 138, + 191, + 481, + 213 + ], + "type": "text", + "content": "3. Baron, E., Zimerman, I., Wolf, L.: 2-d ssm: A general spatial layer for visual transformers. arXiv preprint arXiv:2306.06635 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 213, + 481, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 213, + 481, + 245 + ], + "spans": [ + { + "bbox": [ + 138, + 213, + 481, + 245 + ], + "type": "text", + "content": "4. Barsoum, E., Kender, J., Liu, Z.: Hp-gan: Probabilistic 3d human motion prediction via gan. In: Proceedings of the IEEE conference on computer vision and pattern recognition workshops. pp. 1418-1427 (2018)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 245, + 481, + 288 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 245, + 481, + 288 + ], + "spans": [ + { + "bbox": [ + 138, + 245, + 481, + 288 + ], + "type": "text", + "content": "5. Bhattacharya, U., Rewkowski, N., Banerjee, A., Guhan, P., Bera, A., Manocha, D.: Text2gestures: A transformer-based network for generating emotive body gestures for virtual agents. In: 2021 IEEE virtual reality and 3D user interfaces (VR). pp. 1-10. IEEE (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 289, + 481, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 289, + 481, + 320 + ], + "spans": [ + { + "bbox": [ + 138, + 289, + 481, + 320 + ], + "type": "text", + "content": "6. Chen, X., Jiang, B., Liu, W., Huang, Z., Fu, B., Chen, T., Yu, G.: Executing your commands via motion diffusion in latent space. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18000-18010 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 320, + 481, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 320, + 481, + 342 + ], + "spans": [ + { + "bbox": [ + 138, + 320, + 481, + 342 + ], + "type": "text", + "content": "7. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems 34, 8780-8794 (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 342, + 481, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 342, + 481, + 385 + ], + "spans": [ + { + "bbox": [ + 138, + 342, + 481, + 385 + ], + "type": "text", + "content": "8. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. In: International Conference on Learning Representations (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 385, + 481, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 385, + 481, + 418 + ], + "spans": [ + { + "bbox": [ + 138, + 385, + 481, + 418 + ], + "type": "text", + "content": "9. Ghosh, A., Cheema, N., Oguz, C., Theobalt, C., Slusallek, P.: Synthesis of compositional animations from textual descriptions. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 1396-1406 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 418, + 481, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 418, + 481, + 460 + ], + "spans": [ + { + "bbox": [ + 138, + 418, + 481, + 460 + ], + "type": "text", + "content": "0. Gong, K., Lian, D., Chang, H., Guo, C., Jiang, Z., Zuo, X., Mi, M.B., Wang, X.: Tm2d: Bimodality driven 3d dance generation via music-text integration. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9942-9952 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 460, + 481, + 494 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 460, + 481, + 494 + ], + "spans": [ + { + "bbox": [ + 138, + 460, + 481, + 494 + ], + "type": "text", + "content": "1. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial nets. Advances in neural information processing systems 27 (2014)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 494, + 481, + 515 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 494, + 481, + 515 + ], + "spans": [ + { + "bbox": [ + 138, + 494, + 481, + 515 + ], + "type": "text", + "content": "2. Gu, A., Dao, T.: Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 515, + 481, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 515, + 481, + 546 + ], + "spans": [ + { + "bbox": [ + 138, + 515, + 481, + 546 + ], + "type": "text", + "content": "3. Gu, A., Goel, K., Gupta, A., Ré, C.: On the parameterization and initialization of diagonal state space models. Advances in Neural Information Processing Systems 35, 35971-35983 (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 547, + 481, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 547, + 481, + 568 + ], + "spans": [ + { + "bbox": [ + 138, + 547, + 481, + 568 + ], + "type": "text", + "content": "4. Gu, A., Goel, K., Ré, C.: Efficiently modeling long sequences with structured state spaces. arXiv preprint arXiv:2111.00396 (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 568, + 481, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 568, + 481, + 601 + ], + "spans": [ + { + "bbox": [ + 138, + 568, + 481, + 601 + ], + "type": "text", + "content": "5. Gu, A., Johnson, I., Goel, K., Saab, K., Dao, T., Rudra, A., Ré, C.: Combining recurrent, convolutional, and continuous-time models with linear state space layers. Advances in neural information processing systems 34, 572-585 (2021)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 601, + 481, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 601, + 481, + 633 + ], + "spans": [ + { + "bbox": [ + 138, + 601, + 481, + 633 + ], + "type": "text", + "content": "6. Gulrajani, I., Ahmed, F., Arjovsky, M., Dumoulin, V., Courville, A.C.: Improved training of wasserstein gans. Advances in neural information processing systems 30 (2017)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 633, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 633, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 633, + 481, + 665 + ], + "type": "text", + "content": "7. Guo, C., Zou, S., Zuo, X., Wang, S., Ji, W., Li, X., Cheng, L.: Generating diverse and natural 3d human motions from text. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5152-5161 (2022)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 383, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 383, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 383, + 91, + 447, + 100 + ], + "type": "text", + "content": "Motion Mamba" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "text", + "content": "18. Guo, C., Zuo, X., Wang, S., Zou, S., Sun, Q., Deng, A., Gong, M., Cheng, L.: Action2motion: Conditioned generation of 3d human motions. In: Proceedings of the 28th ACM International Conference on Multimedia. pp. 2021-2029 (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 481, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 481, + 183 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 481, + 183 + ], + "type": "text", + "content": "19. Gupta, A., Gu, A., Berant, J.: Diagonal state spaces are as effective as structured state spaces. Advances in Neural Information Processing Systems 35, 22982-22994 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 184, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 184, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 184, + 481, + 205 + ], + "type": "text", + "content": "20. Harvey, F.G., Yurick, M., Nowrouzezahrai, D., Pal, C.: Robust motion in-between. ACM Transactions on Graphics (TOG) 39(4), 60-1 (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 206, + 481, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 206, + 481, + 239 + ], + "spans": [ + { + "bbox": [ + 132, + 206, + 481, + 239 + ], + "type": "text", + "content": "21. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems 30 (2017)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 239, + 481, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 239, + 481, + 262 + ], + "spans": [ + { + "bbox": [ + 132, + 239, + 481, + 262 + ], + "type": "text", + "content": "22. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems 33, 6840-6851 (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 262, + 481, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 262, + 481, + 296 + ], + "spans": [ + { + "bbox": [ + 132, + 262, + 481, + 296 + ], + "type": "text", + "content": "23. Hopfield, J.J.: Neural networks and physical systems with emergent collective computational abilities. Proceedings of the national academy of sciences 79(8), 2554-2558 (1982)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 297, + 481, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 297, + 481, + 329 + ], + "spans": [ + { + "bbox": [ + 132, + 297, + 481, + 329 + ], + "type": "text", + "content": "24. Jiang, B., Chen, X., Liu, W., Yu, J., Yu, G., Chen, T.: Motiongpt: Human motion as a foreign language. Advances in Neural Information Processing Systems 36 (2024)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 330, + 481, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 330, + 481, + 351 + ], + "spans": [ + { + "bbox": [ + 132, + 330, + 481, + 351 + ], + "type": "text", + "content": "25. Kalman, R.E.: A New Approach to Linear Filtering and Prediction Problems. Journal of Basic Engineering 82(1), 35-45 (1960)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 352, + 481, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 352, + 481, + 363 + ], + "spans": [ + { + "bbox": [ + 132, + 352, + 481, + 363 + ], + "type": "text", + "content": "26. Kingma, D.P., Welling, M.: Auto-encoding variational bayes. stat 1050, 1 (2014)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 364, + 481, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 364, + 481, + 385 + ], + "spans": [ + { + "bbox": [ + 132, + 364, + 481, + 385 + ], + "type": "text", + "content": "27. Kramer, M.A.: Nonlinear principal component analysis using autoassociative neural networks. AIChE journal 37(2), 233-243 (1991)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 386, + 481, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 386, + 481, + 408 + ], + "spans": [ + { + "bbox": [ + 132, + 386, + 481, + 408 + ], + "type": "text", + "content": "28. Li, S., Singh, H., Grover, A.: Mamba-nd: Selective state space modeling for multidimensional data. arXiv preprint arXiv:2402.05892 (2024)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 409, + 481, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 409, + 481, + 441 + ], + "spans": [ + { + "bbox": [ + 132, + 409, + 481, + 441 + ], + "type": "text", + "content": "29. Li, Y., Cai, T., Zhang, Y., Chen, D., Dey, D.: What makes convolutional models great on long sequence modeling? In: The Eleventh International Conference on Learning Representations (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 442, + 481, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 442, + 481, + 464 + ], + "spans": [ + { + "bbox": [ + 132, + 442, + 481, + 464 + ], + "type": "text", + "content": "30. Lin, X., Amer, M.R.: Human motion modeling using dvgans. arXiv preprint arXiv:1804.10652 (2018)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 464, + 481, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 464, + 481, + 487 + ], + "spans": [ + { + "bbox": [ + 132, + 464, + 481, + 487 + ], + "type": "text", + "content": "31. Liu, Y., Tian, Y., Zhao, Y., Yu, H., Xie, L., Wang, Y., Ye, Q., Liu, Y.: Vmamba: Visual state space model. arXiv preprint arXiv:2401.10166 (2024)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 487, + 481, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 487, + 481, + 520 + ], + "spans": [ + { + "bbox": [ + 132, + 487, + 481, + 520 + ], + "type": "text", + "content": "32. Liu, Z., Mao, H., Wu, C.Y., Feichtenhofer, C., Darrell, T., Xie, S.: A convnet for the 2020s. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11976-11986 (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 521, + 481, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 521, + 481, + 553 + ], + "spans": [ + { + "bbox": [ + 132, + 521, + 481, + 553 + ], + "type": "text", + "content": "33. Mahmood, N., Ghorbani, N., Troje, N.F., Pons-Moll, G., Black, M.J.: Amass: Archive of motion capture as surface shapes. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5442-5451 (2019)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 554, + 481, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 554, + 481, + 597 + ], + "spans": [ + { + "bbox": [ + 132, + 554, + 481, + 597 + ], + "type": "text", + "content": "34. Pavlakos, G., Choutas, V., Ghorbani, N., Bolkart, T., Osman, A.A.A., Tzionas, D., Black, M.J.: Expressive body capture: 3d hands, face, and body from a single image. In: Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR) (2019)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 599, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 599, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 599, + 481, + 632 + ], + "type": "text", + "content": "35. Petrovich, M., Black, M.J., Varol, G.: Action-conditioned 3D human motion synthesis with transformer VAE. In: International Conference on Computer Vision (ICCV) (2021)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "type": "text", + "content": "36. Petrovich, M., Black, M.J., Varol, G.: Temos: Generating diverse human motions from textual descriptions. In: European Conference on Computer Vision. pp. 480-497. Springer (2022)" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 255, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 255, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 255, + 101 + ], + "type": "text", + "content": "Zeyu and Akide et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 666 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "37. Petrovich, M., Black, M.J., Varol, G.: TEMOS: Generating diverse human motions from textual descriptions. In: European Conference on Computer Vision (ECCV) (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 481, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 481, + 172 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 481, + 172 + ], + "type": "text", + "content": "38. Plappert, M., Mandery, C., Asfour, T.: The kit motion-language dataset. Big data 4(4), 236-252 (2016)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 481, + 205 + ], + "type": "text", + "content": "39. Plappert, M., Mandery, C., Asfour, T.: Learning a bidirectional mapping between human whole-body motion and natural language using deep recurrent neural networks. Robotics and Autonomous Systems 109, 13-26 (2018)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 205, + 481, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 481, + 247 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 481, + 247 + ], + "type": "text", + "content": "40. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 249, + 481, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 481, + 282 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 481, + 282 + ], + "type": "text", + "content": "41. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 282, + 481, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 282, + 481, + 315 + ], + "spans": [ + { + "bbox": [ + 132, + 282, + 481, + 315 + ], + "type": "text", + "content": "42. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: International Conference on Medical image computing and computer-assisted intervention. pp. 234-241. Springer (2015)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 315, + 481, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 481, + 336 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 481, + 336 + ], + "type": "text", + "content": "43. Rumelhart, D.E., Hinton, G.E., Williams, R.J.: Learning internal representations by error propagation. Parallel Distributed Processing pp. 318-362 (1986)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 336, + 481, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 481, + 369 + ], + "type": "text", + "content": "44. Sahara, C., Ho, J., Chan, W., Salimans, T., Fleet, D.J., Norouzi, M.: Image superresolution via iterative refinement. IEEE Transactions on Pattern Analysis and Machine Intelligence (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 369, + 481, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 481, + 403 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 481, + 403 + ], + "type": "text", + "content": "45. Shi, X., Chen, Z., Wang, H., Yeung, D.Y., Wong, W.K., Woo, W.c.: Convolutional LSTM network: A machine learning approach for precipitation nowcasting. Advances in neural information processing systems 28 (2015)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 403, + 481, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 481, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 481, + 435 + ], + "type": "text", + "content": "46. Smith, J., De Mello, S., Kautz, J., Linderman, S., Byeon, W.: Convolutional state space models for long-range spatiotemporal modeling. Advances in Neural Information Processing Systems 36 (2024)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 435, + 481, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 481, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 481, + 468 + ], + "type": "text", + "content": "47. Sohl-Dickstein, J., Weiss, E., Maheswaranathan, N., Ganguli, S.: Deep unsupervised learning using nonequilibrium thermodynamics. In: International conference on machine learning. pp. 2256-2265. PMLR (2015)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 468, + 481, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 481, + 501 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 481, + 501 + ], + "type": "text", + "content": "48. Tevet, G., Gordon, B., Hertz, A., Bermano, A.H., Cohen-Or, D.: Motionclip: Exposing human motion generation to clip space. In: European Conference on Computer Vision. pp. 358-374. Springer (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 501, + 481, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 481, + 534 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 481, + 534 + ], + "type": "text", + "content": "49. Tevet, G., Raab, S., Gordon, B., Shafir, Y., Cohen-or, D., Bermano, A.H.: Human motion diffusion model. In: The Eleventh International Conference on Learning Representations (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 534, + 481, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 481, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 481, + 555 + ], + "type": "text", + "content": "50. Van Den Oord, A., Vinyals, O., et al.: Neural discrete representation learning. Advances in neural information processing systems 30 (2017)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 555, + 481, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 481, + 588 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 481, + 588 + ], + "type": "text", + "content": "51. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 588, + 481, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 588, + 481, + 611 + ], + "spans": [ + { + "bbox": [ + 132, + 588, + 481, + 611 + ], + "type": "text", + "content": "52. Wang, J., Yan, J.N., Gu, A., Rush, A.M.: Pretraining without attention. arXiv preprint arXiv:2212.10544 (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 611, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 481, + 632 + ], + "type": "text", + "content": "53. Yan, J.N., Gu, J., Rush, A.M.: Diffusion models without attention. arXiv preprint arXiv:2311.18257 (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "text", + "content": "54. Zhang, M., Cai, Z., Pan, L., Hong, F., Guo, X., Yang, L., Liu, Z.: Motiondiffuse: Text-driven human motion generation with diffusion model. IEEE Transactions on Pattern Analysis and Machine Intelligence (2024)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 383, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 383, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 383, + 91, + 447, + 100 + ], + "type": "text", + "content": "Motion Mamba" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 216 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "text", + "content": "55. Zhang, Y., Black, M.J., Tang, S.: We are more than our joints: Predicting how 3d bodies move. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3372-3382 (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 480, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 480, + 183 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 480, + 183 + ], + "type": "text", + "content": "56. Zhong, C., Hu, L., Zhang, Z., Xia, S.: Attt2m: Text-driven human motion generation with multi-perspective attention mechanism. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 509-519 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 480, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 480, + 216 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 480, + 216 + ], + "type": "text", + "content": "57. Zhu, L., Liao, B., Zhang, Q., Wang, X., Liu, W., Wang, X.: Vision mamba: Efficient visual representation learning with bidirectional state space model. arXiv preprint arXiv:2401.09417 (2024)" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 255, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 255, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 255, + 102 + ], + "type": "text", + "content": "Zeyu and Akide et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Motion and Structure from Event-based Normal Flow/2dddd958-d905-447a-8afe-98428ea7b185_content_list.json b/2024/Motion and Structure from Event-based Normal Flow/2dddd958-d905-447a-8afe-98428ea7b185_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..ff377a40d78dbfe97430b78b49bcf9f3628dc9c3 --- /dev/null +++ b/2024/Motion and Structure from Event-based Normal Flow/2dddd958-d905-447a-8afe-98428ea7b185_content_list.json @@ -0,0 +1,1909 @@ +[ + { + "type": "text", + "text": "Motion and Structure from Event-based Normal Flow", + "text_level": 1, + "bbox": [ + 271, + 140, + 730, + 183 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhongyang Ren $^{1*}$ , Bangyan Liao $^{2*}$ , Delei Kong $^{1}$ , Jinghang Li $^{1}$ , Peidong Liu $^{2}$ , Laurent Kneip $^{3}$ , Guillermo Gallego $^{4}$ , and Yi Zhou $^{1*}$", + "bbox": [ + 235, + 210, + 766, + 243 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ School of Robotics, Hunan University", + "bbox": [ + 367, + 253, + 635, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ School of Engineering, Westlake University", + "bbox": [ + 349, + 268, + 651, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 School of Information Science and Technology, ShanghaiTech University", + "bbox": [ + 254, + 282, + 746, + 297 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "TU Berlin, ECDF, SCIoI Excellence Cluster and Robotics Institute Germany", + "bbox": [ + 238, + 296, + 764, + 310 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Recovering the camera motion and scene geometry from visual data is a fundamental problem in computer vision. Its success in conventional (frame-based) vision is attributed to the maturity of feature extraction, data association and multi-view geometry. The emergence of asynchronous (event-based) cameras calls for new approaches that use raw event data as input to solve this fundamental problem. State-of-the-art solutions typically infer data association implicitly by iteratively reversing the event data generation process. However, the nonlinear nature of these methods limits their applicability in real-time tasks, and the constant-motion assumption leads to unstable results under agile motion. To this end, we reformulate the problem in a way that aligns better with the differential working principle of event cameras. We show that event-based normal flow can be used, via the proposed geometric error term, as an alternative to the full (optical) flow in solving a family of geometric problems that involve instantaneous first-order kinematics and scene geometry. Furthermore, we develop a fast linear solver and a continuous-time nonlinear solver on top of the proposed geometric error term. Experiments on both synthetic and real data show the superiority of our linear solver in terms of accuracy and efficiency, and its practicality as an initializer for previous nonlinear solvers. Besides, our continuous-time non-linear solver exhibits exceptional capabilities in accommodating sudden variations in motion since it does not rely on the constant-motion assumption. Our project page can be found at https://nail-hnu.github.io/EvLinearSolver/.", + "bbox": [ + 259, + 347, + 743, + 680 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Event cameras $\\cdot$ Geometric model fitting $\\cdot$ Normal flow", + "bbox": [ + 261, + 693, + 718, + 707 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 755, + 375, + 771 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Neuromorphic event-based cameras are bio-inspired sensors that report only brightness changes at each pixel location asynchronously. Their high temporal", + "bbox": [ + 212, + 786, + 787, + 816 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* equal contribution; $\\boxtimes$ corresponding author (eeyzhou@hnu.edu.cn).", + "bbox": [ + 230, + 825, + 692, + 840 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "resolution enables them to sense at exactly the same rate as the scene dynamics, and thus, event cameras are suitable for dealing with robotic perception [33,38, 39,41,43], navigation [21,34,35,42], and control [6,9] tasks that involve agile motion. A general and fundamental problem among these tasks is to recover the camera's ego-motion and scene geometry from event data. One of the bottlenecks for solving this kind of problems at very high speeds is the lack of methods for efficient and robust data association.", + "bbox": [ + 212, + 146, + 787, + 252 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Unlike its standard-vision counterparts, an event camera cannot observe visual information (i.e. intensity changes) when motion is parallel to edges. Such a partial observability of event data makes optical flow estimation (as well as feature detection and matching) an ill-posed problem. To this end, many existing methods [14,17,19,22,30] for solving the problem of motion-model fitting on event data propose specific metrics that assess event-data association established in either an implicit manner or an explicitly model-driven style. The best fit can be retrieved by iteratively optimizing the data association, often expressed in terms of event alignment. However, we argue two flaws of these methods: $(i)$ The success of these non-linear solutions requires a proper initialization. While a brute-force search may be applicable, it is computationally expensive for real-time application, specially if the dimension of the parameter is high. Efficient and deterministic initialization methods are lacking and urgently needed. $(ii)$ The common assumption of constant velocity only holds within a short time interval. Nevertheless, a time window that is too narrow may contain insufficient number of events for a valid model fitting. The heuristic of using a constant number of events may violate the constant velocity assumption, especially in the case of sudden changes in speed.", + "bbox": [ + 212, + 253, + 787, + 526 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we rethink the problem formulation in a way that is more consistent with the differential working principle of event cameras. Unlike existing pipelines that use events occurring within a time interval to resolve the relative motion, we attempt to recover the structure and instantaneous first-order kinematic parameters involved in the motion field equation, using as input an instantaneous observation, namely event-based normal flow. To this end, we explore effective constraints between normal-flow observations and model parameters for solving a family of motion-and-structure estimation problems, such as the estimation of depth and full optical flow, instantaneous angular velocity and linear velocity, and differential homography (see Fig. 1).", + "bbox": [ + 212, + 529, + 787, + 680 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The contributions of this paper can be summarized as follows:", + "bbox": [ + 240, + 681, + 687, + 696 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- A normal flow constraint with a geometric connection to event data. It overcomes the partial observability issue (i.e., errors caused by using normal flow as replacement for full optical flow) in general problems of geometric model fitting when normal flow is used as input.", + "- Two solvers for a family of motion-and-structure estimation problems using as input sparse event-based normal flow. Our linear solver (can be used with RANSAC) leads to closed-form and deterministic solutions that can be used as an initialization to existing nonlinear methods. Our nonlinear solver," + ], + "bbox": [ + 225, + 715, + 784, + 837 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Z. Ren et al.", + "bbox": [ + 271, + 114, + 357, + 127 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/d1f0b4c57510e255c6b64fbfb4fe1cca27618059d35bbcd12fc8f37b680f0a43.jpg", + "image_caption": [ + "Fig. 1: Overview of our event-based motion estimation approach. Normal flow vectors are computed from the raw event data and are used as input to the two solvers. The result of the linear solver consists of fitted motion models $\\{\\pmb{\\theta}_1, \\pmb{\\theta}_2, \\dots\\}$ at discrete time instants, which can be used to initialize the continuous-time nonlinear solver." + ], + "image_footnote": [], + "bbox": [ + 233, + 143, + 767, + 292 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "under a continuous-time formulation, can handle sudden speed variations, and thus, is free of the constant-motion assumption.", + "bbox": [ + 240, + 388, + 782, + 417 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "- A thorough evaluation on the two proposed solvers, comparing against state-of-the-art methods, and also including an investigation on the performance of existing non-linear methods initialized using our linear solver.", + "bbox": [ + 223, + 417, + 782, + 463 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Outline: The rest of the paper is organized as follows. We first discuss the related work in Sec. 2. Then a compact review of preliminary knowledge is given in Sec. 3. Our approach is presented in Sec. 4, including the normal-flow constraint, the linear models of a family of motion-and-structure estimation problems, and the proposed linear and nonlinear solvers. Experimental evaluation is conducted in Sec. 5 and conclusion are drawn in Sec. 6.", + "bbox": [ + 212, + 472, + 787, + 564 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 585, + 387, + 602 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Event-based Geometric Model Fitting", + "text_level": 1, + "bbox": [ + 214, + 618, + 576, + 633 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The major body of literature on this specific topic follows an iterative scheme, which typically has its own special metric evaluating quantitatively the association between events, and solve the model fitting as a nonlinear optimization problem. Among the first works is the Contrast Maximization method [13, 14], which provides a unified framework for solving a family of motion-model fitting problems. It searches the optimal model parameters that ultimately lead to a sharp image of warped events (IWE). The goodness of fit is evaluated by the alignment of all involved events along point trajectories on the image plane, and such alignment is assessed by the strength of the contours of the IWE. The computational complexity of CMax is high because of iteratively updating of IWE, whose cost scales with the spatial resolution and the amount of event data.", + "bbox": [ + 212, + 643, + 787, + 809 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To lessen the dependency on IWE, Nunes et al. proposes an entropy minimization framework (EMin) [30] that replaces the dispersion metric of [14] with", + "bbox": [ + 214, + 809, + 785, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Motion and Structure from Event-based Normal Flow", + "bbox": [ + 372, + 114, + 732, + 127 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "a similarity measurement between events, enabling to evaluate the goodness of fit in a higher dimensional space. EMin is, still, too expensive for real-time application. Liu et al. [22] propose a two-step iterative method for the problem of 3-D rotation estimation using events as input. Given a rotation estimate, the first step determines event correspondences by looking for the nearest neighbour in the spatio-temporal domain. The second step employs a spatio-temporal registration which aligns corresponding events by minimizing the sum of squared geodesic residual. The least-square nature of the energy function in the second step leads to a computation friendly performance. Gu et al. [17] present a probabilistic approach for spatio-temporal alignment of events. The aligned events at a particular pixel are modeled as a spatio-temporal Poisson point process and the goodness of fit can be quantified using the likelihood of the data. The optimal model is retrieved when the likelihood is maximized. More recently, an event-to-map alignment scheme is proposed in [19], which adopts the time surface (TS) map as a reference for event alignment. The proposed TS loss function drives later events to get increasingly closer to pixel positions of corresponding earlier events, thus, leading to a progressively sharp TS map as iterations proceed.", + "bbox": [ + 212, + 146, + 787, + 404 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In summary, all the above-mentioned methods are nonlinear, and thus, a proper initialization is necessary. Although a brute-force search is widely adopted, sometimes it can be too expensive to locate the convergence basin in parameter space. Moreover, [14,30] may come across a convergence to false-positive minima, known as event collapse [36, 37]. Hence, a quick and deterministic solver (e.g., a closed-form solution) is desirable. Besides, most of these methods assume a constant-motion model, which limits their applicability to scenarios of sudden speed variations. In addition, we have noticed that a recent work [16] can jointly estimate the line parameters and the camera's linear velocities. However, this method is not a general one due to the dependence on line structures and the usage of an IMU. Thus, it is not within the scope of our discussion.", + "bbox": [ + 212, + 404, + 787, + 570 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Motion and Structure from Normal Flow", + "text_level": 1, + "bbox": [ + 215, + 597, + 602, + 612 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Due to the aperture problem, a local estimate of optical flow recovers only the partial component of the full flow vector along the direction of image gradient. Preliminary knowledge about normal flow can be found in the supplementary material. Recovering motion and structure from normal flow on intensity images is a well-studied topic. Its main body of literature typically proposes a qualitatively deterministic solution by which a set (even infinite) of applicable ego motions are determined via an analysis of normal flow classification [3, 11, 20]. Only recently, researches devised ways to formulate the Cheirality constraint as a differentiable term and use it in a continuous optimization approach [1] or as a loss function for training an artificial neural network [31].", + "bbox": [ + 212, + 627, + 787, + 779 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In event-based perception and navigation, similar methodologies as the frame-based ones are also witnessed, where normal flow computed from event data is used as input to determine the Focus of Expansion (FoE) for drones' course estimation [10] and obstacle avoidance [5].", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Z. Ren et al.", + "bbox": [ + 271, + 114, + 357, + 127 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Most of these heuristic methods, while effective, are cumbersome for practical application. In contrast, our method delivers a deterministic way to retrieve motion and structure from normal flow. Additionally, it is noticed that normal flow can replace full flow in some specific problems of geometric model fitting, such as affine models [26, 27]. In this way, the geometric model fitting problem can be easily solved in closed-form. Nevertheless, these methods do not consider the observability bias in normal flows, and thus, such naive substitution may lead to inaccurate model fitting. To this end, we present a robust linear solver based on a carefully-designed geometric measurement model that can overcome the partial observability of event-based normal flow. It is worth mentioning that a recent work [24] also determines instantaneous linear velocity using as input the normal flow. Our difference is seen in the scope of applicability. The usage of normal flow is restricted in the motion flow equation of [24], while our paper extends it to a wider scope (i.e., a family of geometric model fitting problems) by combining normal flow with e.g., differential epipolar geometry, differential homography, etc.", + "bbox": [ + 212, + 146, + 787, + 388 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Preliminaries", + "text_level": 1, + "bbox": [ + 215, + 409, + 380, + 425 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To make this paper self-contained, some preliminary knowledge is revisited. Due to space limitations, here we only list essential concepts. A more detailed review of preliminary knowledge can be found in the supplementary material.", + "bbox": [ + 212, + 439, + 787, + 486 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Motion Field. Consider a perspective camera moving in a static environment, with instantaneous angular and linear velocities in the camera frame $\\boldsymbol{\\omega} = (\\omega_{x},\\omega_{y},\\omega_{z})^{\\top}$ and $\\pmb {\\nu} = (\\nu_{x},\\nu_{y},\\nu_{z})^{\\top}$ , respectively. The motion field at pixel $\\mathbf{x} = (x,y)^{\\top}$ produced by projecting the 3D motion at point with depth $Z(\\mathbf{x})$ is", + "bbox": [ + 212, + 505, + 787, + 568 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {u} = \\frac {1}{Z (\\mathbf {x})} \\mathbf {A} (\\mathbf {x}) \\boldsymbol {\\nu} + \\mathbf {B} (\\mathbf {x}) \\boldsymbol {\\omega}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 403, + 575, + 785, + 608 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{A}(\\mathbf{x})$ and $\\mathbf{B}(\\mathbf{x})$ are $2\\times 3$ matrices that solely depend on $\\mathbf{x}$ . Equation (1) is the motion field equation [40, Ch.8] [23]. Optical flow in rigid scenes aims at estimating the motion field from visual data, which is only possible in regions that do not suffer from the aperture problem.", + "bbox": [ + 212, + 614, + 787, + 676 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Brightness constancy is the assumption stating that the true motion trajectories $\\mathbf{x}(t)$ followed by objects on the image plane are those at which brightness is constant: $I(\\mathbf{x}(t), t) = \\mathrm{const}$ . Differentiating in time (total derivative) gives the equation at $(\\mathbf{x}, t)$ : $\\nabla I \\cdot \\dot{\\mathbf{x}} + \\partial_t I = 0$ , where $\\dot{\\mathbf{x}} = d\\mathbf{x} / dt$ is the velocity of image point $\\mathbf{x}(t)$ . Given $I$ , this equation provides one constraint per point $(\\mathbf{x}, t)$ to estimate $\\dot{\\mathbf{x}}$ , now called optical flow. Without additional equations, only the normal component of the optical flow can be determined [40],", + "bbox": [ + 212, + 696, + 787, + 803 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {n} = - \\frac {\\partial_ {t} I}{\\| \\nabla I \\| ^ {2}} \\nabla I. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 439, + 811, + 785, + 844 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Motion and Structure from Event-based Normal Flow", + "bbox": [ + 372, + 114, + 732, + 127 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Normal Flow. Decomposing the optical flow $\\mathbf{u} = \\mathbf{u}_{\\perp} + \\mathbf{u}_{\\parallel}$ into its normal and parallel components to the local image gradient $(\\nabla I)$ , the normal flow is $\\mathbf{n} = \\mathbf{u}_{\\perp}$ , and the dot-product with $\\mathbf{n}$ gives", + "bbox": [ + 215, + 146, + 787, + 191 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {n} ^ {\\top} \\mathbf {u} = \\| \\mathbf {n} \\| ^ {2}, \\quad \\text {o r e q u i v a l e n t l y} \\quad (\\mathbf {u} - \\mathbf {n}) \\cdot \\mathbf {n} = 0. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 323, + 202, + 785, + 219 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Differential Homography. (Continuous Homography Matrix [25, Ch.5]). If all scene points lie on a plane, their coordinates are related to their image projections via a homography transformation (which encodes the information of the scene plane $\\{\\mathbf{N}, d\\}$ and the camera pose). In this case, the motion field (1) can be rewritten to adopt the form in homogeneous coordinates", + "bbox": [ + 214, + 241, + 782, + 316 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {u}} (\\mathbf {x}) = \\left(\\mathbf {1} - \\hat {\\mathbf {x}} \\mathbf {e} _ {3} ^ {\\top}\\right) \\mathbf {H} _ {d} \\hat {\\mathbf {x}}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 416, + 328, + 784, + 345 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathbf{H}_d\\dot{=} -\\left([ \\boldsymbol {\\omega}]_{\\times} + d^{-1}\\boldsymbol {\\nu}\\mathbf{N}^{\\top}\\right)$ denotes the differential homography matrix $([\\cdot ]_{\\times}$ refers to the cross-product, skew-symmetric matrix), $\\mathbf{1}$ is the identity matrix, $\\hat{\\mathbf{x}} = (x,y,1)^{\\top}$ , $\\hat{\\mathbf{u}} = (u_x,u_y,0)^{\\top}$ and $\\mathbf{e}_3 = (0,0,1)^{\\top}$ .", + "bbox": [ + 215, + 357, + 782, + 404 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Differential Epipolar Geometry. (Continuous Epipolar Geometry [25, Ch.5]). The epipolar geometry examines the distance between a potential match to the determined epipolar line, and can be expressed as $\\hat{\\mathbf{x}}^{\\prime \\top}\\mathbf{E}\\hat{\\mathbf{x}} = 0$ . The differential epipolar geometry can be derived analogously, providing a relationship between $\\mathbf{x},\\mathbf{u}$ and the camera velocities $\\pmb {\\nu},\\pmb{\\omega}$ . In homogeneous coordinates:", + "bbox": [ + 214, + 425, + 787, + 501 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {u}} (\\mathbf {x}) ^ {\\top} [ \\boldsymbol {\\nu} ] _ {\\times} \\hat {\\mathbf {x}} - \\hat {\\mathbf {x}} ^ {\\top} \\mathbf {s} \\hat {\\mathbf {x}} = 0, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 413, + 512, + 784, + 529 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathbf{s} = \\frac{1}{2} ([\\pmb {\\nu}]_{\\times}[\\pmb {\\omega}]_{\\times} + [\\pmb {\\omega}]_{\\times}[\\pmb {\\nu}]_{\\times})$ is a $3\\times 3$ symmetric matrix.", + "bbox": [ + 215, + 540, + 671, + 558 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4 Methodology", + "text_level": 1, + "bbox": [ + 215, + 579, + 380, + 597 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Figure 1 summarizes our approach, grounded upon event-based normal flow. In this section, we first disclose the normal-flow constraint that can overcome the partial observability of normal flow when replacing full optical flow with normal flow in general tasks of geometric model fitting (Sec. 4.1). By combining the constraint and the theory in Sec. 3, we derive linear equations for a family of motion-and-structure estimation problems (Sec. 4.2). To solve these problems, we present a linear solver and a continuous-time nonlinear solver (Sec. 4.3).", + "bbox": [ + 214, + 611, + 787, + 717 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1 Normal Flow Constraint", + "text_level": 1, + "bbox": [ + 215, + 739, + 465, + 753 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Normal flow can be computed easily from the event data, for example by fitting planes locally to the space-time \"cloud\" of events on image space [2, 32]. Specifically, assuming that the event data $e_k \\doteq \\{\\mathbf{x}_k, t_k\\}$ spans locally a surface in the spatio-temporal domain, the surface can be defined by a function $\\Sigma_e: \\mathbb{R}^2 \\to \\mathbb{R}; \\mathbf{x}_k \\mapsto t_k$ that maps the pixel coordinates to the latest event's", + "bbox": [ + 214, + 763, + 787, + 842 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Z. Ren et al.", + "bbox": [ + 271, + 114, + 357, + 127 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4b6f04504e5b1e1da6053b87c1bffd9164d3a1fd887b5441ec3b879c7da47058.jpg", + "image_caption": [ + "Fig. 2: Toy example of applying the proposed normal-flow constraint. (a): A 2D registration task using as input either optical flow or normal flow. The groundtruth displacement is defined by a global flow of $(1.732, -1)^{\\top}$ . (b)-(d): The loss landscape obtained using different geometric measurements in the registration task. The red dot denotes the resulting displacement, and the green cross the groundtruth displacement. The bottom-left regions in (b)-(d) display the corresponding registration results." + ], + "image_footnote": [], + "bbox": [ + 215, + 143, + 341, + 253 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2d17edd895e022ec3c07138772e36bd4fa5c35063e614e26fea68de5c2b3b612.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 341, + 143, + 477, + 253 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/be2ae15f15e7913c2af0fcdd3c71127563bfd3f69bee8e501cce550b212cf651.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 478, + 143, + 632, + 253 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/52aecd3ac01157809c0e7563adeadf90b1b1db4698056f539adf29834c99358c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 633, + 143, + 785, + 253 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "timestamp [2, 5]. The normal flow is parallel to the spatial derivative of $\\Sigma_{e}$ , namely $\\nabla \\Sigma_{e}(\\mathbf{x})$ . More specifically, it can be calculated by", + "bbox": [ + 214, + 377, + 782, + 409 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {n} (\\mathbf {x}) = \\frac {\\nabla \\Sigma_ {e} (\\mathbf {x})}{\\left\\| \\nabla \\Sigma_ {e} (\\mathbf {x}) \\right\\| ^ {2}}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 426, + 420, + 785, + 455 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Readers may refer to the supplementary material for a more detailed derivation. Surely there are other methods to extract event-based normal flow and the solvers we develop in this section are independent of them, i.e., our solvers can also be fed with the extracted normal flow data from such methods.", + "bbox": [ + 214, + 465, + 782, + 525 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Event-based normal flow cannot be used directly as a replacement for the full flow in (1) to solve for the motion parameters. However, we can leverage the relationship between both flows (3) to define an effective geometric error criterion that circumvents the partial observability of normal flow. If the flow $\\mathbf{u}$ depends on some unknown (\"state\") variables $\\pmb{\\theta}$ , we propose estimating them based on how well the normal flow constraint (3) is satisfied, e.g., by minimizing some function of the error terms", + "bbox": [ + 214, + 526, + 784, + 630 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {n} (\\mathbf {x}) ^ {\\top} \\mathbf {u} (\\mathbf {x}; \\boldsymbol {\\theta}) - \\| \\mathbf {n} (\\mathbf {x}) \\| ^ {2} \\doteq \\epsilon_ {\\mathrm {n f}} \\in \\mathbb {R}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 372, + 643, + 784, + 662 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To illustrate the validity of the proposed normal-flow constraint over the naive way of simply using normal flow to replace full optical flow, we use a 2D registration task as toy example. As shown in Fig. 2, the task consists of registering the solid-line triangle to the dashed-line one using as input either optical flow or normal flow. We observe that using the proposed normal-flow constraint as a geometric measurement leads to a registration result that is almost as accurate as the one given by the full optical flow. However, due to the partial observability of normal flow, simply using it as a replacement for full optical flow leads to a biased registration result, as shown in Fig. 2c. As we see in all following experiments, similar conclusions can be made in general problems of geometric model fitting.", + "bbox": [ + 214, + 672, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Motion and Structure from Event-based Normal Flow", + "bbox": [ + 372, + 114, + 732, + 127 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2 Models of Motion and Structure", + "text_level": 1, + "bbox": [ + 215, + 146, + 529, + 161 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Based on the derivation in Sec. 4.1, we establish linear models to a family of motion and structure estimation problems using the event-based normal flow as input. In particular, we first discuss how to recover optical flow and depth by introducing the normal-flow constraint into the differential epipolar geometry. Then we show that the combination of the normal-flow constraint with the motion flow equation (1) results in solutions to the problem of instantaneous angular velocity estimation, and 6-DoF motion estimation given depth as prior, respectively. Finally, we show how to compute the differential homography.", + "bbox": [ + 212, + 170, + 787, + 292 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Optical Flow. Given as input the accurate normal flow, and camera velocities, the full optical flow $\\mathbf{u}(\\mathbf{x})$ at pixel location $\\mathbf{x}$ can be calculated by solving the following linear system obtained from the combination of the normal-flow constraint (3) and the differential epipolar geometry (5):", + "bbox": [ + 214, + 313, + 787, + 376 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\binom {\\mathbf {n} (\\mathbf {x}) ^ {\\top}} {- \\hat {\\mathbf {x}} ^ {\\top} [ \\boldsymbol {\\nu} ] _ {\\times}} \\mathbf {u} (\\mathbf {x}) = \\binom {\\| \\mathbf {n} (\\mathbf {x}) \\| ^ {2}} {\\hat {\\mathbf {x}} ^ {\\top} \\mathbf {s} \\hat {\\mathbf {x}}}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 385, + 385, + 785, + 421 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Depth Estimation. Similarly, if the normal flow and camera velocities are known, the scene depth $Z(\\mathbf{x})$ can be determined by combining the normal-flow constraint (3) with the motion field eq. (1), yielding the closed-form solution:", + "bbox": [ + 214, + 440, + 787, + 487 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nZ (\\mathbf {x}) = \\frac {\\mathbf {n} (\\mathbf {x}) ^ {\\top} \\mathbf {A} (\\mathbf {x}) \\boldsymbol {\\nu}}{\\| \\mathbf {n} (\\mathbf {x}) \\| ^ {2} - \\mathbf {n} (\\mathbf {x}) ^ {\\top} \\mathbf {B} (\\mathbf {x}) \\boldsymbol {\\omega}}. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 383, + 497, + 785, + 532 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Angular Velocity Estimation. Assuming that an event camera undergoes a pure rotational motion, the optical flow $\\mathbf{u}$ in (1) can be simplified as", + "bbox": [ + 214, + 549, + 787, + 580 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {u} (\\mathbf {x}) = \\mathbf {B} (\\mathbf {x}) \\boldsymbol {\\omega}. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 447, + 592, + 785, + 608 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "By substituting (10) into (3), we have the linear equation in $\\omega$ :", + "bbox": [ + 215, + 618, + 674, + 633 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {n} (\\mathbf {x}) ^ {\\top} \\mathbf {B} (\\mathbf {x}) \\omega = \\| \\mathbf {n} (\\mathbf {x}) \\| ^ {2}. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 413, + 643, + 785, + 662 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Since each normal flow provides only one constraint, the minimal problem requires three normal flow vectors as input.", + "bbox": [ + 214, + 672, + 785, + 704 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6-DoF Motion Estimation. Assume the event camera undergoes a general 6-DoF motion with known depth of the observed scene, we can solve the first-order kinematics, namely instantaneous linear and angular velocities, by solving the following linear system obtained from the normal-flow constraint (3) and the motion-flow equation (1):", + "bbox": [ + 214, + 724, + 787, + 801 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {n} (\\mathbf {x}) ^ {\\top} \\mathbf {D} (\\mathbf {x}) \\left[ \\begin{array}{l} \\boldsymbol {\\nu} \\\\ \\boldsymbol {\\omega} \\end{array} \\right] = \\| \\mathbf {n} (\\mathbf {x}) \\| ^ {2}, \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 401, + 811, + 785, + 844 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Z. Ren et al.", + "bbox": [ + 271, + 114, + 357, + 127 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\mathbf{D}(\\mathbf{x}) = \\left[\\frac{1}{Z(\\mathbf{x})}\\mathbf{A}(\\mathbf{x})\\mathbf{B}(\\mathbf{x})\\right]$ is the $2\\times 6$ feature sensitivity matrix [4, 7]. The minimal problem in this context requires six normal flow measurements to determine the 6-DoF motion parameters.", + "bbox": [ + 212, + 143, + 782, + 196 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Motion and Structure Estimation in Planar Scenes. For a moving camera that observes a planar scene, the motion and structure parameters are compatibly encoded in a homography matrix. In our case, we attempt to recover the differential homography $\\mathbf{h}_d$ (the vector form of $\\mathbf{H}_d$ in (4)) because the instantaneous linear and angular velocities are considered. By expanding (4) and rearranging equations, we have", + "bbox": [ + 212, + 215, + 782, + 306 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {u} (\\mathbf {x}) = \\mathbf {C} (\\mathbf {x}) \\mathbf {h} _ {d}, \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 446, + 316, + 784, + 330 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\mathbf{C}(\\mathbf{x})$ is a $2 \\times 9$ matrix that solely depends on the pixel location $\\mathbf{x}$ . Subsequently, substituting (3) into (13), gives:", + "bbox": [ + 212, + 338, + 782, + 369 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {n} (\\mathbf {x}) ^ {\\top} \\mathbf {C} (\\mathbf {x}) \\mathbf {h} _ {d} = \\| \\mathbf {n} (\\mathbf {x}) \\| ^ {2}. \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 410, + 375, + 784, + 393 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "As discussed in [44], retrieving the motion and structure parameters from the decomposition of $\\mathbf{h}_d$ is not trivial and beyond the scope of this paper. Therefore, we only focus on the numerical solution of $\\mathbf{h}_d$ .", + "bbox": [ + 212, + 400, + 782, + 445 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.3 Our Solvers", + "text_level": 1, + "bbox": [ + 214, + 465, + 359, + 478 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "According to the motion-and-structure models derived in Sec. 4.2, we have established certain constraints between the model parameters and the normal flow observations. This allows us to simply develop linear solvers for each specific problem. Besides, we further propose a continuous-time nonlinear solver that can effectively handle the presence of sudden variations in motion and does not rely on the constant-motion assumption. To convey our solvers concisely, we generalize the motion-and-structure models into a unified form, as", + "bbox": [ + 212, + 486, + 784, + 592 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {n} (\\mathbf {x}) ^ {\\top} \\mathbf {O} (\\mathbf {x}) \\boldsymbol {\\theta} = \\| \\mathbf {n} (\\mathbf {x}) \\| ^ {2}, \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 415, + 599, + 784, + 616 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\mathbf{O}(\\mathbf{x})$ is the system matrix and $\\theta$ are the motion/structure parameters.", + "bbox": [ + 212, + 625, + 779, + 640 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Linear Solver. Given as input a set (e.g., K) of sparse normal-flow estimates obtained at the same time, our linear solver works out the following (overdetermined) linear system of equations:", + "bbox": [ + 212, + 657, + 782, + 704 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\left( \\begin{array}{c} \\mathbf {n} \\left(\\mathbf {x} _ {1}\\right) ^ {\\top} \\mathbf {O} \\left(\\mathbf {x} _ {1}\\right) \\\\ \\vdots \\\\ \\mathbf {n} \\left(\\mathbf {x} _ {\\mathrm {K}}\\right) ^ {\\top} \\mathbf {O} \\left(\\mathbf {x} _ {\\mathrm {K}}\\right) \\end{array} \\right) \\boldsymbol {\\theta} = \\left( \\begin{array}{c} \\| \\mathbf {n} \\left(\\mathbf {x} _ {1}\\right) \\| ^ {2} \\\\ \\vdots \\\\ \\| \\mathbf {n} \\left(\\mathbf {x} _ {\\mathrm {K}}\\right) \\| ^ {2} \\end{array} \\right). \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 370, + 710, + 784, + 768 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "For robust estimation, RANSAC [12] is adopted to deal with noise and outliers.", + "bbox": [ + 212, + 773, + 782, + 790 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Motion and Structure from Event-based Normal Flow", + "bbox": [ + 370, + 114, + 732, + 127 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_footnote", + "text": "$^{5}$ $\\mathbf{O}(\\mathbf{x})$ degenerates to 1 in the problem of optical flow estimation.", + "bbox": [ + 217, + 796, + 658, + 811 + ], + "page_idx": 8 + }, + { + "type": "page_footnote", + "text": "An approximation can be made based on the assumption of constant camera motion. In this way, the normal flow estimates can be collected within a short time interval.", + "bbox": [ + 220, + 811, + 782, + 839 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Continuous-Time Nonlinear Solver. To handle the case of a sudden variation in speed that may violate the constant-motion assumption, we further propose a continuous-time nonlinear solver. In particular, we employ cubic B-splines to represent the continuous-time model parameters as", + "bbox": [ + 212, + 146, + 782, + 205 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\theta} (u (t)) = \\sum_ {i = 0} ^ {3} B _ {i, 3} (u (t)) \\mathbf {c} _ {i}, \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 406, + 219, + 785, + 258 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "where $B_{i,k}$ denotes the basis function, $i$ the index of control points, $k (= 3)$ the order of the spline and $\\mathbf{c}_i \\in \\mathbb{R}^N$ the corresponding control point defined in the $N$ -dimensional space of model parameters. Examples of cubic B-splines applied to event camera motion estimation are found in [18, 28]. The function $u(\\cdot)$ is a normalization operator, which transfers time $t$ to the spline's parameter domain by means of basis translation [8]. In this way, the model parameters at any time instant $t$ can be obtained via an interpolation using the known control points.", + "bbox": [ + 212, + 268, + 784, + 375 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "The formulation of the continuous-time model fitting problem goes as follows. The unknown variable consists of all the involved control points, $\\mathcal{X} \\doteq [\\mathbf{c}_0, \\mathbf{c}_1, \\dots, \\mathbf{c}_n]$ , where $n$ denotes the total number of control points. By introducing the continuous-time representation of model parameters to (15), we finally create the following objective function:", + "bbox": [ + 212, + 376, + 784, + 450 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\mathcal {X}} ^ {*} = \\arg \\min _ {\\boldsymbol {\\mathcal {X}}} \\sum_ {e _ {k} \\in \\mathcal {E}} \\left\\| \\mathbf {n} \\left(\\mathbf {x} _ {k}\\right) ^ {\\top} \\mathbf {O} \\left(\\mathbf {x} _ {k}\\right) \\boldsymbol {\\theta} \\left(u \\left(t _ {k}\\right)\\right) - \\| \\mathbf {n} \\left(\\mathbf {x} _ {k}\\right) \\| ^ {2} \\right\\| ^ {2}, \\tag {18}\n$$\n", + "text_format": "latex", + "bbox": [ + 297, + 467, + 785, + 500 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "where $\\mathcal{E}$ denotes the event set in which the norm flow observation is available for every involved event $e_k$ . Note that this continuous-time formulation aligns better with the asynchronous nature of event-based normal flow, enabling a state estimation over a larger temporal window that may lead to a better signal-to-noise ratio. The linear solver can be used to initialize the continuous-time nonlinear solver. An illustrative explanation of our solvers is given in Fig. 1.", + "bbox": [ + 212, + 508, + 782, + 599 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5 Experiments", + "text_level": 1, + "bbox": [ + 215, + 623, + 375, + 641 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "To justify the effectiveness of the proposed algorithms, experiments are conducted on both simulation and real data. First, we introduce the datasets used and the applied metrics for evaluation (Sec. 5.1). Then we report the results of our linear solver and continuous-time nonlinear solver (Sec. 5.2 and Sec. 5.3). Finally, a discussion on the computational complexity and the limitation of our method is provided in Sec. 5.4.", + "bbox": [ + 212, + 655, + 784, + 746 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5.1 Experimental Setup", + "text_level": 1, + "bbox": [ + 215, + 768, + 426, + 784 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Our evaluation utilizes four datasets, including the indoor event-camera dataset (ECD) captured with a DAVIS240C camera and motion-capture system for groundtruth, and a high-resolution dataset with two rotational motion sequences", + "bbox": [ + 212, + 794, + 782, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Z. Ren et al.", + "bbox": [ + 271, + 114, + 357, + 127 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/6d5184372b1a0e4636446f974281d9e865262d9d32e175502a23e5068bfac67b.jpg", + "image_caption": [ + "(a) Estimated optical flow over a map of event accumulation." + ], + "image_footnote": [], + "bbox": [ + 217, + 145, + 415, + 248 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/b484b8710ef572824360f0a0292a6cea7e4fcfc3d925bc5dcd9f1dbedecd369b.jpg", + "image_caption": [ + "(b) Reconstruction result displayed over groundtruth structures (a v-shape roof) and viewed from different perspectives." + ], + "image_footnote": [], + "bbox": [ + 423, + 143, + 602, + 246 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/dde771b4898659557cdb422227caec1f8c8fce581a2f70c55816fd436d16d6d3.jpg", + "image_caption": [ + "Fig. 3: Qualitative result of optical flow and depth on sequence three_wall Translation of our synthetic data. Note that our method returns sparse results." + ], + "image_footnote": [], + "bbox": [ + 604, + 143, + 784, + 247 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "using an iniVation DAVIS346 camera and Xsens IMU. The VECTor dataset, captured with a Prophesee Gen 3 camera and an accompanying RGB-D depth camera, is used for 6-DoF motion tracking. Additionally, a synthetic dataset generated in Blender is employed for motion-and-structure pattern analysis. Evaluation metrics for kinematics estimation include average error (AE) and root mean squared error (RMSE), while differential homography estimation is assessed using the Frobenius norm. Best results are in bold, and second-best are underlined. The supplementary material provides detailed descriptions of datasets and metrics, as well as implementation details of our methods.", + "bbox": [ + 212, + 335, + 787, + 472 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "5.2 Evaluation of the Linear Solver", + "text_level": 1, + "bbox": [ + 214, + 491, + 519, + 505 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "In the following evaluation, we qualitatively and quantitatively assess the performance of our linear method (Ours) in each specific task, and compare against different state-of-the-art (SOTA) methods (CMax [14] and Pro-STR [19]) and their customized version initialized using our linear solver.", + "bbox": [ + 212, + 512, + 785, + 571 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Optical Flow and Depth Estimation. Due to a lack of benchmark for extensive evaluation, we qualitatively evaluate our method on the tasks of optical flow and depth estimation. As illustrated in Fig. 3a, our linear solver demonstrates the capability to accurately reconstruct full flow using normal flow and known motion parameters. An example of the depth estimation result is given in Fig. 3b, showing that the recovered structure aligns well with the groundtruth.", + "bbox": [ + 212, + 590, + 787, + 681 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Angular Velocity Estimation. As shown in Tab. 1, we see that our linear solver outperforms CMax and Pro-STR. Besides, we find the combination of our linear solver and CMax leads to the best result, indicating that the proposed linear solver is a complementary method to existing nonlinear methods.", + "bbox": [ + 212, + 700, + 787, + 761 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Differential Homography Estimation. The result is consistent with that in the above task. Table 2 reveals that our linear method outperforms CMax and Pro-STR in terms of numerical accuracy. The combination of our linear solver and state-of-the-art nonlinear solver leads to the best result.", + "bbox": [ + 212, + 779, + 787, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Motion and Structure from Event-based Normal Flow", + "bbox": [ + 372, + 114, + 730, + 127 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/8b7702183272a27efcd7157645f94b84df46160219da6cde800b334df3814271.jpg", + "table_caption": [ + "Table 1: Evaluation of our linear solver on the task of angular velocity estimation." + ], + "table_footnote": [], + "table_body": "
MethodshapesRotation [29]groundRotation (our data)patternsRotation (our data)
eω(deg/s)RMSEω(deg/s)eω(deg/s)RMSEω(deg/s)eω(deg/s)RMSEω(deg/s)
CMax [14]15.6026.389.4815.585.959.07
Pro-STR [19]90.61145.8131.0744.7410.4711.88
Ours15.3623.983.308.014.706.08
Ours+CMax [14]10.1214.292.853.740.350.73
Ours+Pro-STR [19]19.2340.449.1317.074.315.23
MethoddynamicRotation [29]boxesRotation (our data)cubesRotation (our data)
eω(deg/s)RMSEω(deg/s)eω(deg/s)RMSEω(deg/s)eω(deg/s)RMSEω(deg/s)
CMax [14]16.1426.488.1813.566.7710.62
Pro-STR [19]54.6485.6322.7232.0511.0712.79
Ours9.4118.062.604.782.313.02
Ours+CMax [14]7.7110.852.643.800.490.77
Ours+Pro-STR [19]18.8732.653.495.872.042.76
", + "bbox": [ + 228, + 193, + 782, + 380 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/298324b7af0d69de2ce3e06b90bfbacbba88d7ee4dfebf6d38a6a0c4e312b05b.jpg", + "table_caption": [ + "Table 2: Evaluation of our linear solver for differential homography estimation." + ], + "table_footnote": [], + "table_body": "
Methodboard Slow [15]patterns rotationpatterns 6 dof
F-NormF-NormF-Norm
CMax [14]1.161.331.27
Pro-STR [19]2.071.221.37
Ours0.240.200.43
Ours+CMax [14]0.230.150.41
Ours+Pro-STR [19]0.190.170.42
", + "bbox": [ + 295, + 459, + 710, + 545 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/c1f86fcb6b160ce345ee4f36711bd686c2c77260d9a5055dbc7d1967b9d4ff75.jpg", + "table_caption": [ + "Table 3: Evaluation of our linear solver on the task of 6-DoF motion tracking." + ], + "table_footnote": [], + "table_body": "
Methodcorner Slow [15]patterns _6 dof (our data)
eω(deg/s)RMSEω(deg/s)ev(m/s)RMSEv(m/s)eω(deg/s)RMSEω(deg/s)ev(m/s)RMSEv(m/s)
CMax [14]12.3914.390.340.4112.9415.681.021.23
Pro-STR [19]5.197.840.130.2211.2113.491.331.49
Ours1.311.790.010.020.560.890.120.16
Ours+CMax [14]2.192.820.110.170.550.870.180.30
Ours+Pro-STR [19]1.061.340.040.051.071.560.120.15
Methodmountain_normal [15]cubes _6 dof (our data)
eω(deg/s)RMSEω(deg/s)ev(m/s)RMSEv(m/s)eω(deg/s)RMSEω(deg/s)ev(m/s)RMSEv(m/s)
CMax [14]11.0114.170.440.5210.7414.330.781.05
Pro-STR [19]5.2310.530.140.259.7511.401.291.44
Ours1.281.830.070.090.791.020.110.14
Ours+CMax [14]3.846.300.210.301.141.720.120.14
Ours+Pro-STR [19]1.181.580.060.091.051.330.110.14
", + "bbox": [ + 228, + 625, + 782, + 816 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Z. Ren et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/ed1bf464dcf21f8e30c6fc27c6c4a7026b849de293306d384b183408bec5592c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 238, + 146, + 763, + 287 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/9dbb303a4ba45eb7fccc2c30a6525065060729a1d1244f7888c7ce80fb08dfcd.jpg", + "image_caption": [ + "(a) Results of the continuous-time angular velocity estimator." + ], + "image_footnote": [], + "bbox": [ + 243, + 301, + 759, + 335 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/80db2d36fcde1c23580460abe1e6b7dc7d0034b32ea1f45813e038fc04168c64.jpg", + "image_caption": [ + "(b) Comparison against two state-of-the-art nonlinear solutions using data segments that involve sudden variations of angular velocity.", + "Fig. 4: Angular velocity estimation using the continuous-time nonlinear solver (Ours)." + ], + "image_footnote": [], + "bbox": [ + 243, + 337, + 761, + 435 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "6-DoF Motion Tracking. Given known depth, we recover the linear and angular camera velocities on the task of 6-DoF motion estimation. As shown in Tab. 3, our linear solver reports more accurate results than both CMax and Pro-STR. In addition, we surprisingly find that the linear solver can sometimes outperform the combination of itself with the two nonlinear solvers. This is, according to our investigation, attributed to the violation of constant-motion assumption in some sequences, which is held for the two nonlinear solvers.", + "bbox": [ + 212, + 518, + 787, + 625 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5.3 Evaluation of the Continuous-Time Solver", + "text_level": 1, + "bbox": [ + 215, + 647, + 609, + 662 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "To justify the claimed advantage of the proposed continuous-time nonlinear solver, we use the exemplary task of angular velocity estimation. The sequence used (shapesRotation [29]) captures a purely rotating event camera, and some parts of the data involves aggressive motion. As illustrated in Fig. 4a, our nonlinear solver gives accurate estimation results with a high stability. Moreover, we select three segments that witness a sudden variation of angular speed within a short time interval and compare our continuous-time nonlinear solver against CMax and Pro-STR. Note that all of them are initialized by our linear method for a fair comparison. As Fig. 4b shows, our results are more consistent with the groundtruth angular velocity than the other two methods, which are less accurate due to the dependence on the constant-motion assumption.", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Motion and Structure from Event-based Normal Flow", + "bbox": [ + 372, + 114, + 730, + 127 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/9656dabb1e96a819276ad7d926f7181243e5a58113f2f83a3e7e730899958425.jpg", + "table_caption": [ + "Table 4: Runtime of 6-DoF tracking on sequence mountain_normal (640×480 px)." + ], + "table_footnote": [], + "table_body": "
ComponentRuntime
Normal flow extraction20 ms (for extracting 2000 normal flow vectors)
Linear solver25.97 ms
CT solver28.29 ms
", + "bbox": [ + 282, + 171, + 723, + 224 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5.4 Discussion", + "text_level": 1, + "bbox": [ + 215, + 250, + 349, + 263 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Computational Complexity. The computational complexity of the linear solver depends on the specific task of model fitting. For optical flow and depth estimation, it is linearly proportional to the size of the input event flow, and thus, the complexity is $\\mathcal{O}(N)$ . For the other tasks, the time cost predominantly hinges on the total number of RANSAC iterations and the computations performed per iteration. In each iteration, we typically solve a linear overdetermined system via QR decomposition at a complexity of $\\mathcal{O}(cN^2)$ , where $c$ is the cardinality of a minimal problem (e.g., $c = 3$ for the problem of angular velocity estimation).", + "bbox": [ + 212, + 271, + 787, + 393 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Runtime. All experiments are conducted on a standard computer with Intel i7-13700F CPU, 16G memory and implemented with MATLAB (version R2022a). Using the 6-DoF tracking task as an example, we measure the runtime of the proposed linear solver and continuous-time (CT) solver, as reported in Tab. 4.", + "bbox": [ + 212, + 410, + 785, + 472 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Limitations. To make it clear, we emphasize that our method, like other compared methods, does not resolve the motion and structure parameters simultaneously. It should also be noted that the success of our method relies heavily on the quality of the input normal flow. Particularly, our normal flow extractor (including some other open source solutions for computing normal flow from event data) may fail in the presence of densely repetitive textures, leading to a failure of the modeling fitting scheme. To assess the sensitivity of our method to the quality of event-based normal flow, we propose an analysis on numerical stability using simulation data in the supplementary material.", + "bbox": [ + 212, + 489, + 787, + 626 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusions", + "text_level": 1, + "bbox": [ + 215, + 645, + 370, + 662 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "This work presents a solution to the general problem of geometric model fitting with event data. Considering the differential working principle of event cameras, our method establishes specific constraints between instantaneous motion-and-structure parameters and event-based normal flow. Experiments show that the proposed linear solver leads to closed-form and deterministic solutions, which can be used to initialize existing state-of-the-art nonlinear methods. The proposed continuous-time nonlinear solver, furthermore, demonstrates the capability to handle sudden speed variations, obviating the dependence on the constant-motion assumption. Finally, we hope this work can inspire future research on event-based geometric model fitting, a paramount topic to many downstream techniques, such as event-based 3-D vision, SLAM, and scene segmentation.", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Z. Ren et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 215, + 143, + 392, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "We thank Dr. Yi Yu for proof reading. This work was supported by the National Key Research and Development Project of China under Grant 2023YFB4706600. Funded by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) under Germany's Excellence Strategy - EXC 2002/1 \"Science of Intelligence\" - project number 390523135.", + "bbox": [ + 212, + 180, + 787, + 256 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 282, + 321, + 299 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Barranco, F., Fermüller, C., Aloimonos, Y., Ros, E.: Joint direct estimation of 3D geometry and 3D motion using spatio temporal gradients. Pattern Recog. 113, 107759 (2021). https://doi.org/10.1016/j.patcog.2020.107759", + "2. Benosman, R., Clercq, C., Lagorce, X., Ieng, S.H., Bartolozzi, C.: Event-based visual flow. IEEE Trans. Neural Netw. Learn. Syst. 25(2), 407-417 (2014). https://doi.org/10.1109/TNNLS.2013.2273537", + "3. Brodsky, T., Fermüller, C., Aloimonos, Y.: Structure from motion: Beyond the epipolar constraint. Int. J. Comput. Vis. 37, 231-258 (2000)", + "4. Bryner, S., Gallego, G., Rebecq, H., Scaramuzza, D.: Event-based, direct camera tracking from a photometric 3D map using nonlinear optimization. In: IEEE Int. Conf. Robot. Autom. (ICRA) (2019). https://doi.org/10.1109/ICRA.2019.8794255", + "5. Clady, X., Clercq, C., Ieng, S.H., Houseini, F., Randazzo, M., Natale, L., Bartolozzi, C., Benosman, R.: Asynchronous visual event-based time-to-contact. Front. Neurosci. 8(9) (2014). https://doi.org/10.3389/fnins.2014.00009", + "6. Conradt, J., Cook, M., Berner, R., Lichtsteiner, P., Douglas, R.J., Delbruck, T.: A pencil balancing robot using a pair of AER dynamic vision sensors. In: IEEE Int. Symp. Circuits Syst. (ISCAS). pp. 781-784 (2009). https://doi.org/10.1109/ISCAS.2009.5117867", + "7. Corke, P.: Robotics, Vision and Control: Fundamental Algorithms in MATLAB. Springer Tracts in Advanced Robotics, Springer (2017). https://doi.org/10.1007/978-3-319-54413-7", + "8. De Boor, C.: On calculating with b-splines. Journal of Approximation theory 6(1), 50-62 (1972)", + "9. Delbruck, T., Lang, M.: Robotic goalie with 3ms reaction time at $4\\%$ CPU load using event-based dynamic vision sensor. Front. Neurosci. 7, 223 (2013). https://doi.org/10.3389/fnins.2013.00223", + "0. Dinaux, R., Wessendorp, N., Dupeyroux, J., De Croon, G.C.: Faith: Fast iterative half-plane focus of expansion estimation using optic flow. IEEE Robotics and Automation Letters 6(4), 7627-7634 (2021)", + "1. Fermüller, C.: Passive navigation as a pattern recognition problem. Int. J. Comput. Vis. 14(2), 147-158 (1995)", + "12. Fischler, M.A., Bolles, R.C.: Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Commun. ACM 24(6), 381-395 (1981). https://doi.org/10.1145/358669.358692", + "13. Gallego, G., Gehrig, M., Scaramuzza, D.: Focus is all you need: Loss functions for event-based vision. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR). pp. 12272-12281 (2019). https://doi.org/10.1109/CVPR.2019.01256" + ], + "bbox": [ + 218, + 316, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Motion and Structure from Event-based Normal Flow", + "bbox": [ + 370, + 114, + 730, + 127 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "14. Gallego, G., Rebecq, H., Scaramuzza, D.: A unifying contrast maximization framework for event cameras, with applications to motion, depth, and optical flow estimation. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR). pp. 3867-3876 (2018). https://doi.org/10.1109/CVPR.2018.00407", + "15. Gao, L., Liang, Y., Yang, J., Wu, S., Wang, C., Chen, J., Kneip, L.: Vector: A versatile event-centric benchmark for multi-sensor slam. IEEE Robotics and Automation Letters 7(3), 8217-8224 (2022). https://doi.org/10.1109/LRA.2022.3186770", + "16. Gao, L., Su, H., Gehrig, D., Cannici, M., Scaramuzza, D., Kneip, L.: A 5-point minimal solver for event camera relative motion estimation. In: Int. Conf. Comput. Vis. (ICCV). pp. 8015-8025 (2023). https://doi.org/10.1109/ICCV51070.2023.00739", + "17. Gu, C., Learned-Miller, E., Sheldon, D., Gallego, G., Bideau, P.: The spatiotemporal Poisson point process: A simple model for the alignment of event camera data. In: Int. Conf. Comput. Vis. (ICCV). pp. 13495-13504 (2021). https://doi.org/10.1109/ICCV48922.2021.01324", + "18. Guo, S., Gallego, G.: CMax-SLAM: Event-based rotational-motion bundle adjustment and SLAM system using contrast maximization. IEEE Trans. Robot. pp. 1-20 (2024). https://doi.org/10.1109/TRO.2024.3378443", + "19. Huang, X., Zhang, Y., Xiong, Z.: Progressive spatio-temporal alignment for efficient event-based motion estimation. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR). pp. 1537-1546 (2023)", + "20. Ji, H., Fermuller, C.: A 3D shape constraint on video. IEEE Trans. Pattern Anal. Mach. Intell. 28(6), 1018-1023 (2006)", + "21. Kim, H., Leutenegger, S., Davison, A.J.: Real-time 3D reconstruction and 6-DoF tracking with an event camera. In: Eur. Conf. Comput. Vis. (ECCV). pp. 349-364 (2016). https://doi.org/\\protect{10.1007/978-3-319-46466-4_21}", + "22. Liu, D., Parra, A., Chin, T.J.: Spatiotemporal registration for event-based visual odometry. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR). pp. 4935-4944 (2021). https://doi.org/10.1109/CVPR46437.2021.00490", + "23. Longuet-Higgins, H.C., Prazdny, K.: The interpretation of a moving retinal image. Proceedings of the Royal Society of London. Series B. Biological Sciences 208(1173), 385-397 (1980). https://doi.org/10.1098/rspb.1980.0057", + "24. Lu, X., Zhou, Y., Niu, J., Zhong, S., Shen, S.: Event-based visual inertial velometer (2024)", + "25. Ma, Y., Soatto, S., Košecká, J., Sastry, S.: An Invitation to 3-D Vision: From Images to Geometric Models. Springer (2004)", + "26. Meyer, F., Bouthemy, P.: Estimation of time-to-collision maps from first order motion models and normal flows. In: 1992 11th IAPR International Conference on Pattern Recognition. vol. 1, pp. 78-82. IEEE Computer Society (1992)", + "27. Meyer, F.G.: Time-to-collision from first-order models of the motion field. IEEE transactions on robotics and automation 10(6), 792-798 (1994)", + "28. Mueggler, E., Gallego, G., Rebecq, H., Scaramuzza, D.: Continuous-time visual-inertial odometry for event cameras. IEEE Trans. Robot. 34(6), 1425-1440 (Dec 2018). https://doi.org/10.1109/tro.2018.2858287", + "29. Mueggler, E., Rebecq, H., Gallego, G., Delbruck, T., Scaramuzza, D.: The event-camera dataset and simulator: Event-based data for pose estimation, visual odometry, and SLAM. Int. J. Robot. Research 36(2), 142-149 (2017). https://doi.org/10.1177/0278364917691115", + "30. Nunes, U.M., Demiris, Y.: Entropy minimisation framework for event-based vision model estimation. In: Eur. Conf. Comput. Vis. (ECCV). pp. 161-176 (2020). https://doi.org/10.1007/978-3-030-58558-7_10" + ], + "bbox": [ + 217, + 146, + 785, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Z. Ren et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "31. Parameshwara, C.M., Hari, G., Fermüller, C., Sanket, N.J., Aloimonos, Y.: Diff-posednet: direct differentiable camera pose estimation. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR). pp. 6845-6854 (2022)", + "32. Pijnacker Hordijk, B.J., Scheper, K.Y., De Croon, G.C.: Vertical landing for micro air vehicles using event-based optical flow. J. Field Robot. 35(1), 69-90 (2018)", + "33. Rebecq, H., Gallego, G., Scaramuzza, D.: EMVS: Event-based multi-view stereo. In: British Mach. Vis. Conf. (BMVC) (2016). https://doi.org/10.5244/C.30.63", + "34. Rebecq, H., Horstschaefer, T., Scaramuzza, D.: Real-time visual-inertial odometry for event cameras using keyframe-based nonlinear optimization. In: British Mach. Vis. Conf. (BMVC) (2017)", + "35. Rebecq, H., Horstschäfer, T., Gallego, G., Scaramuzza, D.: EVO: A geometric approach to event-based 6-DOF parallel tracking and mapping in real-time. IEEE Robot. Autom. Lett. 2(2), 593-600 (2017). https://doi.org/10.1109/LRA.2016.2645143", + "36. Shiba, S., Aoki, Y., Gallego, G.: Event collapse in contrast maximization frameworks. Sensors 22(14), 1-20 (2022). https://doi.org/10.3390/s22145190", + "37. Shiba, S., Aoki, Y., Gallego, G.: A fast geometric regularizer to mitigate event collapse in the contrast maximization framework. Adv. Intell. Syst. p. 2200251 (2022). https://doi.org/10.1002/aisy.202200251", + "38. Shiba, S., Aoki, Y., Gallego, G.: Secrets of event-based optical flow. In: Eur. Conf. Comput. Vis. (ECCV). pp. 628-645 (2022). https://doi.org/10.1007/978-3-031-19797-0 36", + "39. Stoffregen, T., Gallego, G., Drummond, T., Kleeman, L., Scaramuzza, D.: Event-based motion segmentation by motion compensation. In: Int. Conf. Comput. Vis. (ICCV). pp. 7243-7252 (2019). https://doi.org/10.1109/ICCV.2019.00734", + "40. Trucco, E., Verri, A.: Introductory Techniques for 3-D Computer Vision. Prentice Hall PTR, Upper Saddle River, NJ, USA (1998)", + "41. Zhou, Y., Gallego, G., Rebecq, H., Kneip, L., Li, H., Scaramuzza, D.: Semidense 3D reconstruction with a stereo event camera. In: Eur. Conf. Comput. Vis. (ECCV). pp. 242-258 (2018). https://doi.org/10.1007/978-3-030-01246-5_15", + "42. Zhou, Y., Gallego, G., Shen, S.: Event-based stereo visual odometry. IEEE Trans. Robot. 37(5), 1433-1450 (2021). https://doi.org/10.1109/TRO.2021.3062252", + "43. Zhu, A.Z., Chen, Y., Daniilidis, K.: Realtime time synchronized event-based stereo. In: Eur. Conf. Comput. Vis. (ECCV). pp. 438-452 (2018). https://doi.org/10.1007/978-3-030-01231-1_27", + "44. Zhuang, B., Tran, Q.H.: Image stitching and rectification for hand-held cameras. In: Eur. Conf. Comput. Vis. (ECCV) (2020). https://doi.org/10.1007/978-3-030-58571-6_15" + ], + "bbox": [ + 212, + 146, + 787, + 674 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Motion and Structure from Event-based Normal Flow", + "bbox": [ + 372, + 114, + 730, + 127 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/Motion and Structure from Event-based Normal Flow/2dddd958-d905-447a-8afe-98428ea7b185_model.json b/2024/Motion and Structure from Event-based Normal Flow/2dddd958-d905-447a-8afe-98428ea7b185_model.json new file mode 100644 index 0000000000000000000000000000000000000000..05e0a12a9ada5f590c317a4c23a0a0d54758f7d9 --- /dev/null +++ b/2024/Motion and Structure from Event-based Normal Flow/2dddd958-d905-447a-8afe-98428ea7b185_model.json @@ -0,0 +1,2445 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.272, + 0.141, + 0.731, + 0.184 + ], + "angle": 0, + "content": "Motion and Structure from Event-based Normal Flow" + }, + { + "type": "text", + "bbox": [ + 0.236, + 0.212, + 0.767, + 0.244 + ], + "angle": 0, + "content": "Zhongyang Ren\\(^{1*}\\), Bangyan Liao\\(^{2*}\\), Delei Kong\\(^{1}\\), Jinghang Li\\(^{1}\\), Peidong Liu\\(^{2}\\), Laurent Kneip\\(^{3}\\), Guillermo Gallego\\(^{4}\\), and Yi Zhou\\(^{1*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.368, + 0.255, + 0.636, + 0.269 + ], + "angle": 0, + "content": "\\(^{1}\\) School of Robotics, Hunan University" + }, + { + "type": "text", + "bbox": [ + 0.35, + 0.27, + 0.652, + 0.284 + ], + "angle": 0, + "content": "\\(^{2}\\) School of Engineering, Westlake University" + }, + { + "type": "text", + "bbox": [ + 0.255, + 0.283, + 0.748, + 0.298 + ], + "angle": 0, + "content": "3 School of Information Science and Technology, ShanghaiTech University" + }, + { + "type": "text", + "bbox": [ + 0.239, + 0.297, + 0.765, + 0.311 + ], + "angle": 0, + "content": "TU Berlin, ECDF, SCIoI Excellence Cluster and Robotics Institute Germany" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.348, + 0.744, + 0.681 + ], + "angle": 0, + "content": "Abstract. Recovering the camera motion and scene geometry from visual data is a fundamental problem in computer vision. Its success in conventional (frame-based) vision is attributed to the maturity of feature extraction, data association and multi-view geometry. The emergence of asynchronous (event-based) cameras calls for new approaches that use raw event data as input to solve this fundamental problem. State-of-the-art solutions typically infer data association implicitly by iteratively reversing the event data generation process. However, the nonlinear nature of these methods limits their applicability in real-time tasks, and the constant-motion assumption leads to unstable results under agile motion. To this end, we reformulate the problem in a way that aligns better with the differential working principle of event cameras. We show that event-based normal flow can be used, via the proposed geometric error term, as an alternative to the full (optical) flow in solving a family of geometric problems that involve instantaneous first-order kinematics and scene geometry. Furthermore, we develop a fast linear solver and a continuous-time nonlinear solver on top of the proposed geometric error term. Experiments on both synthetic and real data show the superiority of our linear solver in terms of accuracy and efficiency, and its practicality as an initializer for previous nonlinear solvers. Besides, our continuous-time non-linear solver exhibits exceptional capabilities in accommodating sudden variations in motion since it does not rely on the constant-motion assumption. Our project page can be found at https://nail-hnu.github.io/EvLinearSolver/." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.694, + 0.719, + 0.708 + ], + "angle": 0, + "content": "Keywords: Event cameras \\(\\cdot\\) Geometric model fitting \\(\\cdot\\) Normal flow" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.756, + 0.377, + 0.772 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.787, + 0.788, + 0.817 + ], + "angle": 0, + "content": "Neuromorphic event-based cameras are bio-inspired sensors that report only brightness changes at each pixel location asynchronously. Their high temporal" + }, + { + "type": "page_footnote", + "bbox": [ + 0.232, + 0.826, + 0.694, + 0.841 + ], + "angle": 0, + "content": "* equal contribution; \\( \\boxtimes \\) corresponding author (eeyzhou@hnu.edu.cn)." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.128 + ], + "angle": 0, + "content": "Z. Ren et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.253 + ], + "angle": 0, + "content": "resolution enables them to sense at exactly the same rate as the scene dynamics, and thus, event cameras are suitable for dealing with robotic perception [33,38, 39,41,43], navigation [21,34,35,42], and control [6,9] tasks that involve agile motion. A general and fundamental problem among these tasks is to recover the camera's ego-motion and scene geometry from event data. One of the bottlenecks for solving this kind of problems at very high speeds is the lack of methods for efficient and robust data association." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.255, + 0.788, + 0.527 + ], + "angle": 0, + "content": "Unlike its standard-vision counterparts, an event camera cannot observe visual information (i.e. intensity changes) when motion is parallel to edges. Such a partial observability of event data makes optical flow estimation (as well as feature detection and matching) an ill-posed problem. To this end, many existing methods [14,17,19,22,30] for solving the problem of motion-model fitting on event data propose specific metrics that assess event-data association established in either an implicit manner or an explicitly model-driven style. The best fit can be retrieved by iteratively optimizing the data association, often expressed in terms of event alignment. However, we argue two flaws of these methods: \\((i)\\) The success of these non-linear solutions requires a proper initialization. While a brute-force search may be applicable, it is computationally expensive for real-time application, specially if the dimension of the parameter is high. Efficient and deterministic initialization methods are lacking and urgently needed. \\((ii)\\) The common assumption of constant velocity only holds within a short time interval. Nevertheless, a time window that is too narrow may contain insufficient number of events for a valid model fitting. The heuristic of using a constant number of events may violate the constant velocity assumption, especially in the case of sudden changes in speed." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.53, + 0.788, + 0.681 + ], + "angle": 0, + "content": "In this paper, we rethink the problem formulation in a way that is more consistent with the differential working principle of event cameras. Unlike existing pipelines that use events occurring within a time interval to resolve the relative motion, we attempt to recover the structure and instantaneous first-order kinematic parameters involved in the motion field equation, using as input an instantaneous observation, namely event-based normal flow. To this end, we explore effective constraints between normal-flow observations and model parameters for solving a family of motion-and-structure estimation problems, such as the estimation of depth and full optical flow, instantaneous angular velocity and linear velocity, and differential homography (see Fig. 1)." + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.682, + 0.689, + 0.698 + ], + "angle": 0, + "content": "The contributions of this paper can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.717, + 0.785, + 0.775 + ], + "angle": 0, + "content": "- A normal flow constraint with a geometric connection to event data. It overcomes the partial observability issue (i.e., errors caused by using normal flow as replacement for full optical flow) in general problems of geometric model fitting when normal flow is used as input." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.78, + 0.785, + 0.838 + ], + "angle": 0, + "content": "- Two solvers for a family of motion-and-structure estimation problems using as input sparse event-based normal flow. Our linear solver (can be used with RANSAC) leads to closed-form and deterministic solutions that can be used as an initialization to existing nonlinear methods. Our nonlinear solver," + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.717, + 0.785, + 0.838 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.373, + 0.115, + 0.733, + 0.128 + ], + "angle": 0, + "content": "Motion and Structure from Event-based Normal Flow" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "image", + "bbox": [ + 0.235, + 0.144, + 0.768, + 0.294 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.302, + 0.788, + 0.358 + ], + "angle": 0, + "content": "Fig. 1: Overview of our event-based motion estimation approach. Normal flow vectors are computed from the raw event data and are used as input to the two solvers. The result of the linear solver consists of fitted motion models \\(\\{\\pmb{\\theta}_1, \\pmb{\\theta}_2, \\dots\\}\\) at discrete time instants, which can be used to initialize the continuous-time nonlinear solver." + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.389, + 0.784, + 0.418 + ], + "angle": 0, + "content": "under a continuous-time formulation, can handle sudden speed variations, and thus, is free of the constant-motion assumption." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.419, + 0.784, + 0.464 + ], + "angle": 0, + "content": "- A thorough evaluation on the two proposed solvers, comparing against state-of-the-art methods, and also including an investigation on the performance of existing non-linear methods initialized using our linear solver." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.473, + 0.788, + 0.565 + ], + "angle": 0, + "content": "Outline: The rest of the paper is organized as follows. We first discuss the related work in Sec. 2. Then a compact review of preliminary knowledge is given in Sec. 3. Our approach is presented in Sec. 4, including the normal-flow constraint, the linear models of a family of motion-and-structure estimation problems, and the proposed linear and nonlinear solvers. Experimental evaluation is conducted in Sec. 5 and conclusion are drawn in Sec. 6." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.587, + 0.388, + 0.603 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.619, + 0.577, + 0.635 + ], + "angle": 0, + "content": "2.1 Event-based Geometric Model Fitting" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.788, + 0.81 + ], + "angle": 0, + "content": "The major body of literature on this specific topic follows an iterative scheme, which typically has its own special metric evaluating quantitatively the association between events, and solve the model fitting as a nonlinear optimization problem. Among the first works is the Contrast Maximization method [13, 14], which provides a unified framework for solving a family of motion-model fitting problems. It searches the optimal model parameters that ultimately lead to a sharp image of warped events (IWE). The goodness of fit is evaluated by the alignment of all involved events along point trajectories on the image plane, and such alignment is assessed by the strength of the contours of the IWE. The computational complexity of CMax is high because of iteratively updating of IWE, whose cost scales with the spatial resolution and the amount of event data." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "To lessen the dependency on IWE, Nunes et al. proposes an entropy minimization framework (EMin) [30] that replaces the dispersion metric of [14] with" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.128 + ], + "angle": 0, + "content": "Z. Ren et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.405 + ], + "angle": 0, + "content": "a similarity measurement between events, enabling to evaluate the goodness of fit in a higher dimensional space. EMin is, still, too expensive for real-time application. Liu et al. [22] propose a two-step iterative method for the problem of 3-D rotation estimation using events as input. Given a rotation estimate, the first step determines event correspondences by looking for the nearest neighbour in the spatio-temporal domain. The second step employs a spatio-temporal registration which aligns corresponding events by minimizing the sum of squared geodesic residual. The least-square nature of the energy function in the second step leads to a computation friendly performance. Gu et al. [17] present a probabilistic approach for spatio-temporal alignment of events. The aligned events at a particular pixel are modeled as a spatio-temporal Poisson point process and the goodness of fit can be quantified using the likelihood of the data. The optimal model is retrieved when the likelihood is maximized. More recently, an event-to-map alignment scheme is proposed in [19], which adopts the time surface (TS) map as a reference for event alignment. The proposed TS loss function drives later events to get increasingly closer to pixel positions of corresponding earlier events, thus, leading to a progressively sharp TS map as iterations proceed." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.405, + 0.788, + 0.571 + ], + "angle": 0, + "content": "In summary, all the above-mentioned methods are nonlinear, and thus, a proper initialization is necessary. Although a brute-force search is widely adopted, sometimes it can be too expensive to locate the convergence basin in parameter space. Moreover, [14,30] may come across a convergence to false-positive minima, known as event collapse [36, 37]. Hence, a quick and deterministic solver (e.g., a closed-form solution) is desirable. Besides, most of these methods assume a constant-motion model, which limits their applicability to scenarios of sudden speed variations. In addition, we have noticed that a recent work [16] can jointly estimate the line parameters and the camera's linear velocities. However, this method is not a general one due to the dependence on line structures and the usage of an IMU. Thus, it is not within the scope of our discussion." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.598, + 0.603, + 0.613 + ], + "angle": 0, + "content": "2.2 Motion and Structure from Normal Flow" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.628, + 0.788, + 0.78 + ], + "angle": 0, + "content": "Due to the aperture problem, a local estimate of optical flow recovers only the partial component of the full flow vector along the direction of image gradient. Preliminary knowledge about normal flow can be found in the supplementary material. Recovering motion and structure from normal flow on intensity images is a well-studied topic. Its main body of literature typically proposes a qualitatively deterministic solution by which a set (even infinite) of applicable ego motions are determined via an analysis of normal flow classification [3, 11, 20]. Only recently, researches devised ways to formulate the Cheirality constraint as a differentiable term and use it in a continuous optimization approach [1] or as a loss function for training an artificial neural network [31]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.842 + ], + "angle": 0, + "content": "In event-based perception and navigation, similar methodologies as the frame-based ones are also witnessed, where normal flow computed from event data is used as input to determine the Focus of Expansion (FoE) for drones' course estimation [10] and obstacle avoidance [5]." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.373, + 0.115, + 0.733, + 0.128 + ], + "angle": 0, + "content": "Motion and Structure from Event-based Normal Flow" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.389 + ], + "angle": 0, + "content": "Most of these heuristic methods, while effective, are cumbersome for practical application. In contrast, our method delivers a deterministic way to retrieve motion and structure from normal flow. Additionally, it is noticed that normal flow can replace full flow in some specific problems of geometric model fitting, such as affine models [26, 27]. In this way, the geometric model fitting problem can be easily solved in closed-form. Nevertheless, these methods do not consider the observability bias in normal flows, and thus, such naive substitution may lead to inaccurate model fitting. To this end, we present a robust linear solver based on a carefully-designed geometric measurement model that can overcome the partial observability of event-based normal flow. It is worth mentioning that a recent work [24] also determines instantaneous linear velocity using as input the normal flow. Our difference is seen in the scope of applicability. The usage of normal flow is restricted in the motion flow equation of [24], while our paper extends it to a wider scope (i.e., a family of geometric model fitting problems) by combining normal flow with e.g., differential epipolar geometry, differential homography, etc." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.41, + 0.382, + 0.426 + ], + "angle": 0, + "content": "3 Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.44, + 0.788, + 0.487 + ], + "angle": 0, + "content": "To make this paper self-contained, some preliminary knowledge is revisited. Due to space limitations, here we only list essential concepts. A more detailed review of preliminary knowledge can be found in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.506, + 0.788, + 0.569 + ], + "angle": 0, + "content": "Motion Field. Consider a perspective camera moving in a static environment, with instantaneous angular and linear velocities in the camera frame \\(\\boldsymbol{\\omega} = (\\omega_{x},\\omega_{y},\\omega_{z})^{\\top}\\) and \\(\\pmb {\\nu} = (\\nu_{x},\\nu_{y},\\nu_{z})^{\\top}\\), respectively. The motion field at pixel \\(\\mathbf{x} = (x,y)^{\\top}\\) produced by projecting the 3D motion at point with depth \\(Z(\\mathbf{x})\\) is" + }, + { + "type": "equation", + "bbox": [ + 0.405, + 0.577, + 0.787, + 0.609 + ], + "angle": 0, + "content": "\\[\n\\mathbf {u} = \\frac {1}{Z (\\mathbf {x})} \\mathbf {A} (\\mathbf {x}) \\boldsymbol {\\nu} + \\mathbf {B} (\\mathbf {x}) \\boldsymbol {\\omega}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.615, + 0.788, + 0.678 + ], + "angle": 0, + "content": "where \\(\\mathbf{A}(\\mathbf{x})\\) and \\(\\mathbf{B}(\\mathbf{x})\\) are \\(2\\times 3\\) matrices that solely depend on \\(\\mathbf{x}\\). Equation (1) is the motion field equation [40, Ch.8] [23]. Optical flow in rigid scenes aims at estimating the motion field from visual data, which is only possible in regions that do not suffer from the aperture problem." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.697, + 0.788, + 0.804 + ], + "angle": 0, + "content": "Brightness constancy is the assumption stating that the true motion trajectories \\(\\mathbf{x}(t)\\) followed by objects on the image plane are those at which brightness is constant: \\(I(\\mathbf{x}(t), t) = \\mathrm{const}\\). Differentiating in time (total derivative) gives the equation at \\((\\mathbf{x}, t)\\): \\(\\nabla I \\cdot \\dot{\\mathbf{x}} + \\partial_t I = 0\\), where \\(\\dot{\\mathbf{x}} = d\\mathbf{x} / dt\\) is the velocity of image point \\(\\mathbf{x}(t)\\). Given \\(I\\), this equation provides one constraint per point \\((\\mathbf{x}, t)\\) to estimate \\(\\dot{\\mathbf{x}}\\), now called optical flow. Without additional equations, only the normal component of the optical flow can be determined [40]," + }, + { + "type": "equation", + "bbox": [ + 0.441, + 0.813, + 0.787, + 0.845 + ], + "angle": 0, + "content": "\\[\n\\mathbf {n} = - \\frac {\\partial_ {t} I}{\\| \\nabla I \\| ^ {2}} \\nabla I. \\tag {2}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.128 + ], + "angle": 0, + "content": "Z. Ren et al." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.788, + 0.193 + ], + "angle": 0, + "content": "Normal Flow. Decomposing the optical flow \\(\\mathbf{u} = \\mathbf{u}_{\\perp} + \\mathbf{u}_{\\parallel}\\) into its normal and parallel components to the local image gradient \\((\\nabla I)\\), the normal flow is \\(\\mathbf{n} = \\mathbf{u}_{\\perp}\\), and the dot-product with \\(\\mathbf{n}\\) gives" + }, + { + "type": "equation", + "bbox": [ + 0.325, + 0.203, + 0.786, + 0.22 + ], + "angle": 0, + "content": "\\[\n\\mathbf {n} ^ {\\top} \\mathbf {u} = \\| \\mathbf {n} \\| ^ {2}, \\quad \\text {o r e q u i v a l e n t l y} \\quad (\\mathbf {u} - \\mathbf {n}) \\cdot \\mathbf {n} = 0. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.242, + 0.784, + 0.318 + ], + "angle": 0, + "content": "Differential Homography. (Continuous Homography Matrix [25, Ch.5]). If all scene points lie on a plane, their coordinates are related to their image projections via a homography transformation (which encodes the information of the scene plane \\(\\{\\mathbf{N}, d\\}\\) and the camera pose). In this case, the motion field (1) can be rewritten to adopt the form in homogeneous coordinates" + }, + { + "type": "equation", + "bbox": [ + 0.418, + 0.329, + 0.785, + 0.347 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {u}} (\\mathbf {x}) = \\left(\\mathbf {1} - \\hat {\\mathbf {x}} \\mathbf {e} _ {3} ^ {\\top}\\right) \\mathbf {H} _ {d} \\hat {\\mathbf {x}}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.358, + 0.784, + 0.405 + ], + "angle": 0, + "content": "where \\(\\mathbf{H}_d\\dot{=} -\\left([ \\boldsymbol {\\omega}]_{\\times} + d^{-1}\\boldsymbol {\\nu}\\mathbf{N}^{\\top}\\right)\\) denotes the differential homography matrix \\(([\\cdot ]_{\\times}\\) refers to the cross-product, skew-symmetric matrix), \\(\\mathbf{1}\\) is the identity matrix, \\(\\hat{\\mathbf{x}} = (x,y,1)^{\\top}\\), \\(\\hat{\\mathbf{u}} = (u_x,u_y,0)^{\\top}\\) and \\(\\mathbf{e}_3 = (0,0,1)^{\\top}\\)." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.426, + 0.788, + 0.502 + ], + "angle": 0, + "content": "Differential Epipolar Geometry. (Continuous Epipolar Geometry [25, Ch.5]). The epipolar geometry examines the distance between a potential match to the determined epipolar line, and can be expressed as \\(\\hat{\\mathbf{x}}^{\\prime \\top}\\mathbf{E}\\hat{\\mathbf{x}} = 0\\). The differential epipolar geometry can be derived analogously, providing a relationship between \\(\\mathbf{x},\\mathbf{u}\\) and the camera velocities \\(\\pmb {\\nu},\\pmb{\\omega}\\). In homogeneous coordinates:" + }, + { + "type": "equation", + "bbox": [ + 0.414, + 0.513, + 0.785, + 0.53 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {u}} (\\mathbf {x}) ^ {\\top} [ \\boldsymbol {\\nu} ] _ {\\times} \\hat {\\mathbf {x}} - \\hat {\\mathbf {x}} ^ {\\top} \\mathbf {s} \\hat {\\mathbf {x}} = 0, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.541, + 0.673, + 0.559 + ], + "angle": 0, + "content": "where \\(\\mathbf{s} = \\frac{1}{2} ([\\pmb {\\nu}]_{\\times}[\\pmb {\\omega}]_{\\times} + [\\pmb {\\omega}]_{\\times}[\\pmb {\\nu}]_{\\times})\\) is a \\(3\\times 3\\) symmetric matrix." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.58, + 0.381, + 0.598 + ], + "angle": 0, + "content": "4 Methodology" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.612, + 0.788, + 0.718 + ], + "angle": 0, + "content": "Figure 1 summarizes our approach, grounded upon event-based normal flow. In this section, we first disclose the normal-flow constraint that can overcome the partial observability of normal flow when replacing full optical flow with normal flow in general tasks of geometric model fitting (Sec. 4.1). By combining the constraint and the theory in Sec. 3, we derive linear equations for a family of motion-and-structure estimation problems (Sec. 4.2). To solve these problems, we present a linear solver and a continuous-time nonlinear solver (Sec. 4.3)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.74, + 0.466, + 0.755 + ], + "angle": 0, + "content": "4.1 Normal Flow Constraint" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.765, + 0.788, + 0.843 + ], + "angle": 0, + "content": "Normal flow can be computed easily from the event data, for example by fitting planes locally to the space-time \"cloud\" of events on image space [2, 32]. Specifically, assuming that the event data \\( e_k \\doteq \\{\\mathbf{x}_k, t_k\\} \\) spans locally a surface in the spatio-temporal domain, the surface can be defined by a function \\( \\Sigma_e: \\mathbb{R}^2 \\to \\mathbb{R}; \\mathbf{x}_k \\mapsto t_k \\) that maps the pixel coordinates to the latest event's" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.373, + 0.115, + 0.733, + 0.128 + ], + "angle": 0, + "content": "Motion and Structure from Event-based Normal Flow" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.217, + 0.144, + 0.342, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.343, + 0.144, + 0.478, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.48, + 0.144, + 0.633, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.635, + 0.144, + 0.787, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.264, + 0.788, + 0.349 + ], + "angle": 0, + "content": "Fig. 2: Toy example of applying the proposed normal-flow constraint. (a): A 2D registration task using as input either optical flow or normal flow. The groundtruth displacement is defined by a global flow of \\((1.732, -1)^{\\top}\\). (b)-(d): The loss landscape obtained using different geometric measurements in the registration task. The red dot denotes the resulting displacement, and the green cross the groundtruth displacement. The bottom-left regions in (b)-(d) display the corresponding registration results." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.378, + 0.784, + 0.41 + ], + "angle": 0, + "content": "timestamp [2, 5]. The normal flow is parallel to the spatial derivative of \\(\\Sigma_{e}\\), namely \\(\\nabla \\Sigma_{e}(\\mathbf{x})\\). More specifically, it can be calculated by" + }, + { + "type": "equation", + "bbox": [ + 0.427, + 0.421, + 0.786, + 0.456 + ], + "angle": 0, + "content": "\\[\n\\mathbf {n} (\\mathbf {x}) = \\frac {\\nabla \\Sigma_ {e} (\\mathbf {x})}{\\left\\| \\nabla \\Sigma_ {e} (\\mathbf {x}) \\right\\| ^ {2}}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.466, + 0.784, + 0.526 + ], + "angle": 0, + "content": "Readers may refer to the supplementary material for a more detailed derivation. Surely there are other methods to extract event-based normal flow and the solvers we develop in this section are independent of them, i.e., our solvers can also be fed with the extracted normal flow data from such methods." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.527, + 0.785, + 0.631 + ], + "angle": 0, + "content": "Event-based normal flow cannot be used directly as a replacement for the full flow in (1) to solve for the motion parameters. However, we can leverage the relationship between both flows (3) to define an effective geometric error criterion that circumvents the partial observability of normal flow. If the flow \\(\\mathbf{u}\\) depends on some unknown (\"state\") variables \\(\\pmb{\\theta}\\), we propose estimating them based on how well the normal flow constraint (3) is satisfied, e.g., by minimizing some function of the error terms" + }, + { + "type": "equation", + "bbox": [ + 0.373, + 0.644, + 0.785, + 0.663 + ], + "angle": 0, + "content": "\\[\n\\mathbf {n} (\\mathbf {x}) ^ {\\top} \\mathbf {u} (\\mathbf {x}; \\boldsymbol {\\theta}) - \\| \\mathbf {n} (\\mathbf {x}) \\| ^ {2} \\doteq \\epsilon_ {\\mathrm {n f}} \\in \\mathbb {R}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.674, + 0.788, + 0.841 + ], + "angle": 0, + "content": "To illustrate the validity of the proposed normal-flow constraint over the naive way of simply using normal flow to replace full optical flow, we use a 2D registration task as toy example. As shown in Fig. 2, the task consists of registering the solid-line triangle to the dashed-line one using as input either optical flow or normal flow. We observe that using the proposed normal-flow constraint as a geometric measurement leads to a registration result that is almost as accurate as the one given by the full optical flow. However, due to the partial observability of normal flow, simply using it as a replacement for full optical flow leads to a biased registration result, as shown in Fig. 2c. As we see in all following experiments, similar conclusions can be made in general problems of geometric model fitting." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.128 + ], + "angle": 0, + "content": "Z. Ren et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.531, + 0.162 + ], + "angle": 0, + "content": "4.2 Models of Motion and Structure" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.171, + 0.788, + 0.294 + ], + "angle": 0, + "content": "Based on the derivation in Sec. 4.1, we establish linear models to a family of motion and structure estimation problems using the event-based normal flow as input. In particular, we first discuss how to recover optical flow and depth by introducing the normal-flow constraint into the differential epipolar geometry. Then we show that the combination of the normal-flow constraint with the motion flow equation (1) results in solutions to the problem of instantaneous angular velocity estimation, and 6-DoF motion estimation given depth as prior, respectively. Finally, we show how to compute the differential homography." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.314, + 0.789, + 0.377 + ], + "angle": 0, + "content": "Optical Flow. Given as input the accurate normal flow, and camera velocities, the full optical flow \\(\\mathbf{u}(\\mathbf{x})\\) at pixel location \\(\\mathbf{x}\\) can be calculated by solving the following linear system obtained from the combination of the normal-flow constraint (3) and the differential epipolar geometry (5):" + }, + { + "type": "equation", + "bbox": [ + 0.387, + 0.386, + 0.787, + 0.422 + ], + "angle": 0, + "content": "\\[\n\\binom {\\mathbf {n} (\\mathbf {x}) ^ {\\top}} {- \\hat {\\mathbf {x}} ^ {\\top} [ \\boldsymbol {\\nu} ] _ {\\times}} \\mathbf {u} (\\mathbf {x}) = \\binom {\\| \\mathbf {n} (\\mathbf {x}) \\| ^ {2}} {\\hat {\\mathbf {x}} ^ {\\top} \\mathbf {s} \\hat {\\mathbf {x}}}. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.441, + 0.788, + 0.488 + ], + "angle": 0, + "content": "Depth Estimation. Similarly, if the normal flow and camera velocities are known, the scene depth \\( Z(\\mathbf{x}) \\) can be determined by combining the normal-flow constraint (3) with the motion field eq. (1), yielding the closed-form solution:" + }, + { + "type": "equation", + "bbox": [ + 0.385, + 0.498, + 0.787, + 0.533 + ], + "angle": 0, + "content": "\\[\nZ (\\mathbf {x}) = \\frac {\\mathbf {n} (\\mathbf {x}) ^ {\\top} \\mathbf {A} (\\mathbf {x}) \\boldsymbol {\\nu}}{\\| \\mathbf {n} (\\mathbf {x}) \\| ^ {2} - \\mathbf {n} (\\mathbf {x}) ^ {\\top} \\mathbf {B} (\\mathbf {x}) \\boldsymbol {\\omega}}. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.55, + 0.788, + 0.582 + ], + "angle": 0, + "content": "Angular Velocity Estimation. Assuming that an event camera undergoes a pure rotational motion, the optical flow \\(\\mathbf{u}\\) in (1) can be simplified as" + }, + { + "type": "equation", + "bbox": [ + 0.449, + 0.593, + 0.786, + 0.609 + ], + "angle": 0, + "content": "\\[\n\\mathbf {u} (\\mathbf {x}) = \\mathbf {B} (\\mathbf {x}) \\boldsymbol {\\omega}. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.619, + 0.675, + 0.635 + ], + "angle": 0, + "content": "By substituting (10) into (3), we have the linear equation in \\(\\omega\\):" + }, + { + "type": "equation", + "bbox": [ + 0.414, + 0.645, + 0.786, + 0.663 + ], + "angle": 0, + "content": "\\[\n\\mathbf {n} (\\mathbf {x}) ^ {\\top} \\mathbf {B} (\\mathbf {x}) \\omega = \\| \\mathbf {n} (\\mathbf {x}) \\| ^ {2}. \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.673, + 0.787, + 0.705 + ], + "angle": 0, + "content": "Since each normal flow provides only one constraint, the minimal problem requires three normal flow vectors as input." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.725, + 0.788, + 0.802 + ], + "angle": 0, + "content": "6-DoF Motion Estimation. Assume the event camera undergoes a general 6-DoF motion with known depth of the observed scene, we can solve the first-order kinematics, namely instantaneous linear and angular velocities, by solving the following linear system obtained from the normal-flow constraint (3) and the motion-flow equation (1):" + }, + { + "type": "equation", + "bbox": [ + 0.403, + 0.812, + 0.787, + 0.845 + ], + "angle": 0, + "content": "\\[\n\\mathbf {n} (\\mathbf {x}) ^ {\\top} \\mathbf {D} (\\mathbf {x}) \\left[ \\begin{array}{l} \\boldsymbol {\\nu} \\\\ \\boldsymbol {\\omega} \\end{array} \\right] = \\| \\mathbf {n} (\\mathbf {x}) \\| ^ {2}, \\tag {12}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.372, + 0.115, + 0.733, + 0.128 + ], + "angle": 0, + "content": "Motion and Structure from Event-based Normal Flow" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.145, + 0.784, + 0.198 + ], + "angle": 0, + "content": "where \\(\\mathbf{D}(\\mathbf{x}) = \\left[\\frac{1}{Z(\\mathbf{x})}\\mathbf{A}(\\mathbf{x})\\mathbf{B}(\\mathbf{x})\\right]\\) is the \\(2\\times 6\\) feature sensitivity matrix [4, 7]. The minimal problem in this context requires six normal flow measurements to determine the 6-DoF motion parameters." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.216, + 0.784, + 0.307 + ], + "angle": 0, + "content": "Motion and Structure Estimation in Planar Scenes. For a moving camera that observes a planar scene, the motion and structure parameters are compatibly encoded in a homography matrix. In our case, we attempt to recover the differential homography \\(\\mathbf{h}_d\\) (the vector form of \\(\\mathbf{H}_d\\) in (4)) because the instantaneous linear and angular velocities are considered. By expanding (4) and rearranging equations, we have" + }, + { + "type": "equation", + "bbox": [ + 0.447, + 0.317, + 0.785, + 0.332 + ], + "angle": 0, + "content": "\\[\n\\mathbf {u} (\\mathbf {x}) = \\mathbf {C} (\\mathbf {x}) \\mathbf {h} _ {d}, \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.339, + 0.784, + 0.37 + ], + "angle": 0, + "content": "where \\(\\mathbf{C}(\\mathbf{x})\\) is a \\(2 \\times 9\\) matrix that solely depends on the pixel location \\(\\mathbf{x}\\). Subsequently, substituting (3) into (13), gives:" + }, + { + "type": "equation", + "bbox": [ + 0.411, + 0.376, + 0.785, + 0.394 + ], + "angle": 0, + "content": "\\[\n\\mathbf {n} (\\mathbf {x}) ^ {\\top} \\mathbf {C} (\\mathbf {x}) \\mathbf {h} _ {d} = \\| \\mathbf {n} (\\mathbf {x}) \\| ^ {2}. \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.401, + 0.784, + 0.446 + ], + "angle": 0, + "content": "As discussed in [44], retrieving the motion and structure parameters from the decomposition of \\(\\mathbf{h}_d\\) is not trivial and beyond the scope of this paper. Therefore, we only focus on the numerical solution of \\(\\mathbf{h}_d\\)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.466, + 0.36, + 0.479 + ], + "angle": 0, + "content": "4.3 Our Solvers" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.487, + 0.785, + 0.593 + ], + "angle": 0, + "content": "According to the motion-and-structure models derived in Sec. 4.2, we have established certain constraints between the model parameters and the normal flow observations. This allows us to simply develop linear solvers for each specific problem. Besides, we further propose a continuous-time nonlinear solver that can effectively handle the presence of sudden variations in motion and does not rely on the constant-motion assumption. To convey our solvers concisely, we generalize the motion-and-structure models into a unified form, as" + }, + { + "type": "equation", + "bbox": [ + 0.416, + 0.601, + 0.785, + 0.617 + ], + "angle": 0, + "content": "\\[\n\\mathbf {n} (\\mathbf {x}) ^ {\\top} \\mathbf {O} (\\mathbf {x}) \\boldsymbol {\\theta} = \\| \\mathbf {n} (\\mathbf {x}) \\| ^ {2}, \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.625, + 0.78, + 0.641 + ], + "angle": 0, + "content": "where \\(\\mathbf{O}(\\mathbf{x})\\) is the system matrix and \\(\\theta\\) are the motion/structure parameters." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.784, + 0.705 + ], + "angle": 0, + "content": "Linear Solver. Given as input a set (e.g., K) of sparse normal-flow estimates obtained at the same time, our linear solver works out the following (overdetermined) linear system of equations:" + }, + { + "type": "equation", + "bbox": [ + 0.372, + 0.712, + 0.785, + 0.769 + ], + "angle": 0, + "content": "\\[\n\\left( \\begin{array}{c} \\mathbf {n} \\left(\\mathbf {x} _ {1}\\right) ^ {\\top} \\mathbf {O} \\left(\\mathbf {x} _ {1}\\right) \\\\ \\vdots \\\\ \\mathbf {n} \\left(\\mathbf {x} _ {\\mathrm {K}}\\right) ^ {\\top} \\mathbf {O} \\left(\\mathbf {x} _ {\\mathrm {K}}\\right) \\end{array} \\right) \\boldsymbol {\\theta} = \\left( \\begin{array}{c} \\| \\mathbf {n} \\left(\\mathbf {x} _ {1}\\right) \\| ^ {2} \\\\ \\vdots \\\\ \\| \\mathbf {n} \\left(\\mathbf {x} _ {\\mathrm {K}}\\right) \\| ^ {2} \\end{array} \\right). \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.775, + 0.783, + 0.791 + ], + "angle": 0, + "content": "For robust estimation, RANSAC [12] is adopted to deal with noise and outliers." + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.797, + 0.66, + 0.812 + ], + "angle": 0, + "content": "\\(^{5}\\) \\(\\mathbf{O}(\\mathbf{x})\\) degenerates to 1 in the problem of optical flow estimation." + }, + { + "type": "page_footnote", + "bbox": [ + 0.222, + 0.813, + 0.784, + 0.84 + ], + "angle": 0, + "content": "An approximation can be made based on the assumption of constant camera motion. In this way, the normal flow estimates can be collected within a short time interval." + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.797, + 0.784, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.128 + ], + "angle": 0, + "content": "Z. Ren et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.207 + ], + "angle": 0, + "content": "Continuous-Time Nonlinear Solver. To handle the case of a sudden variation in speed that may violate the constant-motion assumption, we further propose a continuous-time nonlinear solver. In particular, we employ cubic B-splines to represent the continuous-time model parameters as" + }, + { + "type": "equation", + "bbox": [ + 0.408, + 0.22, + 0.786, + 0.26 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\theta} (u (t)) = \\sum_ {i = 0} ^ {3} B _ {i, 3} (u (t)) \\mathbf {c} _ {i}, \\tag {17}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.27, + 0.785, + 0.376 + ], + "angle": 0, + "content": "where \\( B_{i,k} \\) denotes the basis function, \\( i \\) the index of control points, \\( k (= 3) \\) the order of the spline and \\( \\mathbf{c}_i \\in \\mathbb{R}^N \\) the corresponding control point defined in the \\( N \\)-dimensional space of model parameters. Examples of cubic B-splines applied to event camera motion estimation are found in [18, 28]. The function \\( u(\\cdot) \\) is a normalization operator, which transfers time \\( t \\) to the spline's parameter domain by means of basis translation [8]. In this way, the model parameters at any time instant \\( t \\) can be obtained via an interpolation using the known control points." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.377, + 0.785, + 0.452 + ], + "angle": 0, + "content": "The formulation of the continuous-time model fitting problem goes as follows. The unknown variable consists of all the involved control points, \\(\\mathcal{X} \\doteq [\\mathbf{c}_0, \\mathbf{c}_1, \\dots, \\mathbf{c}_n]\\), where \\(n\\) denotes the total number of control points. By introducing the continuous-time representation of model parameters to (15), we finally create the following objective function:" + }, + { + "type": "equation", + "bbox": [ + 0.299, + 0.468, + 0.786, + 0.5 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\mathcal {X}} ^ {*} = \\arg \\min _ {\\boldsymbol {\\mathcal {X}}} \\sum_ {e _ {k} \\in \\mathcal {E}} \\left\\| \\mathbf {n} \\left(\\mathbf {x} _ {k}\\right) ^ {\\top} \\mathbf {O} \\left(\\mathbf {x} _ {k}\\right) \\boldsymbol {\\theta} \\left(u \\left(t _ {k}\\right)\\right) - \\| \\mathbf {n} \\left(\\mathbf {x} _ {k}\\right) \\| ^ {2} \\right\\| ^ {2}, \\tag {18}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.51, + 0.784, + 0.601 + ], + "angle": 0, + "content": "where \\(\\mathcal{E}\\) denotes the event set in which the norm flow observation is available for every involved event \\(e_k\\). Note that this continuous-time formulation aligns better with the asynchronous nature of event-based normal flow, enabling a state estimation over a larger temporal window that may lead to a better signal-to-noise ratio. The linear solver can be used to initialize the continuous-time nonlinear solver. An illustrative explanation of our solvers is given in Fig. 1." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.624, + 0.376, + 0.642 + ], + "angle": 0, + "content": "5 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.656, + 0.785, + 0.747 + ], + "angle": 0, + "content": "To justify the effectiveness of the proposed algorithms, experiments are conducted on both simulation and real data. First, we introduce the datasets used and the applied metrics for evaluation (Sec. 5.1). Then we report the results of our linear solver and continuous-time nonlinear solver (Sec. 5.2 and Sec. 5.3). Finally, a discussion on the computational complexity and the limitation of our method is provided in Sec. 5.4." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.77, + 0.427, + 0.785 + ], + "angle": 0, + "content": "5.1 Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.784, + 0.841 + ], + "angle": 0, + "content": "Our evaluation utilizes four datasets, including the indoor event-camera dataset (ECD) captured with a DAVIS240C camera and motion-capture system for groundtruth, and a high-resolution dataset with two rotational motion sequences" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.373, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Motion and Structure from Event-based Normal Flow" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.146, + 0.416, + 0.249 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.218, + 0.25, + 0.415, + 0.271 + ], + "angle": 0, + "content": "(a) Estimated optical flow over a map of event accumulation." + }, + { + "type": "image", + "bbox": [ + 0.424, + 0.144, + 0.603, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.425, + 0.25, + 0.785, + 0.272 + ], + "angle": 0, + "content": "(b) Reconstruction result displayed over groundtruth structures (a v-shape roof) and viewed from different perspectives." + }, + { + "type": "image", + "bbox": [ + 0.606, + 0.144, + 0.785, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.283, + 0.788, + 0.311 + ], + "angle": 0, + "content": "Fig. 3: Qualitative result of optical flow and depth on sequence three_wall Translation of our synthetic data. Note that our method returns sparse results." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.337, + 0.789, + 0.473 + ], + "angle": 0, + "content": "using an iniVation DAVIS346 camera and Xsens IMU. The VECTor dataset, captured with a Prophesee Gen 3 camera and an accompanying RGB-D depth camera, is used for 6-DoF motion tracking. Additionally, a synthetic dataset generated in Blender is employed for motion-and-structure pattern analysis. Evaluation metrics for kinematics estimation include average error (AE) and root mean squared error (RMSE), while differential homography estimation is assessed using the Frobenius norm. Best results are in bold, and second-best are underlined. The supplementary material provides detailed descriptions of datasets and metrics, as well as implementation details of our methods." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.492, + 0.52, + 0.506 + ], + "angle": 0, + "content": "5.2 Evaluation of the Linear Solver" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.513, + 0.786, + 0.573 + ], + "angle": 0, + "content": "In the following evaluation, we qualitatively and quantitatively assess the performance of our linear method (Ours) in each specific task, and compare against different state-of-the-art (SOTA) methods (CMax [14] and Pro-STR [19]) and their customized version initialized using our linear solver." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.592, + 0.788, + 0.683 + ], + "angle": 0, + "content": "Optical Flow and Depth Estimation. Due to a lack of benchmark for extensive evaluation, we qualitatively evaluate our method on the tasks of optical flow and depth estimation. As illustrated in Fig. 3a, our linear solver demonstrates the capability to accurately reconstruct full flow using normal flow and known motion parameters. An example of the depth estimation result is given in Fig. 3b, showing that the recovered structure aligns well with the groundtruth." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.701, + 0.788, + 0.762 + ], + "angle": 0, + "content": "Angular Velocity Estimation. As shown in Tab. 1, we see that our linear solver outperforms CMax and Pro-STR. Besides, we find the combination of our linear solver and CMax leads to the best result, indicating that the proposed linear solver is a complementary method to existing nonlinear methods." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Differential Homography Estimation. The result is consistent with that in the above task. Table 2 reveals that our linear method outperforms CMax and Pro-STR in terms of numerical accuracy. The combination of our linear solver and state-of-the-art nonlinear solver leads to the best result." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.127 + ], + "angle": 0, + "content": "Z. Ren et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.225, + 0.168, + 0.777, + 0.181 + ], + "angle": 0, + "content": "Table 1: Evaluation of our linear solver on the task of angular velocity estimation." + }, + { + "type": "table", + "bbox": [ + 0.229, + 0.194, + 0.784, + 0.381 + ], + "angle": 0, + "content": "
MethodshapesRotation [29]groundRotation (our data)patternsRotation (our data)
eω(deg/s)RMSEω(deg/s)eω(deg/s)RMSEω(deg/s)eω(deg/s)RMSEω(deg/s)
CMax [14]15.6026.389.4815.585.959.07
Pro-STR [19]90.61145.8131.0744.7410.4711.88
Ours15.3623.983.308.014.706.08
Ours+CMax [14]10.1214.292.853.740.350.73
Ours+Pro-STR [19]19.2340.449.1317.074.315.23
MethoddynamicRotation [29]boxesRotation (our data)cubesRotation (our data)
eω(deg/s)RMSEω(deg/s)eω(deg/s)RMSEω(deg/s)eω(deg/s)RMSEω(deg/s)
CMax [14]16.1426.488.1813.566.7710.62
Pro-STR [19]54.6485.6322.7232.0511.0712.79
Ours9.4118.062.604.782.313.02
Ours+CMax [14]7.7110.852.643.800.490.77
Ours+Pro-STR [19]18.8732.653.495.872.042.76
" + }, + { + "type": "table_caption", + "bbox": [ + 0.235, + 0.433, + 0.767, + 0.447 + ], + "angle": 0, + "content": "Table 2: Evaluation of our linear solver for differential homography estimation." + }, + { + "type": "table", + "bbox": [ + 0.297, + 0.46, + 0.712, + 0.546 + ], + "angle": 0, + "content": "
Methodboard Slow [15]patterns rotationpatterns 6 dof
F-NormF-NormF-Norm
CMax [14]1.161.331.27
Pro-STR [19]2.071.221.37
Ours0.240.200.43
Ours+CMax [14]0.230.150.41
Ours+Pro-STR [19]0.190.170.42
" + }, + { + "type": "table_caption", + "bbox": [ + 0.236, + 0.598, + 0.763, + 0.613 + ], + "angle": 0, + "content": "Table 3: Evaluation of our linear solver on the task of 6-DoF motion tracking." + }, + { + "type": "table", + "bbox": [ + 0.23, + 0.625, + 0.783, + 0.817 + ], + "angle": 0, + "content": "
Methodcorner Slow [15]patterns _6 dof (our data)
eω(deg/s)RMSEω(deg/s)ev(m/s)RMSEv(m/s)eω(deg/s)RMSEω(deg/s)ev(m/s)RMSEv(m/s)
CMax [14]12.3914.390.340.4112.9415.681.021.23
Pro-STR [19]5.197.840.130.2211.2113.491.331.49
Ours1.311.790.010.020.560.890.120.16
Ours+CMax [14]2.192.820.110.170.550.870.180.30
Ours+Pro-STR [19]1.061.340.040.051.071.560.120.15
Methodmountain_normal [15]cubes _6 dof (our data)
eω(deg/s)RMSEω(deg/s)ev(m/s)RMSEv(m/s)eω(deg/s)RMSEω(deg/s)ev(m/s)RMSEv(m/s)
CMax [14]11.0114.170.440.5210.7414.330.781.05
Pro-STR [19]5.2310.530.140.259.7511.401.291.44
Ours1.281.830.070.090.791.020.110.14
Ours+CMax [14]3.846.300.210.301.141.720.120.14
Ours+Pro-STR [19]1.181.580.060.091.051.330.110.14
" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.373, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Motion and Structure from Event-based Normal Flow" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.24, + 0.147, + 0.764, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.325, + 0.291, + 0.68, + 0.302 + ], + "angle": 0, + "content": "(a) Results of the continuous-time angular velocity estimator." + }, + { + "type": "image", + "bbox": [ + 0.244, + 0.303, + 0.761, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.244, + 0.338, + 0.762, + 0.436 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.239, + 0.441, + 0.763, + 0.463 + ], + "angle": 0, + "content": "(b) Comparison against two state-of-the-art nonlinear solutions using data segments that involve sudden variations of angular velocity." + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.474, + 0.784, + 0.489 + ], + "angle": 0, + "content": "Fig. 4: Angular velocity estimation using the continuous-time nonlinear solver (Ours)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.519, + 0.788, + 0.625 + ], + "angle": 0, + "content": "6-DoF Motion Tracking. Given known depth, we recover the linear and angular camera velocities on the task of 6-DoF motion estimation. As shown in Tab. 3, our linear solver reports more accurate results than both CMax and Pro-STR. In addition, we surprisingly find that the linear solver can sometimes outperform the combination of itself with the two nonlinear solvers. This is, according to our investigation, attributed to the violation of constant-motion assumption in some sequences, which is held for the two nonlinear solvers." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.648, + 0.61, + 0.663 + ], + "angle": 0, + "content": "5.3 Evaluation of the Continuous-Time Solver" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.789, + 0.842 + ], + "angle": 0, + "content": "To justify the claimed advantage of the proposed continuous-time nonlinear solver, we use the exemplary task of angular velocity estimation. The sequence used (shapesRotation [29]) captures a purely rotating event camera, and some parts of the data involves aggressive motion. As illustrated in Fig. 4a, our nonlinear solver gives accurate estimation results with a high stability. Moreover, we select three segments that witness a sudden variation of angular speed within a short time interval and compare our continuous-time nonlinear solver against CMax and Pro-STR. Note that all of them are initialized by our linear method for a fair comparison. As Fig. 4b shows, our results are more consistent with the groundtruth angular velocity than the other two methods, which are less accurate due to the dependence on the constant-motion assumption." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.127 + ], + "angle": 0, + "content": "Z. Ren et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.221, + 0.145, + 0.78, + 0.16 + ], + "angle": 0, + "content": "Table 4: Runtime of 6-DoF tracking on sequence mountain_normal (640×480 px)." + }, + { + "type": "table", + "bbox": [ + 0.283, + 0.172, + 0.725, + 0.226 + ], + "angle": 0, + "content": "
ComponentRuntime
Normal flow extraction20 ms (for extracting 2000 normal flow vectors)
Linear solver25.97 ms
CT solver28.29 ms
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.251, + 0.351, + 0.265 + ], + "angle": 0, + "content": "5.4 Discussion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.272, + 0.788, + 0.394 + ], + "angle": 0, + "content": "Computational Complexity. The computational complexity of the linear solver depends on the specific task of model fitting. For optical flow and depth estimation, it is linearly proportional to the size of the input event flow, and thus, the complexity is \\(\\mathcal{O}(N)\\). For the other tasks, the time cost predominantly hinges on the total number of RANSAC iterations and the computations performed per iteration. In each iteration, we typically solve a linear overdetermined system via QR decomposition at a complexity of \\(\\mathcal{O}(cN^2)\\), where \\(c\\) is the cardinality of a minimal problem (e.g., \\(c = 3\\) for the problem of angular velocity estimation)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.411, + 0.787, + 0.473 + ], + "angle": 0, + "content": "Runtime. All experiments are conducted on a standard computer with Intel i7-13700F CPU, 16G memory and implemented with MATLAB (version R2022a). Using the 6-DoF tracking task as an example, we measure the runtime of the proposed linear solver and continuous-time (CT) solver, as reported in Tab. 4." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.491, + 0.788, + 0.627 + ], + "angle": 0, + "content": "Limitations. To make it clear, we emphasize that our method, like other compared methods, does not resolve the motion and structure parameters simultaneously. It should also be noted that the success of our method relies heavily on the quality of the input normal flow. Particularly, our normal flow extractor (including some other open source solutions for computing normal flow from event data) may fail in the presence of densely repetitive textures, leading to a failure of the modeling fitting scheme. To assess the sensitivity of our method to the quality of event-based normal flow, we propose an analysis on numerical stability using simulation data in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.646, + 0.371, + 0.663 + ], + "angle": 0, + "content": "6 Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.841 + ], + "angle": 0, + "content": "This work presents a solution to the general problem of geometric model fitting with event data. Considering the differential working principle of event cameras, our method establishes specific constraints between instantaneous motion-and-structure parameters and event-based normal flow. Experiments show that the proposed linear solver leads to closed-form and deterministic solutions, which can be used to initialize existing state-of-the-art nonlinear methods. The proposed continuous-time nonlinear solver, furthermore, demonstrates the capability to handle sudden speed variations, obviating the dependence on the constant-motion assumption. Finally, we hope this work can inspire future research on event-based geometric model fitting, a paramount topic to many downstream techniques, such as event-based 3-D vision, SLAM, and scene segmentation." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.372, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Motion and Structure from Event-based Normal Flow" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.393, + 0.163 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.181, + 0.788, + 0.257 + ], + "angle": 0, + "content": "We thank Dr. Yi Yu for proof reading. This work was supported by the National Key Research and Development Project of China under Grant 2023YFB4706600. Funded by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) under Germany's Excellence Strategy - EXC 2002/1 \"Science of Intelligence\" - project number 390523135." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.284, + 0.323, + 0.3 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.318, + 0.787, + 0.36 + ], + "angle": 0, + "content": "1. Barranco, F., Fermüller, C., Aloimonos, Y., Ros, E.: Joint direct estimation of 3D geometry and 3D motion using spatio temporal gradients. Pattern Recog. 113, 107759 (2021). https://doi.org/10.1016/j.patcog.2020.107759" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.362, + 0.787, + 0.403 + ], + "angle": 0, + "content": "2. Benosman, R., Clercq, C., Lagorce, X., Ieng, S.H., Bartolozzi, C.: Event-based visual flow. IEEE Trans. Neural Netw. Learn. Syst. 25(2), 407-417 (2014). https://doi.org/10.1109/TNNLS.2013.2273537" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.405, + 0.787, + 0.432 + ], + "angle": 0, + "content": "3. Brodsky, T., Fermüller, C., Aloimonos, Y.: Structure from motion: Beyond the epipolar constraint. Int. J. Comput. Vis. 37, 231-258 (2000)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.434, + 0.787, + 0.473 + ], + "angle": 0, + "content": "4. Bryner, S., Gallego, G., Rebecq, H., Scaramuzza, D.: Event-based, direct camera tracking from a photometric 3D map using nonlinear optimization. In: IEEE Int. Conf. Robot. Autom. (ICRA) (2019). https://doi.org/10.1109/ICRA.2019.8794255" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.476, + 0.787, + 0.515 + ], + "angle": 0, + "content": "5. Clady, X., Clercq, C., Ieng, S.H., Houseini, F., Randazzo, M., Natale, L., Bartolozzi, C., Benosman, R.: Asynchronous visual event-based time-to-contact. Front. Neurosci. 8(9) (2014). https://doi.org/10.3389/fnins.2014.00009" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.517, + 0.787, + 0.571 + ], + "angle": 0, + "content": "6. Conradt, J., Cook, M., Berner, R., Lichtsteiner, P., Douglas, R.J., Delbruck, T.: A pencil balancing robot using a pair of AER dynamic vision sensors. In: IEEE Int. Symp. Circuits Syst. (ISCAS). pp. 781-784 (2009). https://doi.org/10.1109/ISCAS.2009.5117867" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.573, + 0.787, + 0.613 + ], + "angle": 0, + "content": "7. Corke, P.: Robotics, Vision and Control: Fundamental Algorithms in MATLAB. Springer Tracts in Advanced Robotics, Springer (2017). https://doi.org/10.1007/978-3-319-54413-7" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.616, + 0.787, + 0.642 + ], + "angle": 0, + "content": "8. De Boor, C.: On calculating with b-splines. Journal of Approximation theory 6(1), 50-62 (1972)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.644, + 0.787, + 0.685 + ], + "angle": 0, + "content": "9. Delbruck, T., Lang, M.: Robotic goalie with 3ms reaction time at \\(4\\%\\) CPU load using event-based dynamic vision sensor. Front. Neurosci. 7, 223 (2013). https://doi.org/10.3389/fnins.2013.00223" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.687, + 0.787, + 0.728 + ], + "angle": 0, + "content": "0. Dinaux, R., Wessendorp, N., Dupeyroux, J., De Croon, G.C.: Faith: Fast iterative half-plane focus of expansion estimation using optic flow. IEEE Robotics and Automation Letters 6(4), 7627-7634 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.729, + 0.787, + 0.755 + ], + "angle": 0, + "content": "1. Fermüller, C.: Passive navigation as a pattern recognition problem. Int. J. Comput. Vis. 14(2), 147-158 (1995)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.757, + 0.787, + 0.797 + ], + "angle": 0, + "content": "12. Fischler, M.A., Bolles, R.C.: Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Commun. ACM 24(6), 381-395 (1981). https://doi.org/10.1145/358669.358692" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.799, + 0.787, + 0.84 + ], + "angle": 0, + "content": "13. Gallego, G., Gehrig, M., Scaramuzza, D.: Focus is all you need: Loss functions for event-based vision. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR). pp. 12272-12281 (2019). https://doi.org/10.1109/CVPR.2019.01256" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.318, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.127 + ], + "angle": 0, + "content": "Z. Ren et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.203 + ], + "angle": 0, + "content": "14. Gallego, G., Rebecq, H., Scaramuzza, D.: A unifying contrast maximization framework for event cameras, with applications to motion, depth, and optical flow estimation. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR). pp. 3867-3876 (2018). https://doi.org/10.1109/CVPR.2018.00407" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.203, + 0.787, + 0.244 + ], + "angle": 0, + "content": "15. Gao, L., Liang, Y., Yang, J., Wu, S., Wang, C., Chen, J., Kneip, L.: Vector: A versatile event-centric benchmark for multi-sensor slam. IEEE Robotics and Automation Letters 7(3), 8217-8224 (2022). https://doi.org/10.1109/LRA.2022.3186770" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.244, + 0.787, + 0.297 + ], + "angle": 0, + "content": "16. Gao, L., Su, H., Gehrig, D., Cannici, M., Scaramuzza, D., Kneip, L.: A 5-point minimal solver for event camera relative motion estimation. In: Int. Conf. Comput. Vis. (ICCV). pp. 8015-8025 (2023). https://doi.org/10.1109/ICCV51070.2023.00739" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.298, + 0.787, + 0.354 + ], + "angle": 0, + "content": "17. Gu, C., Learned-Miller, E., Sheldon, D., Gallego, G., Bideau, P.: The spatiotemporal Poisson point process: A simple model for the alignment of event camera data. In: Int. Conf. Comput. Vis. (ICCV). pp. 13495-13504 (2021). https://doi.org/10.1109/ICCV48922.2021.01324" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.354, + 0.787, + 0.393 + ], + "angle": 0, + "content": "18. Guo, S., Gallego, G.: CMax-SLAM: Event-based rotational-motion bundle adjustment and SLAM system using contrast maximization. IEEE Trans. Robot. pp. 1-20 (2024). https://doi.org/10.1109/TRO.2024.3378443" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.393, + 0.787, + 0.434 + ], + "angle": 0, + "content": "19. Huang, X., Zhang, Y., Xiong, Z.: Progressive spatio-temporal alignment for efficient event-based motion estimation. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR). pp. 1537-1546 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.434, + 0.787, + 0.461 + ], + "angle": 0, + "content": "20. Ji, H., Fermuller, C.: A 3D shape constraint on video. IEEE Trans. Pattern Anal. Mach. Intell. 28(6), 1018-1023 (2006)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.461, + 0.787, + 0.502 + ], + "angle": 0, + "content": "21. Kim, H., Leutenegger, S., Davison, A.J.: Real-time 3D reconstruction and 6-DoF tracking with an event camera. In: Eur. Conf. Comput. Vis. (ECCV). pp. 349-364 (2016). https://doi.org/\\protect{10.1007/978-3-319-46466-4_21}" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.502, + 0.787, + 0.543 + ], + "angle": 0, + "content": "22. Liu, D., Parra, A., Chin, T.J.: Spatiotemporal registration for event-based visual odometry. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR). pp. 4935-4944 (2021). https://doi.org/10.1109/CVPR46437.2021.00490" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.543, + 0.787, + 0.583 + ], + "angle": 0, + "content": "23. Longuet-Higgins, H.C., Prazdny, K.: The interpretation of a moving retinal image. Proceedings of the Royal Society of London. Series B. Biological Sciences 208(1173), 385-397 (1980). https://doi.org/10.1098/rspb.1980.0057" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.583, + 0.787, + 0.61 + ], + "angle": 0, + "content": "24. Lu, X., Zhou, Y., Niu, J., Zhong, S., Shen, S.: Event-based visual inertial velometer (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.61, + 0.787, + 0.637 + ], + "angle": 0, + "content": "25. Ma, Y., Soatto, S., Košecká, J., Sastry, S.: An Invitation to 3-D Vision: From Images to Geometric Models. Springer (2004)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.637, + 0.787, + 0.678 + ], + "angle": 0, + "content": "26. Meyer, F., Bouthemy, P.: Estimation of time-to-collision maps from first order motion models and normal flows. In: 1992 11th IAPR International Conference on Pattern Recognition. vol. 1, pp. 78-82. IEEE Computer Society (1992)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.678, + 0.787, + 0.704 + ], + "angle": 0, + "content": "27. Meyer, F.G.: Time-to-collision from first-order models of the motion field. IEEE transactions on robotics and automation 10(6), 792-798 (1994)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.704, + 0.787, + 0.744 + ], + "angle": 0, + "content": "28. Mueggler, E., Gallego, G., Rebecq, H., Scaramuzza, D.: Continuous-time visual-inertial odometry for event cameras. IEEE Trans. Robot. 34(6), 1425-1440 (Dec 2018). https://doi.org/10.1109/tro.2018.2858287" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.744, + 0.787, + 0.799 + ], + "angle": 0, + "content": "29. Mueggler, E., Rebecq, H., Gallego, G., Delbruck, T., Scaramuzza, D.: The event-camera dataset and simulator: Event-based data for pose estimation, visual odometry, and SLAM. Int. J. Robot. Research 36(2), 142-149 (2017). https://doi.org/10.1177/0278364917691115" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.799, + 0.787, + 0.841 + ], + "angle": 0, + "content": "30. Nunes, U.M., Demiris, Y.: Entropy minimisation framework for event-based vision model estimation. In: Eur. Conf. Comput. Vis. (ECCV). pp. 161-176 (2020). https://doi.org/10.1007/978-3-030-58558-7_10" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.373, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Motion and Structure from Event-based Normal Flow" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "31. Parameshwara, C.M., Hari, G., Fermüller, C., Sanket, N.J., Aloimonos, Y.: Diff-posednet: direct differentiable camera pose estimation. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR). pp. 6845-6854 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.787, + 0.217 + ], + "angle": 0, + "content": "32. Pijnacker Hordijk, B.J., Scheper, K.Y., De Croon, G.C.: Vertical landing for micro air vehicles using event-based optical flow. J. Field Robot. 35(1), 69-90 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.218, + 0.786, + 0.244 + ], + "angle": 0, + "content": "33. Rebecq, H., Gallego, G., Scaramuzza, D.: EMVS: Event-based multi-view stereo. In: British Mach. Vis. Conf. (BMVC) (2016). https://doi.org/10.5244/C.30.63" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.246, + 0.785, + 0.286 + ], + "angle": 0, + "content": "34. Rebecq, H., Horstschaefer, T., Scaramuzza, D.: Real-time visual-inertial odometry for event cameras using keyframe-based nonlinear optimization. In: British Mach. Vis. Conf. (BMVC) (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.288, + 0.787, + 0.34 + ], + "angle": 0, + "content": "35. Rebecq, H., Horstschäfer, T., Gallego, G., Scaramuzza, D.: EVO: A geometric approach to event-based 6-DOF parallel tracking and mapping in real-time. IEEE Robot. Autom. Lett. 2(2), 593-600 (2017). https://doi.org/10.1109/LRA.2016.2645143" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.342, + 0.786, + 0.37 + ], + "angle": 0, + "content": "36. Shiba, S., Aoki, Y., Gallego, G.: Event collapse in contrast maximization frameworks. Sensors 22(14), 1-20 (2022). https://doi.org/10.3390/s22145190" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.371, + 0.786, + 0.411 + ], + "angle": 0, + "content": "37. Shiba, S., Aoki, Y., Gallego, G.: A fast geometric regularizer to mitigate event collapse in the contrast maximization framework. Adv. Intell. Syst. p. 2200251 (2022). https://doi.org/10.1002/aisy.202200251" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.412, + 0.786, + 0.452 + ], + "angle": 0, + "content": "38. Shiba, S., Aoki, Y., Gallego, G.: Secrets of event-based optical flow. In: Eur. Conf. Comput. Vis. (ECCV). pp. 628-645 (2022). https://doi.org/10.1007/978-3-031-19797-0 36" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.786, + 0.493 + ], + "angle": 0, + "content": "39. Stoffregen, T., Gallego, G., Drummond, T., Kleeman, L., Scaramuzza, D.: Event-based motion segmentation by motion compensation. In: Int. Conf. Comput. Vis. (ICCV). pp. 7243-7252 (2019). https://doi.org/10.1109/ICCV.2019.00734" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.495, + 0.786, + 0.522 + ], + "angle": 0, + "content": "40. Trucco, E., Verri, A.: Introductory Techniques for 3-D Computer Vision. Prentice Hall PTR, Upper Saddle River, NJ, USA (1998)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.523, + 0.786, + 0.564 + ], + "angle": 0, + "content": "41. Zhou, Y., Gallego, G., Rebecq, H., Kneip, L., Li, H., Scaramuzza, D.: Semidense 3D reconstruction with a stereo event camera. In: Eur. Conf. Comput. Vis. (ECCV). pp. 242-258 (2018). https://doi.org/10.1007/978-3-030-01246-5_15" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.565, + 0.786, + 0.592 + ], + "angle": 0, + "content": "42. Zhou, Y., Gallego, G., Shen, S.: Event-based stereo visual odometry. IEEE Trans. Robot. 37(5), 1433-1450 (2021). https://doi.org/10.1109/TRO.2021.3062252" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.593, + 0.786, + 0.632 + ], + "angle": 0, + "content": "43. Zhu, A.Z., Chen, Y., Daniilidis, K.: Realtime time synchronized event-based stereo. In: Eur. Conf. Comput. Vis. (ECCV). pp. 438-452 (2018). https://doi.org/10.1007/978-3-030-01231-1_27" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.633, + 0.786, + 0.675 + ], + "angle": 0, + "content": "44. Zhuang, B., Tran, Q.H.: Image stitching and rectification for hand-held cameras. In: Eur. Conf. Comput. Vis. (ECCV) (2020). https://doi.org/10.1007/978-3-030-58571-6_15" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.675 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Motion and Structure from Event-based Normal Flow/2dddd958-d905-447a-8afe-98428ea7b185_origin.pdf b/2024/Motion and Structure from Event-based Normal Flow/2dddd958-d905-447a-8afe-98428ea7b185_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..4aff3bdf960b8b68e619b967050c7f299c15c04b --- /dev/null +++ b/2024/Motion and Structure from Event-based Normal Flow/2dddd958-d905-447a-8afe-98428ea7b185_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ead6bd314cb92ee8490786a3e1790c0014bc972bb80aea3a2e8155c40bd3b3b3 +size 5710742 diff --git a/2024/Motion and Structure from Event-based Normal Flow/full.md b/2024/Motion and Structure from Event-based Normal Flow/full.md new file mode 100644 index 0000000000000000000000000000000000000000..539391ab342b1743b98688bc3da4326b422e261d --- /dev/null +++ b/2024/Motion and Structure from Event-based Normal Flow/full.md @@ -0,0 +1,343 @@ +# Motion and Structure from Event-based Normal Flow + +Zhongyang Ren $^{1*}$ , Bangyan Liao $^{2*}$ , Delei Kong $^{1}$ , Jinghang Li $^{1}$ , Peidong Liu $^{2}$ , Laurent Kneip $^{3}$ , Guillermo Gallego $^{4}$ , and Yi Zhou $^{1*}$ + +$^{1}$ School of Robotics, Hunan University + +$^{2}$ School of Engineering, Westlake University + +3 School of Information Science and Technology, ShanghaiTech University + +TU Berlin, ECDF, SCIoI Excellence Cluster and Robotics Institute Germany + +Abstract. Recovering the camera motion and scene geometry from visual data is a fundamental problem in computer vision. Its success in conventional (frame-based) vision is attributed to the maturity of feature extraction, data association and multi-view geometry. The emergence of asynchronous (event-based) cameras calls for new approaches that use raw event data as input to solve this fundamental problem. State-of-the-art solutions typically infer data association implicitly by iteratively reversing the event data generation process. However, the nonlinear nature of these methods limits their applicability in real-time tasks, and the constant-motion assumption leads to unstable results under agile motion. To this end, we reformulate the problem in a way that aligns better with the differential working principle of event cameras. We show that event-based normal flow can be used, via the proposed geometric error term, as an alternative to the full (optical) flow in solving a family of geometric problems that involve instantaneous first-order kinematics and scene geometry. Furthermore, we develop a fast linear solver and a continuous-time nonlinear solver on top of the proposed geometric error term. Experiments on both synthetic and real data show the superiority of our linear solver in terms of accuracy and efficiency, and its practicality as an initializer for previous nonlinear solvers. Besides, our continuous-time non-linear solver exhibits exceptional capabilities in accommodating sudden variations in motion since it does not rely on the constant-motion assumption. Our project page can be found at https://nail-hnu.github.io/EvLinearSolver/. + +Keywords: Event cameras $\cdot$ Geometric model fitting $\cdot$ Normal flow + +# 1 Introduction + +Neuromorphic event-based cameras are bio-inspired sensors that report only brightness changes at each pixel location asynchronously. Their high temporal + +resolution enables them to sense at exactly the same rate as the scene dynamics, and thus, event cameras are suitable for dealing with robotic perception [33,38, 39,41,43], navigation [21,34,35,42], and control [6,9] tasks that involve agile motion. A general and fundamental problem among these tasks is to recover the camera's ego-motion and scene geometry from event data. One of the bottlenecks for solving this kind of problems at very high speeds is the lack of methods for efficient and robust data association. + +Unlike its standard-vision counterparts, an event camera cannot observe visual information (i.e. intensity changes) when motion is parallel to edges. Such a partial observability of event data makes optical flow estimation (as well as feature detection and matching) an ill-posed problem. To this end, many existing methods [14,17,19,22,30] for solving the problem of motion-model fitting on event data propose specific metrics that assess event-data association established in either an implicit manner or an explicitly model-driven style. The best fit can be retrieved by iteratively optimizing the data association, often expressed in terms of event alignment. However, we argue two flaws of these methods: $(i)$ The success of these non-linear solutions requires a proper initialization. While a brute-force search may be applicable, it is computationally expensive for real-time application, specially if the dimension of the parameter is high. Efficient and deterministic initialization methods are lacking and urgently needed. $(ii)$ The common assumption of constant velocity only holds within a short time interval. Nevertheless, a time window that is too narrow may contain insufficient number of events for a valid model fitting. The heuristic of using a constant number of events may violate the constant velocity assumption, especially in the case of sudden changes in speed. + +In this paper, we rethink the problem formulation in a way that is more consistent with the differential working principle of event cameras. Unlike existing pipelines that use events occurring within a time interval to resolve the relative motion, we attempt to recover the structure and instantaneous first-order kinematic parameters involved in the motion field equation, using as input an instantaneous observation, namely event-based normal flow. To this end, we explore effective constraints between normal-flow observations and model parameters for solving a family of motion-and-structure estimation problems, such as the estimation of depth and full optical flow, instantaneous angular velocity and linear velocity, and differential homography (see Fig. 1). + +The contributions of this paper can be summarized as follows: + +- A normal flow constraint with a geometric connection to event data. It overcomes the partial observability issue (i.e., errors caused by using normal flow as replacement for full optical flow) in general problems of geometric model fitting when normal flow is used as input. +- Two solvers for a family of motion-and-structure estimation problems using as input sparse event-based normal flow. Our linear solver (can be used with RANSAC) leads to closed-form and deterministic solutions that can be used as an initialization to existing nonlinear methods. Our nonlinear solver, + +![](images/d1f0b4c57510e255c6b64fbfb4fe1cca27618059d35bbcd12fc8f37b680f0a43.jpg) +Fig. 1: Overview of our event-based motion estimation approach. Normal flow vectors are computed from the raw event data and are used as input to the two solvers. The result of the linear solver consists of fitted motion models $\{\pmb{\theta}_1, \pmb{\theta}_2, \dots\}$ at discrete time instants, which can be used to initialize the continuous-time nonlinear solver. + +under a continuous-time formulation, can handle sudden speed variations, and thus, is free of the constant-motion assumption. + +- A thorough evaluation on the two proposed solvers, comparing against state-of-the-art methods, and also including an investigation on the performance of existing non-linear methods initialized using our linear solver. + +Outline: The rest of the paper is organized as follows. We first discuss the related work in Sec. 2. Then a compact review of preliminary knowledge is given in Sec. 3. Our approach is presented in Sec. 4, including the normal-flow constraint, the linear models of a family of motion-and-structure estimation problems, and the proposed linear and nonlinear solvers. Experimental evaluation is conducted in Sec. 5 and conclusion are drawn in Sec. 6. + +# 2 Related Work + +# 2.1 Event-based Geometric Model Fitting + +The major body of literature on this specific topic follows an iterative scheme, which typically has its own special metric evaluating quantitatively the association between events, and solve the model fitting as a nonlinear optimization problem. Among the first works is the Contrast Maximization method [13, 14], which provides a unified framework for solving a family of motion-model fitting problems. It searches the optimal model parameters that ultimately lead to a sharp image of warped events (IWE). The goodness of fit is evaluated by the alignment of all involved events along point trajectories on the image plane, and such alignment is assessed by the strength of the contours of the IWE. The computational complexity of CMax is high because of iteratively updating of IWE, whose cost scales with the spatial resolution and the amount of event data. + +To lessen the dependency on IWE, Nunes et al. proposes an entropy minimization framework (EMin) [30] that replaces the dispersion metric of [14] with + +a similarity measurement between events, enabling to evaluate the goodness of fit in a higher dimensional space. EMin is, still, too expensive for real-time application. Liu et al. [22] propose a two-step iterative method for the problem of 3-D rotation estimation using events as input. Given a rotation estimate, the first step determines event correspondences by looking for the nearest neighbour in the spatio-temporal domain. The second step employs a spatio-temporal registration which aligns corresponding events by minimizing the sum of squared geodesic residual. The least-square nature of the energy function in the second step leads to a computation friendly performance. Gu et al. [17] present a probabilistic approach for spatio-temporal alignment of events. The aligned events at a particular pixel are modeled as a spatio-temporal Poisson point process and the goodness of fit can be quantified using the likelihood of the data. The optimal model is retrieved when the likelihood is maximized. More recently, an event-to-map alignment scheme is proposed in [19], which adopts the time surface (TS) map as a reference for event alignment. The proposed TS loss function drives later events to get increasingly closer to pixel positions of corresponding earlier events, thus, leading to a progressively sharp TS map as iterations proceed. + +In summary, all the above-mentioned methods are nonlinear, and thus, a proper initialization is necessary. Although a brute-force search is widely adopted, sometimes it can be too expensive to locate the convergence basin in parameter space. Moreover, [14,30] may come across a convergence to false-positive minima, known as event collapse [36, 37]. Hence, a quick and deterministic solver (e.g., a closed-form solution) is desirable. Besides, most of these methods assume a constant-motion model, which limits their applicability to scenarios of sudden speed variations. In addition, we have noticed that a recent work [16] can jointly estimate the line parameters and the camera's linear velocities. However, this method is not a general one due to the dependence on line structures and the usage of an IMU. Thus, it is not within the scope of our discussion. + +# 2.2 Motion and Structure from Normal Flow + +Due to the aperture problem, a local estimate of optical flow recovers only the partial component of the full flow vector along the direction of image gradient. Preliminary knowledge about normal flow can be found in the supplementary material. Recovering motion and structure from normal flow on intensity images is a well-studied topic. Its main body of literature typically proposes a qualitatively deterministic solution by which a set (even infinite) of applicable ego motions are determined via an analysis of normal flow classification [3, 11, 20]. Only recently, researches devised ways to formulate the Cheirality constraint as a differentiable term and use it in a continuous optimization approach [1] or as a loss function for training an artificial neural network [31]. + +In event-based perception and navigation, similar methodologies as the frame-based ones are also witnessed, where normal flow computed from event data is used as input to determine the Focus of Expansion (FoE) for drones' course estimation [10] and obstacle avoidance [5]. + +Most of these heuristic methods, while effective, are cumbersome for practical application. In contrast, our method delivers a deterministic way to retrieve motion and structure from normal flow. Additionally, it is noticed that normal flow can replace full flow in some specific problems of geometric model fitting, such as affine models [26, 27]. In this way, the geometric model fitting problem can be easily solved in closed-form. Nevertheless, these methods do not consider the observability bias in normal flows, and thus, such naive substitution may lead to inaccurate model fitting. To this end, we present a robust linear solver based on a carefully-designed geometric measurement model that can overcome the partial observability of event-based normal flow. It is worth mentioning that a recent work [24] also determines instantaneous linear velocity using as input the normal flow. Our difference is seen in the scope of applicability. The usage of normal flow is restricted in the motion flow equation of [24], while our paper extends it to a wider scope (i.e., a family of geometric model fitting problems) by combining normal flow with e.g., differential epipolar geometry, differential homography, etc. + +# 3 Preliminaries + +To make this paper self-contained, some preliminary knowledge is revisited. Due to space limitations, here we only list essential concepts. A more detailed review of preliminary knowledge can be found in the supplementary material. + +Motion Field. Consider a perspective camera moving in a static environment, with instantaneous angular and linear velocities in the camera frame $\boldsymbol{\omega} = (\omega_{x},\omega_{y},\omega_{z})^{\top}$ and $\pmb {\nu} = (\nu_{x},\nu_{y},\nu_{z})^{\top}$ , respectively. The motion field at pixel $\mathbf{x} = (x,y)^{\top}$ produced by projecting the 3D motion at point with depth $Z(\mathbf{x})$ is + +$$ +\mathbf {u} = \frac {1}{Z (\mathbf {x})} \mathbf {A} (\mathbf {x}) \boldsymbol {\nu} + \mathbf {B} (\mathbf {x}) \boldsymbol {\omega}, \tag {1} +$$ + +where $\mathbf{A}(\mathbf{x})$ and $\mathbf{B}(\mathbf{x})$ are $2\times 3$ matrices that solely depend on $\mathbf{x}$ . Equation (1) is the motion field equation [40, Ch.8] [23]. Optical flow in rigid scenes aims at estimating the motion field from visual data, which is only possible in regions that do not suffer from the aperture problem. + +Brightness constancy is the assumption stating that the true motion trajectories $\mathbf{x}(t)$ followed by objects on the image plane are those at which brightness is constant: $I(\mathbf{x}(t), t) = \mathrm{const}$ . Differentiating in time (total derivative) gives the equation at $(\mathbf{x}, t)$ : $\nabla I \cdot \dot{\mathbf{x}} + \partial_t I = 0$ , where $\dot{\mathbf{x}} = d\mathbf{x} / dt$ is the velocity of image point $\mathbf{x}(t)$ . Given $I$ , this equation provides one constraint per point $(\mathbf{x}, t)$ to estimate $\dot{\mathbf{x}}$ , now called optical flow. Without additional equations, only the normal component of the optical flow can be determined [40], + +$$ +\mathbf {n} = - \frac {\partial_ {t} I}{\| \nabla I \| ^ {2}} \nabla I. \tag {2} +$$ + +Normal Flow. Decomposing the optical flow $\mathbf{u} = \mathbf{u}_{\perp} + \mathbf{u}_{\parallel}$ into its normal and parallel components to the local image gradient $(\nabla I)$ , the normal flow is $\mathbf{n} = \mathbf{u}_{\perp}$ , and the dot-product with $\mathbf{n}$ gives + +$$ +\mathbf {n} ^ {\top} \mathbf {u} = \| \mathbf {n} \| ^ {2}, \quad \text {o r e q u i v a l e n t l y} \quad (\mathbf {u} - \mathbf {n}) \cdot \mathbf {n} = 0. \tag {3} +$$ + +Differential Homography. (Continuous Homography Matrix [25, Ch.5]). If all scene points lie on a plane, their coordinates are related to their image projections via a homography transformation (which encodes the information of the scene plane $\{\mathbf{N}, d\}$ and the camera pose). In this case, the motion field (1) can be rewritten to adopt the form in homogeneous coordinates + +$$ +\hat {\mathbf {u}} (\mathbf {x}) = \left(\mathbf {1} - \hat {\mathbf {x}} \mathbf {e} _ {3} ^ {\top}\right) \mathbf {H} _ {d} \hat {\mathbf {x}}, \tag {4} +$$ + +where $\mathbf{H}_d\dot{=} -\left([ \boldsymbol {\omega}]_{\times} + d^{-1}\boldsymbol {\nu}\mathbf{N}^{\top}\right)$ denotes the differential homography matrix $([\cdot ]_{\times}$ refers to the cross-product, skew-symmetric matrix), $\mathbf{1}$ is the identity matrix, $\hat{\mathbf{x}} = (x,y,1)^{\top}$ , $\hat{\mathbf{u}} = (u_x,u_y,0)^{\top}$ and $\mathbf{e}_3 = (0,0,1)^{\top}$ . + +Differential Epipolar Geometry. (Continuous Epipolar Geometry [25, Ch.5]). The epipolar geometry examines the distance between a potential match to the determined epipolar line, and can be expressed as $\hat{\mathbf{x}}^{\prime \top}\mathbf{E}\hat{\mathbf{x}} = 0$ . The differential epipolar geometry can be derived analogously, providing a relationship between $\mathbf{x},\mathbf{u}$ and the camera velocities $\pmb {\nu},\pmb{\omega}$ . In homogeneous coordinates: + +$$ +\hat {\mathbf {u}} (\mathbf {x}) ^ {\top} [ \boldsymbol {\nu} ] _ {\times} \hat {\mathbf {x}} - \hat {\mathbf {x}} ^ {\top} \mathbf {s} \hat {\mathbf {x}} = 0, \tag {5} +$$ + +where $\mathbf{s} = \frac{1}{2} ([\pmb {\nu}]_{\times}[\pmb {\omega}]_{\times} + [\pmb {\omega}]_{\times}[\pmb {\nu}]_{\times})$ is a $3\times 3$ symmetric matrix. + +# 4 Methodology + +Figure 1 summarizes our approach, grounded upon event-based normal flow. In this section, we first disclose the normal-flow constraint that can overcome the partial observability of normal flow when replacing full optical flow with normal flow in general tasks of geometric model fitting (Sec. 4.1). By combining the constraint and the theory in Sec. 3, we derive linear equations for a family of motion-and-structure estimation problems (Sec. 4.2). To solve these problems, we present a linear solver and a continuous-time nonlinear solver (Sec. 4.3). + +# 4.1 Normal Flow Constraint + +Normal flow can be computed easily from the event data, for example by fitting planes locally to the space-time "cloud" of events on image space [2, 32]. Specifically, assuming that the event data $e_k \doteq \{\mathbf{x}_k, t_k\}$ spans locally a surface in the spatio-temporal domain, the surface can be defined by a function $\Sigma_e: \mathbb{R}^2 \to \mathbb{R}; \mathbf{x}_k \mapsto t_k$ that maps the pixel coordinates to the latest event's + +![](images/4b6f04504e5b1e1da6053b87c1bffd9164d3a1fd887b5441ec3b879c7da47058.jpg) +Fig. 2: Toy example of applying the proposed normal-flow constraint. (a): A 2D registration task using as input either optical flow or normal flow. The groundtruth displacement is defined by a global flow of $(1.732, -1)^{\top}$ . (b)-(d): The loss landscape obtained using different geometric measurements in the registration task. The red dot denotes the resulting displacement, and the green cross the groundtruth displacement. The bottom-left regions in (b)-(d) display the corresponding registration results. + +![](images/2d17edd895e022ec3c07138772e36bd4fa5c35063e614e26fea68de5c2b3b612.jpg) + +![](images/be2ae15f15e7913c2af0fcdd3c71127563bfd3f69bee8e501cce550b212cf651.jpg) + +![](images/52aecd3ac01157809c0e7563adeadf90b1b1db4698056f539adf29834c99358c.jpg) + +timestamp [2, 5]. The normal flow is parallel to the spatial derivative of $\Sigma_{e}$ , namely $\nabla \Sigma_{e}(\mathbf{x})$ . More specifically, it can be calculated by + +$$ +\mathbf {n} (\mathbf {x}) = \frac {\nabla \Sigma_ {e} (\mathbf {x})}{\left\| \nabla \Sigma_ {e} (\mathbf {x}) \right\| ^ {2}}. \tag {6} +$$ + +Readers may refer to the supplementary material for a more detailed derivation. Surely there are other methods to extract event-based normal flow and the solvers we develop in this section are independent of them, i.e., our solvers can also be fed with the extracted normal flow data from such methods. + +Event-based normal flow cannot be used directly as a replacement for the full flow in (1) to solve for the motion parameters. However, we can leverage the relationship between both flows (3) to define an effective geometric error criterion that circumvents the partial observability of normal flow. If the flow $\mathbf{u}$ depends on some unknown ("state") variables $\pmb{\theta}$ , we propose estimating them based on how well the normal flow constraint (3) is satisfied, e.g., by minimizing some function of the error terms + +$$ +\mathbf {n} (\mathbf {x}) ^ {\top} \mathbf {u} (\mathbf {x}; \boldsymbol {\theta}) - \| \mathbf {n} (\mathbf {x}) \| ^ {2} \doteq \epsilon_ {\mathrm {n f}} \in \mathbb {R}. \tag {7} +$$ + +To illustrate the validity of the proposed normal-flow constraint over the naive way of simply using normal flow to replace full optical flow, we use a 2D registration task as toy example. As shown in Fig. 2, the task consists of registering the solid-line triangle to the dashed-line one using as input either optical flow or normal flow. We observe that using the proposed normal-flow constraint as a geometric measurement leads to a registration result that is almost as accurate as the one given by the full optical flow. However, due to the partial observability of normal flow, simply using it as a replacement for full optical flow leads to a biased registration result, as shown in Fig. 2c. As we see in all following experiments, similar conclusions can be made in general problems of geometric model fitting. + +# 4.2 Models of Motion and Structure + +Based on the derivation in Sec. 4.1, we establish linear models to a family of motion and structure estimation problems using the event-based normal flow as input. In particular, we first discuss how to recover optical flow and depth by introducing the normal-flow constraint into the differential epipolar geometry. Then we show that the combination of the normal-flow constraint with the motion flow equation (1) results in solutions to the problem of instantaneous angular velocity estimation, and 6-DoF motion estimation given depth as prior, respectively. Finally, we show how to compute the differential homography. + +Optical Flow. Given as input the accurate normal flow, and camera velocities, the full optical flow $\mathbf{u}(\mathbf{x})$ at pixel location $\mathbf{x}$ can be calculated by solving the following linear system obtained from the combination of the normal-flow constraint (3) and the differential epipolar geometry (5): + +$$ +\binom {\mathbf {n} (\mathbf {x}) ^ {\top}} {- \hat {\mathbf {x}} ^ {\top} [ \boldsymbol {\nu} ] _ {\times}} \mathbf {u} (\mathbf {x}) = \binom {\| \mathbf {n} (\mathbf {x}) \| ^ {2}} {\hat {\mathbf {x}} ^ {\top} \mathbf {s} \hat {\mathbf {x}}}. \tag {8} +$$ + +Depth Estimation. Similarly, if the normal flow and camera velocities are known, the scene depth $Z(\mathbf{x})$ can be determined by combining the normal-flow constraint (3) with the motion field eq. (1), yielding the closed-form solution: + +$$ +Z (\mathbf {x}) = \frac {\mathbf {n} (\mathbf {x}) ^ {\top} \mathbf {A} (\mathbf {x}) \boldsymbol {\nu}}{\| \mathbf {n} (\mathbf {x}) \| ^ {2} - \mathbf {n} (\mathbf {x}) ^ {\top} \mathbf {B} (\mathbf {x}) \boldsymbol {\omega}}. \tag {9} +$$ + +Angular Velocity Estimation. Assuming that an event camera undergoes a pure rotational motion, the optical flow $\mathbf{u}$ in (1) can be simplified as + +$$ +\mathbf {u} (\mathbf {x}) = \mathbf {B} (\mathbf {x}) \boldsymbol {\omega}. \tag {10} +$$ + +By substituting (10) into (3), we have the linear equation in $\omega$ : + +$$ +\mathbf {n} (\mathbf {x}) ^ {\top} \mathbf {B} (\mathbf {x}) \omega = \| \mathbf {n} (\mathbf {x}) \| ^ {2}. \tag {11} +$$ + +Since each normal flow provides only one constraint, the minimal problem requires three normal flow vectors as input. + +6-DoF Motion Estimation. Assume the event camera undergoes a general 6-DoF motion with known depth of the observed scene, we can solve the first-order kinematics, namely instantaneous linear and angular velocities, by solving the following linear system obtained from the normal-flow constraint (3) and the motion-flow equation (1): + +$$ +\mathbf {n} (\mathbf {x}) ^ {\top} \mathbf {D} (\mathbf {x}) \left[ \begin{array}{l} \boldsymbol {\nu} \\ \boldsymbol {\omega} \end{array} \right] = \| \mathbf {n} (\mathbf {x}) \| ^ {2}, \tag {12} +$$ + +where $\mathbf{D}(\mathbf{x}) = \left[\frac{1}{Z(\mathbf{x})}\mathbf{A}(\mathbf{x})\mathbf{B}(\mathbf{x})\right]$ is the $2\times 6$ feature sensitivity matrix [4, 7]. The minimal problem in this context requires six normal flow measurements to determine the 6-DoF motion parameters. + +Motion and Structure Estimation in Planar Scenes. For a moving camera that observes a planar scene, the motion and structure parameters are compatibly encoded in a homography matrix. In our case, we attempt to recover the differential homography $\mathbf{h}_d$ (the vector form of $\mathbf{H}_d$ in (4)) because the instantaneous linear and angular velocities are considered. By expanding (4) and rearranging equations, we have + +$$ +\mathbf {u} (\mathbf {x}) = \mathbf {C} (\mathbf {x}) \mathbf {h} _ {d}, \tag {13} +$$ + +where $\mathbf{C}(\mathbf{x})$ is a $2 \times 9$ matrix that solely depends on the pixel location $\mathbf{x}$ . Subsequently, substituting (3) into (13), gives: + +$$ +\mathbf {n} (\mathbf {x}) ^ {\top} \mathbf {C} (\mathbf {x}) \mathbf {h} _ {d} = \| \mathbf {n} (\mathbf {x}) \| ^ {2}. \tag {14} +$$ + +As discussed in [44], retrieving the motion and structure parameters from the decomposition of $\mathbf{h}_d$ is not trivial and beyond the scope of this paper. Therefore, we only focus on the numerical solution of $\mathbf{h}_d$ . + +# 4.3 Our Solvers + +According to the motion-and-structure models derived in Sec. 4.2, we have established certain constraints between the model parameters and the normal flow observations. This allows us to simply develop linear solvers for each specific problem. Besides, we further propose a continuous-time nonlinear solver that can effectively handle the presence of sudden variations in motion and does not rely on the constant-motion assumption. To convey our solvers concisely, we generalize the motion-and-structure models into a unified form, as + +$$ +\mathbf {n} (\mathbf {x}) ^ {\top} \mathbf {O} (\mathbf {x}) \boldsymbol {\theta} = \| \mathbf {n} (\mathbf {x}) \| ^ {2}, \tag {15} +$$ + +where $\mathbf{O}(\mathbf{x})$ is the system matrix and $\theta$ are the motion/structure parameters. + +Linear Solver. Given as input a set (e.g., K) of sparse normal-flow estimates obtained at the same time, our linear solver works out the following (overdetermined) linear system of equations: + +$$ +\left( \begin{array}{c} \mathbf {n} \left(\mathbf {x} _ {1}\right) ^ {\top} \mathbf {O} \left(\mathbf {x} _ {1}\right) \\ \vdots \\ \mathbf {n} \left(\mathbf {x} _ {\mathrm {K}}\right) ^ {\top} \mathbf {O} \left(\mathbf {x} _ {\mathrm {K}}\right) \end{array} \right) \boldsymbol {\theta} = \left( \begin{array}{c} \| \mathbf {n} \left(\mathbf {x} _ {1}\right) \| ^ {2} \\ \vdots \\ \| \mathbf {n} \left(\mathbf {x} _ {\mathrm {K}}\right) \| ^ {2} \end{array} \right). \tag {16} +$$ + +For robust estimation, RANSAC [12] is adopted to deal with noise and outliers. + +Continuous-Time Nonlinear Solver. To handle the case of a sudden variation in speed that may violate the constant-motion assumption, we further propose a continuous-time nonlinear solver. In particular, we employ cubic B-splines to represent the continuous-time model parameters as + +$$ +\boldsymbol {\theta} (u (t)) = \sum_ {i = 0} ^ {3} B _ {i, 3} (u (t)) \mathbf {c} _ {i}, \tag {17} +$$ + +where $B_{i,k}$ denotes the basis function, $i$ the index of control points, $k (= 3)$ the order of the spline and $\mathbf{c}_i \in \mathbb{R}^N$ the corresponding control point defined in the $N$ -dimensional space of model parameters. Examples of cubic B-splines applied to event camera motion estimation are found in [18, 28]. The function $u(\cdot)$ is a normalization operator, which transfers time $t$ to the spline's parameter domain by means of basis translation [8]. In this way, the model parameters at any time instant $t$ can be obtained via an interpolation using the known control points. + +The formulation of the continuous-time model fitting problem goes as follows. The unknown variable consists of all the involved control points, $\mathcal{X} \doteq [\mathbf{c}_0, \mathbf{c}_1, \dots, \mathbf{c}_n]$ , where $n$ denotes the total number of control points. By introducing the continuous-time representation of model parameters to (15), we finally create the following objective function: + +$$ +\boldsymbol {\mathcal {X}} ^ {*} = \arg \min _ {\boldsymbol {\mathcal {X}}} \sum_ {e _ {k} \in \mathcal {E}} \left\| \mathbf {n} \left(\mathbf {x} _ {k}\right) ^ {\top} \mathbf {O} \left(\mathbf {x} _ {k}\right) \boldsymbol {\theta} \left(u \left(t _ {k}\right)\right) - \| \mathbf {n} \left(\mathbf {x} _ {k}\right) \| ^ {2} \right\| ^ {2}, \tag {18} +$$ + +where $\mathcal{E}$ denotes the event set in which the norm flow observation is available for every involved event $e_k$ . Note that this continuous-time formulation aligns better with the asynchronous nature of event-based normal flow, enabling a state estimation over a larger temporal window that may lead to a better signal-to-noise ratio. The linear solver can be used to initialize the continuous-time nonlinear solver. An illustrative explanation of our solvers is given in Fig. 1. + +# 5 Experiments + +To justify the effectiveness of the proposed algorithms, experiments are conducted on both simulation and real data. First, we introduce the datasets used and the applied metrics for evaluation (Sec. 5.1). Then we report the results of our linear solver and continuous-time nonlinear solver (Sec. 5.2 and Sec. 5.3). Finally, a discussion on the computational complexity and the limitation of our method is provided in Sec. 5.4. + +# 5.1 Experimental Setup + +Our evaluation utilizes four datasets, including the indoor event-camera dataset (ECD) captured with a DAVIS240C camera and motion-capture system for groundtruth, and a high-resolution dataset with two rotational motion sequences + +![](images/6d5184372b1a0e4636446f974281d9e865262d9d32e175502a23e5068bfac67b.jpg) +(a) Estimated optical flow over a map of event accumulation. + +![](images/b484b8710ef572824360f0a0292a6cea7e4fcfc3d925bc5dcd9f1dbedecd369b.jpg) +(b) Reconstruction result displayed over groundtruth structures (a v-shape roof) and viewed from different perspectives. + +![](images/dde771b4898659557cdb422227caec1f8c8fce581a2f70c55816fd436d16d6d3.jpg) +Fig. 3: Qualitative result of optical flow and depth on sequence three_wall Translation of our synthetic data. Note that our method returns sparse results. + +using an iniVation DAVIS346 camera and Xsens IMU. The VECTor dataset, captured with a Prophesee Gen 3 camera and an accompanying RGB-D depth camera, is used for 6-DoF motion tracking. Additionally, a synthetic dataset generated in Blender is employed for motion-and-structure pattern analysis. Evaluation metrics for kinematics estimation include average error (AE) and root mean squared error (RMSE), while differential homography estimation is assessed using the Frobenius norm. Best results are in bold, and second-best are underlined. The supplementary material provides detailed descriptions of datasets and metrics, as well as implementation details of our methods. + +# 5.2 Evaluation of the Linear Solver + +In the following evaluation, we qualitatively and quantitatively assess the performance of our linear method (Ours) in each specific task, and compare against different state-of-the-art (SOTA) methods (CMax [14] and Pro-STR [19]) and their customized version initialized using our linear solver. + +Optical Flow and Depth Estimation. Due to a lack of benchmark for extensive evaluation, we qualitatively evaluate our method on the tasks of optical flow and depth estimation. As illustrated in Fig. 3a, our linear solver demonstrates the capability to accurately reconstruct full flow using normal flow and known motion parameters. An example of the depth estimation result is given in Fig. 3b, showing that the recovered structure aligns well with the groundtruth. + +Angular Velocity Estimation. As shown in Tab. 1, we see that our linear solver outperforms CMax and Pro-STR. Besides, we find the combination of our linear solver and CMax leads to the best result, indicating that the proposed linear solver is a complementary method to existing nonlinear methods. + +Differential Homography Estimation. The result is consistent with that in the above task. Table 2 reveals that our linear method outperforms CMax and Pro-STR in terms of numerical accuracy. The combination of our linear solver and state-of-the-art nonlinear solver leads to the best result. + +Table 1: Evaluation of our linear solver on the task of angular velocity estimation. + +
MethodshapesRotation [29]groundRotation (our data)patternsRotation (our data)
eω(deg/s)RMSEω(deg/s)eω(deg/s)RMSEω(deg/s)eω(deg/s)RMSEω(deg/s)
CMax [14]15.6026.389.4815.585.959.07
Pro-STR [19]90.61145.8131.0744.7410.4711.88
Ours15.3623.983.308.014.706.08
Ours+CMax [14]10.1214.292.853.740.350.73
Ours+Pro-STR [19]19.2340.449.1317.074.315.23
MethoddynamicRotation [29]boxesRotation (our data)cubesRotation (our data)
eω(deg/s)RMSEω(deg/s)eω(deg/s)RMSEω(deg/s)eω(deg/s)RMSEω(deg/s)
CMax [14]16.1426.488.1813.566.7710.62
Pro-STR [19]54.6485.6322.7232.0511.0712.79
Ours9.4118.062.604.782.313.02
Ours+CMax [14]7.7110.852.643.800.490.77
Ours+Pro-STR [19]18.8732.653.495.872.042.76
+ +Table 2: Evaluation of our linear solver for differential homography estimation. + +
Methodboard Slow [15]patterns rotationpatterns 6 dof
F-NormF-NormF-Norm
CMax [14]1.161.331.27
Pro-STR [19]2.071.221.37
Ours0.240.200.43
Ours+CMax [14]0.230.150.41
Ours+Pro-STR [19]0.190.170.42
+ +Table 3: Evaluation of our linear solver on the task of 6-DoF motion tracking. + +
Methodcorner Slow [15]patterns _6 dof (our data)
eω(deg/s)RMSEω(deg/s)ev(m/s)RMSEv(m/s)eω(deg/s)RMSEω(deg/s)ev(m/s)RMSEv(m/s)
CMax [14]12.3914.390.340.4112.9415.681.021.23
Pro-STR [19]5.197.840.130.2211.2113.491.331.49
Ours1.311.790.010.020.560.890.120.16
Ours+CMax [14]2.192.820.110.170.550.870.180.30
Ours+Pro-STR [19]1.061.340.040.051.071.560.120.15
Methodmountain_normal [15]cubes _6 dof (our data)
eω(deg/s)RMSEω(deg/s)ev(m/s)RMSEv(m/s)eω(deg/s)RMSEω(deg/s)ev(m/s)RMSEv(m/s)
CMax [14]11.0114.170.440.5210.7414.330.781.05
Pro-STR [19]5.2310.530.140.259.7511.401.291.44
Ours1.281.830.070.090.791.020.110.14
Ours+CMax [14]3.846.300.210.301.141.720.120.14
Ours+Pro-STR [19]1.181.580.060.091.051.330.110.14
+ +![](images/ed1bf464dcf21f8e30c6fc27c6c4a7026b849de293306d384b183408bec5592c.jpg) + +![](images/9dbb303a4ba45eb7fccc2c30a6525065060729a1d1244f7888c7ce80fb08dfcd.jpg) +(a) Results of the continuous-time angular velocity estimator. + +![](images/80db2d36fcde1c23580460abe1e6b7dc7d0034b32ea1f45813e038fc04168c64.jpg) +(b) Comparison against two state-of-the-art nonlinear solutions using data segments that involve sudden variations of angular velocity. +Fig. 4: Angular velocity estimation using the continuous-time nonlinear solver (Ours). + +6-DoF Motion Tracking. Given known depth, we recover the linear and angular camera velocities on the task of 6-DoF motion estimation. As shown in Tab. 3, our linear solver reports more accurate results than both CMax and Pro-STR. In addition, we surprisingly find that the linear solver can sometimes outperform the combination of itself with the two nonlinear solvers. This is, according to our investigation, attributed to the violation of constant-motion assumption in some sequences, which is held for the two nonlinear solvers. + +# 5.3 Evaluation of the Continuous-Time Solver + +To justify the claimed advantage of the proposed continuous-time nonlinear solver, we use the exemplary task of angular velocity estimation. The sequence used (shapesRotation [29]) captures a purely rotating event camera, and some parts of the data involves aggressive motion. As illustrated in Fig. 4a, our nonlinear solver gives accurate estimation results with a high stability. Moreover, we select three segments that witness a sudden variation of angular speed within a short time interval and compare our continuous-time nonlinear solver against CMax and Pro-STR. Note that all of them are initialized by our linear method for a fair comparison. As Fig. 4b shows, our results are more consistent with the groundtruth angular velocity than the other two methods, which are less accurate due to the dependence on the constant-motion assumption. + +Table 4: Runtime of 6-DoF tracking on sequence mountain_normal (640×480 px). + +
ComponentRuntime
Normal flow extraction20 ms (for extracting 2000 normal flow vectors)
Linear solver25.97 ms
CT solver28.29 ms
+ +# 5.4 Discussion + +Computational Complexity. The computational complexity of the linear solver depends on the specific task of model fitting. For optical flow and depth estimation, it is linearly proportional to the size of the input event flow, and thus, the complexity is $\mathcal{O}(N)$ . For the other tasks, the time cost predominantly hinges on the total number of RANSAC iterations and the computations performed per iteration. In each iteration, we typically solve a linear overdetermined system via QR decomposition at a complexity of $\mathcal{O}(cN^2)$ , where $c$ is the cardinality of a minimal problem (e.g., $c = 3$ for the problem of angular velocity estimation). + +Runtime. All experiments are conducted on a standard computer with Intel i7-13700F CPU, 16G memory and implemented with MATLAB (version R2022a). Using the 6-DoF tracking task as an example, we measure the runtime of the proposed linear solver and continuous-time (CT) solver, as reported in Tab. 4. + +Limitations. To make it clear, we emphasize that our method, like other compared methods, does not resolve the motion and structure parameters simultaneously. It should also be noted that the success of our method relies heavily on the quality of the input normal flow. Particularly, our normal flow extractor (including some other open source solutions for computing normal flow from event data) may fail in the presence of densely repetitive textures, leading to a failure of the modeling fitting scheme. To assess the sensitivity of our method to the quality of event-based normal flow, we propose an analysis on numerical stability using simulation data in the supplementary material. + +# 6 Conclusions + +This work presents a solution to the general problem of geometric model fitting with event data. Considering the differential working principle of event cameras, our method establishes specific constraints between instantaneous motion-and-structure parameters and event-based normal flow. Experiments show that the proposed linear solver leads to closed-form and deterministic solutions, which can be used to initialize existing state-of-the-art nonlinear methods. The proposed continuous-time nonlinear solver, furthermore, demonstrates the capability to handle sudden speed variations, obviating the dependence on the constant-motion assumption. Finally, we hope this work can inspire future research on event-based geometric model fitting, a paramount topic to many downstream techniques, such as event-based 3-D vision, SLAM, and scene segmentation. + +# Acknowledgments + +We thank Dr. Yi Yu for proof reading. This work was supported by the National Key Research and Development Project of China under Grant 2023YFB4706600. Funded by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) under Germany's Excellence Strategy - EXC 2002/1 "Science of Intelligence" - project number 390523135. + +# References + +1. Barranco, F., Fermüller, C., Aloimonos, Y., Ros, E.: Joint direct estimation of 3D geometry and 3D motion using spatio temporal gradients. Pattern Recog. 113, 107759 (2021). https://doi.org/10.1016/j.patcog.2020.107759 +2. Benosman, R., Clercq, C., Lagorce, X., Ieng, S.H., Bartolozzi, C.: Event-based visual flow. IEEE Trans. Neural Netw. Learn. Syst. 25(2), 407-417 (2014). https://doi.org/10.1109/TNNLS.2013.2273537 +3. Brodsky, T., Fermüller, C., Aloimonos, Y.: Structure from motion: Beyond the epipolar constraint. Int. J. Comput. Vis. 37, 231-258 (2000) +4. Bryner, S., Gallego, G., Rebecq, H., Scaramuzza, D.: Event-based, direct camera tracking from a photometric 3D map using nonlinear optimization. In: IEEE Int. Conf. Robot. Autom. (ICRA) (2019). https://doi.org/10.1109/ICRA.2019.8794255 +5. Clady, X., Clercq, C., Ieng, S.H., Houseini, F., Randazzo, M., Natale, L., Bartolozzi, C., Benosman, R.: Asynchronous visual event-based time-to-contact. Front. Neurosci. 8(9) (2014). https://doi.org/10.3389/fnins.2014.00009 +6. Conradt, J., Cook, M., Berner, R., Lichtsteiner, P., Douglas, R.J., Delbruck, T.: A pencil balancing robot using a pair of AER dynamic vision sensors. In: IEEE Int. Symp. Circuits Syst. (ISCAS). pp. 781-784 (2009). https://doi.org/10.1109/ISCAS.2009.5117867 +7. Corke, P.: Robotics, Vision and Control: Fundamental Algorithms in MATLAB. Springer Tracts in Advanced Robotics, Springer (2017). https://doi.org/10.1007/978-3-319-54413-7 +8. De Boor, C.: On calculating with b-splines. Journal of Approximation theory 6(1), 50-62 (1972) +9. Delbruck, T., Lang, M.: Robotic goalie with 3ms reaction time at $4\%$ CPU load using event-based dynamic vision sensor. Front. Neurosci. 7, 223 (2013). https://doi.org/10.3389/fnins.2013.00223 +0. Dinaux, R., Wessendorp, N., Dupeyroux, J., De Croon, G.C.: Faith: Fast iterative half-plane focus of expansion estimation using optic flow. IEEE Robotics and Automation Letters 6(4), 7627-7634 (2021) +1. Fermüller, C.: Passive navigation as a pattern recognition problem. Int. J. Comput. Vis. 14(2), 147-158 (1995) +12. Fischler, M.A., Bolles, R.C.: Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Commun. ACM 24(6), 381-395 (1981). https://doi.org/10.1145/358669.358692 +13. Gallego, G., Gehrig, M., Scaramuzza, D.: Focus is all you need: Loss functions for event-based vision. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR). pp. 12272-12281 (2019). https://doi.org/10.1109/CVPR.2019.01256 + +14. Gallego, G., Rebecq, H., Scaramuzza, D.: A unifying contrast maximization framework for event cameras, with applications to motion, depth, and optical flow estimation. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR). pp. 3867-3876 (2018). https://doi.org/10.1109/CVPR.2018.00407 +15. Gao, L., Liang, Y., Yang, J., Wu, S., Wang, C., Chen, J., Kneip, L.: Vector: A versatile event-centric benchmark for multi-sensor slam. IEEE Robotics and Automation Letters 7(3), 8217-8224 (2022). https://doi.org/10.1109/LRA.2022.3186770 +16. Gao, L., Su, H., Gehrig, D., Cannici, M., Scaramuzza, D., Kneip, L.: A 5-point minimal solver for event camera relative motion estimation. In: Int. Conf. Comput. Vis. (ICCV). pp. 8015-8025 (2023). https://doi.org/10.1109/ICCV51070.2023.00739 +17. Gu, C., Learned-Miller, E., Sheldon, D., Gallego, G., Bideau, P.: The spatiotemporal Poisson point process: A simple model for the alignment of event camera data. In: Int. Conf. Comput. Vis. (ICCV). pp. 13495-13504 (2021). https://doi.org/10.1109/ICCV48922.2021.01324 +18. Guo, S., Gallego, G.: CMax-SLAM: Event-based rotational-motion bundle adjustment and SLAM system using contrast maximization. IEEE Trans. Robot. pp. 1-20 (2024). https://doi.org/10.1109/TRO.2024.3378443 +19. Huang, X., Zhang, Y., Xiong, Z.: Progressive spatio-temporal alignment for efficient event-based motion estimation. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR). pp. 1537-1546 (2023) +20. Ji, H., Fermuller, C.: A 3D shape constraint on video. IEEE Trans. Pattern Anal. Mach. Intell. 28(6), 1018-1023 (2006) +21. Kim, H., Leutenegger, S., Davison, A.J.: Real-time 3D reconstruction and 6-DoF tracking with an event camera. In: Eur. Conf. Comput. Vis. (ECCV). pp. 349-364 (2016). https://doi.org/\protect{10.1007/978-3-319-46466-4_21} +22. Liu, D., Parra, A., Chin, T.J.: Spatiotemporal registration for event-based visual odometry. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR). pp. 4935-4944 (2021). https://doi.org/10.1109/CVPR46437.2021.00490 +23. Longuet-Higgins, H.C., Prazdny, K.: The interpretation of a moving retinal image. Proceedings of the Royal Society of London. Series B. Biological Sciences 208(1173), 385-397 (1980). https://doi.org/10.1098/rspb.1980.0057 +24. Lu, X., Zhou, Y., Niu, J., Zhong, S., Shen, S.: Event-based visual inertial velometer (2024) +25. Ma, Y., Soatto, S., Košecká, J., Sastry, S.: An Invitation to 3-D Vision: From Images to Geometric Models. Springer (2004) +26. Meyer, F., Bouthemy, P.: Estimation of time-to-collision maps from first order motion models and normal flows. In: 1992 11th IAPR International Conference on Pattern Recognition. vol. 1, pp. 78-82. IEEE Computer Society (1992) +27. Meyer, F.G.: Time-to-collision from first-order models of the motion field. IEEE transactions on robotics and automation 10(6), 792-798 (1994) +28. Mueggler, E., Gallego, G., Rebecq, H., Scaramuzza, D.: Continuous-time visual-inertial odometry for event cameras. IEEE Trans. Robot. 34(6), 1425-1440 (Dec 2018). https://doi.org/10.1109/tro.2018.2858287 +29. Mueggler, E., Rebecq, H., Gallego, G., Delbruck, T., Scaramuzza, D.: The event-camera dataset and simulator: Event-based data for pose estimation, visual odometry, and SLAM. Int. J. Robot. Research 36(2), 142-149 (2017). https://doi.org/10.1177/0278364917691115 +30. Nunes, U.M., Demiris, Y.: Entropy minimisation framework for event-based vision model estimation. In: Eur. Conf. Comput. Vis. (ECCV). pp. 161-176 (2020). https://doi.org/10.1007/978-3-030-58558-7_10 + +31. Parameshwara, C.M., Hari, G., Fermüller, C., Sanket, N.J., Aloimonos, Y.: Diff-posednet: direct differentiable camera pose estimation. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR). pp. 6845-6854 (2022) +32. Pijnacker Hordijk, B.J., Scheper, K.Y., De Croon, G.C.: Vertical landing for micro air vehicles using event-based optical flow. J. Field Robot. 35(1), 69-90 (2018) +33. Rebecq, H., Gallego, G., Scaramuzza, D.: EMVS: Event-based multi-view stereo. In: British Mach. Vis. Conf. (BMVC) (2016). https://doi.org/10.5244/C.30.63 +34. Rebecq, H., Horstschaefer, T., Scaramuzza, D.: Real-time visual-inertial odometry for event cameras using keyframe-based nonlinear optimization. In: British Mach. Vis. Conf. (BMVC) (2017) +35. Rebecq, H., Horstschäfer, T., Gallego, G., Scaramuzza, D.: EVO: A geometric approach to event-based 6-DOF parallel tracking and mapping in real-time. IEEE Robot. Autom. Lett. 2(2), 593-600 (2017). https://doi.org/10.1109/LRA.2016.2645143 +36. Shiba, S., Aoki, Y., Gallego, G.: Event collapse in contrast maximization frameworks. Sensors 22(14), 1-20 (2022). https://doi.org/10.3390/s22145190 +37. Shiba, S., Aoki, Y., Gallego, G.: A fast geometric regularizer to mitigate event collapse in the contrast maximization framework. Adv. Intell. Syst. p. 2200251 (2022). https://doi.org/10.1002/aisy.202200251 +38. Shiba, S., Aoki, Y., Gallego, G.: Secrets of event-based optical flow. In: Eur. Conf. Comput. Vis. (ECCV). pp. 628-645 (2022). https://doi.org/10.1007/978-3-031-19797-0 36 +39. Stoffregen, T., Gallego, G., Drummond, T., Kleeman, L., Scaramuzza, D.: Event-based motion segmentation by motion compensation. In: Int. Conf. Comput. Vis. (ICCV). pp. 7243-7252 (2019). https://doi.org/10.1109/ICCV.2019.00734 +40. Trucco, E., Verri, A.: Introductory Techniques for 3-D Computer Vision. Prentice Hall PTR, Upper Saddle River, NJ, USA (1998) +41. Zhou, Y., Gallego, G., Rebecq, H., Kneip, L., Li, H., Scaramuzza, D.: Semidense 3D reconstruction with a stereo event camera. In: Eur. Conf. Comput. Vis. (ECCV). pp. 242-258 (2018). https://doi.org/10.1007/978-3-030-01246-5_15 +42. Zhou, Y., Gallego, G., Shen, S.: Event-based stereo visual odometry. IEEE Trans. Robot. 37(5), 1433-1450 (2021). https://doi.org/10.1109/TRO.2021.3062252 +43. Zhu, A.Z., Chen, Y., Daniilidis, K.: Realtime time synchronized event-based stereo. In: Eur. Conf. Comput. Vis. (ECCV). pp. 438-452 (2018). https://doi.org/10.1007/978-3-030-01231-1_27 +44. Zhuang, B., Tran, Q.H.: Image stitching and rectification for hand-held cameras. In: Eur. Conf. Comput. Vis. (ECCV) (2020). https://doi.org/10.1007/978-3-030-58571-6_15 \ No newline at end of file diff --git a/2024/Motion and Structure from Event-based Normal Flow/images.zip b/2024/Motion and Structure from Event-based Normal Flow/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..293429c02dd39378c920e95fd05ec8e63b2b8b19 --- /dev/null +++ b/2024/Motion and Structure from Event-based Normal Flow/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6d4be248968328dd4f318c8c26f3705cbe9c4bfa0093156ab4e82e24f9a85024 +size 452098 diff --git a/2024/Motion and Structure from Event-based Normal Flow/layout.json b/2024/Motion and Structure from Event-based Normal Flow/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..44f3ade6da80359ae497af7bb66152246c9577c9 --- /dev/null +++ b/2024/Motion and Structure from Event-based Normal Flow/layout.json @@ -0,0 +1,9283 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 166, + 111, + 447, + 145 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 111, + 447, + 145 + ], + "spans": [ + { + "bbox": [ + 166, + 111, + 447, + 145 + ], + "type": "text", + "content": "Motion and Structure from Event-based Normal Flow" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 144, + 167, + 469, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 144, + 167, + 469, + 193 + ], + "spans": [ + { + "bbox": [ + 144, + 167, + 469, + 193 + ], + "type": "text", + "content": "Zhongyang Ren" + }, + { + "bbox": [ + 144, + 167, + 469, + 193 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 144, + 167, + 469, + 193 + ], + "type": "text", + "content": ", Bangyan Liao" + }, + { + "bbox": [ + 144, + 167, + 469, + 193 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 144, + 167, + 469, + 193 + ], + "type": "text", + "content": ", Delei Kong" + }, + { + "bbox": [ + 144, + 167, + 469, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 144, + 167, + 469, + 193 + ], + "type": "text", + "content": ", Jinghang Li" + }, + { + "bbox": [ + 144, + 167, + 469, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 144, + 167, + 469, + 193 + ], + "type": "text", + "content": ", Peidong Liu" + }, + { + "bbox": [ + 144, + 167, + 469, + 193 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 144, + 167, + 469, + 193 + ], + "type": "text", + "content": ", Laurent Kneip" + }, + { + "bbox": [ + 144, + 167, + 469, + 193 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 144, + 167, + 469, + 193 + ], + "type": "text", + "content": ", Guillermo Gallego" + }, + { + "bbox": [ + 144, + 167, + 469, + 193 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 144, + 167, + 469, + 193 + ], + "type": "text", + "content": ", and Yi Zhou" + }, + { + "bbox": [ + 144, + 167, + 469, + 193 + ], + "type": "inline_equation", + "content": "^{1*}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 225, + 201, + 389, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 201, + 389, + 213 + ], + "spans": [ + { + "bbox": [ + 225, + 201, + 389, + 213 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 225, + 201, + 389, + 213 + ], + "type": "text", + "content": " School of Robotics, Hunan University" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 214, + 213, + 399, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 214, + 213, + 399, + 224 + ], + "spans": [ + { + "bbox": [ + 214, + 213, + 399, + 224 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 214, + 213, + 399, + 224 + ], + "type": "text", + "content": " School of Engineering, Westlake University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 156, + 224, + 457, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 156, + 224, + 457, + 236 + ], + "spans": [ + { + "bbox": [ + 156, + 224, + 457, + 236 + ], + "type": "text", + "content": "3 School of Information Science and Technology, ShanghaiTech University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 146, + 235, + 468, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 235, + 468, + 246 + ], + "spans": [ + { + "bbox": [ + 146, + 235, + 468, + 246 + ], + "type": "text", + "content": "TU Berlin, ECDF, SCIoI Excellence Cluster and Robotics Institute Germany" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 159, + 275, + 455, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 275, + 455, + 539 + ], + "spans": [ + { + "bbox": [ + 159, + 275, + 455, + 539 + ], + "type": "text", + "content": "Abstract. Recovering the camera motion and scene geometry from visual data is a fundamental problem in computer vision. Its success in conventional (frame-based) vision is attributed to the maturity of feature extraction, data association and multi-view geometry. The emergence of asynchronous (event-based) cameras calls for new approaches that use raw event data as input to solve this fundamental problem. State-of-the-art solutions typically infer data association implicitly by iteratively reversing the event data generation process. However, the nonlinear nature of these methods limits their applicability in real-time tasks, and the constant-motion assumption leads to unstable results under agile motion. To this end, we reformulate the problem in a way that aligns better with the differential working principle of event cameras. We show that event-based normal flow can be used, via the proposed geometric error term, as an alternative to the full (optical) flow in solving a family of geometric problems that involve instantaneous first-order kinematics and scene geometry. Furthermore, we develop a fast linear solver and a continuous-time nonlinear solver on top of the proposed geometric error term. Experiments on both synthetic and real data show the superiority of our linear solver in terms of accuracy and efficiency, and its practicality as an initializer for previous nonlinear solvers. Besides, our continuous-time non-linear solver exhibits exceptional capabilities in accommodating sudden variations in motion since it does not rely on the constant-motion assumption. Our project page can be found at https://nail-hnu.github.io/EvLinearSolver/." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 160, + 549, + 440, + 560 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 549, + 440, + 560 + ], + "spans": [ + { + "bbox": [ + 160, + 549, + 440, + 560 + ], + "type": "text", + "content": "Keywords: Event cameras " + }, + { + "bbox": [ + 160, + 549, + 440, + 560 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 549, + 440, + 560 + ], + "type": "text", + "content": " Geometric model fitting " + }, + { + "bbox": [ + 160, + 549, + 440, + 560 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 549, + 440, + 560 + ], + "type": "text", + "content": " Normal flow" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 598, + 230, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 598, + 230, + 611 + ], + "spans": [ + { + "bbox": [ + 132, + 598, + 230, + 611 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 623, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 623, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 623, + 482, + 647 + ], + "type": "text", + "content": "Neuromorphic event-based cameras are bio-inspired sensors that report only brightness changes at each pixel location asynchronously. Their high temporal" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 141, + 654, + 424, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 654, + 424, + 666 + ], + "spans": [ + { + "bbox": [ + 141, + 654, + 424, + 666 + ], + "type": "text", + "content": "* equal contribution; " + }, + { + "bbox": [ + 141, + 654, + 424, + 666 + ], + "type": "inline_equation", + "content": "\\boxtimes" + }, + { + "bbox": [ + 141, + 654, + 424, + 666 + ], + "type": "text", + "content": " corresponding author (eeyzhou@hnu.edu.cn)." + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": "resolution enables them to sense at exactly the same rate as the scene dynamics, and thus, event cameras are suitable for dealing with robotic perception [33,38, 39,41,43], navigation [21,34,35,42], and control [6,9] tasks that involve agile motion. A general and fundamental problem among these tasks is to recover the camera's ego-motion and scene geometry from event data. One of the bottlenecks for solving this kind of problems at very high speeds is the lack of methods for efficient and robust data association." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 201, + 482, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 201, + 482, + 417 + ], + "spans": [ + { + "bbox": [ + 130, + 201, + 482, + 417 + ], + "type": "text", + "content": "Unlike its standard-vision counterparts, an event camera cannot observe visual information (i.e. intensity changes) when motion is parallel to edges. Such a partial observability of event data makes optical flow estimation (as well as feature detection and matching) an ill-posed problem. To this end, many existing methods [14,17,19,22,30] for solving the problem of motion-model fitting on event data propose specific metrics that assess event-data association established in either an implicit manner or an explicitly model-driven style. The best fit can be retrieved by iteratively optimizing the data association, often expressed in terms of event alignment. However, we argue two flaws of these methods: " + }, + { + "bbox": [ + 130, + 201, + 482, + 417 + ], + "type": "inline_equation", + "content": "(i)" + }, + { + "bbox": [ + 130, + 201, + 482, + 417 + ], + "type": "text", + "content": " The success of these non-linear solutions requires a proper initialization. While a brute-force search may be applicable, it is computationally expensive for real-time application, specially if the dimension of the parameter is high. Efficient and deterministic initialization methods are lacking and urgently needed. " + }, + { + "bbox": [ + 130, + 201, + 482, + 417 + ], + "type": "inline_equation", + "content": "(ii)" + }, + { + "bbox": [ + 130, + 201, + 482, + 417 + ], + "type": "text", + "content": " The common assumption of constant velocity only holds within a short time interval. Nevertheless, a time window that is too narrow may contain insufficient number of events for a valid model fitting. The heuristic of using a constant number of events may violate the constant velocity assumption, especially in the case of sudden changes in speed." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 419, + 482, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 419, + 482, + 539 + ], + "spans": [ + { + "bbox": [ + 130, + 419, + 482, + 539 + ], + "type": "text", + "content": "In this paper, we rethink the problem formulation in a way that is more consistent with the differential working principle of event cameras. Unlike existing pipelines that use events occurring within a time interval to resolve the relative motion, we attempt to recover the structure and instantaneous first-order kinematic parameters involved in the motion field equation, using as input an instantaneous observation, namely event-based normal flow. To this end, we explore effective constraints between normal-flow observations and model parameters for solving a family of motion-and-structure estimation problems, such as the estimation of depth and full optical flow, instantaneous angular velocity and linear velocity, and differential homography (see Fig. 1)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 147, + 540, + 421, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 540, + 421, + 552 + ], + "spans": [ + { + "bbox": [ + 147, + 540, + 421, + 552 + ], + "type": "text", + "content": "The contributions of this paper can be summarized as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 567, + 480, + 663 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 138, + 567, + 480, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 567, + 480, + 613 + ], + "spans": [ + { + "bbox": [ + 138, + 567, + 480, + 613 + ], + "type": "text", + "content": "- A normal flow constraint with a geometric connection to event data. It overcomes the partial observability issue (i.e., errors caused by using normal flow as replacement for full optical flow) in general problems of geometric model fitting when normal flow is used as input." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 617, + 480, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 617, + 480, + 663 + ], + "spans": [ + { + "bbox": [ + 138, + 617, + 480, + 663 + ], + "type": "text", + "content": "- Two solvers for a family of motion-and-structure estimation problems using as input sparse event-based normal flow. Our linear solver (can be used with RANSAC) leads to closed-form and deterministic solutions that can be used as an initialization to existing nonlinear methods. Our nonlinear solver," + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "text", + "content": "Z. Ren et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 143, + 114, + 470, + 232 + ], + "blocks": [ + { + "bbox": [ + 143, + 114, + 470, + 232 + ], + "lines": [ + { + "bbox": [ + 143, + 114, + 470, + 232 + ], + "spans": [ + { + "bbox": [ + 143, + 114, + 470, + 232 + ], + "type": "image", + "image_path": "d1f0b4c57510e255c6b64fbfb4fe1cca27618059d35bbcd12fc8f37b680f0a43.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 239, + 482, + 283 + ], + "lines": [ + { + "bbox": [ + 131, + 239, + 482, + 283 + ], + "spans": [ + { + "bbox": [ + 131, + 239, + 482, + 283 + ], + "type": "text", + "content": "Fig. 1: Overview of our event-based motion estimation approach. Normal flow vectors are computed from the raw event data and are used as input to the two solvers. The result of the linear solver consists of fitted motion models " + }, + { + "bbox": [ + 131, + 239, + 482, + 283 + ], + "type": "inline_equation", + "content": "\\{\\pmb{\\theta}_1, \\pmb{\\theta}_2, \\dots\\}" + }, + { + "bbox": [ + 131, + 239, + 482, + 283 + ], + "type": "text", + "content": " at discrete time instants, which can be used to initialize the continuous-time nonlinear solver." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 147, + 308, + 479, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 308, + 479, + 331 + ], + "spans": [ + { + "bbox": [ + 147, + 308, + 479, + 331 + ], + "type": "text", + "content": "under a continuous-time formulation, can handle sudden speed variations, and thus, is free of the constant-motion assumption." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 137, + 331, + 479, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 331, + 479, + 367 + ], + "spans": [ + { + "bbox": [ + 137, + 331, + 479, + 367 + ], + "type": "text", + "content": "- A thorough evaluation on the two proposed solvers, comparing against state-of-the-art methods, and also including an investigation on the performance of existing non-linear methods initialized using our linear solver." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 374, + 482, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 374, + 482, + 447 + ], + "spans": [ + { + "bbox": [ + 130, + 374, + 482, + 447 + ], + "type": "text", + "content": "Outline: The rest of the paper is organized as follows. We first discuss the related work in Sec. 2. Then a compact review of preliminary knowledge is given in Sec. 3. Our approach is presented in Sec. 4, including the normal-flow constraint, the linear models of a family of motion-and-structure estimation problems, and the proposed linear and nonlinear solvers. Experimental evaluation is conducted in Sec. 5 and conclusion are drawn in Sec. 6." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 464, + 237, + 477 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 464, + 237, + 477 + ], + "spans": [ + { + "bbox": [ + 132, + 464, + 237, + 477 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 490, + 353, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 490, + 353, + 502 + ], + "spans": [ + { + "bbox": [ + 131, + 490, + 353, + 502 + ], + "type": "text", + "content": "2.1 Event-based Geometric Model Fitting" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 510, + 482, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 641 + ], + "type": "text", + "content": "The major body of literature on this specific topic follows an iterative scheme, which typically has its own special metric evaluating quantitatively the association between events, and solve the model fitting as a nonlinear optimization problem. Among the first works is the Contrast Maximization method [13, 14], which provides a unified framework for solving a family of motion-model fitting problems. It searches the optimal model parameters that ultimately lead to a sharp image of warped events (IWE). The goodness of fit is evaluated by the alignment of all involved events along point trajectories on the image plane, and such alignment is assessed by the strength of the contours of the IWE. The computational complexity of CMax is high because of iteratively updating of IWE, whose cost scales with the spatial resolution and the amount of event data." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "content": "To lessen the dependency on IWE, Nunes et al. proposes an entropy minimization framework (EMin) [30] that replaces the dispersion metric of [14] with" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 228, + 91, + 448, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 91, + 448, + 101 + ], + "spans": [ + { + "bbox": [ + 228, + 91, + 448, + 101 + ], + "type": "text", + "content": "Motion and Structure from Event-based Normal Flow" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 320 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 320 + ], + "type": "text", + "content": "a similarity measurement between events, enabling to evaluate the goodness of fit in a higher dimensional space. EMin is, still, too expensive for real-time application. Liu et al. [22] propose a two-step iterative method for the problem of 3-D rotation estimation using events as input. Given a rotation estimate, the first step determines event correspondences by looking for the nearest neighbour in the spatio-temporal domain. The second step employs a spatio-temporal registration which aligns corresponding events by minimizing the sum of squared geodesic residual. The least-square nature of the energy function in the second step leads to a computation friendly performance. Gu et al. [17] present a probabilistic approach for spatio-temporal alignment of events. The aligned events at a particular pixel are modeled as a spatio-temporal Poisson point process and the goodness of fit can be quantified using the likelihood of the data. The optimal model is retrieved when the likelihood is maximized. More recently, an event-to-map alignment scheme is proposed in [19], which adopts the time surface (TS) map as a reference for event alignment. The proposed TS loss function drives later events to get increasingly closer to pixel positions of corresponding earlier events, thus, leading to a progressively sharp TS map as iterations proceed." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 320, + 482, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 320, + 482, + 452 + ], + "spans": [ + { + "bbox": [ + 130, + 320, + 482, + 452 + ], + "type": "text", + "content": "In summary, all the above-mentioned methods are nonlinear, and thus, a proper initialization is necessary. Although a brute-force search is widely adopted, sometimes it can be too expensive to locate the convergence basin in parameter space. Moreover, [14,30] may come across a convergence to false-positive minima, known as event collapse [36, 37]. Hence, a quick and deterministic solver (e.g., a closed-form solution) is desirable. Besides, most of these methods assume a constant-motion model, which limits their applicability to scenarios of sudden speed variations. In addition, we have noticed that a recent work [16] can jointly estimate the line parameters and the camera's linear velocities. However, this method is not a general one due to the dependence on line structures and the usage of an IMU. Thus, it is not within the scope of our discussion." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 473, + 369, + 485 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 473, + 369, + 485 + ], + "spans": [ + { + "bbox": [ + 132, + 473, + 369, + 485 + ], + "type": "text", + "content": "2.2 Motion and Structure from Normal Flow" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 497, + 482, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 497, + 482, + 617 + ], + "spans": [ + { + "bbox": [ + 130, + 497, + 482, + 617 + ], + "type": "text", + "content": "Due to the aperture problem, a local estimate of optical flow recovers only the partial component of the full flow vector along the direction of image gradient. Preliminary knowledge about normal flow can be found in the supplementary material. Recovering motion and structure from normal flow on intensity images is a well-studied topic. Its main body of literature typically proposes a qualitatively deterministic solution by which a set (even infinite) of applicable ego motions are determined via an analysis of normal flow classification [3, 11, 20]. Only recently, researches devised ways to formulate the Cheirality constraint as a differentiable term and use it in a continuous optimization approach [1] or as a loss function for training an artificial neural network [31]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "In event-based perception and navigation, similar methodologies as the frame-based ones are also witnessed, where normal flow computed from event data is used as input to determine the Focus of Expansion (FoE) for drones' course estimation [10] and obstacle avoidance [5]." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "text", + "content": "Z. Ren et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "type": "text", + "content": "Most of these heuristic methods, while effective, are cumbersome for practical application. In contrast, our method delivers a deterministic way to retrieve motion and structure from normal flow. Additionally, it is noticed that normal flow can replace full flow in some specific problems of geometric model fitting, such as affine models [26, 27]. In this way, the geometric model fitting problem can be easily solved in closed-form. Nevertheless, these methods do not consider the observability bias in normal flows, and thus, such naive substitution may lead to inaccurate model fitting. To this end, we present a robust linear solver based on a carefully-designed geometric measurement model that can overcome the partial observability of event-based normal flow. It is worth mentioning that a recent work [24] also determines instantaneous linear velocity using as input the normal flow. Our difference is seen in the scope of applicability. The usage of normal flow is restricted in the motion flow equation of [24], while our paper extends it to a wider scope (i.e., a family of geometric model fitting problems) by combining normal flow with e.g., differential epipolar geometry, differential homography, etc." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 324, + 233, + 337 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 324, + 233, + 337 + ], + "spans": [ + { + "bbox": [ + 132, + 324, + 233, + 337 + ], + "type": "text", + "content": "3 Preliminaries" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 348, + 482, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 348, + 482, + 385 + ], + "spans": [ + { + "bbox": [ + 130, + 348, + 482, + 385 + ], + "type": "text", + "content": "To make this paper self-contained, some preliminary knowledge is revisited. Due to space limitations, here we only list essential concepts. A more detailed review of preliminary knowledge can be found in the supplementary material." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 400, + 482, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 400, + 482, + 450 + ], + "spans": [ + { + "bbox": [ + 130, + 400, + 482, + 450 + ], + "type": "text", + "content": "Motion Field. Consider a perspective camera moving in a static environment, with instantaneous angular and linear velocities in the camera frame " + }, + { + "bbox": [ + 130, + 400, + 482, + 450 + ], + "type": "inline_equation", + "content": "\\boldsymbol{\\omega} = (\\omega_{x},\\omega_{y},\\omega_{z})^{\\top}" + }, + { + "bbox": [ + 130, + 400, + 482, + 450 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 400, + 482, + 450 + ], + "type": "inline_equation", + "content": "\\pmb {\\nu} = (\\nu_{x},\\nu_{y},\\nu_{z})^{\\top}" + }, + { + "bbox": [ + 130, + 400, + 482, + 450 + ], + "type": "text", + "content": ", respectively. The motion field at pixel " + }, + { + "bbox": [ + 130, + 400, + 482, + 450 + ], + "type": "inline_equation", + "content": "\\mathbf{x} = (x,y)^{\\top}" + }, + { + "bbox": [ + 130, + 400, + 482, + 450 + ], + "type": "text", + "content": " produced by projecting the 3D motion at point with depth " + }, + { + "bbox": [ + 130, + 400, + 482, + 450 + ], + "type": "inline_equation", + "content": "Z(\\mathbf{x})" + }, + { + "bbox": [ + 130, + 400, + 482, + 450 + ], + "type": "text", + "content": " is" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 247, + 456, + 481, + 482 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 247, + 456, + 481, + 482 + ], + "spans": [ + { + "bbox": [ + 247, + 456, + 481, + 482 + ], + "type": "interline_equation", + "content": "\\mathbf {u} = \\frac {1}{Z (\\mathbf {x})} \\mathbf {A} (\\mathbf {x}) \\boldsymbol {\\nu} + \\mathbf {B} (\\mathbf {x}) \\boldsymbol {\\omega}, \\tag {1}", + "image_path": "14d1a1924e3db36cd19c40defe5ad04c21ae4df56c082a41e2c38c3f1d8206d5.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 487, + 482, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 487, + 482, + 536 + ], + "spans": [ + { + "bbox": [ + 130, + 487, + 482, + 536 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 487, + 482, + 536 + ], + "type": "inline_equation", + "content": "\\mathbf{A}(\\mathbf{x})" + }, + { + "bbox": [ + 130, + 487, + 482, + 536 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 487, + 482, + 536 + ], + "type": "inline_equation", + "content": "\\mathbf{B}(\\mathbf{x})" + }, + { + "bbox": [ + 130, + 487, + 482, + 536 + ], + "type": "text", + "content": " are " + }, + { + "bbox": [ + 130, + 487, + 482, + 536 + ], + "type": "inline_equation", + "content": "2\\times 3" + }, + { + "bbox": [ + 130, + 487, + 482, + 536 + ], + "type": "text", + "content": " matrices that solely depend on " + }, + { + "bbox": [ + 130, + 487, + 482, + 536 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 130, + 487, + 482, + 536 + ], + "type": "text", + "content": ". Equation (1) is the motion field equation [40, Ch.8] [23]. Optical flow in rigid scenes aims at estimating the motion field from visual data, which is only possible in regions that do not suffer from the aperture problem." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 552, + 482, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 552, + 482, + 636 + ], + "spans": [ + { + "bbox": [ + 130, + 552, + 482, + 636 + ], + "type": "text", + "content": "Brightness constancy is the assumption stating that the true motion trajectories " + }, + { + "bbox": [ + 130, + 552, + 482, + 636 + ], + "type": "inline_equation", + "content": "\\mathbf{x}(t)" + }, + { + "bbox": [ + 130, + 552, + 482, + 636 + ], + "type": "text", + "content": " followed by objects on the image plane are those at which brightness is constant: " + }, + { + "bbox": [ + 130, + 552, + 482, + 636 + ], + "type": "inline_equation", + "content": "I(\\mathbf{x}(t), t) = \\mathrm{const}" + }, + { + "bbox": [ + 130, + 552, + 482, + 636 + ], + "type": "text", + "content": ". Differentiating in time (total derivative) gives the equation at " + }, + { + "bbox": [ + 130, + 552, + 482, + 636 + ], + "type": "inline_equation", + "content": "(\\mathbf{x}, t)" + }, + { + "bbox": [ + 130, + 552, + 482, + 636 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 130, + 552, + 482, + 636 + ], + "type": "inline_equation", + "content": "\\nabla I \\cdot \\dot{\\mathbf{x}} + \\partial_t I = 0" + }, + { + "bbox": [ + 130, + 552, + 482, + 636 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 552, + 482, + 636 + ], + "type": "inline_equation", + "content": "\\dot{\\mathbf{x}} = d\\mathbf{x} / dt" + }, + { + "bbox": [ + 130, + 552, + 482, + 636 + ], + "type": "text", + "content": " is the velocity of image point " + }, + { + "bbox": [ + 130, + 552, + 482, + 636 + ], + "type": "inline_equation", + "content": "\\mathbf{x}(t)" + }, + { + "bbox": [ + 130, + 552, + 482, + 636 + ], + "type": "text", + "content": ". Given " + }, + { + "bbox": [ + 130, + 552, + 482, + 636 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 130, + 552, + 482, + 636 + ], + "type": "text", + "content": ", this equation provides one constraint per point " + }, + { + "bbox": [ + 130, + 552, + 482, + 636 + ], + "type": "inline_equation", + "content": "(\\mathbf{x}, t)" + }, + { + "bbox": [ + 130, + 552, + 482, + 636 + ], + "type": "text", + "content": " to estimate " + }, + { + "bbox": [ + 130, + 552, + 482, + 636 + ], + "type": "inline_equation", + "content": "\\dot{\\mathbf{x}}" + }, + { + "bbox": [ + 130, + 552, + 482, + 636 + ], + "type": "text", + "content": ", now called optical flow. Without additional equations, only the normal component of the optical flow can be determined [40]," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 269, + 643, + 481, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 269, + 643, + 481, + 669 + ], + "spans": [ + { + "bbox": [ + 269, + 643, + 481, + 669 + ], + "type": "interline_equation", + "content": "\\mathbf {n} = - \\frac {\\partial_ {t} I}{\\| \\nabla I \\| ^ {2}} \\nabla I. \\tag {2}", + "image_path": "e2a924edca2fd5dbb74b2d439c473d1548111cc2ed6894662d9d9bfad13a05e1.jpg" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 228, + 91, + 448, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 91, + 448, + 101 + ], + "spans": [ + { + "bbox": [ + 228, + 91, + 448, + 101 + ], + "type": "text", + "content": "Motion and Structure from Event-based Normal Flow" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 152 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 152 + ], + "type": "text", + "content": "Normal Flow. Decomposing the optical flow " + }, + { + "bbox": [ + 132, + 116, + 482, + 152 + ], + "type": "inline_equation", + "content": "\\mathbf{u} = \\mathbf{u}_{\\perp} + \\mathbf{u}_{\\parallel}" + }, + { + "bbox": [ + 132, + 116, + 482, + 152 + ], + "type": "text", + "content": " into its normal and parallel components to the local image gradient " + }, + { + "bbox": [ + 132, + 116, + 482, + 152 + ], + "type": "inline_equation", + "content": "(\\nabla I)" + }, + { + "bbox": [ + 132, + 116, + 482, + 152 + ], + "type": "text", + "content": ", the normal flow is " + }, + { + "bbox": [ + 132, + 116, + 482, + 152 + ], + "type": "inline_equation", + "content": "\\mathbf{n} = \\mathbf{u}_{\\perp}" + }, + { + "bbox": [ + 132, + 116, + 482, + 152 + ], + "type": "text", + "content": ", and the dot-product with " + }, + { + "bbox": [ + 132, + 116, + 482, + 152 + ], + "type": "inline_equation", + "content": "\\mathbf{n}" + }, + { + "bbox": [ + 132, + 116, + 482, + 152 + ], + "type": "text", + "content": " gives" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 198, + 160, + 481, + 174 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 160, + 481, + 174 + ], + "spans": [ + { + "bbox": [ + 198, + 160, + 481, + 174 + ], + "type": "interline_equation", + "content": "\\mathbf {n} ^ {\\top} \\mathbf {u} = \\| \\mathbf {n} \\| ^ {2}, \\quad \\text {o r e q u i v a l e n t l y} \\quad (\\mathbf {u} - \\mathbf {n}) \\cdot \\mathbf {n} = 0. \\tag {3}", + "image_path": "99b2a9d26deeada459da62feb52c432b0fa3e1e050de3d549dd9c7576af51a6e.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 191, + 479, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 191, + 479, + 251 + ], + "spans": [ + { + "bbox": [ + 131, + 191, + 479, + 251 + ], + "type": "text", + "content": "Differential Homography. (Continuous Homography Matrix [25, Ch.5]). If all scene points lie on a plane, their coordinates are related to their image projections via a homography transformation (which encodes the information of the scene plane " + }, + { + "bbox": [ + 131, + 191, + 479, + 251 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{N}, d\\}" + }, + { + "bbox": [ + 131, + 191, + 479, + 251 + ], + "type": "text", + "content": " and the camera pose). In this case, the motion field (1) can be rewritten to adopt the form in homogeneous coordinates" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 255, + 260, + 480, + 274 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 260, + 480, + 274 + ], + "spans": [ + { + "bbox": [ + 255, + 260, + 480, + 274 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {u}} (\\mathbf {x}) = \\left(\\mathbf {1} - \\hat {\\mathbf {x}} \\mathbf {e} _ {3} ^ {\\top}\\right) \\mathbf {H} _ {d} \\hat {\\mathbf {x}}, \\tag {4}", + "image_path": "b144b1601754c3ab254abc646de23930ccf12392876b16140242be3d8f571bf2.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 283, + 479, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 283, + 479, + 320 + ], + "spans": [ + { + "bbox": [ + 132, + 283, + 479, + 320 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 132, + 283, + 479, + 320 + ], + "type": "inline_equation", + "content": "\\mathbf{H}_d\\dot{=} -\\left([ \\boldsymbol {\\omega}]_{\\times} + d^{-1}\\boldsymbol {\\nu}\\mathbf{N}^{\\top}\\right)" + }, + { + "bbox": [ + 132, + 283, + 479, + 320 + ], + "type": "text", + "content": " denotes the differential homography matrix " + }, + { + "bbox": [ + 132, + 283, + 479, + 320 + ], + "type": "inline_equation", + "content": "([\\cdot ]_{\\times}" + }, + { + "bbox": [ + 132, + 283, + 479, + 320 + ], + "type": "text", + "content": " refers to the cross-product, skew-symmetric matrix), " + }, + { + "bbox": [ + 132, + 283, + 479, + 320 + ], + "type": "inline_equation", + "content": "\\mathbf{1}" + }, + { + "bbox": [ + 132, + 283, + 479, + 320 + ], + "type": "text", + "content": " is the identity matrix, " + }, + { + "bbox": [ + 132, + 283, + 479, + 320 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}} = (x,y,1)^{\\top}" + }, + { + "bbox": [ + 132, + 283, + 479, + 320 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 132, + 283, + 479, + 320 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{u}} = (u_x,u_y,0)^{\\top}" + }, + { + "bbox": [ + 132, + 283, + 479, + 320 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 283, + 479, + 320 + ], + "type": "inline_equation", + "content": "\\mathbf{e}_3 = (0,0,1)^{\\top}" + }, + { + "bbox": [ + 132, + 283, + 479, + 320 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 337, + 482, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 337, + 482, + 397 + ], + "spans": [ + { + "bbox": [ + 131, + 337, + 482, + 397 + ], + "type": "text", + "content": "Differential Epipolar Geometry. (Continuous Epipolar Geometry [25, Ch.5]). The epipolar geometry examines the distance between a potential match to the determined epipolar line, and can be expressed as " + }, + { + "bbox": [ + 131, + 337, + 482, + 397 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}^{\\prime \\top}\\mathbf{E}\\hat{\\mathbf{x}} = 0" + }, + { + "bbox": [ + 131, + 337, + 482, + 397 + ], + "type": "text", + "content": ". The differential epipolar geometry can be derived analogously, providing a relationship between " + }, + { + "bbox": [ + 131, + 337, + 482, + 397 + ], + "type": "inline_equation", + "content": "\\mathbf{x},\\mathbf{u}" + }, + { + "bbox": [ + 131, + 337, + 482, + 397 + ], + "type": "text", + "content": " and the camera velocities " + }, + { + "bbox": [ + 131, + 337, + 482, + 397 + ], + "type": "inline_equation", + "content": "\\pmb {\\nu},\\pmb{\\omega}" + }, + { + "bbox": [ + 131, + 337, + 482, + 397 + ], + "type": "text", + "content": ". In homogeneous coordinates:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 253, + 406, + 480, + 419 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 253, + 406, + 480, + 419 + ], + "spans": [ + { + "bbox": [ + 253, + 406, + 480, + 419 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {u}} (\\mathbf {x}) ^ {\\top} [ \\boldsymbol {\\nu} ] _ {\\times} \\hat {\\mathbf {x}} - \\hat {\\mathbf {x}} ^ {\\top} \\mathbf {s} \\hat {\\mathbf {x}} = 0, \\tag {5}", + "image_path": "406270399b7d0d07044723365553f4b93ef416225ddd1a342f398d29256f8908.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 428, + 411, + 442 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 428, + 411, + 442 + ], + "spans": [ + { + "bbox": [ + 132, + 428, + 411, + 442 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 132, + 428, + 411, + 442 + ], + "type": "inline_equation", + "content": "\\mathbf{s} = \\frac{1}{2} ([\\pmb {\\nu}]_{\\times}[\\pmb {\\omega}]_{\\times} + [\\pmb {\\omega}]_{\\times}[\\pmb {\\nu}]_{\\times})" + }, + { + "bbox": [ + 132, + 428, + 411, + 442 + ], + "type": "text", + "content": " is a " + }, + { + "bbox": [ + 132, + 428, + 411, + 442 + ], + "type": "inline_equation", + "content": "3\\times 3" + }, + { + "bbox": [ + 132, + 428, + 411, + 442 + ], + "type": "text", + "content": " symmetric matrix." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 459, + 233, + 473 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 459, + 233, + 473 + ], + "spans": [ + { + "bbox": [ + 132, + 459, + 233, + 473 + ], + "type": "text", + "content": "4 Methodology" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 484, + 482, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 484, + 482, + 568 + ], + "spans": [ + { + "bbox": [ + 131, + 484, + 482, + 568 + ], + "type": "text", + "content": "Figure 1 summarizes our approach, grounded upon event-based normal flow. In this section, we first disclose the normal-flow constraint that can overcome the partial observability of normal flow when replacing full optical flow with normal flow in general tasks of geometric model fitting (Sec. 4.1). By combining the constraint and the theory in Sec. 3, we derive linear equations for a family of motion-and-structure estimation problems (Sec. 4.2). To solve these problems, we present a linear solver and a continuous-time nonlinear solver (Sec. 4.3)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 586, + 285, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 586, + 285, + 597 + ], + "spans": [ + { + "bbox": [ + 132, + 586, + 285, + 597 + ], + "type": "text", + "content": "4.1 Normal Flow Constraint" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 605, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 605, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 131, + 605, + 482, + 667 + ], + "type": "text", + "content": "Normal flow can be computed easily from the event data, for example by fitting planes locally to the space-time \"cloud\" of events on image space [2, 32]. Specifically, assuming that the event data " + }, + { + "bbox": [ + 131, + 605, + 482, + 667 + ], + "type": "inline_equation", + "content": "e_k \\doteq \\{\\mathbf{x}_k, t_k\\}" + }, + { + "bbox": [ + 131, + 605, + 482, + 667 + ], + "type": "text", + "content": " spans locally a surface in the spatio-temporal domain, the surface can be defined by a function " + }, + { + "bbox": [ + 131, + 605, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\Sigma_e: \\mathbb{R}^2 \\to \\mathbb{R}; \\mathbf{x}_k \\mapsto t_k" + }, + { + "bbox": [ + 131, + 605, + 482, + 667 + ], + "type": "text", + "content": " that maps the pixel coordinates to the latest event's" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "text", + "content": "Z. Ren et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 114, + 209, + 201 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 209, + 201 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 209, + 201 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 209, + 201 + ], + "type": "image", + "image_path": "4b6f04504e5b1e1da6053b87c1bffd9164d3a1fd887b5441ec3b879c7da47058.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 209, + 482, + 276 + ], + "lines": [ + { + "bbox": [ + 131, + 209, + 482, + 276 + ], + "spans": [ + { + "bbox": [ + 131, + 209, + 482, + 276 + ], + "type": "text", + "content": "Fig. 2: Toy example of applying the proposed normal-flow constraint. (a): A 2D registration task using as input either optical flow or normal flow. The groundtruth displacement is defined by a global flow of " + }, + { + "bbox": [ + 131, + 209, + 482, + 276 + ], + "type": "inline_equation", + "content": "(1.732, -1)^{\\top}" + }, + { + "bbox": [ + 131, + 209, + 482, + 276 + ], + "type": "text", + "content": ". (b)-(d): The loss landscape obtained using different geometric measurements in the registration task. The red dot denotes the resulting displacement, and the green cross the groundtruth displacement. The bottom-left regions in (b)-(d) display the corresponding registration results." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 209, + 114, + 292, + 201 + ], + "blocks": [ + { + "bbox": [ + 209, + 114, + 292, + 201 + ], + "lines": [ + { + "bbox": [ + 209, + 114, + 292, + 201 + ], + "spans": [ + { + "bbox": [ + 209, + 114, + 292, + 201 + ], + "type": "image", + "image_path": "2d17edd895e022ec3c07138772e36bd4fa5c35063e614e26fea68de5c2b3b612.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 293, + 114, + 387, + 201 + ], + "blocks": [ + { + "bbox": [ + 293, + 114, + 387, + 201 + ], + "lines": [ + { + "bbox": [ + 293, + 114, + 387, + 201 + ], + "spans": [ + { + "bbox": [ + 293, + 114, + 387, + 201 + ], + "type": "image", + "image_path": "be2ae15f15e7913c2af0fcdd3c71127563bfd3f69bee8e501cce550b212cf651.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 388, + 114, + 481, + 201 + ], + "blocks": [ + { + "bbox": [ + 388, + 114, + 481, + 201 + ], + "lines": [ + { + "bbox": [ + 388, + 114, + 481, + 201 + ], + "spans": [ + { + "bbox": [ + 388, + 114, + 481, + 201 + ], + "type": "image", + "image_path": "52aecd3ac01157809c0e7563adeadf90b1b1db4698056f539adf29834c99358c.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 299, + 479, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 299, + 479, + 324 + ], + "spans": [ + { + "bbox": [ + 131, + 299, + 479, + 324 + ], + "type": "text", + "content": "timestamp [2, 5]. The normal flow is parallel to the spatial derivative of " + }, + { + "bbox": [ + 131, + 299, + 479, + 324 + ], + "type": "inline_equation", + "content": "\\Sigma_{e}" + }, + { + "bbox": [ + 131, + 299, + 479, + 324 + ], + "type": "text", + "content": ", namely " + }, + { + "bbox": [ + 131, + 299, + 479, + 324 + ], + "type": "inline_equation", + "content": "\\nabla \\Sigma_{e}(\\mathbf{x})" + }, + { + "bbox": [ + 131, + 299, + 479, + 324 + ], + "type": "text", + "content": ". More specifically, it can be calculated by" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 261, + 333, + 481, + 361 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 261, + 333, + 481, + 361 + ], + "spans": [ + { + "bbox": [ + 261, + 333, + 481, + 361 + ], + "type": "interline_equation", + "content": "\\mathbf {n} (\\mathbf {x}) = \\frac {\\nabla \\Sigma_ {e} (\\mathbf {x})}{\\left\\| \\nabla \\Sigma_ {e} (\\mathbf {x}) \\right\\| ^ {2}}. \\tag {6}", + "image_path": "9e3a9d1f51ec58085db0f057f2986b779dd4de851def8108ac36e2bf37989733.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 369, + 479, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 369, + 479, + 416 + ], + "spans": [ + { + "bbox": [ + 131, + 369, + 479, + 416 + ], + "type": "text", + "content": "Readers may refer to the supplementary material for a more detailed derivation. Surely there are other methods to extract event-based normal flow and the solvers we develop in this section are independent of them, i.e., our solvers can also be fed with the extracted normal flow data from such methods." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 417, + 480, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 417, + 480, + 499 + ], + "spans": [ + { + "bbox": [ + 131, + 417, + 480, + 499 + ], + "type": "text", + "content": "Event-based normal flow cannot be used directly as a replacement for the full flow in (1) to solve for the motion parameters. However, we can leverage the relationship between both flows (3) to define an effective geometric error criterion that circumvents the partial observability of normal flow. If the flow " + }, + { + "bbox": [ + 131, + 417, + 480, + 499 + ], + "type": "inline_equation", + "content": "\\mathbf{u}" + }, + { + "bbox": [ + 131, + 417, + 480, + 499 + ], + "type": "text", + "content": " depends on some unknown (\"state\") variables " + }, + { + "bbox": [ + 131, + 417, + 480, + 499 + ], + "type": "inline_equation", + "content": "\\pmb{\\theta}" + }, + { + "bbox": [ + 131, + 417, + 480, + 499 + ], + "type": "text", + "content": ", we propose estimating them based on how well the normal flow constraint (3) is satisfied, e.g., by minimizing some function of the error terms" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 228, + 510, + 480, + 525 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 510, + 480, + 525 + ], + "spans": [ + { + "bbox": [ + 228, + 510, + 480, + 525 + ], + "type": "interline_equation", + "content": "\\mathbf {n} (\\mathbf {x}) ^ {\\top} \\mathbf {u} (\\mathbf {x}; \\boldsymbol {\\theta}) - \\| \\mathbf {n} (\\mathbf {x}) \\| ^ {2} \\doteq \\epsilon_ {\\mathrm {n f}} \\in \\mathbb {R}. \\tag {7}", + "image_path": "d8c3598570f894324dd3b0e63dcd2d8500b03cfa70bea2c0dbe0c4c6b50c7129.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 533, + 482, + 666 + ], + "type": "text", + "content": "To illustrate the validity of the proposed normal-flow constraint over the naive way of simply using normal flow to replace full optical flow, we use a 2D registration task as toy example. As shown in Fig. 2, the task consists of registering the solid-line triangle to the dashed-line one using as input either optical flow or normal flow. We observe that using the proposed normal-flow constraint as a geometric measurement leads to a registration result that is almost as accurate as the one given by the full optical flow. However, due to the partial observability of normal flow, simply using it as a replacement for full optical flow leads to a biased registration result, as shown in Fig. 2c. As we see in all following experiments, similar conclusions can be made in general problems of geometric model fitting." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 228, + 91, + 448, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 91, + 448, + 101 + ], + "spans": [ + { + "bbox": [ + 228, + 91, + 448, + 101 + ], + "type": "text", + "content": "Motion and Structure from Event-based Normal Flow" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 324, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 324, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 324, + 128 + ], + "type": "text", + "content": "4.2 Models of Motion and Structure" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 135, + 482, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 135, + 482, + 232 + ], + "spans": [ + { + "bbox": [ + 130, + 135, + 482, + 232 + ], + "type": "text", + "content": "Based on the derivation in Sec. 4.1, we establish linear models to a family of motion and structure estimation problems using the event-based normal flow as input. In particular, we first discuss how to recover optical flow and depth by introducing the normal-flow constraint into the differential epipolar geometry. Then we show that the combination of the normal-flow constraint with the motion flow equation (1) results in solutions to the problem of instantaneous angular velocity estimation, and 6-DoF motion estimation given depth as prior, respectively. Finally, we show how to compute the differential homography." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 248, + 482, + 298 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 248, + 482, + 298 + ], + "spans": [ + { + "bbox": [ + 131, + 248, + 482, + 298 + ], + "type": "text", + "content": "Optical Flow. Given as input the accurate normal flow, and camera velocities, the full optical flow " + }, + { + "bbox": [ + 131, + 248, + 482, + 298 + ], + "type": "inline_equation", + "content": "\\mathbf{u}(\\mathbf{x})" + }, + { + "bbox": [ + 131, + 248, + 482, + 298 + ], + "type": "text", + "content": " at pixel location " + }, + { + "bbox": [ + 131, + 248, + 482, + 298 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 131, + 248, + 482, + 298 + ], + "type": "text", + "content": " can be calculated by solving the following linear system obtained from the combination of the normal-flow constraint (3) and the differential epipolar geometry (5):" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 236, + 305, + 481, + 334 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 236, + 305, + 481, + 334 + ], + "spans": [ + { + "bbox": [ + 236, + 305, + 481, + 334 + ], + "type": "interline_equation", + "content": "\\binom {\\mathbf {n} (\\mathbf {x}) ^ {\\top}} {- \\hat {\\mathbf {x}} ^ {\\top} [ \\boldsymbol {\\nu} ] _ {\\times}} \\mathbf {u} (\\mathbf {x}) = \\binom {\\| \\mathbf {n} (\\mathbf {x}) \\| ^ {2}} {\\hat {\\mathbf {x}} ^ {\\top} \\mathbf {s} \\hat {\\mathbf {x}}}. \\tag {8}", + "image_path": "dbc3a6c276829028d66f1628a862d458e07b5352499f93c9168af84f09d98836.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 349, + 482, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 349, + 482, + 386 + ], + "spans": [ + { + "bbox": [ + 131, + 349, + 482, + 386 + ], + "type": "text", + "content": "Depth Estimation. Similarly, if the normal flow and camera velocities are known, the scene depth " + }, + { + "bbox": [ + 131, + 349, + 482, + 386 + ], + "type": "inline_equation", + "content": "Z(\\mathbf{x})" + }, + { + "bbox": [ + 131, + 349, + 482, + 386 + ], + "type": "text", + "content": " can be determined by combining the normal-flow constraint (3) with the motion field eq. (1), yielding the closed-form solution:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 235, + 394, + 481, + 422 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 394, + 481, + 422 + ], + "spans": [ + { + "bbox": [ + 235, + 394, + 481, + 422 + ], + "type": "interline_equation", + "content": "Z (\\mathbf {x}) = \\frac {\\mathbf {n} (\\mathbf {x}) ^ {\\top} \\mathbf {A} (\\mathbf {x}) \\boldsymbol {\\nu}}{\\| \\mathbf {n} (\\mathbf {x}) \\| ^ {2} - \\mathbf {n} (\\mathbf {x}) ^ {\\top} \\mathbf {B} (\\mathbf {x}) \\boldsymbol {\\omega}}. \\tag {9}", + "image_path": "b83e65ef1dd301ef452dc49c0faeff8bf352ae6659e6c9e75e4a6dd37a640eee.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 435, + 482, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 435, + 482, + 460 + ], + "spans": [ + { + "bbox": [ + 131, + 435, + 482, + 460 + ], + "type": "text", + "content": "Angular Velocity Estimation. Assuming that an event camera undergoes a pure rotational motion, the optical flow " + }, + { + "bbox": [ + 131, + 435, + 482, + 460 + ], + "type": "inline_equation", + "content": "\\mathbf{u}" + }, + { + "bbox": [ + 131, + 435, + 482, + 460 + ], + "type": "text", + "content": " in (1) can be simplified as" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 274, + 469, + 481, + 482 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 274, + 469, + 481, + 482 + ], + "spans": [ + { + "bbox": [ + 274, + 469, + 481, + 482 + ], + "type": "interline_equation", + "content": "\\mathbf {u} (\\mathbf {x}) = \\mathbf {B} (\\mathbf {x}) \\boldsymbol {\\omega}. \\tag {10}", + "image_path": "c4bd077b7dae82a2f33be04f430c924efc4084c6ddcc1cebbdb8ea04c4fde470.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 490, + 413, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 490, + 413, + 502 + ], + "spans": [ + { + "bbox": [ + 132, + 490, + 413, + 502 + ], + "type": "text", + "content": "By substituting (10) into (3), we have the linear equation in " + }, + { + "bbox": [ + 132, + 490, + 413, + 502 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 132, + 490, + 413, + 502 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 253, + 510, + 481, + 525 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 253, + 510, + 481, + 525 + ], + "spans": [ + { + "bbox": [ + 253, + 510, + 481, + 525 + ], + "type": "interline_equation", + "content": "\\mathbf {n} (\\mathbf {x}) ^ {\\top} \\mathbf {B} (\\mathbf {x}) \\omega = \\| \\mathbf {n} (\\mathbf {x}) \\| ^ {2}. \\tag {11}", + "image_path": "fa158d9a91995d0ab26d4f681a117d255f817bbf4edff96bd7f34c0fba734808.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 533, + 481, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 533, + 481, + 558 + ], + "spans": [ + { + "bbox": [ + 131, + 533, + 481, + 558 + ], + "type": "text", + "content": "Since each normal flow provides only one constraint, the minimal problem requires three normal flow vectors as input." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 574, + 482, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 574, + 482, + 635 + ], + "spans": [ + { + "bbox": [ + 131, + 574, + 482, + 635 + ], + "type": "text", + "content": "6-DoF Motion Estimation. Assume the event camera undergoes a general 6-DoF motion with known depth of the observed scene, we can solve the first-order kinematics, namely instantaneous linear and angular velocities, by solving the following linear system obtained from the normal-flow constraint (3) and the motion-flow equation (1):" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 246, + 643, + 481, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 246, + 643, + 481, + 669 + ], + "spans": [ + { + "bbox": [ + 246, + 643, + 481, + 669 + ], + "type": "interline_equation", + "content": "\\mathbf {n} (\\mathbf {x}) ^ {\\top} \\mathbf {D} (\\mathbf {x}) \\left[ \\begin{array}{l} \\boldsymbol {\\nu} \\\\ \\boldsymbol {\\omega} \\end{array} \\right] = \\| \\mathbf {n} (\\mathbf {x}) \\| ^ {2}, \\tag {12}", + "image_path": "fa3d2741161b8f84706a5e426388a16c27236ae95cc472d8b40d852d62e88270.jpg" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "text", + "content": "Z. Ren et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 114, + 479, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 114, + 479, + 156 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 479, + 156 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 114, + 479, + 156 + ], + "type": "inline_equation", + "content": "\\mathbf{D}(\\mathbf{x}) = \\left[\\frac{1}{Z(\\mathbf{x})}\\mathbf{A}(\\mathbf{x})\\mathbf{B}(\\mathbf{x})\\right]" + }, + { + "bbox": [ + 130, + 114, + 479, + 156 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 130, + 114, + 479, + 156 + ], + "type": "inline_equation", + "content": "2\\times 6" + }, + { + "bbox": [ + 130, + 114, + 479, + 156 + ], + "type": "text", + "content": " feature sensitivity matrix [4, 7]. The minimal problem in this context requires six normal flow measurements to determine the 6-DoF motion parameters." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 171, + 479, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 171, + 479, + 243 + ], + "spans": [ + { + "bbox": [ + 130, + 171, + 479, + 243 + ], + "type": "text", + "content": "Motion and Structure Estimation in Planar Scenes. For a moving camera that observes a planar scene, the motion and structure parameters are compatibly encoded in a homography matrix. In our case, we attempt to recover the differential homography " + }, + { + "bbox": [ + 130, + 171, + 479, + 243 + ], + "type": "inline_equation", + "content": "\\mathbf{h}_d" + }, + { + "bbox": [ + 130, + 171, + 479, + 243 + ], + "type": "text", + "content": " (the vector form of " + }, + { + "bbox": [ + 130, + 171, + 479, + 243 + ], + "type": "inline_equation", + "content": "\\mathbf{H}_d" + }, + { + "bbox": [ + 130, + 171, + 479, + 243 + ], + "type": "text", + "content": " in (4)) because the instantaneous linear and angular velocities are considered. By expanding (4) and rearranging equations, we have" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 273, + 251, + 480, + 262 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 273, + 251, + 480, + 262 + ], + "spans": [ + { + "bbox": [ + 273, + 251, + 480, + 262 + ], + "type": "interline_equation", + "content": "\\mathbf {u} (\\mathbf {x}) = \\mathbf {C} (\\mathbf {x}) \\mathbf {h} _ {d}, \\tag {13}", + "image_path": "a6ca37ccb65990acf19316587d0e31ad47d0aafeb5c294f57335a44fea23eb94.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 268, + 479, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 268, + 479, + 293 + ], + "spans": [ + { + "bbox": [ + 130, + 268, + 479, + 293 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 268, + 479, + 293 + ], + "type": "inline_equation", + "content": "\\mathbf{C}(\\mathbf{x})" + }, + { + "bbox": [ + 130, + 268, + 479, + 293 + ], + "type": "text", + "content": " is a " + }, + { + "bbox": [ + 130, + 268, + 479, + 293 + ], + "type": "inline_equation", + "content": "2 \\times 9" + }, + { + "bbox": [ + 130, + 268, + 479, + 293 + ], + "type": "text", + "content": " matrix that solely depends on the pixel location " + }, + { + "bbox": [ + 130, + 268, + 479, + 293 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 130, + 268, + 479, + 293 + ], + "type": "text", + "content": ". Subsequently, substituting (3) into (13), gives:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 251, + 297, + 480, + 312 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 251, + 297, + 480, + 312 + ], + "spans": [ + { + "bbox": [ + 251, + 297, + 480, + 312 + ], + "type": "interline_equation", + "content": "\\mathbf {n} (\\mathbf {x}) ^ {\\top} \\mathbf {C} (\\mathbf {x}) \\mathbf {h} _ {d} = \\| \\mathbf {n} (\\mathbf {x}) \\| ^ {2}. \\tag {14}", + "image_path": "6fe0cf0f7b5529acf9e0388e11de71fca7507affb98e7dce1a70f1708deee2ec.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 317, + 479, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 317, + 479, + 353 + ], + "spans": [ + { + "bbox": [ + 130, + 317, + 479, + 353 + ], + "type": "text", + "content": "As discussed in [44], retrieving the motion and structure parameters from the decomposition of " + }, + { + "bbox": [ + 130, + 317, + 479, + 353 + ], + "type": "inline_equation", + "content": "\\mathbf{h}_d" + }, + { + "bbox": [ + 130, + 317, + 479, + 353 + ], + "type": "text", + "content": " is not trivial and beyond the scope of this paper. Therefore, we only focus on the numerical solution of " + }, + { + "bbox": [ + 130, + 317, + 479, + 353 + ], + "type": "inline_equation", + "content": "\\mathbf{h}_d" + }, + { + "bbox": [ + 130, + 317, + 479, + 353 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 369, + 220, + 379 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 369, + 220, + 379 + ], + "spans": [ + { + "bbox": [ + 131, + 369, + 220, + 379 + ], + "type": "text", + "content": "4.3 Our Solvers" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 385, + 480, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 385, + 480, + 469 + ], + "spans": [ + { + "bbox": [ + 130, + 385, + 480, + 469 + ], + "type": "text", + "content": "According to the motion-and-structure models derived in Sec. 4.2, we have established certain constraints between the model parameters and the normal flow observations. This allows us to simply develop linear solvers for each specific problem. Besides, we further propose a continuous-time nonlinear solver that can effectively handle the presence of sudden variations in motion and does not rely on the constant-motion assumption. To convey our solvers concisely, we generalize the motion-and-structure models into a unified form, as" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 254, + 475, + 480, + 488 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 254, + 475, + 480, + 488 + ], + "spans": [ + { + "bbox": [ + 254, + 475, + 480, + 488 + ], + "type": "interline_equation", + "content": "\\mathbf {n} (\\mathbf {x}) ^ {\\top} \\mathbf {O} (\\mathbf {x}) \\boldsymbol {\\theta} = \\| \\mathbf {n} (\\mathbf {x}) \\| ^ {2}, \\tag {15}", + "image_path": "75b62ebac3e24dd9b454dd76caaead4fd5a659dbe0f57a2fb336fdcda01fe02b.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 495, + 477, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 495, + 477, + 507 + ], + "spans": [ + { + "bbox": [ + 130, + 495, + 477, + 507 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 495, + 477, + 507 + ], + "type": "inline_equation", + "content": "\\mathbf{O}(\\mathbf{x})" + }, + { + "bbox": [ + 130, + 495, + 477, + 507 + ], + "type": "text", + "content": " is the system matrix and " + }, + { + "bbox": [ + 130, + 495, + 477, + 507 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 130, + 495, + 477, + 507 + ], + "type": "text", + "content": " are the motion/structure parameters." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 521, + 479, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 479, + 558 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 479, + 558 + ], + "type": "text", + "content": "Linear Solver. Given as input a set (e.g., K) of sparse normal-flow estimates obtained at the same time, our linear solver works out the following (overdetermined) linear system of equations:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 227, + 563, + 480, + 609 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 227, + 563, + 480, + 609 + ], + "spans": [ + { + "bbox": [ + 227, + 563, + 480, + 609 + ], + "type": "interline_equation", + "content": "\\left( \\begin{array}{c} \\mathbf {n} \\left(\\mathbf {x} _ {1}\\right) ^ {\\top} \\mathbf {O} \\left(\\mathbf {x} _ {1}\\right) \\\\ \\vdots \\\\ \\mathbf {n} \\left(\\mathbf {x} _ {\\mathrm {K}}\\right) ^ {\\top} \\mathbf {O} \\left(\\mathbf {x} _ {\\mathrm {K}}\\right) \\end{array} \\right) \\boldsymbol {\\theta} = \\left( \\begin{array}{c} \\| \\mathbf {n} \\left(\\mathbf {x} _ {1}\\right) \\| ^ {2} \\\\ \\vdots \\\\ \\| \\mathbf {n} \\left(\\mathbf {x} _ {\\mathrm {K}}\\right) \\| ^ {2} \\end{array} \\right). \\tag {16}", + "image_path": "d8c121f06b6172aa5507328e5c4f02c034ac68c6c4381dfc1d17208d2aea2e79.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 613, + 479, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 613, + 479, + 626 + ], + "spans": [ + { + "bbox": [ + 130, + 613, + 479, + 626 + ], + "type": "text", + "content": "For robust estimation, RANSAC [12] is adopted to deal with noise and outliers." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 227, + 91, + 448, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 227, + 91, + 448, + 101 + ], + "spans": [ + { + "bbox": [ + 227, + 91, + 448, + 101 + ], + "type": "text", + "content": "Motion and Structure from Event-based Normal Flow" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 631, + 403, + 643 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 631, + 403, + 643 + ], + "spans": [ + { + "bbox": [ + 133, + 631, + 403, + 643 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 133, + 631, + 403, + 643 + ], + "type": "inline_equation", + "content": "\\mathbf{O}(\\mathbf{x})" + }, + { + "bbox": [ + 133, + 631, + 403, + 643 + ], + "type": "text", + "content": " degenerates to 1 in the problem of optical flow estimation." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 135, + 643, + 479, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 643, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 135, + 643, + 479, + 665 + ], + "type": "text", + "content": "An approximation can be made based on the assumption of constant camera motion. In this way, the normal flow estimates can be collected within a short time interval." + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": "Continuous-Time Nonlinear Solver. To handle the case of a sudden variation in speed that may violate the constant-motion assumption, we further propose a continuous-time nonlinear solver. In particular, we employ cubic B-splines to represent the continuous-time model parameters as" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 249, + 174, + 481, + 205 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 249, + 174, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 249, + 174, + 481, + 205 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\theta} (u (t)) = \\sum_ {i = 0} ^ {3} B _ {i, 3} (u (t)) \\mathbf {c} _ {i}, \\tag {17}", + "image_path": "c0a2e4983adb25a21c9b8e1cc3d09b6aedac8005b9b647e95ede13468505c38e.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 213, + 480, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 213, + 480, + 297 + ], + "spans": [ + { + "bbox": [ + 130, + 213, + 480, + 297 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 213, + 480, + 297 + ], + "type": "inline_equation", + "content": "B_{i,k}" + }, + { + "bbox": [ + 130, + 213, + 480, + 297 + ], + "type": "text", + "content": " denotes the basis function, " + }, + { + "bbox": [ + 130, + 213, + 480, + 297 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 213, + 480, + 297 + ], + "type": "text", + "content": " the index of control points, " + }, + { + "bbox": [ + 130, + 213, + 480, + 297 + ], + "type": "inline_equation", + "content": "k (= 3)" + }, + { + "bbox": [ + 130, + 213, + 480, + 297 + ], + "type": "text", + "content": " the order of the spline and " + }, + { + "bbox": [ + 130, + 213, + 480, + 297 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_i \\in \\mathbb{R}^N" + }, + { + "bbox": [ + 130, + 213, + 480, + 297 + ], + "type": "text", + "content": " the corresponding control point defined in the " + }, + { + "bbox": [ + 130, + 213, + 480, + 297 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 213, + 480, + 297 + ], + "type": "text", + "content": "-dimensional space of model parameters. Examples of cubic B-splines applied to event camera motion estimation are found in [18, 28]. The function " + }, + { + "bbox": [ + 130, + 213, + 480, + 297 + ], + "type": "inline_equation", + "content": "u(\\cdot)" + }, + { + "bbox": [ + 130, + 213, + 480, + 297 + ], + "type": "text", + "content": " is a normalization operator, which transfers time " + }, + { + "bbox": [ + 130, + 213, + 480, + 297 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 213, + 480, + 297 + ], + "type": "text", + "content": " to the spline's parameter domain by means of basis translation [8]. In this way, the model parameters at any time instant " + }, + { + "bbox": [ + 130, + 213, + 480, + 297 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 213, + 480, + 297 + ], + "type": "text", + "content": " can be obtained via an interpolation using the known control points." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 298, + 480, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 298, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 130, + 298, + 480, + 357 + ], + "type": "text", + "content": "The formulation of the continuous-time model fitting problem goes as follows. The unknown variable consists of all the involved control points, " + }, + { + "bbox": [ + 130, + 298, + 480, + 357 + ], + "type": "inline_equation", + "content": "\\mathcal{X} \\doteq [\\mathbf{c}_0, \\mathbf{c}_1, \\dots, \\mathbf{c}_n]" + }, + { + "bbox": [ + 130, + 298, + 480, + 357 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 298, + 480, + 357 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 298, + 480, + 357 + ], + "type": "text", + "content": " denotes the total number of control points. By introducing the continuous-time representation of model parameters to (15), we finally create the following objective function:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 182, + 370, + 481, + 396 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 182, + 370, + 481, + 396 + ], + "spans": [ + { + "bbox": [ + 182, + 370, + 481, + 396 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\mathcal {X}} ^ {*} = \\arg \\min _ {\\boldsymbol {\\mathcal {X}}} \\sum_ {e _ {k} \\in \\mathcal {E}} \\left\\| \\mathbf {n} \\left(\\mathbf {x} _ {k}\\right) ^ {\\top} \\mathbf {O} \\left(\\mathbf {x} _ {k}\\right) \\boldsymbol {\\theta} \\left(u \\left(t _ {k}\\right)\\right) - \\| \\mathbf {n} \\left(\\mathbf {x} _ {k}\\right) \\| ^ {2} \\right\\| ^ {2}, \\tag {18}", + "image_path": "65a3c42c0fa0ca45ce9ebf9493bb57b1921530da4612454fff67d40b217f4e03.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 403, + 479, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 403, + 479, + 475 + ], + "spans": [ + { + "bbox": [ + 130, + 403, + 479, + 475 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 403, + 479, + 475 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 130, + 403, + 479, + 475 + ], + "type": "text", + "content": " denotes the event set in which the norm flow observation is available for every involved event " + }, + { + "bbox": [ + 130, + 403, + 479, + 475 + ], + "type": "inline_equation", + "content": "e_k" + }, + { + "bbox": [ + 130, + 403, + 479, + 475 + ], + "type": "text", + "content": ". Note that this continuous-time formulation aligns better with the asynchronous nature of event-based normal flow, enabling a state estimation over a larger temporal window that may lead to a better signal-to-noise ratio. The linear solver can be used to initialize the continuous-time nonlinear solver. An illustrative explanation of our solvers is given in Fig. 1." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 494, + 230, + 508 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 494, + 230, + 508 + ], + "spans": [ + { + "bbox": [ + 132, + 494, + 230, + 508 + ], + "type": "text", + "content": "5 Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 519, + 480, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 519, + 480, + 591 + ], + "spans": [ + { + "bbox": [ + 130, + 519, + 480, + 591 + ], + "type": "text", + "content": "To justify the effectiveness of the proposed algorithms, experiments are conducted on both simulation and real data. First, we introduce the datasets used and the applied metrics for evaluation (Sec. 5.1). Then we report the results of our linear solver and continuous-time nonlinear solver (Sec. 5.2 and Sec. 5.3). Finally, a discussion on the computational complexity and the limitation of our method is provided in Sec. 5.4." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 609, + 261, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 609, + 261, + 621 + ], + "spans": [ + { + "bbox": [ + 132, + 609, + 261, + 621 + ], + "type": "text", + "content": "5.1 Experimental Setup" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 629, + 479, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 479, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 479, + 666 + ], + "type": "text", + "content": "Our evaluation utilizes four datasets, including the indoor event-camera dataset (ECD) captured with a DAVIS240C camera and motion-capture system for groundtruth, and a high-resolution dataset with two rotational motion sequences" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "text", + "content": "Z. Ren et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 115, + 254, + 197 + ], + "blocks": [ + { + "bbox": [ + 133, + 115, + 254, + 197 + ], + "lines": [ + { + "bbox": [ + 133, + 115, + 254, + 197 + ], + "spans": [ + { + "bbox": [ + 133, + 115, + 254, + 197 + ], + "type": "image", + "image_path": "6d5184372b1a0e4636446f974281d9e865262d9d32e175502a23e5068bfac67b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 133, + 198, + 253, + 214 + ], + "lines": [ + { + "bbox": [ + 133, + 198, + 253, + 214 + ], + "spans": [ + { + "bbox": [ + 133, + 198, + 253, + 214 + ], + "type": "text", + "content": "(a) Estimated optical flow over a map of event accumulation." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 259, + 114, + 369, + 195 + ], + "blocks": [ + { + "bbox": [ + 259, + 114, + 369, + 195 + ], + "lines": [ + { + "bbox": [ + 259, + 114, + 369, + 195 + ], + "spans": [ + { + "bbox": [ + 259, + 114, + 369, + 195 + ], + "type": "image", + "image_path": "b484b8710ef572824360f0a0292a6cea7e4fcfc3d925bc5dcd9f1dbedecd369b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 260, + 198, + 480, + 215 + ], + "lines": [ + { + "bbox": [ + 260, + 198, + 480, + 215 + ], + "spans": [ + { + "bbox": [ + 260, + 198, + 480, + 215 + ], + "type": "text", + "content": "(b) Reconstruction result displayed over groundtruth structures (a v-shape roof) and viewed from different perspectives." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 370, + 114, + 480, + 196 + ], + "blocks": [ + { + "bbox": [ + 370, + 114, + 480, + 196 + ], + "lines": [ + { + "bbox": [ + 370, + 114, + 480, + 196 + ], + "spans": [ + { + "bbox": [ + 370, + 114, + 480, + 196 + ], + "type": "image", + "image_path": "dde771b4898659557cdb422227caec1f8c8fce581a2f70c55816fd436d16d6d3.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 224, + 482, + 246 + ], + "lines": [ + { + "bbox": [ + 130, + 224, + 482, + 246 + ], + "spans": [ + { + "bbox": [ + 130, + 224, + 482, + 246 + ], + "type": "text", + "content": "Fig. 3: Qualitative result of optical flow and depth on sequence three_wall Translation of our synthetic data. Note that our method returns sparse results." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 266, + 482, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 266, + 482, + 374 + ], + "spans": [ + { + "bbox": [ + 130, + 266, + 482, + 374 + ], + "type": "text", + "content": "using an iniVation DAVIS346 camera and Xsens IMU. The VECTor dataset, captured with a Prophesee Gen 3 camera and an accompanying RGB-D depth camera, is used for 6-DoF motion tracking. Additionally, a synthetic dataset generated in Blender is employed for motion-and-structure pattern analysis. Evaluation metrics for kinematics estimation include average error (AE) and root mean squared error (RMSE), while differential homography estimation is assessed using the Frobenius norm. Best results are in bold, and second-best are underlined. The supplementary material provides detailed descriptions of datasets and metrics, as well as implementation details of our methods." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 389, + 318, + 400 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 389, + 318, + 400 + ], + "spans": [ + { + "bbox": [ + 131, + 389, + 318, + 400 + ], + "type": "text", + "content": "5.2 Evaluation of the Linear Solver" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 406, + 481, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 406, + 481, + 453 + ], + "spans": [ + { + "bbox": [ + 130, + 406, + 481, + 453 + ], + "type": "text", + "content": "In the following evaluation, we qualitatively and quantitatively assess the performance of our linear method (Ours) in each specific task, and compare against different state-of-the-art (SOTA) methods (CMax [14] and Pro-STR [19]) and their customized version initialized using our linear solver." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 468, + 482, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 468, + 482, + 540 + ], + "spans": [ + { + "bbox": [ + 130, + 468, + 482, + 540 + ], + "type": "text", + "content": "Optical Flow and Depth Estimation. Due to a lack of benchmark for extensive evaluation, we qualitatively evaluate our method on the tasks of optical flow and depth estimation. As illustrated in Fig. 3a, our linear solver demonstrates the capability to accurately reconstruct full flow using normal flow and known motion parameters. An example of the depth estimation result is given in Fig. 3b, showing that the recovered structure aligns well with the groundtruth." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 555, + 482, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 555, + 482, + 603 + ], + "spans": [ + { + "bbox": [ + 130, + 555, + 482, + 603 + ], + "type": "text", + "content": "Angular Velocity Estimation. As shown in Tab. 1, we see that our linear solver outperforms CMax and Pro-STR. Besides, we find the combination of our linear solver and CMax leads to the best result, indicating that the proposed linear solver is a complementary method to existing nonlinear methods." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 617, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 665 + ], + "type": "text", + "content": "Differential Homography Estimation. The result is consistent with that in the above task. Table 2 reveals that our linear method outperforms CMax and Pro-STR in terms of numerical accuracy. The combination of our linear solver and state-of-the-art nonlinear solver leads to the best result." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 228, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 228, + 91, + 447, + 101 + ], + "type": "text", + "content": "Motion and Structure from Event-based Normal Flow" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 140, + 153, + 479, + 301 + ], + "blocks": [ + { + "bbox": [ + 137, + 133, + 475, + 143 + ], + "lines": [ + { + "bbox": [ + 137, + 133, + 475, + 143 + ], + "spans": [ + { + "bbox": [ + 137, + 133, + 475, + 143 + ], + "type": "text", + "content": "Table 1: Evaluation of our linear solver on the task of angular velocity estimation." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 140, + 153, + 479, + 301 + ], + "lines": [ + { + "bbox": [ + 140, + 153, + 479, + 301 + ], + "spans": [ + { + "bbox": [ + 140, + 153, + 479, + 301 + ], + "type": "table", + "html": "
MethodshapesRotation [29]groundRotation (our data)patternsRotation (our data)
eω(deg/s)RMSEω(deg/s)eω(deg/s)RMSEω(deg/s)eω(deg/s)RMSEω(deg/s)
CMax [14]15.6026.389.4815.585.959.07
Pro-STR [19]90.61145.8131.0744.7410.4711.88
Ours15.3623.983.308.014.706.08
Ours+CMax [14]10.1214.292.853.740.350.73
Ours+Pro-STR [19]19.2340.449.1317.074.315.23
MethoddynamicRotation [29]boxesRotation (our data)cubesRotation (our data)
eω(deg/s)RMSEω(deg/s)eω(deg/s)RMSEω(deg/s)eω(deg/s)RMSEω(deg/s)
CMax [14]16.1426.488.1813.566.7710.62
Pro-STR [19]54.6485.6322.7232.0511.0712.79
Ours9.4118.062.604.782.313.02
Ours+CMax [14]7.7110.852.643.800.490.77
Ours+Pro-STR [19]18.8732.653.495.872.042.76
", + "image_path": "8b7702183272a27efcd7157645f94b84df46160219da6cde800b334df3814271.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 181, + 364, + 435, + 432 + ], + "blocks": [ + { + "bbox": [ + 143, + 342, + 469, + 354 + ], + "lines": [ + { + "bbox": [ + 143, + 342, + 469, + 354 + ], + "spans": [ + { + "bbox": [ + 143, + 342, + 469, + 354 + ], + "type": "text", + "content": "Table 2: Evaluation of our linear solver for differential homography estimation." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 181, + 364, + 435, + 432 + ], + "lines": [ + { + "bbox": [ + 181, + 364, + 435, + 432 + ], + "spans": [ + { + "bbox": [ + 181, + 364, + 435, + 432 + ], + "type": "table", + "html": "
Methodboard Slow [15]patterns rotationpatterns 6 dof
F-NormF-NormF-Norm
CMax [14]1.161.331.27
Pro-STR [19]2.071.221.37
Ours0.240.200.43
Ours+CMax [14]0.230.150.41
Ours+Pro-STR [19]0.190.170.42
", + "image_path": "298324b7af0d69de2ce3e06b90bfbacbba88d7ee4dfebf6d38a6a0c4e312b05b.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 140, + 495, + 479, + 647 + ], + "blocks": [ + { + "bbox": [ + 144, + 473, + 466, + 485 + ], + "lines": [ + { + "bbox": [ + 144, + 473, + 466, + 485 + ], + "spans": [ + { + "bbox": [ + 144, + 473, + 466, + 485 + ], + "type": "text", + "content": "Table 3: Evaluation of our linear solver on the task of 6-DoF motion tracking." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 140, + 495, + 479, + 647 + ], + "lines": [ + { + "bbox": [ + 140, + 495, + 479, + 647 + ], + "spans": [ + { + "bbox": [ + 140, + 495, + 479, + 647 + ], + "type": "table", + "html": "
Methodcorner Slow [15]patterns _6 dof (our data)
eω(deg/s)RMSEω(deg/s)ev(m/s)RMSEv(m/s)eω(deg/s)RMSEω(deg/s)ev(m/s)RMSEv(m/s)
CMax [14]12.3914.390.340.4112.9415.681.021.23
Pro-STR [19]5.197.840.130.2211.2113.491.331.49
Ours1.311.790.010.020.560.890.120.16
Ours+CMax [14]2.192.820.110.170.550.870.180.30
Ours+Pro-STR [19]1.061.340.040.051.071.560.120.15
Methodmountain_normal [15]cubes _6 dof (our data)
eω(deg/s)RMSEω(deg/s)ev(m/s)RMSEv(m/s)eω(deg/s)RMSEω(deg/s)ev(m/s)RMSEv(m/s)
CMax [14]11.0114.170.440.5210.7414.330.781.05
Pro-STR [19]5.2310.530.140.259.7511.401.291.44
Ours1.281.830.070.090.791.020.110.14
Ours+CMax [14]3.846.300.210.301.141.720.120.14
Ours+Pro-STR [19]1.181.580.060.091.051.330.110.14
", + "image_path": "c1f86fcb6b160ce345ee4f36711bd686c2c77260d9a5055dbc7d1967b9d4ff75.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "Z. Ren et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 146, + 116, + 467, + 228 + ], + "blocks": [ + { + "bbox": [ + 146, + 116, + 467, + 228 + ], + "lines": [ + { + "bbox": [ + 146, + 116, + 467, + 228 + ], + "spans": [ + { + "bbox": [ + 146, + 116, + 467, + 228 + ], + "type": "image", + "image_path": "ed1bf464dcf21f8e30c6fc27c6c4a7026b849de293306d384b183408bec5592c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 149, + 239, + 465, + 266 + ], + "blocks": [ + { + "bbox": [ + 198, + 230, + 416, + 239 + ], + "lines": [ + { + "bbox": [ + 198, + 230, + 416, + 239 + ], + "spans": [ + { + "bbox": [ + 198, + 230, + 416, + 239 + ], + "type": "text", + "content": "(a) Results of the continuous-time angular velocity estimator." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 149, + 239, + 465, + 266 + ], + "lines": [ + { + "bbox": [ + 149, + 239, + 465, + 266 + ], + "spans": [ + { + "bbox": [ + 149, + 239, + 465, + 266 + ], + "type": "image", + "image_path": "9dbb303a4ba45eb7fccc2c30a6525065060729a1d1244f7888c7ce80fb08dfcd.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 149, + 267, + 466, + 345 + ], + "blocks": [ + { + "bbox": [ + 149, + 267, + 466, + 345 + ], + "lines": [ + { + "bbox": [ + 149, + 267, + 466, + 345 + ], + "spans": [ + { + "bbox": [ + 149, + 267, + 466, + 345 + ], + "type": "image", + "image_path": "80db2d36fcde1c23580460abe1e6b7dc7d0034b32ea1f45813e038fc04168c64.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 146, + 349, + 466, + 366 + ], + "lines": [ + { + "bbox": [ + 146, + 349, + 466, + 366 + ], + "spans": [ + { + "bbox": [ + 146, + 349, + 466, + 366 + ], + "type": "text", + "content": "(b) Comparison against two state-of-the-art nonlinear solutions using data segments that involve sudden variations of angular velocity." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 132, + 375, + 479, + 387 + ], + "lines": [ + { + "bbox": [ + 132, + 375, + 479, + 387 + ], + "spans": [ + { + "bbox": [ + 132, + 375, + 479, + 387 + ], + "type": "text", + "content": "Fig. 4: Angular velocity estimation using the continuous-time nonlinear solver (Ours)." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 411, + 482, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 411, + 482, + 495 + ], + "spans": [ + { + "bbox": [ + 130, + 411, + 482, + 495 + ], + "type": "text", + "content": "6-DoF Motion Tracking. Given known depth, we recover the linear and angular camera velocities on the task of 6-DoF motion estimation. As shown in Tab. 3, our linear solver reports more accurate results than both CMax and Pro-STR. In addition, we surprisingly find that the linear solver can sometimes outperform the combination of itself with the two nonlinear solvers. This is, according to our investigation, attributed to the violation of constant-motion assumption in some sequences, which is held for the two nonlinear solvers." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 513, + 373, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 513, + 373, + 525 + ], + "spans": [ + { + "bbox": [ + 132, + 513, + 373, + 525 + ], + "type": "text", + "content": "5.3 Evaluation of the Continuous-Time Solver" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "To justify the claimed advantage of the proposed continuous-time nonlinear solver, we use the exemplary task of angular velocity estimation. The sequence used (shapesRotation [29]) captures a purely rotating event camera, and some parts of the data involves aggressive motion. As illustrated in Fig. 4a, our nonlinear solver gives accurate estimation results with a high stability. Moreover, we select three segments that witness a sudden variation of angular speed within a short time interval and compare our continuous-time nonlinear solver against CMax and Pro-STR. Note that all of them are initialized by our linear method for a fair comparison. As Fig. 4b shows, our results are more consistent with the groundtruth angular velocity than the other two methods, which are less accurate due to the dependence on the constant-motion assumption." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 228, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 228, + 91, + 447, + 101 + ], + "type": "text", + "content": "Motion and Structure from Event-based Normal Flow" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 173, + 136, + 443, + 178 + ], + "blocks": [ + { + "bbox": [ + 135, + 114, + 477, + 126 + ], + "lines": [ + { + "bbox": [ + 135, + 114, + 477, + 126 + ], + "spans": [ + { + "bbox": [ + 135, + 114, + 477, + 126 + ], + "type": "text", + "content": "Table 4: Runtime of 6-DoF tracking on sequence mountain_normal (640×480 px)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 173, + 136, + 443, + 178 + ], + "lines": [ + { + "bbox": [ + 173, + 136, + 443, + 178 + ], + "spans": [ + { + "bbox": [ + 173, + 136, + 443, + 178 + ], + "type": "table", + "html": "
ComponentRuntime
Normal flow extraction20 ms (for extracting 2000 normal flow vectors)
Linear solver25.97 ms
CT solver28.29 ms
", + "image_path": "9656dabb1e96a819276ad7d926f7181243e5a58113f2f83a3e7e730899958425.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 198, + 214, + 209 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 198, + 214, + 209 + ], + "spans": [ + { + "bbox": [ + 132, + 198, + 214, + 209 + ], + "type": "text", + "content": "5.4 Discussion" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 215, + 482, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 215, + 482, + 312 + ], + "spans": [ + { + "bbox": [ + 130, + 215, + 482, + 312 + ], + "type": "text", + "content": "Computational Complexity. The computational complexity of the linear solver depends on the specific task of model fitting. For optical flow and depth estimation, it is linearly proportional to the size of the input event flow, and thus, the complexity is " + }, + { + "bbox": [ + 130, + 215, + 482, + 312 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(N)" + }, + { + "bbox": [ + 130, + 215, + 482, + 312 + ], + "type": "text", + "content": ". For the other tasks, the time cost predominantly hinges on the total number of RANSAC iterations and the computations performed per iteration. In each iteration, we typically solve a linear overdetermined system via QR decomposition at a complexity of " + }, + { + "bbox": [ + 130, + 215, + 482, + 312 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(cN^2)" + }, + { + "bbox": [ + 130, + 215, + 482, + 312 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 215, + 482, + 312 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 215, + 482, + 312 + ], + "type": "text", + "content": " is the cardinality of a minimal problem (e.g., " + }, + { + "bbox": [ + 130, + 215, + 482, + 312 + ], + "type": "inline_equation", + "content": "c = 3" + }, + { + "bbox": [ + 130, + 215, + 482, + 312 + ], + "type": "text", + "content": " for the problem of angular velocity estimation)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 325, + 481, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 325, + 481, + 374 + ], + "spans": [ + { + "bbox": [ + 130, + 325, + 481, + 374 + ], + "type": "text", + "content": "Runtime. All experiments are conducted on a standard computer with Intel i7-13700F CPU, 16G memory and implemented with MATLAB (version R2022a). Using the 6-DoF tracking task as an example, we measure the runtime of the proposed linear solver and continuous-time (CT) solver, as reported in Tab. 4." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 388, + 482, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 388, + 482, + 496 + ], + "spans": [ + { + "bbox": [ + 130, + 388, + 482, + 496 + ], + "type": "text", + "content": "Limitations. To make it clear, we emphasize that our method, like other compared methods, does not resolve the motion and structure parameters simultaneously. It should also be noted that the success of our method relies heavily on the quality of the input normal flow. Particularly, our normal flow extractor (including some other open source solutions for computing normal flow from event data) may fail in the presence of densely repetitive textures, leading to a failure of the modeling fitting scheme. To assess the sensitivity of our method to the quality of event-based normal flow, we propose an analysis on numerical stability using simulation data in the supplementary material." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 511, + 227, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 511, + 227, + 525 + ], + "spans": [ + { + "bbox": [ + 132, + 511, + 227, + 525 + ], + "type": "text", + "content": "6 Conclusions" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "This work presents a solution to the general problem of geometric model fitting with event data. Considering the differential working principle of event cameras, our method establishes specific constraints between instantaneous motion-and-structure parameters and event-based normal flow. Experiments show that the proposed linear solver leads to closed-form and deterministic solutions, which can be used to initialize existing state-of-the-art nonlinear methods. The proposed continuous-time nonlinear solver, furthermore, demonstrates the capability to handle sudden speed variations, obviating the dependence on the constant-motion assumption. Finally, we hope this work can inspire future research on event-based geometric model fitting, a paramount topic to many downstream techniques, such as event-based 3-D vision, SLAM, and scene segmentation." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "Z. Ren et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 240, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 240, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 240, + 129 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 143, + 482, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 143, + 482, + 203 + ], + "spans": [ + { + "bbox": [ + 130, + 143, + 482, + 203 + ], + "type": "text", + "content": "We thank Dr. Yi Yu for proof reading. This work was supported by the National Key Research and Development Project of China under Grant 2023YFB4706600. Funded by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) under Germany's Excellence Strategy - EXC 2002/1 \"Science of Intelligence\" - project number 390523135." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 224, + 197, + 237 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 224, + 197, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 224, + 197, + 237 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 251, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 138, + 251, + 481, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 251, + 481, + 285 + ], + "spans": [ + { + "bbox": [ + 138, + 251, + 481, + 285 + ], + "type": "text", + "content": "1. Barranco, F., Fermüller, C., Aloimonos, Y., Ros, E.: Joint direct estimation of 3D geometry and 3D motion using spatio temporal gradients. Pattern Recog. 113, 107759 (2021). https://doi.org/10.1016/j.patcog.2020.107759" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 286, + 481, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 286, + 481, + 319 + ], + "spans": [ + { + "bbox": [ + 138, + 286, + 481, + 319 + ], + "type": "text", + "content": "2. Benosman, R., Clercq, C., Lagorce, X., Ieng, S.H., Bartolozzi, C.: Event-based visual flow. IEEE Trans. Neural Netw. Learn. Syst. 25(2), 407-417 (2014). https://doi.org/10.1109/TNNLS.2013.2273537" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 320, + 481, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 320, + 481, + 342 + ], + "spans": [ + { + "bbox": [ + 138, + 320, + 481, + 342 + ], + "type": "text", + "content": "3. Brodsky, T., Fermüller, C., Aloimonos, Y.: Structure from motion: Beyond the epipolar constraint. Int. J. Comput. Vis. 37, 231-258 (2000)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 343, + 481, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 343, + 481, + 374 + ], + "spans": [ + { + "bbox": [ + 138, + 343, + 481, + 374 + ], + "type": "text", + "content": "4. Bryner, S., Gallego, G., Rebecq, H., Scaramuzza, D.: Event-based, direct camera tracking from a photometric 3D map using nonlinear optimization. In: IEEE Int. Conf. Robot. Autom. (ICRA) (2019). https://doi.org/10.1109/ICRA.2019.8794255" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 376, + 481, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 376, + 481, + 407 + ], + "spans": [ + { + "bbox": [ + 138, + 376, + 481, + 407 + ], + "type": "text", + "content": "5. Clady, X., Clercq, C., Ieng, S.H., Houseini, F., Randazzo, M., Natale, L., Bartolozzi, C., Benosman, R.: Asynchronous visual event-based time-to-contact. Front. Neurosci. 8(9) (2014). https://doi.org/10.3389/fnins.2014.00009" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 409, + 481, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 409, + 481, + 452 + ], + "spans": [ + { + "bbox": [ + 138, + 409, + 481, + 452 + ], + "type": "text", + "content": "6. Conradt, J., Cook, M., Berner, R., Lichtsteiner, P., Douglas, R.J., Delbruck, T.: A pencil balancing robot using a pair of AER dynamic vision sensors. In: IEEE Int. Symp. Circuits Syst. (ISCAS). pp. 781-784 (2009). https://doi.org/10.1109/ISCAS.2009.5117867" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 453, + 481, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 453, + 481, + 485 + ], + "spans": [ + { + "bbox": [ + 138, + 453, + 481, + 485 + ], + "type": "text", + "content": "7. Corke, P.: Robotics, Vision and Control: Fundamental Algorithms in MATLAB. Springer Tracts in Advanced Robotics, Springer (2017). https://doi.org/10.1007/978-3-319-54413-7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 487, + 481, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 487, + 481, + 508 + ], + "spans": [ + { + "bbox": [ + 138, + 487, + 481, + 508 + ], + "type": "text", + "content": "8. De Boor, C.: On calculating with b-splines. Journal of Approximation theory 6(1), 50-62 (1972)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 510, + 481, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 510, + 481, + 542 + ], + "spans": [ + { + "bbox": [ + 138, + 510, + 481, + 542 + ], + "type": "text", + "content": "9. Delbruck, T., Lang, M.: Robotic goalie with 3ms reaction time at " + }, + { + "bbox": [ + 138, + 510, + 481, + 542 + ], + "type": "inline_equation", + "content": "4\\%" + }, + { + "bbox": [ + 138, + 510, + 481, + 542 + ], + "type": "text", + "content": " CPU load using event-based dynamic vision sensor. Front. Neurosci. 7, 223 (2013). https://doi.org/10.3389/fnins.2013.00223" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 544, + 481, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 544, + 481, + 576 + ], + "spans": [ + { + "bbox": [ + 138, + 544, + 481, + 576 + ], + "type": "text", + "content": "0. Dinaux, R., Wessendorp, N., Dupeyroux, J., De Croon, G.C.: Faith: Fast iterative half-plane focus of expansion estimation using optic flow. IEEE Robotics and Automation Letters 6(4), 7627-7634 (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 577, + 481, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 577, + 481, + 597 + ], + "spans": [ + { + "bbox": [ + 138, + 577, + 481, + 597 + ], + "type": "text", + "content": "1. Fermüller, C.: Passive navigation as a pattern recognition problem. Int. J. Comput. Vis. 14(2), 147-158 (1995)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 599, + 481, + 631 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 599, + 481, + 631 + ], + "spans": [ + { + "bbox": [ + 134, + 599, + 481, + 631 + ], + "type": "text", + "content": "12. Fischler, M.A., Bolles, R.C.: Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Commun. ACM 24(6), 381-395 (1981). https://doi.org/10.1145/358669.358692" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 134, + 632, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 632, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 632, + 481, + 665 + ], + "type": "text", + "content": "13. Gallego, G., Gehrig, M., Scaramuzza, D.: Focus is all you need: Loss functions for event-based vision. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR). pp. 12272-12281 (2019). https://doi.org/10.1109/CVPR.2019.01256" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 227, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 227, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 227, + 91, + 447, + 101 + ], + "type": "text", + "content": "Motion and Structure from Event-based Normal Flow" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "type": "text", + "content": "14. Gallego, G., Rebecq, H., Scaramuzza, D.: A unifying contrast maximization framework for event cameras, with applications to motion, depth, and optical flow estimation. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR). pp. 3867-3876 (2018). https://doi.org/10.1109/CVPR.2018.00407" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 160, + 481, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 160, + 481, + 193 + ], + "spans": [ + { + "bbox": [ + 133, + 160, + 481, + 193 + ], + "type": "text", + "content": "15. Gao, L., Liang, Y., Yang, J., Wu, S., Wang, C., Chen, J., Kneip, L.: Vector: A versatile event-centric benchmark for multi-sensor slam. IEEE Robotics and Automation Letters 7(3), 8217-8224 (2022). https://doi.org/10.1109/LRA.2022.3186770" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 193, + 481, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 193, + 481, + 235 + ], + "spans": [ + { + "bbox": [ + 133, + 193, + 481, + 235 + ], + "type": "text", + "content": "16. Gao, L., Su, H., Gehrig, D., Cannici, M., Scaramuzza, D., Kneip, L.: A 5-point minimal solver for event camera relative motion estimation. In: Int. Conf. Comput. Vis. (ICCV). pp. 8015-8025 (2023). https://doi.org/10.1109/ICCV51070.2023.00739" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 236, + 481, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 236, + 481, + 280 + ], + "spans": [ + { + "bbox": [ + 133, + 236, + 481, + 280 + ], + "type": "text", + "content": "17. Gu, C., Learned-Miller, E., Sheldon, D., Gallego, G., Bideau, P.: The spatiotemporal Poisson point process: A simple model for the alignment of event camera data. In: Int. Conf. Comput. Vis. (ICCV). pp. 13495-13504 (2021). https://doi.org/10.1109/ICCV48922.2021.01324" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 280, + 481, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 280, + 481, + 311 + ], + "spans": [ + { + "bbox": [ + 133, + 280, + 481, + 311 + ], + "type": "text", + "content": "18. Guo, S., Gallego, G.: CMax-SLAM: Event-based rotational-motion bundle adjustment and SLAM system using contrast maximization. IEEE Trans. Robot. pp. 1-20 (2024). https://doi.org/10.1109/TRO.2024.3378443" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 311, + 481, + 343 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 311, + 481, + 343 + ], + "spans": [ + { + "bbox": [ + 133, + 311, + 481, + 343 + ], + "type": "text", + "content": "19. Huang, X., Zhang, Y., Xiong, Z.: Progressive spatio-temporal alignment for efficient event-based motion estimation. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR). pp. 1537-1546 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 343, + 481, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 343, + 481, + 365 + ], + "spans": [ + { + "bbox": [ + 133, + 343, + 481, + 365 + ], + "type": "text", + "content": "20. Ji, H., Fermuller, C.: A 3D shape constraint on video. IEEE Trans. Pattern Anal. Mach. Intell. 28(6), 1018-1023 (2006)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 133, + 365, + 481, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 365, + 481, + 397 + ], + "spans": [ + { + "bbox": [ + 133, + 365, + 481, + 397 + ], + "type": "text", + "content": "21. Kim, H., Leutenegger, S., Davison, A.J.: Real-time 3D reconstruction and 6-DoF tracking with an event camera. In: Eur. Conf. Comput. Vis. (ECCV). pp. 349-364 (2016). https://doi.org/\\protect{10.1007/978-3-319-46466-4_21}" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 133, + 397, + 481, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 397, + 481, + 430 + ], + "spans": [ + { + "bbox": [ + 133, + 397, + 481, + 430 + ], + "type": "text", + "content": "22. Liu, D., Parra, A., Chin, T.J.: Spatiotemporal registration for event-based visual odometry. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR). pp. 4935-4944 (2021). https://doi.org/10.1109/CVPR46437.2021.00490" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 133, + 430, + 481, + 461 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 430, + 481, + 461 + ], + "spans": [ + { + "bbox": [ + 133, + 430, + 481, + 461 + ], + "type": "text", + "content": "23. Longuet-Higgins, H.C., Prazdny, K.: The interpretation of a moving retinal image. Proceedings of the Royal Society of London. Series B. Biological Sciences 208(1173), 385-397 (1980). https://doi.org/10.1098/rspb.1980.0057" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 133, + 461, + 481, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 461, + 481, + 483 + ], + "spans": [ + { + "bbox": [ + 133, + 461, + 481, + 483 + ], + "type": "text", + "content": "24. Lu, X., Zhou, Y., Niu, J., Zhong, S., Shen, S.: Event-based visual inertial velometer (2024)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 133, + 483, + 481, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 483, + 481, + 504 + ], + "spans": [ + { + "bbox": [ + 133, + 483, + 481, + 504 + ], + "type": "text", + "content": "25. Ma, Y., Soatto, S., Košecká, J., Sastry, S.: An Invitation to 3-D Vision: From Images to Geometric Models. Springer (2004)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 133, + 504, + 481, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 504, + 481, + 536 + ], + "spans": [ + { + "bbox": [ + 133, + 504, + 481, + 536 + ], + "type": "text", + "content": "26. Meyer, F., Bouthemy, P.: Estimation of time-to-collision maps from first order motion models and normal flows. In: 1992 11th IAPR International Conference on Pattern Recognition. vol. 1, pp. 78-82. IEEE Computer Society (1992)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 536, + 481, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 536, + 481, + 557 + ], + "spans": [ + { + "bbox": [ + 133, + 536, + 481, + 557 + ], + "type": "text", + "content": "27. Meyer, F.G.: Time-to-collision from first-order models of the motion field. IEEE transactions on robotics and automation 10(6), 792-798 (1994)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 133, + 557, + 481, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 557, + 481, + 589 + ], + "spans": [ + { + "bbox": [ + 133, + 557, + 481, + 589 + ], + "type": "text", + "content": "28. Mueggler, E., Gallego, G., Rebecq, H., Scaramuzza, D.: Continuous-time visual-inertial odometry for event cameras. IEEE Trans. Robot. 34(6), 1425-1440 (Dec 2018). https://doi.org/10.1109/tro.2018.2858287" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 133, + 589, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 589, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 133, + 589, + 481, + 632 + ], + "type": "text", + "content": "29. Mueggler, E., Rebecq, H., Gallego, G., Delbruck, T., Scaramuzza, D.: The event-camera dataset and simulator: Event-based data for pose estimation, visual odometry, and SLAM. Int. J. Robot. Research 36(2), 142-149 (2017). https://doi.org/10.1177/0278364917691115" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 133, + 632, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 632, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 632, + 481, + 666 + ], + "type": "text", + "content": "30. Nunes, U.M., Demiris, Y.: Entropy minimisation framework for event-based vision model estimation. In: Eur. Conf. Comput. Vis. (ECCV). pp. 161-176 (2020). https://doi.org/10.1007/978-3-030-58558-7_10" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "Z. Ren et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 534 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "31. Parameshwara, C.M., Hari, G., Fermüller, C., Sanket, N.J., Aloimonos, Y.: Diff-posednet: direct differentiable camera pose estimation. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR). pp. 6845-6854 (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 481, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 481, + 171 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 481, + 171 + ], + "type": "text", + "content": "32. Pijnacker Hordijk, B.J., Scheper, K.Y., De Croon, G.C.: Vertical landing for micro air vehicles using event-based optical flow. J. Field Robot. 35(1), 69-90 (2018)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 481, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 481, + 193 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 481, + 193 + ], + "type": "text", + "content": "33. Rebecq, H., Gallego, G., Scaramuzza, D.: EMVS: Event-based multi-view stereo. In: British Mach. Vis. Conf. (BMVC) (2016). https://doi.org/10.5244/C.30.63" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 194, + 480, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 480, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 480, + 226 + ], + "type": "text", + "content": "34. Rebecq, H., Horstschaefer, T., Scaramuzza, D.: Real-time visual-inertial odometry for event cameras using keyframe-based nonlinear optimization. In: British Mach. Vis. Conf. (BMVC) (2017)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 228, + 481, + 269 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 228, + 481, + 269 + ], + "spans": [ + { + "bbox": [ + 132, + 228, + 481, + 269 + ], + "type": "text", + "content": "35. Rebecq, H., Horstschäfer, T., Gallego, G., Scaramuzza, D.: EVO: A geometric approach to event-based 6-DOF parallel tracking and mapping in real-time. IEEE Robot. Autom. Lett. 2(2), 593-600 (2017). https://doi.org/10.1109/LRA.2016.2645143" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 270, + 481, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 270, + 481, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 270, + 481, + 293 + ], + "type": "text", + "content": "36. Shiba, S., Aoki, Y., Gallego, G.: Event collapse in contrast maximization frameworks. Sensors 22(14), 1-20 (2022). https://doi.org/10.3390/s22145190" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 293, + 481, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 481, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 481, + 325 + ], + "type": "text", + "content": "37. Shiba, S., Aoki, Y., Gallego, G.: A fast geometric regularizer to mitigate event collapse in the contrast maximization framework. Adv. Intell. Syst. p. 2200251 (2022). https://doi.org/10.1002/aisy.202200251" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 326, + 481, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 481, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 481, + 357 + ], + "type": "text", + "content": "38. Shiba, S., Aoki, Y., Gallego, G.: Secrets of event-based optical flow. In: Eur. Conf. Comput. Vis. (ECCV). pp. 628-645 (2022). https://doi.org/10.1007/978-3-031-19797-0 36" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 358, + 481, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 481, + 390 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 481, + 390 + ], + "type": "text", + "content": "39. Stoffregen, T., Gallego, G., Drummond, T., Kleeman, L., Scaramuzza, D.: Event-based motion segmentation by motion compensation. In: Int. Conf. Comput. Vis. (ICCV). pp. 7243-7252 (2019). https://doi.org/10.1109/ICCV.2019.00734" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 392, + 481, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 481, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 481, + 413 + ], + "type": "text", + "content": "40. Trucco, E., Verri, A.: Introductory Techniques for 3-D Computer Vision. Prentice Hall PTR, Upper Saddle River, NJ, USA (1998)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 133, + 414, + 481, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 414, + 481, + 446 + ], + "spans": [ + { + "bbox": [ + 133, + 414, + 481, + 446 + ], + "type": "text", + "content": "41. Zhou, Y., Gallego, G., Rebecq, H., Kneip, L., Li, H., Scaramuzza, D.: Semidense 3D reconstruction with a stereo event camera. In: Eur. Conf. Comput. Vis. (ECCV). pp. 242-258 (2018). https://doi.org/10.1007/978-3-030-01246-5_15" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 133, + 447, + 481, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 447, + 481, + 468 + ], + "spans": [ + { + "bbox": [ + 133, + 447, + 481, + 468 + ], + "type": "text", + "content": "42. Zhou, Y., Gallego, G., Shen, S.: Event-based stereo visual odometry. IEEE Trans. Robot. 37(5), 1433-1450 (2021). https://doi.org/10.1109/TRO.2021.3062252" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 133, + 469, + 481, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 469, + 481, + 500 + ], + "spans": [ + { + "bbox": [ + 133, + 469, + 481, + 500 + ], + "type": "text", + "content": "43. Zhu, A.Z., Chen, Y., Daniilidis, K.: Realtime time synchronized event-based stereo. In: Eur. Conf. Comput. Vis. (ECCV). pp. 438-452 (2018). https://doi.org/10.1007/978-3-030-01231-1_27" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 501, + 481, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 501, + 481, + 534 + ], + "spans": [ + { + "bbox": [ + 133, + 501, + 481, + 534 + ], + "type": "text", + "content": "44. Zhuang, B., Tran, Q.H.: Image stitching and rectification for hand-held cameras. In: Eur. Conf. Comput. Vis. (ECCV) (2020). https://doi.org/10.1007/978-3-030-58571-6_15" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 228, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 228, + 91, + 447, + 101 + ], + "type": "text", + "content": "Motion and Structure from Event-based Normal Flow" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Motion-Guided Latent Diffusion for Temporally Consistent Real-world Video Super-resolution/dbea7f31-cd1b-433d-b163-70053b8d34db_content_list.json b/2024/Motion-Guided Latent Diffusion for Temporally Consistent Real-world Video Super-resolution/dbea7f31-cd1b-433d-b163-70053b8d34db_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..ab82d50e35e05e6a1d28a69c8aa77a93855ccff2 --- /dev/null +++ b/2024/Motion-Guided Latent Diffusion for Temporally Consistent Real-world Video Super-resolution/dbea7f31-cd1b-433d-b163-70053b8d34db_content_list.json @@ -0,0 +1,2353 @@ +[ + { + "type": "text", + "text": "Motion-Guided Latent Diffusion for Temporally Consistent Real-world Video Super-resolution", + "text_level": 1, + "bbox": [ + 228, + 140, + 774, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xi Yang $^{1,2}$ , Chenhang He $^{1}$ , Jianqi Ma $^{1,2}$ , and Lei Zhang $^{1,2^{\\star}}$", + "bbox": [ + 284, + 218, + 715, + 236 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 The Hong Kong Polytechnic University, Hong Kong SAR, China \n2 OPPO Research Institute, Shenzhen, China \n{xxxxi.yang, jianqi.ma}@connect.polyu.hk, chenhang.he@polyu.edu.hk, cslzhang@comp.polyu.edu.hk", + "bbox": [ + 245, + 246, + 756, + 304 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Real-world low-resolution (LR) videos have diverse and complex degradations, imposing great challenges on video super-resolution (VSR) algorithms to reproduce their high-resolution (HR) counterparts with high quality. Recently, the diffusion models have shown compelling performance in generating realistic details for image restoration tasks. However, the diffusion process has randomness, making it hard to control the contents of restored images. This issue becomes more serious when applying diffusion models to VSR tasks because temporal consistency is crucial to the perceptual quality of videos. In this paper, we propose an effective real-world VSR algorithm by leveraging the strength of pre-trained latent diffusion models. To ensure the content consistency among adjacent frames, we exploit the temporal dynamics in LR videos to guide the diffusion process by optimizing the latent sampling path with a motion-guided loss, ensuring that the generated HR video maintains a coherent and continuous visual flow. To further mitigate the discontinuity of generated details, we insert temporal modules to the decoder and fine-tune it with an innovative sequence-oriented loss. The proposed motion-guided latent diffusion (MGLD) based VSR algorithm achieves significantly better perceptual quality than state-of-the-arts on real-world VSR benchmark datasets, validating the effectiveness of the proposed model design and training strategies. Codes and models are available at https://github.com/IanYeung/MGLD-VSR.", + "bbox": [ + 261, + 338, + 743, + 645 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Video super-resolution $\\cdot$ Motion-guided latent diffusion", + "bbox": [ + 261, + 656, + 718, + 672 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 696, + 375, + 713 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Video super-resolution (VSR) [3] aims at reconstructing a high-resolution (HR) video from a given low-resolution (LR) video sequence. Thanks to the rapid development of deep learning techniques, significant progress has been made for VSR in the past decade. Representative VSR methods include sliding window based methods like EDVR [49], recurrent based methods like BasicVSR [9] and", + "bbox": [ + 212, + 726, + 787, + 805 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding author. This work is supported by the Hong Kong RGC RIF grant (R5001-18) and the PolyU-OPPO Joint Innovation Lab.", + "bbox": [ + 217, + 810, + 785, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7b6e1ee9655944cacfa565d1a1a9b298bcfe8dd926e46e0f9828df0ca4fb1d2c.jpg", + "image_caption": [ + "Fig. 1: The VSR results of 4 consecutive frames in the Sequence 026 from VideoLQ. The top row shows the VSR results by RealBasicVSR [11], the middle row shows the naive VSR results by running StableSR [48] on each frame of the sequence, and the bottom row shows the VSR results produced by our proposed MGLD-VSR method. Our method generates realistic details while achieving good temporal consistency." + ], + "image_footnote": [], + "bbox": [ + 256, + 142, + 751, + 422 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "more recent transformer based methods [30]. However, most of the above works assume simple degradations (e.g., bicubic resizing, downsampling after Gaussian blur) between the LR and HR videos. As a result, the trained VSR models are hard to be generalized to real-world LR videos, whose degradations are unknown and much more complex.", + "bbox": [ + 212, + 535, + 787, + 611 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, real-world VSR [11, 53, 55] has gained attention from researchers for its high potential value in practical applications such as camera-phone video enhancement and online video streaming. Different from traditional VSR tasks, which focus on super-resolving clean LR videos with simple synthetic degradation (e.g., bicubic downsampling), real-world VSR aims to enhance videos with complex and unknown degradations that often appear in real-world low-quality videos. The main challenge lies in how to effectively reproduce video details while suppressing the visual artifacts caused by unknown degradations. Existing real-world VSR algorithms mostly attempt to trade-off between details and artifacts using techniques such as Laplacian pyramid generative adversarial loss [55], input pre-cleaning [11] and hidden state attention [53], etc. Despite these efforts and progresses, their performance on real-world testing sets remains limited due to the insufficient training data and the limited model capacity.", + "bbox": [ + 212, + 611, + 787, + 808 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Since the seminal work of denoising diffusion probabilistic model (DDPM) [20], diffusion-based generative models have achieved great success in image gen", + "bbox": [ + 214, + 809, + 785, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "X. Yang et al.", + "bbox": [ + 271, + 114, + 366, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "eration. In particular, the latent diffusion models (LDMs) [42] have demonstrated compelling results not only in text-to-image generation, but also in downstream tasks such as image editing [24], inpainting [32], colorization [7], etc. Recently, researchers have also attempted to employ the powerful generative priors of LDMs for real-world image restoration tasks [48] and shown promising results. It is thus interesting to investigate whether the latent diffusion priors can be used to improve real-world VSR results. However, this is a non-trivial task because of the intrinsic randomness in the diffusion process. This randomness in turn gives rise to the temporal inconsistencies among video frames in the latent space. Such inconsistencies can be further amplified by the decoder of the LDM. As shown in the middle row of Figure 1, directly employing the LDM based image super-resolution algorithm [48] for VSR will lead to inconsistent details across consecutive frames, disturbing the visual perception of the reproduced video.", + "bbox": [ + 212, + 145, + 787, + 343 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To address the above-mentioned challenges, we propose a Motion-Guided Latent Diffusion (MGLD) model for real-world VSR in this paper, aiming to produce high-quality HR video sequence with good temporal consistency. We propose to incorporate the motion dynamics of input LR video into the generation of SR outputs. This can be done through a guided diffusion process where we can generate conditional samples from a guidance model based on its score function [14, 44]. In particular, we calculate the optical flows among adjacent LR frames, and use the calculated optical flows to warp the latent features of each frame to align the features of neighboring frames. The $l_{1}$ norm of the warping error is employed as the motion-guided loss, and its gradients can be incorporated in the sampling process to update the latent features.", + "bbox": [ + 212, + 343, + 787, + 508 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To overcome the inconsistent details generated by the original decoder of LDM, we insert temporal modules to the decoder and fine-tune it with the ground-truth HR video sequences. We also introduce an innovative sequence-oriented loss to guide the decoder to improve the continuity of the details. As depicted in the bottom row of Figure 1, our proposed MGLD-VSR approach could mitigate the discontinuity of generated details while maintaining natural and visually appealing texture recovery.", + "bbox": [ + 212, + 508, + 787, + 614 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The contributions of this paper are summarized as follows. First, we propose a diffusion sampling process based on motion-guided loss, allowing the temporal dynamics of the input frame to be used in generating temporally consistent latent features. Second, we propose a temporal-aware sequence decoder, along with two sequence-oriented losses, to further enhance the continuity of generated videos. Our proposed MGLD-VSR model achieves highly competitive real-world VSR results, exhibiting perceptually much more realistic details with fewer flickering artifacts than existing state-of-the-arts.", + "bbox": [ + 212, + 614, + 787, + 734 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 760, + 387, + 777 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Video Super-Resolution. The recent advances in VSR research largely attribute to the rapid development of deep-learning technologies [6, 9, 22, 30, 49]. Based on the paradigms, existing deep-learning based VSR algorithms could", + "bbox": [ + 212, + 794, + 785, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "MGLD-VSR", + "bbox": [ + 645, + 114, + 730, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "be roughly classified into two categories. The first category performs VSR in a multiple-input single-output (MISO) fashion. Typical algorithms include DUF [22], EDVR [49], and MuCAN [28], etc. The MISO algorithms are usually sliding-window based, which aggregate information from neighboring frames to generate the super-resolved result of the center frame. The other category of methods adopt a multiple-input multiple-output (MIMO) design, such as BasicVSR [9], BasicVSR++ [10], VRT [30] and RVRT [31]. These methods take multiple LR frames as input and generate their corresponding SR results in one go. Compared with MISO, the amortized running cost of MIMO is usually much lower as it avoids a series of redundant operations. In addition, benefiting from the lower cost, the MIMO methods are able to aggregate information from more frames to improve the performance. We adopt the MIMO design in this work.", + "bbox": [ + 212, + 146, + 787, + 328 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Real-world VSR. Most existing VSR methods are trained on synthetic data with simple degradations (e.g., bicubic resizing, downsampling after Gaussian blur). Due to the synthetic-to-real gap, models trained with such data cannot perform well on real-world videos. To tackle this problem, Yang et al. [55] collected LR-HR video pairs with iPhone cameras to better model real-world degradations. While being effective in some cases, it is relatively labor-intensive and may not generalize well to videos collected by other devices. Later works usually adopt an easier solution: synthesizing more realistic training data by mixing more types of degradations, such as blur, downsampling, noise and video compression [11, 53]. With the improved synthetic training data, the trained models could achieve better super-resolution performance on real-world videos. However, there is still much distribution gap between the synthetic LR data and real-world LR videos. Instead of working on synthesizing more realistic training data, in this work we investigate how to leverage the strong generative priors of pre-trained latent diffusion models to improve the real-world VSR performance.", + "bbox": [ + 212, + 332, + 787, + 559 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Generative Prior for Image Restoration. The highly ill-posed nature of real-world restoration problem brings forward a great challenge on real-world image restoration algorithms. To boost the performance, a series of methods have introduced generative priors into the image restoration process. Early methods usually adopt pre-trained generative adversarial networks (GANs) as the generative prior [8,19,34,46,50]. Although effective, the priors adopted in these methods are usually tailored for specific tasks (e.g., face image restoration), limiting their capability in handling more general natural image restoration problems. Considering the higher generality of diffusion models than GANs in natural image generation, recently researchers have started to investigate how to leverage generative diffusion priors for image restoration, and demonstrated interesting results [17, 48]. In this work, we investigate the possibility of applying latent diffusion priors to real-world VSR, and develop effective techniques to tackle the flickering issue and improve the temporal consistency of generated videos.", + "bbox": [ + 212, + 563, + 787, + 776 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Diffusion for Video Synthesis. Recently, diffusion based models have demonstrated its superiority in the video synthesis [4,5,18,43]. As one of the pioneer works, Make-a-Video [43] leverages a pretrained text-to-image (T2I) model and build a text-to-video (T2V) model by inserting temporal modules into the", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "X. Yang et al.", + "bbox": [ + 271, + 114, + 364, + 128 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ce3187b7ca565a36504f90fd37e72cbb09efd9510150185681d3ea2ce3a260c7.jpg", + "image_caption": [ + "MGLD Sampling Pipeline" + ], + "image_footnote": [], + "bbox": [ + 220, + 162, + 575, + 349 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/02c68e471630b2d143c8f181d77ecc91c3ea73b81ecd4f9887593d44e420770c.jpg", + "image_caption": [ + "Update with Motion Guidance" + ], + "image_footnote": [], + "bbox": [ + 593, + 162, + 763, + 234 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/4eb625b1717bda730056a1a487b4e67015f21241eb72a978c4a2228d3126a7e7.jpg", + "image_caption": [ + "Temporal-aware Sequence Decoder", + "Fig. 2: Overview of the proposed MGLD-VSR framework for real-world VSR. We first estimate the forward and backward optical flow with a pre-trained optical flow estimation network and then employ a motion-guided loss to update the latent diffusion sampling process. The motion-guided loss is computed by summing the forward and backward masked warping error of the latent sequence at each time-step. After a number of $T$ sampling steps, we obtain the generated latent sequence and feed it into the temporal-aware sequence decoder to reconstruct the VSR sequence." + ], + "image_footnote": [], + "bbox": [ + 594, + 250, + 769, + 351 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "T2I model and designing a cascaded architecture. VideoLDM [5] adopts a latent diffusion paradigm and train the T2V model on large-scale paired text-video datasets, greatly improving the generation results. Stable Video Diffusion [4] further improve the results by conducting training on more carefully constructed datasets and introducing more advanced training paradigm. In this work, we leverage the diffusion prior and build a powerful real-world VSR model, which aims to tackle the ill-posedness of real-world VSR problem with diffusion prior and obtain temporally consistent VSR results with rich details.", + "bbox": [ + 212, + 508, + 787, + 630 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Methodology", + "text_level": 1, + "bbox": [ + 214, + 655, + 380, + 672 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given an LR video sequence of $N$ frames $I = \\{I_1, I_2, \\ldots, I_N\\}$ , our goal is to reproduce an HR video sequence $\\hat{I} = \\{\\hat{I}_1, \\hat{I}_2, \\ldots, \\hat{I}_N\\}$ with enhanced visual quality while maintaining the content consistency across adjacent frames. Inspired by the recently developed diffusion-based image restoration methods [17, 48], we propose to harness the generative priors of the pre-trained SD model [42] to improve the real-world VSR performance. First, we propose a motion-guided diffusion sampling process, where the temporal dynamics from the LR sequence are introduced into the diffusion sampling process to enhance the consistency of latent features across frames. Then, we further improve the quality of generated details by designing a temporal-aware sequence decoder and fine-tuning", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "MGLD-VSR", + "bbox": [ + 645, + 114, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "it with a video-oriented loss. The overall framework of our algorithm, namely motion-guided latent diffusion (MGLD), is illustrated in Figure 2.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.1 Latent Diffusion Model", + "text_level": 1, + "bbox": [ + 214, + 196, + 455, + 210 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Diffusion models [21] are a class of generative models which transform Gaussian noise into data samples through an iterative reverse Markovian process. This process can be described by: $q(x_{t}|x_{t - 1})\\coloneqq \\mathcal{N}(x_{t},(1 - \\beta_{t})x_{t - 1},\\beta_{t}I)$ , where the Gaussian noise is iteratively applied to the sample $x_{t - 1}$ at time-step $t - 1$ to produce $x_{t}$ , guided by a variance scheduler $\\{\\beta_t\\}_{t = 1}^T$ along the diffusion process. The key objective of diffusion models is to train a denoising U-Net $\\epsilon_{\\theta}$ parameterized by $\\theta$ and conditioned on the time-step $t$ and an optional signal $y$ . Given the noise target $\\epsilon_{t}$ , the diffusion loss can be defined as: $L(\\theta) = \\mathbb{E}_{t\\sim U(1,T),\\epsilon_t\\sim \\mathcal{N}(0,I)}\\| \\epsilon_t - \\epsilon_\\theta (\\mathbf{x}_t;t,y)\\| _2^2$ .", + "bbox": [ + 212, + 220, + 787, + 358 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Latent Diffusion Models (LDMs) [42] have made a remarkable impact in the field of image generation. These models employ Variational Autoencoders (VAE) to map the images into a latent space, enabling them to be trained on large-scale text-image datasets. This equips LDMs with powerful prior knowledge about natural images, which can be leveraged by downstream tasks such as image restoration and super-resolution [48] to produce images with clear details and high-quality content. However, in the case of VSR, the stochastic nature of diffusion process often leads to the temporal inconsistencies in the restored video sequence. To solve this issue, we propose a MGLD model that achieves high-quality VSR by utilizing guidance during the sampling process and incorporating temporal awareness in the decoder.", + "bbox": [ + 212, + 359, + 787, + 523 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Motion-guided Diffusion Sampling", + "text_level": 1, + "bbox": [ + 214, + 544, + 545, + 559 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "From a score-based perspective [44], the sampling process of a diffusion model involves the application of Langevin Dynamics to both gradient of the data distribution $\\log p(\\mathbf{z})$ and a guidance model $g = G(z)$ , which can be a network [15,38] or an energy model [12]:", + "bbox": [ + 212, + 566, + 787, + 628 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\nabla_ {\\mathbf {z}} \\log p (\\mathbf {z}, g) = \\nabla_ {\\mathbf {z}} \\log p (\\mathbf {z}) + \\nabla_ {\\mathbf {z}} \\log p (g | \\mathbf {z}). \\qquad (1)\n$$\n", + "text_format": "latex", + "bbox": [ + 346, + 636, + 785, + 655 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this work, we introduce a technique that utilizes the temporal dynamics in low-resolution input videos to create latent features that maintain temporal consistency, thereby enhancing VSR results. Specifically, we incorporate an innovative motion-guided module within the sampling process to measure the warping errors of latent features across frames. Given the LR frames, we first compute their optical flow with $\\mathcal{F}$ [47] and downsample the flow map to the dimension of latent features. Given the forward flow $O_{f}$ and backward flow $O_{b}$ , we warp the latent features to their neighbouring frames and compute the accumulated warping error along both directions:", + "bbox": [ + 212, + 662, + 787, + 800 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nE (z) ^ {t} = \\sum_ {i = 1} ^ {N - 1} \\| \\left(\\operatorname {W a r p} \\left(z _ {i} ^ {t}, O _ {b, i}\\right) - z _ {i + 1} ^ {t})\\right) \\| _ {1} + \\sum_ {i = 2} ^ {N} \\| \\left(\\operatorname {W a r p} \\left(z _ {i} ^ {t}, O _ {f, i - 1}\\right) - z _ {i - 1} ^ {t})\\right) \\| _ {1}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 215, + 805, + 785, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "X. Yang et al.", + "bbox": [ + 271, + 114, + 364, + 128 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We empirically find that the presence of occlusion will pose adverse impact on flow estimation, which in turn causes distraction in the sampling process and leads to artifacts in the final VSR results. To solve this issue, we further estimate the occlusion mask $\\mathcal{M}$ in each frame [33] and ignore the contributions in the occluded region. Given the variance $\\sigma_t^2$ of noise scheduler at step $t$ , the motion-guided sampling process at sampling step $t$ can be written as:", + "bbox": [ + 212, + 146, + 782, + 237 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {z} ^ {t} \\leftarrow \\operatorname {D D P M} \\left(z ^ {t + 1}, t\\right) - \\eta \\sigma_ {t} ^ {2} \\nabla_ {z} \\left(\\mathcal {M} \\circ E (z) ^ {t}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 348, + 243, + 785, + 260 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where the DDPM [21] scheduler is applied for diffusion reverse (denoising) process and $\\eta = 10$ is a weighting factor to control the guidance strength. After $T$ iterations, the final denoised latent sequence $z^0$ is fed into the VAE decoder to obtain the final video sequence.", + "bbox": [ + 212, + 266, + 784, + 327 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Temporal-aware Decoder Fine-tuning", + "text_level": 1, + "bbox": [ + 214, + 345, + 570, + 361 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "With the proposed motion-guided diffusion sampling, the VSR model could better exploit the temporal dynamics to generate temporally more consistent HR videos. However, since the guidance is performed in the low-resolution latent space, which is 8 times smaller in both height and width than the output image space, the video reconstruction through a VAE decoder may still lead to temporal discontinuities in the generated details of VSR output. Thus, we introduce a temporal-aware sequence decoder and fine-tune it with ground-truth video sequences to improve the smoothness and accuracy of output videos.", + "bbox": [ + 212, + 366, + 784, + 487 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As shown in Figure 2, we build our sequence decoder on top of the pretrained VAE decoder by incorporating 1D convolutions along the temporal dimension. This design facilitates interleaved spatial-temporal interactions and enables our module to restore details with high continuity while incurring minimal computational cost. We further incorporate the information $e$ from the VAE encoder through a Controllable Feature Warping (CFW) module [48, 60] for better restoration and generation results. To leverage the restorative power of the pretrained VAE, we freeze their original spatial blocks and only update the parameters in the temporal convolution and CFW module.", + "bbox": [ + 212, + 487, + 784, + 622 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We first employ the $\\ell_1$ loss and perceptual loss [23, 58] as the reconstruction loss $L_{\\mathrm{recon}}$ for each frame, then utilize the GAN-loss and frame difference loss that tailored for video sequences. The frame difference loss $L_{\\mathrm{diff}}$ calculates the difference between consecutive predicted frames $\\hat{I}$ and ground truth frames $I^{gt}$ , which is defined as:", + "bbox": [ + 212, + 623, + 784, + 698 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {d i f f}} = \\sum \\| \\left(\\hat {I} _ {i + 1} - \\hat {I} _ {i}\\right) - \\left(I _ {i + 1} ^ {g t} - I _ {i} ^ {g t}\\right) \\| _ {1}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 362, + 704, + 784, + 722 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "It is noteworthy that diffusion models typically generate strong structures. To make the super-resolved video look more natural, we need to apply consistency constraints to those structured regions. We hence introduce a structure weighted consistency loss $L_{\\mathrm{swc}}$ :", + "bbox": [ + 212, + 728, + 782, + 789 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} L _ {\\mathrm {s w c}} = \\sum \\| \\mathcal {M} \\circ (\\mathcal {W} _ {i + 1} (\\operatorname {W a r p} (\\hat {I} _ {i}, O _ {b, i} ^ {g t}) - \\hat {I} _ {i + 1})) \\| _ {1} \\tag {5} \\\\ + \\sum \\| \\mathcal {M} \\circ (\\mathcal {W} _ {i - 1} (\\operatorname {W a r p} (\\hat {I _ {i}}, O _ {f, i} ^ {g t}) - \\hat {I} _ {i - 1})) \\| _ {1}. \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 326, + 795, + 785, + 837 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "MGLD-VSR", + "bbox": [ + 643, + 113, + 730, + 127 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $O_{f}^{gt}$ , $O_{b}^{gt}$ denote the forward and backward flow computed on the ground-truth sequence. This loss helps the network to generate consistent details by modulating the loss through a weighted map $\\mathcal{W} = (1 + wS)$ computed on the ground-truth frames, where $w = 3$ is a weighting factor and $S$ is a structure (edge) map using the Sobel operator. Occlusion masking is also adopted. Given the GAN-based loss $L_{\\mathrm{GAN}}$ and reconstruction loss $L_{\\mathrm{recon}}$ , the total loss for decoder fine-tuning is defined as:", + "bbox": [ + 212, + 145, + 787, + 252 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {v i d e o}} = L _ {\\text {r e c o n}} + \\alpha L _ {\\text {d i f f}} + \\beta L _ {\\text {s w c}} + \\gamma L _ {\\text {G A N}}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 349, + 267, + 785, + 282 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\alpha, \\beta$ and $\\gamma$ are empirically set to 0.5, 0.5 and 0.025.", + "bbox": [ + 214, + 294, + 635, + 310 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 214, + 337, + 375, + 354 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.1 Experimental Settings", + "text_level": 1, + "bbox": [ + 214, + 372, + 447, + 388 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Implementation Details. The training of MGLD-VSR consists of two main stages. In the first stage, we fine-tune the denoising U-Net for latent space diffusion. The weight of the denoising U-Net is initialized from Stable Diffusion V2.1. We insert 1D temporal convolution into the U-Net to help temporal dynamic modeling. We then fix the weight in the denoising U-Net of SD model and train the conditioning and temporal modeling modules. For conditioning, we adopt a small time-aware encoder to encode LR condition and inject it into the denoising U-Net with spatial feature transform operation as in [48]. The training batch size, sequence length and latent image size are set to 24, 6 and $64 \\times 64$ , respectively. In the second stage, we first generate the clean latent sequence with the proposed motion-guided diffusion sampling process, then fine-tune the temporal-aware sequence decoder with the LR sequence, the generated latent sequence and the HR sequence. We fix the VAE decoder and insert temporal modeling layers and CFW modules for training. The training batch size, sequence length and image size are set to 4, 5 and $512 \\times 512$ , respectively.", + "bbox": [ + 212, + 400, + 787, + 626 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our models are trained on 4 NVIDIA A100 GPUs with the PyTorch [40] framework. Adam [26] is chosen as the optimizer. During inference, due to memory constrains, we divide the LR video into multiple sequences to run. To handle output with arbitrary resolution, we apply a progressive patch aggregation sampling algorithm [48]. For each sequence, we run the sampling for 50 steps.", + "bbox": [ + 212, + 627, + 787, + 703 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Training and Testing Datasets. As in previous works [11, 49, 53], we merge the training set and validation set of REDS [37] in training and leave 4 sequences for testing (REDS4). We follow the degradation pipeline of RealBasicVSR [11] to synthesize the training sequence pairs, which involves blur, noise, downsampling and compression degradations. We evaluate our method on both synthetic and real-world datasets. To synthesize LR-HR testing sequence pairs, we apply the RealBasicVSR degradation pipeline on several widely used VSR datasets, including REDS4 [37], UDM10 [56] and SPMCS [45]. REDS4 contains 4 video sequences, each having 100 frames. UDM10 consists of 10 sequences, each", + "bbox": [ + 212, + 704, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "X. Yang et al.", + "bbox": [ + 271, + 114, + 366, + 128 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "having 32 frames. SPMCS has 30 sequences, each having 31 frames. For real-world dataset, we adopt VideoLQ [11] for testing, which contains 50 real-world sequences with complex degradations.", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Evaluation Metrics. On synthetic datasets, we could adopt full-reference metrics to evaluate different real-world VSR methods. The perceptual quality metrics LPIPS [58] and DISTS [16] are selected. Fidelity-oriented metrics PSNR and SSIM are also calculated for reference. Apart from image-based metrics, the video quality assessment metric VMAF [29] is also employed.", + "bbox": [ + 212, + 191, + 784, + 268 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "On real-world datasets, as ground truths are not available, we compute the no-reference image quality metrics NIQE [36] and BRISQUE [35] for evaluation. In addition, we also include a deep learning based image-based metric MUSIQ [25] and a state-of-the-art video quality assessment metric DOVER [52].", + "bbox": [ + 212, + 268, + 785, + 330 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.2 Experimental Results", + "text_level": 1, + "bbox": [ + 215, + 356, + 439, + 371 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "To demonstrate the effectiveness of the our MGLD-VSR algorithm, we compare it with seven state-of-the-art methods, including three real-world image super-resolution models (BSRGAN [57], RealESRGAN [51], StableSR [48]), one classic VSR model (RVRT [31]), and three blind or real-world VSR models (DBVSR [39], RealVSR [55], RealBasicVSR [11].", + "bbox": [ + 212, + 385, + 784, + 460 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Quantitative Comparison. We first show the quantitative comparison on the synthetic datasets and real-world video benchmarks in Table 1. As shown from Table 1, our method achieves the best results in terms of full-reference perceptual based metrics LPIPS and DISTS on all synthetic test datasets, indicating that the proposed method is able to reconstruct high-quality plausible details from sequences degraded with complex degradations. Although methods like DBVSR may achieve better performance in terms of PSNR or SSIM on some datasets, they tend to generate blurry results, as demonstrated in the LPIPS and DISTS metrics. In terms of VMAF, our model achieves the best or second best results. As for real-world VSR dataset VideoLQ, our methods achieve the best performance in terms of NIQE, BRISQUE, and DOVER and the second best results in terms of MUSIQ, which reflects its strong capability to enhance the real-world videos and generate realistic details and textures.", + "bbox": [ + 212, + 462, + 785, + 656 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Qualitative Comparison. To demonstrate the effectiveness of MGLD-VSR, we visualize the VSR results on synthetic datasets in Figure 3 and the results on real-world VideoLQ in Figure 4. One can see that MGLD-VSR can remove complex spatial-variant degradations and generate realistic details, outperforming other state-of-the-art real-world VSR algorithms. On synthetic datasets, as shown in Figure 3, MGLD-VSR can well reconstruct the structure while generating plausible details under complex degradations. On real-world videos, as shown by the example of sequence 008 in Figure 4, MGLD-VSR reconstructs clearer characters and patterns than competing methods. For the example of sequence 013, MGLD-VSR not only reconstructs the structure of the windows, and at the same time generates plausible exterior structure of the brick wall. More visual comparisons can be found in the supplementary materials.", + "bbox": [ + 212, + 657, + 785, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "MGLD-VSR", + "bbox": [ + 645, + 114, + 730, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/55f22dc9c116bf24440aad83758f454974968c85ece5dc625766e7048ebc4643.jpg", + "table_caption": [ + "Table 1: Quantitative comparison of different VSR models evaluated on popular VSR benchmarking datasets. Bold and underline represent the best and second best score, respectively. For synthetic datasets (REDS4, UDM10, SPMCS), we adopt full-reference metrics to evaluate different methods. For real-world video datasets (VideoLQ), we adopt no-reference metrics to compare different methods." + ], + "table_footnote": [], + "table_body": "
DatasetsMetricsBicubicRVRTBSRGANRealESRGANStableSRDBVSRRealVSRRealBasicVSROurs
REDS4PSNR ↑22.1322.2922.0221.3422.0422.2921.7822.3522.46
SSIM ↑0.57300.57590.57020.56470.56830.57600.56630.58460.5723
LPIPS ↓0.62770.62210.48020.45090.45530.62210.62670.41900.3776
DISTS ↓0.36670.35540.21180.17500.18450.35560.36200.15810.1151
VMAF ↑2.12554.800526.241333.995526.69532.60752.590334.988434.3214
UDM10PSNR ↑25.7826.1825.2924.9425.5826.1425.0925.5625.99
SSIM ↑0.75590.76290.73230.74390.75920.76350.75240.74480.7548
LPIPS ↓0.44410.44130.42170.38840.36840.43980.43300.38970.3491
DISTS ↓0.27300.25750.19260.16510.14970.25850.26080.16370.1369
VMAF ↑5.464710.172431.087436.915833.82647.20486.007938.263839.3892
SPMCSPSNR ↑22.2822.5122.1221.6622.2622.5022.0322.4122.66
SSIM ↑0.56910.57530.57300.56940.58440.57640.57380.57360.5960
LPIPS ↓0.55720.54710.45090.44350.43210.54580.54870.44200.3990
DISTS ↓0.35930.34150.23070.21500.23510.34350.34790.21300.1934
VMAF ↑0.96565.353322.482827.693521.80603.75762.702427.451928.7032
VideoLQNIQE ↓9.13196.58884.21044.20364.44426.75046.33313.69153.5346
BRISQUE ↓72.740562.176825.240029.842826.309261.189757.276524.548421.9839
MUSIQ ↑21.678928.420452.221349.838047.804529.003924.218955.969152.7812
DOVER ↑0.36770.49070.71840.73090.71080.49980.52020.74010.7481
", + "bbox": [ + 215, + 226, + 785, + 441 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Temporal Consistency. Compared with single image super-resolution, one major aspect of VSR is the temporal consistency of the generated videos. Following most previous works, in this paper, we adopt the average warping error (WE) [13,27] of the sequence to quantitatively measure the temporal consistency: $\\mathrm{WE} = \\frac{1}{N - 1}\\sum_{i = 1}^{N - 1}\\| \\hat{I}_{i + 1} - \\mathrm{Warp}(\\hat{I}_i,O_{b,i})\\| _1$ , where $\\hat{I}_i$ is the predicted frame and $O_{b,i}$ is the estimated backward optical flow [54] from ground-truth video. On real-world video datasets without ground-truth, the optical flow is computed on the predicted frames.", + "bbox": [ + 212, + 486, + 787, + 608 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "It is worth mentioning that, despite the wide use of WE for temporal consistency evaluation, it may not be able to faithfully reflect the real human perception on the video. For example, WE could be easily attacked by blurring the sequence, resulting much higher WE scores but lower video quality. The WE scores of competing methods are given in Table 2. One can see that our diffusion prior based method MGLD-VSR is not good at this metric because it tends to generate more details and textures, which are not favourable by WE. Nonetheless, the WE of MGLD-VSR is better than StableSR, which is also driven by the latent diffusion priors. To better showcase the effectiveness of our method, we visualize the temporal profiles of the VSR results produced by different algorithms in Figure 5. One can see that MGLD-VSR reconstructs more accurate details while maintaining good temporal consistency.", + "bbox": [ + 212, + 611, + 787, + 792 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "User Study. We also conduct a user study to further examine the effectiveness of the proposed algorithm in comparison with RealBasicVSR [11] and StableSR [48]. The study is conducted in pairs: given the LR video, the subject", + "bbox": [ + 212, + 794, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "X. Yang et al.", + "bbox": [ + 271, + 114, + 364, + 128 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/1a21be49bca64884ee81ee330d398fa5ebaac7c1c601e8e5fe8736961f50206a.jpg", + "image_caption": [ + "Sequence 011 from REDS4" + ], + "image_footnote": [], + "bbox": [ + 220, + 143, + 336, + 194 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/8a42fa41971422baccc438633c04648760f223367e810a1f60bbd666b170b7ba.jpg", + "image_caption": [ + "Bicubic" + ], + "image_footnote": [], + "bbox": [ + 344, + 143, + 411, + 194 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/b3b26876e14df09051d32da7e9b1be2577bbaa8595e4d090fb01bed9148200c5.jpg", + "image_caption": [ + "Real-ESRGAN" + ], + "image_footnote": [], + "bbox": [ + 419, + 143, + 486, + 194 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/09490f3ead1d67dfa12550c5c9f9aea7dd6bac950b5555c0a4f00199911b956d.jpg", + "image_caption": [ + "StableSR" + ], + "image_footnote": [], + "bbox": [ + 495, + 143, + 560, + 194 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/7378eabc63867bcda96fb3c85fcdabed4ad6c519f5ec49d655eb379bbd761c77.jpg", + "image_caption": [ + "RealBasicVSR" + ], + "image_footnote": [], + "bbox": [ + 568, + 143, + 635, + 194 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/159c2189dcbabd27947074e6a56869999ff598cd02d109dc6bbcb69312fd269c.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 642, + 143, + 710, + 195 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/6a8edb8c0dbd6417bf15337a10049a46e926a2bff6602e810f7d5abe14c89214.jpg", + "image_caption": [ + "GT" + ], + "image_footnote": [], + "bbox": [ + 717, + 143, + 785, + 195 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/26f5f2891c7c911908c82f8aaa13f12b673a13cc57814840ab5158e88e1df629.jpg", + "image_caption": [ + "Sequence band from UDM10" + ], + "image_footnote": [], + "bbox": [ + 220, + 205, + 336, + 257 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/1bfc860ed73a6becfed9f2e5255c7facc155d14f32eb369ccbb13c826dd7ea47.jpg", + "image_caption": [ + "Bicubic" + ], + "image_footnote": [], + "bbox": [ + 344, + 205, + 411, + 257 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/9805048d8ad0be3e6f49ffbe4a4794d0dad08f16ea7593523fb3898f170e054b.jpg", + "image_caption": [ + "Real-ESRGAN" + ], + "image_footnote": [], + "bbox": [ + 419, + 205, + 486, + 257 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/a4d544fc5392c0a3d07024189fe4094b2a74177237830c5ce3f75aa91ae924d1.jpg", + "image_caption": [ + "StableSR" + ], + "image_footnote": [], + "bbox": [ + 495, + 205, + 560, + 257 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/e81829d39c8daae6cd1402e573c7aadc63b3dbe8e6578d69e93e8df8e7f1f16d.jpg", + "image_caption": [ + "RealBasicVSR" + ], + "image_footnote": [], + "bbox": [ + 568, + 205, + 635, + 257 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/22ea8ba186e6d2f8f77a1abd7a78e8be7dd12d563e258d7dae4768bfa585194c.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 643, + 205, + 710, + 257 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/dc9b5fcbd53908c7cc836cbb2fff032d8a3f2ae99f7e8dc560a712ee58426030.jpg", + "image_caption": [ + "GT" + ], + "image_footnote": [], + "bbox": [ + 717, + 205, + 785, + 257 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/e892820ceaf4c9765a5508bf22ee3de820fb930065cd95e1b589ef69247898b2.jpg", + "image_caption": [ + "Sequence 000002 from SPMCS" + ], + "image_footnote": [], + "bbox": [ + 220, + 268, + 336, + 319 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/d9f9eaf4732ee295736f2928923672cd5ef09e5fe4f39767958d37fcdd0816ad.jpg", + "image_caption": [ + "Bicubic" + ], + "image_footnote": [], + "bbox": [ + 344, + 268, + 411, + 319 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/2d0c996430a220b3c1625fa3804172d3aea2e0cf06b8c14460137cc8597e3e1c.jpg", + "image_caption": [ + "Real-ESRGAN", + "Fig. 3: Qualitative comparison on synthetic datasets (REDS4, UDM10, SPMCS) for $\\times 4$ VSR. (Zoom-in for best view.)" + ], + "image_footnote": [], + "bbox": [ + 419, + 268, + 486, + 319 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/e001fbe85d63fd5ddc80b8bf5b0989087430cca1b6d75a43733b4c7bd05cb1ca.jpg", + "image_caption": [ + "StableSR" + ], + "image_footnote": [], + "bbox": [ + 495, + 268, + 560, + 319 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/9f4897da176e1b91f6ccaa78375773acaf6f1037050de6131a49339d3f6ec88d.jpg", + "image_caption": [ + "RealBasicVSR" + ], + "image_footnote": [], + "bbox": [ + 568, + 268, + 635, + 319 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/b4fbbdb03af3b88df08c54ec8f8ae09d3f61761a3abb01106b13ff72b1f4bb77.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 642, + 268, + 710, + 319 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/a9a1a880f31d483a04caf95d4e88f63e9edc6cd8380e9f32e5de11bb85587e0b.jpg", + "image_caption": [ + "GT" + ], + "image_footnote": [], + "bbox": [ + 717, + 268, + 785, + 319 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/97aa431f058833bc287cd30d604226b4c806f02bdba49fcea5147bec2d149c5b.jpg", + "table_caption": [ + "Table 2: Temporal consistency comparison of competing VSR methods by using WE $\\downarrow$ . Bold and underline represent the best and second best scores, respectively." + ], + "table_footnote": [], + "table_body": "
DatasetsBicubicRVRTBSRGANRealESRGANStableSRDBVSRRealVSRRealBasicVSROurs
REDS48.576.5010.2511.9111.818.518.6110.0910.08
UDM103.103.064.824.994.803.042.754.334.55
SPMCS3.013.585.187.124.623.584.182.723.87
VideoLQ5.486.387.727.649.335.985.746.767.56
", + "bbox": [ + 217, + 431, + 785, + 497 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "is asked to choose the better one from two HR videos obtained by different methods, considering both the reconstruction quality and the temporal consistency. As shown in Figure 6, our method gains more votes in the competitions.", + "bbox": [ + 212, + 537, + 782, + 583 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.3 Ablation Study", + "text_level": 1, + "bbox": [ + 214, + 604, + 388, + 619 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "In this section, we conduct a series of analyses on the proposed framework, including its Motion-guided Diffusion Sampling (MDS) module and Temporal-aware Sequence Decoder (TSD) module, and employ the VideoLQ dataset for experiments. We adopt WE to measure temporal consistency and no-reference metrics NIQE, BRISQUE, MUSIQ and DOVER to measure perceptual quality.", + "bbox": [ + 212, + 628, + 784, + 703 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Different Components of MGLD-VSR. As can seen from Table 3, the introduction of MDS improves all the metrics over the baseline (i.e., without using MDS and TSD), suggesting that MDS helps not only in reducing the temporal differences (i.e., WE), but also enhancing the perceptual quality of the reconstructed video sequence.", + "bbox": [ + 212, + 704, + 784, + 779 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "The incorporation of MDS aids in maintaining perceptual continuity by ensuring that the restored frames could transition smoothly from one to the next. This contributes to more natural and visually pleasing VSR results. The incorporation of TSD also yields better results across all the metrics compared to", + "bbox": [ + 212, + 779, + 784, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "MGLD-VSR", + "bbox": [ + 645, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/66d1aed951c44d33847292360718af7dba651bcb9d8c3e1d3693f08aeb3682b4.jpg", + "image_caption": [ + "Sequence 008 from VideoLQ" + ], + "image_footnote": [], + "bbox": [ + 217, + 167, + 393, + 271 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/a696e082a10641f9b191d58ce98491f9830d067de57894e4b0de284ebb11a8b5.jpg", + "image_caption": [ + "Bicubic" + ], + "image_footnote": [], + "bbox": [ + 398, + 143, + 483, + 210 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/ca382d079fecafd21112c6abe1bb792379ae77338d04df50b17b97742b2aa06e.jpg", + "image_caption": [ + "BSRGAN" + ], + "image_footnote": [], + "bbox": [ + 496, + 143, + 584, + 210 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/0bcebf50697bdf5a9e9916fd0f55415b6198bbb3cfd2bb5f2c4890f5ae6b2ec1.jpg", + "image_caption": [ + "Real-ESRGAN" + ], + "image_footnote": [], + "bbox": [ + 598, + 143, + 684, + 210 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/a3b03c9d59434da6ff64abd589aca844099c8b7f96b05c19bb17c73f465fb1dc.jpg", + "image_caption": [ + "StableSR" + ], + "image_footnote": [], + "bbox": [ + 699, + 143, + 785, + 210 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/5feba6f652cf468516ad5f3f1a0f0cc68bb7a72c6d7efcdeb7bd915a6def1d3a.jpg", + "image_caption": [ + "Sequence 013 from VideoLQ" + ], + "image_footnote": [], + "bbox": [ + 217, + 337, + 393, + 415 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/489d4622053853c6083303f2bc9794fa27e8414fa17759937c74b9f885f36d02.jpg", + "image_caption": [ + "Bicubic" + ], + "image_footnote": [], + "bbox": [ + 398, + 301, + 483, + 368 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/48127e6a671b247d61e8adb6788c8ca1d73a28f0628d6b83c36fb756179b2ac2.jpg", + "image_caption": [ + "RealVSR" + ], + "image_footnote": [], + "bbox": [ + 398, + 380, + 483, + 445 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/712dcec4f5b3baa9c86dec11dd36c6a3c77f6cfd3eb20942b4881645e8526579.jpg", + "image_caption": [ + "BSRGAN" + ], + "image_footnote": [], + "bbox": [ + 498, + 301, + 583, + 368 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/bc3cf4999da48aa0574ba136619ec6f812d21b4e7f1bf7d02e54ff136187f4ce.jpg", + "image_caption": [ + "DBVSR" + ], + "image_footnote": [], + "bbox": [ + 498, + 380, + 583, + 445 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/da61c9ee62dab88f5a59bff79bd7ff741c8631ae9fcf3858f10d20ab435faae2.jpg", + "image_caption": [ + "Real-ESRGAN" + ], + "image_footnote": [], + "bbox": [ + 598, + 301, + 684, + 368 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/ed52997a91f5dd302afdfa831f62fcc94bf6989c4ced72f051d355c4cd485e60.jpg", + "image_caption": [ + "RealBasicVSR" + ], + "image_footnote": [], + "bbox": [ + 599, + 380, + 684, + 445 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/c0d3c2ffa3de0e90eaff7bfa46f74083b1dc6876dd29fac78000ffe9ed24f811.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 699, + 220, + 785, + 287 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/b46a49a679e57918d39e6bd9e82df7a38e68716fbee0f3df5d93ad6a1d12d718.jpg", + "image_caption": [ + "StableSR" + ], + "image_footnote": [], + "bbox": [ + 699, + 301, + 785, + 368 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/6ace3dbd55c52f74aebd9e603790d6d9a06104909e473232527b96b4d066c890.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 699, + 380, + 785, + 445 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/76cd62080f30b42c8b4c191efbab7bdb3ade7c1707e2a7e67eb2f87edbf7b4bd.jpg", + "image_caption": [ + "Fig. 4: Qualitative comparison on the real-world video dataset (VideoLQ) for $\\times 4$ video SR. (Zoom-in for best view.)" + ], + "image_footnote": [], + "bbox": [ + 217, + 507, + 320, + 554 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/66589e67a61bd62f29bbb514747cd0e51a37359d6c51a3323433769bd1bc7f4e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 321, + 507, + 395, + 554 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/622ca7949d5e8f5666ae9765a52dc45c2f7e9e9c74c5f4bd23bbe66c68f0987a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 398, + 507, + 473, + 554 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/7dcd77593b048342fd1c8e8d7622d1850a294beba46d7e2ff98e4a0311a9e869.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 477, + 507, + 550, + 554 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/f07208c77bcd272c498bb30726320220336bd96c8e0d2262a9e771af03eb22ad.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 553, + 507, + 627, + 554 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/d3bc162df8c76ab72a924e375dffdd8b619f1a618ce5c6a1fce49479cfa9ec00.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 632, + 507, + 705, + 554 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/26afcc10b95caa60126528e6429bbdb47e5eef4b8e06bf4829efcb64be8a2683.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 709, + 507, + 784, + 554 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/7134fc49cba0d0d657c3cb94b509b31a6d10d24d3bdff82a1fc68a82e1c0c756.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 217, + 559, + 320, + 604 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/89927672146a443faa86ceddbacb6bbbd9730af44148e53977bfb4bfeb605361.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 321, + 559, + 395, + 604 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/c76c5094e3a20fd31a9738da9bfa69bb14819e7a30d4c0a93d8d1ac1e29d6941.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 398, + 559, + 473, + 604 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/c6e99f77deaa80429e501bac068865126af5ef8663fcb467a1060c4c8637e1a9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 477, + 559, + 550, + 604 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/efd09aca113e4c94e9574cabfc958431a893fbbcbffae3f92b9cfc394c27d1b7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 553, + 559, + 627, + 604 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/fee8658ada73b209db1beb2c7c6f548706e5001e4849205f645fb8c996656898.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 632, + 559, + 705, + 604 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/dec93126028840fd97e1dfa59a2a5e7130cba18b32f864374f615b07f351d052.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 709, + 559, + 784, + 604 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/b883b80d49db29bf4b07b2bb73b5c39b623578280af3134b3458c93b8e302c3c.jpg", + "image_caption": [ + "Fig. 5: Temporal profiles of competing real-world VSR methods.", + "Fig. 6: User study on 20 videos from VideoLQ evaluated by 10 subjects. MGLD-VSR gains more votes than competing methods." + ], + "image_footnote": [], + "bbox": [ + 218, + 657, + 498, + 757 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/2714013b4093a5f8bc2ffd51805f1956e184ee627ec42971ea1746429b91f188.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 657, + 784, + 757 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "X. Yang et al.", + "bbox": [ + 271, + 114, + 366, + 128 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/b2c5116369287659ad7031dbfb8b1224c306e9e278b2d3abc6683cbb87698ca0.jpg", + "table_caption": [ + "Table 3: Ablation study on the different components of the proposed algorithm. \"MDS\" refers to Motion-guided Diffusion Sampling and \"TSD\" refers to Temporal-aware Sequence Decoder." + ], + "table_footnote": [], + "table_body": "
MDS TSDWE ↓NIQE ↓BRISQUE ↓MUSIQ ↑DOVER ↑
✓9.30694.109825.324951.56770.7412
8.98483.987624.084052.41430.7435
✓8.36793.628122.080652.68740.7468
✓7.54783.534621.983952.78120.7481
", + "bbox": [ + 326, + 198, + 669, + 263 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/774778f79401f7818e2b2b94abf61de0c3f753e8e0c90daa247afcebf06dd8f2.jpg", + "image_caption": [ + "Sequence 002 from VideoLQ" + ], + "image_footnote": [], + "bbox": [ + 230, + 282, + 413, + 390 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/1c67da428bce5bf1d85d6d26285eacfd3f1ad94446e5a853cdeb3029e3ce45fc.jpg", + "image_caption": [ + "0", + "Fig. 7: Impact of guidance strength during sampling. The numbers below the cropped patches indicate the motion guidance weighting factor $\\eta$ ." + ], + "image_footnote": [], + "bbox": [ + 416, + 282, + 485, + 388 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/6e5f2e88a23d0c8f1bec715d642579b62990a6d5df38361d7d2696f0ea29158b.jpg", + "image_caption": [ + "10" + ], + "image_footnote": [], + "bbox": [ + 488, + 282, + 555, + 388 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/57e87a8e5d2af8edab1741764e385b45164c7b5265e43daa99fdcdd7e87ddb6c.jpg", + "image_caption": [ + "20" + ], + "image_footnote": [], + "bbox": [ + 560, + 282, + 627, + 388 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/4d11dc341d7114a89d0f12de62381f3f6b01066fec2f3be970f4353457b6912d.jpg", + "image_caption": [ + "30" + ], + "image_footnote": [], + "bbox": [ + 632, + 282, + 700, + 388 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/5890a77f2e8d82ee4decff05dcb01ccb6098d1bfee84b750dfaaf75858d6c340.jpg", + "image_caption": [ + "40" + ], + "image_footnote": [], + "bbox": [ + 704, + 282, + 771, + 388 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "the baseline. This occurs because TSD is capable of extracting richer contextual information from neighboring frames. When both MDS and TSD are used, the results show further improvements. The full MGLD-VSR approach achieves the best scores on all the metrics, indicating the synergy of the two proposed components for real-world VSR.", + "bbox": [ + 212, + 479, + 784, + 556 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Impact of Flow Network. As shown in Table 4, the flow estimation network in motion guided sampling does not show big influence on the results. This might because the motion guided sampling is applied in low-resolution space.", + "bbox": [ + 212, + 560, + 784, + 606 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Impact of Guidance Strength. During guided sampling process, we introduce a weighting factor to control the guidance strength. As shown in Figure 7, a moderate guidance strength could reduce the artifacts and improve the temporal consistency of the generated results. A large guidance strength will blur the results and degrade the reconstructed details. To balance the generation power and temporal consistency, we set the weighting factor $\\eta = 10$ .", + "bbox": [ + 212, + 609, + 784, + 700 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "The Design of VAE Decoder. To validate the effectiveness of the proposed architecture and loss design for the VAE decoder, we conduct a series of study and the results are shown in Table 5. As can be seen from the table, the introduction of the adaptive temporal modeling module improves the temporal consistency as well as perceptual quality (measured by no-reference metrics), and incorporating the feature from encoder to decoder during decoder fine-tuning further improves the results. By introducing two sequence based losses during fine-tuning, we improve the temporal consistency of the results while maintaining or improving the perceptual quality.", + "bbox": [ + 212, + 704, + 785, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "MGLD-VSR", + "bbox": [ + 643, + 113, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/a95ffbbc21846a1837f4365aa124e4ee593568649492d162fbcd95b396775a35.jpg", + "table_caption": [ + "Table 4: Ablation study on the flow network in motion guided diffusion sampling." + ], + "table_footnote": [], + "table_body": "
Flow NetworkWE ↓NIQE ↓MUSIQ ↑DOVER ↑
SpyNet [41]7.67343.554552.69440.7478
RAFT [47]7.54783.534652.78120.7481
", + "bbox": [ + 354, + 170, + 643, + 213 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/74128d21b83bd577ef27e38089f38c45ba5cc239405323257c9d64d7671883dd.jpg", + "table_caption": [ + "Table 5: Ablation study on the different designs of VAE decoder. \"TM\" refers to the temporal modeling module. \"CFW\" refers to the controllable feature warping module that injects VAE encoder feature into VAE decoder." + ], + "table_footnote": [], + "table_body": "
ModelsTM CFWWE ↓ NIQE ↓ MUSIQ ↑ DOVER ↑LossesLdiffLswcWE ↓ NIQE ↓ MUSIQ ↑ DOVER ↑
(a)9.86704.067150.27070.7360(a)7.94803.535952.7894
(b)✓8.68463.845152.54080.7431(b)✓7.75793.540952.5648
(c)✓8.98423.923652.43860.7426(c)✓7.76883.537652.7456
(d)✓7.54783.534652.78120.7481(d)✓✓7.54783.534652.7812
", + "bbox": [ + 217, + 282, + 781, + 343 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "4.4 Limitations", + "text_level": 1, + "bbox": [ + 215, + 375, + 356, + 387 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Our proposed MGLD-VSR method leverages the diffusion prior from pre-trained SD model, aiming to achieve perceptually realistic and temporally consistent VSR results. Due to the scale of pre-trained SD model, it does not have advantages in terms of number of parameters and inference time compared with those non-diffusion VSR methods (comparisons can be found in the supplementary materials). In addition, since MGLD-VSR will generate more visual details than those non-diffusion VSR methods, it does not have advantages in terms of temporal consistency metrics either. In the future, we can resort to model distillation [2] and more advanced diffusion sampling [1,59] techniques to improve the efficiency of MGLD-VSR, and resort to the recently developed video synthesis models [4] to improve the temporal consistency.", + "bbox": [ + 212, + 406, + 787, + 574 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 604, + 359, + 619 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We presented a motion-guided latent diffusion algorithm, namely MGLD-VSR, for high-quality and temporally consistent real-world VSR. Our method exploited the strong generative prior of pre-trained latent diffusion model to restore video sequence with realistic details. To ensure the content consistency among frames, we developed a motion-guided sampling mechanism, where the motion dynamics of input video were incorporated in the diffusion sampling process. To further improve the continuity of generated details, we introduced a temporal module into the decoder, and fine-tuned it by high quality video sequences with delicately designed sequence-based losses. The fine-tuned sequence decoder mitigated the flickering artifacts and further improved the perceptual quality. Our proposed MGLD-VSR achieved state-of-the-art visual results on the widely adopted real-world VSR benchmarks, validating the great potentials of latent diffusion priors in enhancing real-world video quality.", + "bbox": [ + 212, + 643, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "X. Yang et al.", + "bbox": [ + 271, + 114, + 364, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Lu et al., C.: Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. NeurIPS 35, 5775-5787 (2022)", + "2. Luo et al., S.: Latent consistency models: Synthesizing high-resolution images with few-step inference. arXiv preprint arXiv:2310.04378 (2023)", + "3. Bishop, C.M., Blake, A., Marthi, B.: Super-resolution enhancement of video. In: AISTATS (2003)", + "4. Blattmann, A., Dockhorn, T., Kulal, S., Mendelevitch, D., Kilian, M., Lorenz, D., Levi, Y., English, Z., Voleti, V., Letts, A., et al.: Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127 (2023)", + "5. Blattmann, A., Rombach, R., Ling, H., Dockhorn, T., Kim, S.W., Fidler, S., Kreis, K.: Align your latents: High-resolution video synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22563-22575 (2023)", + "6. Caballero, J., Ledig, C., Aitken, A., Acosta, A., Totz, J., Wang, Z., Shi, W.: Real-time video super-resolution with spatio-temporal networks and motion compensation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4778-4787 (2017)", + "7. Carrillo, H., Clément, M., Bugeau, A., Simo-Serra, E.: Diffusart: Enhancing line art colorization with conditional diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3485-3489 (2023)", + "8. Chan, K.C., Wang, X., Xu, X., Gu, J., Loy, C.C.: Glean: Generative latent bank for large-factor image super-resolution. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 14245-14254 (2021)", + "9. Chan, K.C., Wang, X., Yu, K., Dong, C., Loy, C.C.: Basicvsr: The search for essential components in video super-resolution and beyond. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4947-4956 (2021)", + "10. Chan, K.C., Zhou, S., Xu, X., Loy, C.C.: Basicvsr++: Improving video superresolution with enhanced propagation and alignment. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5972-5981 (2022)", + "11. Chan, K.C., Zhou, S., Xu, X., Loy, C.C.: Investigating tradeoffs in real-world video super-resolution. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5962-5971 (2022)", + "12. Chen, M., Laina, I., Vedaldi, A.: Training-free layout control with cross-attention guidance. arXiv preprint arXiv:2304.03373 (2023)", + "13. Chu, M., Xie, Y., Mayer, J., Leal-Taixe, L., Thuerey, N.: Learning temporal coherence via self-supervision for gan-based video generation. ACM Transactions on Graphics (TOG) 39(4), 75-1 (2020)", + "14. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. In: Ranzato, M., Beygelzimer, A., Dauphin, Y., Liang, P., Vaughan, J.W. (eds.) Advances in Neural Information Processing Systems. vol. 34, pp. 8780-8794. Curran Associates, Inc. (2021), https://proceedings.neurips.cc/paper_files/paper/2021/file/49ad23d1ec9fa4bd8d77d02681df5cfa-Paper.pdf", + "15. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. In: Ranzato, M., Beygelzimer, A., Dauphin, Y., Liang, P., Vaughan, J.W. (eds.) Advances" + ], + "bbox": [ + 218, + 179, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "MGLD-VSR", + "bbox": [ + 645, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "in Neural Information Processing Systems. vol. 34, pp. 8780-8794. Curran Associates, Inc. (2021), https://proceedings.neurips.cc/paper_files/paper/2021/file/49ad23d1ec9fa4bd8d77d02681df5cfa-Paper.pdf", + "16. Ding, K., Ma, K., Wang, S., Simoncelli, E.P.: Image quality assessment: Unifying structure and texture similarity. IEEE transactions on pattern analysis and machine intelligence 44(5), 2567-2581 (2020)", + "17. Fei, B., Lyu, Z., Pan, L., Zhang, J., Yang, W., Luo, T., Zhang, B., Dai, B.: Generative diffusion prior for unified image restoration and enhancement. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9935-9946 (2023)", + "18. Ge, S., Nah, S., Liu, G., Poon, T., Tao, A., Catanzaro, B., Jacobs, D., Huang, J.B., Liu, M.Y., Balaji, Y.: Preserve your own correlation: A noise prior for video diffusion models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 22930-22941 (2023)", + "19. Gu, J., Shen, Y., Zhou, B.: Image processing using multi-code gan prior. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 3012-3021 (2020)", + "20. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in neural information processing systems 33, 6840-6851 (2020)", + "21. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. In: Larochelle, H., Ranzato, M., Hadsell, R., Balcan, M., Lin, H. (eds.) Advances in Neural Information Processing Systems. vol. 33, pp. 6840-6851. Curran Associates, Inc. (2020), https://proceedings.neurips.cc/paper_files/paper/2020/file/4c5bcfec8584af0d967f1ab10179ca4b-Paper.pdf", + "22. Jo, Y., Oh, S.W., Kang, J., Kim, S.J.: Deep video super-resolution network using dynamic upsampling filters without explicit motion compensation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3224-3232 (2018)", + "23. Johnson, J., Alahi, A., Fei-Fei, L.: Perceptual losses for real-time style transfer and super-resolution. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14. pp. 694-711. Springer (2016)", + "24. Kawar, B., Zada, S., Lang, O., Tov, O., Chang, H., Dekel, T., Mosseri, I., Irani, M.: Imagic: Text-based real image editing with diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6007-6017 (2023)", + "25. Ke, J., Wang, Q., Wang, Y., Milanfar, P., Yang, F.: Musiq: Multi-scale image quality transformer. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5148-5157 (2021)", + "26. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014)", + "27. Lai, W.S., Huang, J.B., Wang, O., Shechtman, E., Yumer, E., Yang, M.H.: Learning blind video temporal consistency. In: Proceedings of the European conference on computer vision (ECCV). pp. 170-185 (2018)", + "28. Li, W., Tao, X., Guo, T., Qi, L., Lu, J., Jia, J.: Mucan: Multi-correspondence aggregation network for video super-resolution. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part X 16. pp. 335-351. Springer (2020)", + "29. Li, Z., Bampis, C., Novak, J., Aaron, A., Swanson, K., Moorthy, A., Cock, J.: Vmaf: The journey continues. Netflix Technology Blog 25(1) (2018)" + ], + "bbox": [ + 215, + 147, + 785, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "X. Yang et al.", + "bbox": [ + 271, + 114, + 364, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "30. Liang, J., Cao, J., Fan, Y., Zhang, K., Ranjan, R., Li, Y., Timofte, R., Van Gool, L.: Vrt: A video restoration transformer. arXiv preprint arXiv:2201.12288 (2022)", + "31. Liang, J., Fan, Y., Xiang, X., Ranjan, R., Ilg, E., Green, S., Cao, J., Zhang, K., Timofte, R., Gool, L.V.: Recurrent video restoration transformer with guided deformable attention. Advances in Neural Information Processing Systems 35, 378-393 (2022)", + "32. Lugmayr, A., Danelljan, M., Romero, A., Yu, F., Timofte, R., Van Gool, L.: Repaint: Inpainting using denoising diffusion probabilistic models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11461-11471 (2022)", + "33. Meister, S., Hur, J., Roth, S.: Unflow: Unsupervised learning of optical flow with a bidirectional census loss. In: Proceedings of the AAAI conference on artificial intelligence. vol. 32 (2018)", + "34. Menon, S., Damian, A., Hu, S., Ravi, N., Rudin, C.: Pulse: Self-supervised photoupsampling via latent space exploration of generative models. In: Proceedings of the iee/cvf conference on computer vision and pattern recognition. pp. 2437-2445 (2020)", + "35. Mittal, A., Moorthy, A.K., Bovik, A.C.: No-reference image quality assessment in the spatial domain. IEEE Transactions on image processing 21(12), 4695-4708 (2012)", + "36. Mittal, A., Soundararajan, R., Bovik, A.C.: Making a \"completely blind\" image quality analyzer. IEEE Signal processing letters 20(3), 209-212 (2012)", + "37. Nah, S., Baik, S., Hong, S., Moon, G., Son, S., Timofte, R., Mu Lee, K.: Ntire 2019 challenge on video deblurring and super-resolution: Dataset and study. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. pp. 0-0 (2019)", + "38. Nichol, A.Q., Dhariwal, P., Ramesh, A., Shyam, P., Mishkin, P., Mcgrew, B., Sutskever, I., Chen, M.: GLIDE: Towards photorealistic image generation and editing with text-guided diffusion models. In: Chaudhuri, K., Jegelka, S., Song, L., Szepesvari, C., Niu, G., Sabato, S. (eds.) Proceedings of the 39th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 162, pp. 16784-16804. PMLR (17-23 Jul 2022), https://proceedings.mlr.press/v162/nichol22a.html", + "39. Pan, J., Bai, H., Dong, J., Zhang, J., Tang, J.: Deep blind video super-resolution. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4811-4820 (2021)", + "40. Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., et al.: Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems 32 (2019)", + "41. Ranjan, A., Black, M.J.: Optical flow estimation using a spatial pyramid network. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4161-4170 (2017)", + "42. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models (2021)", + "43. Singer, U., Polyak, A., Hayes, T., Yin, X., An, J., Zhang, S., Hu, Q., Yang, H., Ashual, O., Gafni, O., et al.: Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792 (2022)", + "44. Song, Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score-based generative modeling through stochastic differential equations. In: Interna" + ], + "bbox": [ + 212, + 146, + 785, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "MGLD-VSR", + "bbox": [ + 645, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tional Conference on Learning Representations (2021), https://openreview.net/ forum?id=PxTIG12RRHS", + "45. Tao, X., Gao, H., Liao, R., Wang, J., Jia, J.: Detail-revealing deep video superresolution. In: Proceedings of the IEEE international conference on computer vision. pp. 4472-4480 (2017)", + "46. Tao Yang, Peiran Ren, X.X., Zhang, L.: Gan prior embedded network for blind face restoration in the wild. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2021)", + "47. Teed, Z., Deng, J.: Raft: Recurrent all-pairs field transforms for optical flow. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16. pp. 402-419. Springer (2020)", + "48. Wang, J., Yue, Z., Zhou, S., Chan, K.C., Loy, C.C.: Exploiting diffusion prior for real-world image super-resolution. arXiv preprint arXiv:2305.07015 (2023)", + "49. Wang, X., Chan, K.C., Yu, K., Dong, C., Change Loy, C.: Edvr: Video restoration with enhanced deformable convolutional networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. pp. 0-0 (2019)", + "50. Wang, X., Li, Y., Zhang, H., Shan, Y.: Towards real-world blind face restoration with generative facial prior. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9168-9178 (2021)", + "51. Wang, X., Xie, L., Dong, C., Shan, Y.: Real-esrgan: Training real-world blind superresolution with pure synthetic data. In: International Conference on Computer Vision Workshops (ICCVW) (2021)", + "52. Wu, H., Zhang, E., Liao, L., Chen, C., Hou, J., Wang, A., Sun, W., Yan, Q., Lin, W.: Exploring video quality assessment on user generated contents from aesthetic and technical perspectives. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 20144-20154 (2023)", + "53. Xie, L., Wang, X., Shi, S., Gu, J., Dong, C., Shan, Y.: Mitigating artifacts in real-world video super-resolution models. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 2956-2964 (2023)", + "54. Xu, H., Yang, J., Cai, J., Zhang, J., Tong, X.: High-resolution optical flow from 1d attention and correlation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10498-10507 (2021)", + "55. Yang, X., Xiang, W., Zeng, H., Zhang, L.: Real-world video super-resolution: A benchmark dataset and a decomposition based learning scheme. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4781-4790 (2021)", + "56. Yi, P., Wang, Z., Jiang, K., Jiang, J., Ma, J.: Progressive fusion video superresolution network via exploiting non-local spatio-temporal correlations. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 3106-3115 (2019)", + "57. Zhang, K., Liang, J., Van Gool, L., Timofte, R.: Designing a practical degradation model for deep blind image super-resolution. In: IEEE International Conference on Computer Vision. pp. 4791-4800 (2021)", + "58. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 586-595 (2018)", + "59. Zhao, W., Bai, L., Rao, Y., Zhou, J., Lu, J.: Unipc: A unified predictor-corrector framework for fast sampling of diffusion models. Advances in Neural Information Processing Systems 36 (2024)" + ], + "bbox": [ + 217, + 147, + 784, + 839 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "X. Yang et al.", + "bbox": [ + 271, + 114, + 364, + 128 + ], + "page_idx": 17 + }, + { + "type": "ref_text", + "text": "60. Zhou, S., Chan, K.C., Li, C., Loy, C.C.: Towards robust blind face restoration with codebook lookup transformer. In: NeurIPS (2022)", + "bbox": [ + 215, + 146, + 784, + 175 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "MGLD-VSR", + "bbox": [ + 645, + 114, + 730, + 126 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 18 + } +] \ No newline at end of file diff --git a/2024/Motion-Guided Latent Diffusion for Temporally Consistent Real-world Video Super-resolution/dbea7f31-cd1b-433d-b163-70053b8d34db_model.json b/2024/Motion-Guided Latent Diffusion for Temporally Consistent Real-world Video Super-resolution/dbea7f31-cd1b-433d-b163-70053b8d34db_model.json new file mode 100644 index 0000000000000000000000000000000000000000..f0c09a02aea6d6f87f1a2f9fe0eae6e29b38a1f6 --- /dev/null +++ b/2024/Motion-Guided Latent Diffusion for Temporally Consistent Real-world Video Super-resolution/dbea7f31-cd1b-433d-b163-70053b8d34db_model.json @@ -0,0 +1,3351 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.23, + 0.141, + 0.776, + 0.187 + ], + "angle": 0, + "content": "Motion-Guided Latent Diffusion for Temporally Consistent Real-world Video Super-resolution" + }, + { + "type": "text", + "bbox": [ + 0.285, + 0.219, + 0.717, + 0.237 + ], + "angle": 0, + "content": "Xi Yang\\(^{1,2}\\), Chenhang He\\(^{1}\\), Jianqi Ma\\(^{1,2}\\), and Lei Zhang\\(^{1,2^{\\star}}\\)" + }, + { + "type": "text", + "bbox": [ + 0.246, + 0.247, + 0.757, + 0.305 + ], + "angle": 0, + "content": "1 The Hong Kong Polytechnic University, Hong Kong SAR, China \n2 OPPO Research Institute, Shenzhen, China \n{xxxxi.yang, jianqi.ma}@connect.polyu.hk, chenhang.he@polyu.edu.hk, cslzhang@comp.polyu.edu.hk" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.339, + 0.744, + 0.646 + ], + "angle": 0, + "content": "Abstract. Real-world low-resolution (LR) videos have diverse and complex degradations, imposing great challenges on video super-resolution (VSR) algorithms to reproduce their high-resolution (HR) counterparts with high quality. Recently, the diffusion models have shown compelling performance in generating realistic details for image restoration tasks. However, the diffusion process has randomness, making it hard to control the contents of restored images. This issue becomes more serious when applying diffusion models to VSR tasks because temporal consistency is crucial to the perceptual quality of videos. In this paper, we propose an effective real-world VSR algorithm by leveraging the strength of pre-trained latent diffusion models. To ensure the content consistency among adjacent frames, we exploit the temporal dynamics in LR videos to guide the diffusion process by optimizing the latent sampling path with a motion-guided loss, ensuring that the generated HR video maintains a coherent and continuous visual flow. To further mitigate the discontinuity of generated details, we insert temporal modules to the decoder and fine-tune it with an innovative sequence-oriented loss. The proposed motion-guided latent diffusion (MGLD) based VSR algorithm achieves significantly better perceptual quality than state-of-the-arts on real-world VSR benchmark datasets, validating the effectiveness of the proposed model design and training strategies. Codes and models are available at https://github.com/IanYeung/MGLD-VSR." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.657, + 0.719, + 0.673 + ], + "angle": 0, + "content": "Keywords: Video super-resolution \\(\\cdot\\) Motion-guided latent diffusion" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.697, + 0.377, + 0.714 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.727, + 0.788, + 0.806 + ], + "angle": 0, + "content": "Video super-resolution (VSR) [3] aims at reconstructing a high-resolution (HR) video from a given low-resolution (LR) video sequence. Thanks to the rapid development of deep learning techniques, significant progress has been made for VSR in the past decade. Representative VSR methods include sliding window based methods like EDVR [49], recurrent based methods like BasicVSR [9] and" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.811, + 0.787, + 0.842 + ], + "angle": 0, + "content": "* Corresponding author. This work is supported by the Hong Kong RGC RIF grant (R5001-18) and the PolyU-OPPO Joint Innovation Lab." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.367, + 0.129 + ], + "angle": 0, + "content": "X. Yang et al." + }, + { + "type": "image", + "bbox": [ + 0.258, + 0.143, + 0.752, + 0.424 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.435, + 0.788, + 0.507 + ], + "angle": 0, + "content": "Fig. 1: The VSR results of 4 consecutive frames in the Sequence 026 from VideoLQ. The top row shows the VSR results by RealBasicVSR [11], the middle row shows the naive VSR results by running StableSR [48] on each frame of the sequence, and the bottom row shows the VSR results produced by our proposed MGLD-VSR method. Our method generates realistic details while achieving good temporal consistency." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.536, + 0.788, + 0.612 + ], + "angle": 0, + "content": "more recent transformer based methods [30]. However, most of the above works assume simple degradations (e.g., bicubic resizing, downsampling after Gaussian blur) between the LR and HR videos. As a result, the trained VSR models are hard to be generalized to real-world LR videos, whose degradations are unknown and much more complex." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.612, + 0.788, + 0.809 + ], + "angle": 0, + "content": "Recently, real-world VSR [11, 53, 55] has gained attention from researchers for its high potential value in practical applications such as camera-phone video enhancement and online video streaming. Different from traditional VSR tasks, which focus on super-resolving clean LR videos with simple synthetic degradation (e.g., bicubic downsampling), real-world VSR aims to enhance videos with complex and unknown degradations that often appear in real-world low-quality videos. The main challenge lies in how to effectively reproduce video details while suppressing the visual artifacts caused by unknown degradations. Existing real-world VSR algorithms mostly attempt to trade-off between details and artifacts using techniques such as Laplacian pyramid generative adversarial loss [55], input pre-cleaning [11] and hidden state attention [53], etc. Despite these efforts and progresses, their performance on real-world testing sets remains limited due to the insufficient training data and the limited model capacity." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Since the seminal work of denoising diffusion probabilistic model (DDPM) [20], diffusion-based generative models have achieved great success in image gen" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.646, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "MGLD-VSR" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.788, + 0.344 + ], + "angle": 0, + "content": "eration. In particular, the latent diffusion models (LDMs) [42] have demonstrated compelling results not only in text-to-image generation, but also in downstream tasks such as image editing [24], inpainting [32], colorization [7], etc. Recently, researchers have also attempted to employ the powerful generative priors of LDMs for real-world image restoration tasks [48] and shown promising results. It is thus interesting to investigate whether the latent diffusion priors can be used to improve real-world VSR results. However, this is a non-trivial task because of the intrinsic randomness in the diffusion process. This randomness in turn gives rise to the temporal inconsistencies among video frames in the latent space. Such inconsistencies can be further amplified by the decoder of the LDM. As shown in the middle row of Figure 1, directly employing the LDM based image super-resolution algorithm [48] for VSR will lead to inconsistent details across consecutive frames, disturbing the visual perception of the reproduced video." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.344, + 0.788, + 0.509 + ], + "angle": 0, + "content": "To address the above-mentioned challenges, we propose a Motion-Guided Latent Diffusion (MGLD) model for real-world VSR in this paper, aiming to produce high-quality HR video sequence with good temporal consistency. We propose to incorporate the motion dynamics of input LR video into the generation of SR outputs. This can be done through a guided diffusion process where we can generate conditional samples from a guidance model based on its score function [14, 44]. In particular, we calculate the optical flows among adjacent LR frames, and use the calculated optical flows to warp the latent features of each frame to align the features of neighboring frames. The \\( l_{1} \\) norm of the warping error is employed as the motion-guided loss, and its gradients can be incorporated in the sampling process to update the latent features." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.51, + 0.788, + 0.615 + ], + "angle": 0, + "content": "To overcome the inconsistent details generated by the original decoder of LDM, we insert temporal modules to the decoder and fine-tune it with the ground-truth HR video sequences. We also introduce an innovative sequence-oriented loss to guide the decoder to improve the continuity of the details. As depicted in the bottom row of Figure 1, our proposed MGLD-VSR approach could mitigate the discontinuity of generated details while maintaining natural and visually appealing texture recovery." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.616, + 0.788, + 0.736 + ], + "angle": 0, + "content": "The contributions of this paper are summarized as follows. First, we propose a diffusion sampling process based on motion-guided loss, allowing the temporal dynamics of the input frame to be used in generating temporally consistent latent features. Second, we propose a temporal-aware sequence decoder, along with two sequence-oriented losses, to further enhance the continuity of generated videos. Our proposed MGLD-VSR model achieves highly competitive real-world VSR results, exhibiting perceptually much more realistic details with fewer flickering artifacts than existing state-of-the-arts." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.761, + 0.388, + 0.779 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Video Super-Resolution. The recent advances in VSR research largely attribute to the rapid development of deep-learning technologies [6, 9, 22, 30, 49]. Based on the paradigms, existing deep-learning based VSR algorithms could" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.366, + 0.129 + ], + "angle": 0, + "content": "X. Yang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.329 + ], + "angle": 0, + "content": "be roughly classified into two categories. The first category performs VSR in a multiple-input single-output (MISO) fashion. Typical algorithms include DUF [22], EDVR [49], and MuCAN [28], etc. The MISO algorithms are usually sliding-window based, which aggregate information from neighboring frames to generate the super-resolved result of the center frame. The other category of methods adopt a multiple-input multiple-output (MIMO) design, such as BasicVSR [9], BasicVSR++ [10], VRT [30] and RVRT [31]. These methods take multiple LR frames as input and generate their corresponding SR results in one go. Compared with MISO, the amortized running cost of MIMO is usually much lower as it avoids a series of redundant operations. In addition, benefiting from the lower cost, the MIMO methods are able to aggregate information from more frames to improve the performance. We adopt the MIMO design in this work." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.333, + 0.788, + 0.56 + ], + "angle": 0, + "content": "Real-world VSR. Most existing VSR methods are trained on synthetic data with simple degradations (e.g., bicubic resizing, downsampling after Gaussian blur). Due to the synthetic-to-real gap, models trained with such data cannot perform well on real-world videos. To tackle this problem, Yang et al. [55] collected LR-HR video pairs with iPhone cameras to better model real-world degradations. While being effective in some cases, it is relatively labor-intensive and may not generalize well to videos collected by other devices. Later works usually adopt an easier solution: synthesizing more realistic training data by mixing more types of degradations, such as blur, downsampling, noise and video compression [11, 53]. With the improved synthetic training data, the trained models could achieve better super-resolution performance on real-world videos. However, there is still much distribution gap between the synthetic LR data and real-world LR videos. Instead of working on synthesizing more realistic training data, in this work we investigate how to leverage the strong generative priors of pre-trained latent diffusion models to improve the real-world VSR performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.564, + 0.788, + 0.777 + ], + "angle": 0, + "content": "Generative Prior for Image Restoration. The highly ill-posed nature of real-world restoration problem brings forward a great challenge on real-world image restoration algorithms. To boost the performance, a series of methods have introduced generative priors into the image restoration process. Early methods usually adopt pre-trained generative adversarial networks (GANs) as the generative prior [8,19,34,46,50]. Although effective, the priors adopted in these methods are usually tailored for specific tasks (e.g., face image restoration), limiting their capability in handling more general natural image restoration problems. Considering the higher generality of diffusion models than GANs in natural image generation, recently researchers have started to investigate how to leverage generative diffusion priors for image restoration, and demonstrated interesting results [17, 48]. In this work, we investigate the possibility of applying latent diffusion priors to real-world VSR, and develop effective techniques to tackle the flickering issue and improve the temporal consistency of generated videos." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Diffusion for Video Synthesis. Recently, diffusion based models have demonstrated its superiority in the video synthesis [4,5,18,43]. As one of the pioneer works, Make-a-Video [43] leverages a pretrained text-to-image (T2I) model and build a text-to-video (T2V) model by inserting temporal modules into the" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.646, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "MGLD-VSR" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image_caption", + "bbox": [ + 0.35, + 0.147, + 0.456, + 0.157 + ], + "angle": 0, + "content": "MGLD Sampling Pipeline" + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.163, + 0.576, + 0.35 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.621, + 0.148, + 0.746, + 0.157 + ], + "angle": 0, + "content": "Update with Motion Guidance" + }, + { + "type": "image", + "bbox": [ + 0.594, + 0.163, + 0.764, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.613, + 0.24, + 0.761, + 0.249 + ], + "angle": 0, + "content": "Temporal-aware Sequence Decoder" + }, + { + "type": "image", + "bbox": [ + 0.596, + 0.251, + 0.77, + 0.352 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.381, + 0.788, + 0.478 + ], + "angle": 0, + "content": "Fig. 2: Overview of the proposed MGLD-VSR framework for real-world VSR. We first estimate the forward and backward optical flow with a pre-trained optical flow estimation network and then employ a motion-guided loss to update the latent diffusion sampling process. The motion-guided loss is computed by summing the forward and backward masked warping error of the latent sequence at each time-step. After a number of \\( T \\) sampling steps, we obtain the generated latent sequence and feed it into the temporal-aware sequence decoder to reconstruct the VSR sequence." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.509, + 0.788, + 0.631 + ], + "angle": 0, + "content": "T2I model and designing a cascaded architecture. VideoLDM [5] adopts a latent diffusion paradigm and train the T2V model on large-scale paired text-video datasets, greatly improving the generation results. Stable Video Diffusion [4] further improve the results by conducting training on more carefully constructed datasets and introducing more advanced training paradigm. In this work, we leverage the diffusion prior and build a powerful real-world VSR model, which aims to tackle the ill-posedness of real-world VSR problem with diffusion prior and obtain temporally consistent VSR results with rich details." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.656, + 0.381, + 0.673 + ], + "angle": 0, + "content": "3 Methodology" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Given an LR video sequence of \\(N\\) frames \\(I = \\{I_1, I_2, \\ldots, I_N\\}\\), our goal is to reproduce an HR video sequence \\(\\hat{I} = \\{\\hat{I}_1, \\hat{I}_2, \\ldots, \\hat{I}_N\\}\\) with enhanced visual quality while maintaining the content consistency across adjacent frames. Inspired by the recently developed diffusion-based image restoration methods [17, 48], we propose to harness the generative priors of the pre-trained SD model [42] to improve the real-world VSR performance. First, we propose a motion-guided diffusion sampling process, where the temporal dynamics from the LR sequence are introduced into the diffusion sampling process to enhance the consistency of latent features across frames. Then, we further improve the quality of generated details by designing a temporal-aware sequence decoder and fine-tuning" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.366, + 0.129 + ], + "angle": 0, + "content": "X. Yang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.178 + ], + "angle": 0, + "content": "it with a video-oriented loss. The overall framework of our algorithm, namely motion-guided latent diffusion (MGLD), is illustrated in Figure 2." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.198, + 0.456, + 0.212 + ], + "angle": 0, + "content": "3.1 Latent Diffusion Model" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.221, + 0.788, + 0.359 + ], + "angle": 0, + "content": "Diffusion models [21] are a class of generative models which transform Gaussian noise into data samples through an iterative reverse Markovian process. This process can be described by: \\( q(x_{t}|x_{t - 1})\\coloneqq \\mathcal{N}(x_{t},(1 - \\beta_{t})x_{t - 1},\\beta_{t}I) \\), where the Gaussian noise is iteratively applied to the sample \\( x_{t - 1} \\) at time-step \\( t - 1 \\) to produce \\( x_{t} \\), guided by a variance scheduler \\( \\{\\beta_t\\}_{t = 1}^T \\) along the diffusion process. The key objective of diffusion models is to train a denoising U-Net \\( \\epsilon_{\\theta} \\) parameterized by \\( \\theta \\) and conditioned on the time-step \\( t \\) and an optional signal \\( y \\). Given the noise target \\( \\epsilon_{t} \\), the diffusion loss can be defined as: \\( L(\\theta) = \\mathbb{E}_{t\\sim U(1,T),\\epsilon_t\\sim \\mathcal{N}(0,I)}\\| \\epsilon_t - \\epsilon_\\theta (\\mathbf{x}_t;t,y)\\| _2^2 \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.36, + 0.788, + 0.525 + ], + "angle": 0, + "content": "Latent Diffusion Models (LDMs) [42] have made a remarkable impact in the field of image generation. These models employ Variational Autoencoders (VAE) to map the images into a latent space, enabling them to be trained on large-scale text-image datasets. This equips LDMs with powerful prior knowledge about natural images, which can be leveraged by downstream tasks such as image restoration and super-resolution [48] to produce images with clear details and high-quality content. However, in the case of VSR, the stochastic nature of diffusion process often leads to the temporal inconsistencies in the restored video sequence. To solve this issue, we propose a MGLD model that achieves high-quality VSR by utilizing guidance during the sampling process and incorporating temporal awareness in the decoder." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.545, + 0.547, + 0.56 + ], + "angle": 0, + "content": "3.2 Motion-guided Diffusion Sampling" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.568, + 0.788, + 0.63 + ], + "angle": 0, + "content": "From a score-based perspective [44], the sampling process of a diffusion model involves the application of Langevin Dynamics to both gradient of the data distribution \\(\\log p(\\mathbf{z})\\) and a guidance model \\(g = G(z)\\), which can be a network [15,38] or an energy model [12]:" + }, + { + "type": "equation", + "bbox": [ + 0.348, + 0.637, + 0.786, + 0.656 + ], + "angle": 0, + "content": "\\[\n\\nabla_ {\\mathbf {z}} \\log p (\\mathbf {z}, g) = \\nabla_ {\\mathbf {z}} \\log p (\\mathbf {z}) + \\nabla_ {\\mathbf {z}} \\log p (g | \\mathbf {z}). \\qquad (1)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.664, + 0.788, + 0.801 + ], + "angle": 0, + "content": "In this work, we introduce a technique that utilizes the temporal dynamics in low-resolution input videos to create latent features that maintain temporal consistency, thereby enhancing VSR results. Specifically, we incorporate an innovative motion-guided module within the sampling process to measure the warping errors of latent features across frames. Given the LR frames, we first compute their optical flow with \\(\\mathcal{F}\\) [47] and downsample the flow map to the dimension of latent features. Given the forward flow \\(O_{f}\\) and backward flow \\(O_{b}\\), we warp the latent features to their neighbouring frames and compute the accumulated warping error along both directions:" + }, + { + "type": "equation", + "bbox": [ + 0.217, + 0.806, + 0.787, + 0.842 + ], + "angle": 0, + "content": "\\[\nE (z) ^ {t} = \\sum_ {i = 1} ^ {N - 1} \\| \\left(\\operatorname {W a r p} \\left(z _ {i} ^ {t}, O _ {b, i}\\right) - z _ {i + 1} ^ {t})\\right) \\| _ {1} + \\sum_ {i = 2} ^ {N} \\| \\left(\\operatorname {W a r p} \\left(z _ {i} ^ {t}, O _ {f, i - 1}\\right) - z _ {i - 1} ^ {t})\\right) \\| _ {1}. \\tag {2}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.645, + 0.114, + 0.732, + 0.128 + ], + "angle": 0, + "content": "MGLD-VSR" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.238 + ], + "angle": 0, + "content": "We empirically find that the presence of occlusion will pose adverse impact on flow estimation, which in turn causes distraction in the sampling process and leads to artifacts in the final VSR results. To solve this issue, we further estimate the occlusion mask \\(\\mathcal{M}\\) in each frame [33] and ignore the contributions in the occluded region. Given the variance \\(\\sigma_t^2\\) of noise scheduler at step \\(t\\), the motion-guided sampling process at sampling step \\(t\\) can be written as:" + }, + { + "type": "equation", + "bbox": [ + 0.349, + 0.244, + 0.786, + 0.261 + ], + "angle": 0, + "content": "\\[\n\\hat {z} ^ {t} \\leftarrow \\operatorname {D D P M} \\left(z ^ {t + 1}, t\\right) - \\eta \\sigma_ {t} ^ {2} \\nabla_ {z} \\left(\\mathcal {M} \\circ E (z) ^ {t}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.267, + 0.785, + 0.328 + ], + "angle": 0, + "content": "where the DDPM [21] scheduler is applied for diffusion reverse (denoising) process and \\(\\eta = 10\\) is a weighting factor to control the guidance strength. After \\(T\\) iterations, the final denoised latent sequence \\(z^0\\) is fed into the VAE decoder to obtain the final video sequence." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.346, + 0.571, + 0.362 + ], + "angle": 0, + "content": "3.3 Temporal-aware Decoder Fine-tuning" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.367, + 0.785, + 0.488 + ], + "angle": 0, + "content": "With the proposed motion-guided diffusion sampling, the VSR model could better exploit the temporal dynamics to generate temporally more consistent HR videos. However, since the guidance is performed in the low-resolution latent space, which is 8 times smaller in both height and width than the output image space, the video reconstruction through a VAE decoder may still lead to temporal discontinuities in the generated details of VSR output. Thus, we introduce a temporal-aware sequence decoder and fine-tune it with ground-truth video sequences to improve the smoothness and accuracy of output videos." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.488, + 0.785, + 0.623 + ], + "angle": 0, + "content": "As shown in Figure 2, we build our sequence decoder on top of the pretrained VAE decoder by incorporating 1D convolutions along the temporal dimension. This design facilitates interleaved spatial-temporal interactions and enables our module to restore details with high continuity while incurring minimal computational cost. We further incorporate the information \\( e \\) from the VAE encoder through a Controllable Feature Warping (CFW) module [48, 60] for better restoration and generation results. To leverage the restorative power of the pretrained VAE, we freeze their original spatial blocks and only update the parameters in the temporal convolution and CFW module." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.624, + 0.785, + 0.699 + ], + "angle": 0, + "content": "We first employ the \\(\\ell_1\\) loss and perceptual loss [23, 58] as the reconstruction loss \\(L_{\\mathrm{recon}}\\) for each frame, then utilize the GAN-loss and frame difference loss that tailored for video sequences. The frame difference loss \\(L_{\\mathrm{diff}}\\) calculates the difference between consecutive predicted frames \\(\\hat{I}\\) and ground truth frames \\(I^{gt}\\), which is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.363, + 0.705, + 0.785, + 0.723 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {d i f f}} = \\sum \\| \\left(\\hat {I} _ {i + 1} - \\hat {I} _ {i}\\right) - \\left(I _ {i + 1} ^ {g t} - I _ {i} ^ {g t}\\right) \\| _ {1}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.729, + 0.784, + 0.79 + ], + "angle": 0, + "content": "It is noteworthy that diffusion models typically generate strong structures. To make the super-resolved video look more natural, we need to apply consistency constraints to those structured regions. We hence introduce a structure weighted consistency loss \\( L_{\\mathrm{swc}} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.328, + 0.796, + 0.786, + 0.838 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} L _ {\\mathrm {s w c}} = \\sum \\| \\mathcal {M} \\circ (\\mathcal {W} _ {i + 1} (\\operatorname {W a r p} (\\hat {I} _ {i}, O _ {b, i} ^ {g t}) - \\hat {I} _ {i + 1})) \\| _ {1} \\tag {5} \\\\ + \\sum \\| \\mathcal {M} \\circ (\\mathcal {W} _ {i - 1} (\\operatorname {W a r p} (\\hat {I _ {i}}, O _ {f, i} ^ {g t}) - \\hat {I} _ {i - 1})) \\| _ {1}. \\\\ \\end{array}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.367, + 0.129 + ], + "angle": 0, + "content": "X. Yang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.788, + 0.253 + ], + "angle": 0, + "content": "where \\( O_{f}^{gt} \\), \\( O_{b}^{gt} \\) denote the forward and backward flow computed on the ground-truth sequence. This loss helps the network to generate consistent details by modulating the loss through a weighted map \\( \\mathcal{W} = (1 + wS) \\) computed on the ground-truth frames, where \\( w = 3 \\) is a weighting factor and \\( S \\) is a structure (edge) map using the Sobel operator. Occlusion masking is also adopted. Given the GAN-based loss \\( L_{\\mathrm{GAN}} \\) and reconstruction loss \\( L_{\\mathrm{recon}} \\), the total loss for decoder fine-tuning is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.351, + 0.268, + 0.786, + 0.283 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {v i d e o}} = L _ {\\text {r e c o n}} + \\alpha L _ {\\text {d i f f}} + \\beta L _ {\\text {s w c}} + \\gamma L _ {\\text {G A N}}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.295, + 0.637, + 0.311 + ], + "angle": 0, + "content": "where \\(\\alpha, \\beta\\) and \\(\\gamma\\) are empirically set to 0.5, 0.5 and 0.025." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.338, + 0.377, + 0.355 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.373, + 0.448, + 0.389 + ], + "angle": 0, + "content": "4.1 Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.401, + 0.789, + 0.627 + ], + "angle": 0, + "content": "Implementation Details. The training of MGLD-VSR consists of two main stages. In the first stage, we fine-tune the denoising U-Net for latent space diffusion. The weight of the denoising U-Net is initialized from Stable Diffusion V2.1. We insert 1D temporal convolution into the U-Net to help temporal dynamic modeling. We then fix the weight in the denoising U-Net of SD model and train the conditioning and temporal modeling modules. For conditioning, we adopt a small time-aware encoder to encode LR condition and inject it into the denoising U-Net with spatial feature transform operation as in [48]. The training batch size, sequence length and latent image size are set to 24, 6 and \\(64 \\times 64\\), respectively. In the second stage, we first generate the clean latent sequence with the proposed motion-guided diffusion sampling process, then fine-tune the temporal-aware sequence decoder with the LR sequence, the generated latent sequence and the HR sequence. We fix the VAE decoder and insert temporal modeling layers and CFW modules for training. The training batch size, sequence length and image size are set to 4, 5 and \\(512 \\times 512\\), respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.628, + 0.788, + 0.704 + ], + "angle": 0, + "content": "Our models are trained on 4 NVIDIA A100 GPUs with the PyTorch [40] framework. Adam [26] is chosen as the optimizer. During inference, due to memory constrains, we divide the LR video into multiple sequences to run. To handle output with arbitrary resolution, we apply a progressive patch aggregation sampling algorithm [48]. For each sequence, we run the sampling for 50 steps." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Training and Testing Datasets. As in previous works [11, 49, 53], we merge the training set and validation set of REDS [37] in training and leave 4 sequences for testing (REDS4). We follow the degradation pipeline of RealBasicVSR [11] to synthesize the training sequence pairs, which involves blur, noise, downsampling and compression degradations. We evaluate our method on both synthetic and real-world datasets. To synthesize LR-HR testing sequence pairs, we apply the RealBasicVSR degradation pipeline on several widely used VSR datasets, including REDS4 [37], UDM10 [56] and SPMCS [45]. REDS4 contains 4 video sequences, each having 100 frames. UDM10 consists of 10 sequences, each" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.646, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "MGLD-VSR" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.192 + ], + "angle": 0, + "content": "having 32 frames. SPMCS has 30 sequences, each having 31 frames. For real-world dataset, we adopt VideoLQ [11] for testing, which contains 50 real-world sequences with complex degradations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.193, + 0.785, + 0.269 + ], + "angle": 0, + "content": "Evaluation Metrics. On synthetic datasets, we could adopt full-reference metrics to evaluate different real-world VSR methods. The perceptual quality metrics LPIPS [58] and DISTS [16] are selected. Fidelity-oriented metrics PSNR and SSIM are also calculated for reference. Apart from image-based metrics, the video quality assessment metric VMAF [29] is also employed." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.27, + 0.787, + 0.331 + ], + "angle": 0, + "content": "On real-world datasets, as ground truths are not available, we compute the no-reference image quality metrics NIQE [36] and BRISQUE [35] for evaluation. In addition, we also include a deep learning based image-based metric MUSIQ [25] and a state-of-the-art video quality assessment metric DOVER [52]." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.357, + 0.441, + 0.372 + ], + "angle": 0, + "content": "4.2 Experimental Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.386, + 0.785, + 0.461 + ], + "angle": 0, + "content": "To demonstrate the effectiveness of the our MGLD-VSR algorithm, we compare it with seven state-of-the-art methods, including three real-world image super-resolution models (BSRGAN [57], RealESRGAN [51], StableSR [48]), one classic VSR model (RVRT [31]), and three blind or real-world VSR models (DBVSR [39], RealVSR [55], RealBasicVSR [11]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.463, + 0.786, + 0.657 + ], + "angle": 0, + "content": "Quantitative Comparison. We first show the quantitative comparison on the synthetic datasets and real-world video benchmarks in Table 1. As shown from Table 1, our method achieves the best results in terms of full-reference perceptual based metrics LPIPS and DISTS on all synthetic test datasets, indicating that the proposed method is able to reconstruct high-quality plausible details from sequences degraded with complex degradations. Although methods like DBVSR may achieve better performance in terms of PSNR or SSIM on some datasets, they tend to generate blurry results, as demonstrated in the LPIPS and DISTS metrics. In terms of VMAF, our model achieves the best or second best results. As for real-world VSR dataset VideoLQ, our methods achieve the best performance in terms of NIQE, BRISQUE, and DOVER and the second best results in terms of MUSIQ, which reflects its strong capability to enhance the real-world videos and generate realistic details and textures." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Qualitative Comparison. To demonstrate the effectiveness of MGLD-VSR, we visualize the VSR results on synthetic datasets in Figure 3 and the results on real-world VideoLQ in Figure 4. One can see that MGLD-VSR can remove complex spatial-variant degradations and generate realistic details, outperforming other state-of-the-art real-world VSR algorithms. On synthetic datasets, as shown in Figure 3, MGLD-VSR can well reconstruct the structure while generating plausible details under complex degradations. On real-world videos, as shown by the example of sequence 008 in Figure 4, MGLD-VSR reconstructs clearer characters and patterns than competing methods. For the example of sequence 013, MGLD-VSR not only reconstructs the structure of the windows, and at the same time generates plausible exterior structure of the brick wall. More visual comparisons can be found in the supplementary materials." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.366, + 0.129 + ], + "angle": 0, + "content": "X. Yang et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.215 + ], + "angle": 0, + "content": "Table 1: Quantitative comparison of different VSR models evaluated on popular VSR benchmarking datasets. Bold and underline represent the best and second best score, respectively. For synthetic datasets (REDS4, UDM10, SPMCS), we adopt full-reference metrics to evaluate different methods. For real-world video datasets (VideoLQ), we adopt no-reference metrics to compare different methods." + }, + { + "type": "table", + "bbox": [ + 0.217, + 0.227, + 0.787, + 0.442 + ], + "angle": 0, + "content": "
DatasetsMetricsBicubicRVRTBSRGANRealESRGANStableSRDBVSRRealVSRRealBasicVSROurs
REDS4PSNR ↑22.1322.2922.0221.3422.0422.2921.7822.3522.46
SSIM ↑0.57300.57590.57020.56470.56830.57600.56630.58460.5723
LPIPS ↓0.62770.62210.48020.45090.45530.62210.62670.41900.3776
DISTS ↓0.36670.35540.21180.17500.18450.35560.36200.15810.1151
VMAF ↑2.12554.800526.241333.995526.69532.60752.590334.988434.3214
UDM10PSNR ↑25.7826.1825.2924.9425.5826.1425.0925.5625.99
SSIM ↑0.75590.76290.73230.74390.75920.76350.75240.74480.7548
LPIPS ↓0.44410.44130.42170.38840.36840.43980.43300.38970.3491
DISTS ↓0.27300.25750.19260.16510.14970.25850.26080.16370.1369
VMAF ↑5.464710.172431.087436.915833.82647.20486.007938.263839.3892
SPMCSPSNR ↑22.2822.5122.1221.6622.2622.5022.0322.4122.66
SSIM ↑0.56910.57530.57300.56940.58440.57640.57380.57360.5960
LPIPS ↓0.55720.54710.45090.44350.43210.54580.54870.44200.3990
DISTS ↓0.35930.34150.23070.21500.23510.34350.34790.21300.1934
VMAF ↑0.96565.353322.482827.693521.80603.75762.702427.451928.7032
VideoLQNIQE ↓9.13196.58884.21044.20364.44426.75046.33313.69153.5346
BRISQUE ↓72.740562.176825.240029.842826.309261.189757.276524.548421.9839
MUSIQ ↑21.678928.420452.221349.838047.804529.003924.218955.969152.7812
DOVER ↑0.36770.49070.71840.73090.71080.49980.52020.74010.7481
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.487, + 0.788, + 0.609 + ], + "angle": 0, + "content": "Temporal Consistency. Compared with single image super-resolution, one major aspect of VSR is the temporal consistency of the generated videos. Following most previous works, in this paper, we adopt the average warping error (WE) [13,27] of the sequence to quantitatively measure the temporal consistency: \\(\\mathrm{WE} = \\frac{1}{N - 1}\\sum_{i = 1}^{N - 1}\\| \\hat{I}_{i + 1} - \\mathrm{Warp}(\\hat{I}_i,O_{b,i})\\| _1\\), where \\(\\hat{I}_i\\) is the predicted frame and \\(O_{b,i}\\) is the estimated backward optical flow [54] from ground-truth video. On real-world video datasets without ground-truth, the optical flow is computed on the predicted frames." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.612, + 0.788, + 0.793 + ], + "angle": 0, + "content": "It is worth mentioning that, despite the wide use of WE for temporal consistency evaluation, it may not be able to faithfully reflect the real human perception on the video. For example, WE could be easily attacked by blurring the sequence, resulting much higher WE scores but lower video quality. The WE scores of competing methods are given in Table 2. One can see that our diffusion prior based method MGLD-VSR is not good at this metric because it tends to generate more details and textures, which are not favourable by WE. Nonetheless, the WE of MGLD-VSR is better than StableSR, which is also driven by the latent diffusion priors. To better showcase the effectiveness of our method, we visualize the temporal profiles of the VSR results produced by different algorithms in Figure 5. One can see that MGLD-VSR reconstructs more accurate details while maintaining good temporal consistency." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.842 + ], + "angle": 0, + "content": "User Study. We also conduct a user study to further examine the effectiveness of the proposed algorithm in comparison with RealBasicVSR [11] and StableSR [48]. The study is conducted in pairs: given the LR video, the subject" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.646, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "MGLD-VSR" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.144, + 0.338, + 0.195 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.232, + 0.196, + 0.329, + 0.204 + ], + "angle": 0, + "content": "Sequence 011 from REDS4" + }, + { + "type": "image", + "bbox": [ + 0.346, + 0.144, + 0.413, + 0.195 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.365, + 0.196, + 0.393, + 0.203 + ], + "angle": 0, + "content": "Bicubic" + }, + { + "type": "image", + "bbox": [ + 0.42, + 0.144, + 0.488, + 0.195 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.426, + 0.196, + 0.481, + 0.204 + ], + "angle": 0, + "content": "Real-ESRGAN" + }, + { + "type": "image", + "bbox": [ + 0.496, + 0.144, + 0.562, + 0.195 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.196, + 0.545, + 0.204 + ], + "angle": 0, + "content": "StableSR" + }, + { + "type": "image", + "bbox": [ + 0.57, + 0.144, + 0.637, + 0.195 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.577, + 0.196, + 0.629, + 0.204 + ], + "angle": 0, + "content": "RealBasicVSR" + }, + { + "type": "image", + "bbox": [ + 0.643, + 0.144, + 0.712, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.668, + 0.197, + 0.686, + 0.204 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.718, + 0.144, + 0.786, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.745, + 0.197, + 0.759, + 0.203 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image", + "bbox": [ + 0.222, + 0.207, + 0.338, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.228, + 0.259, + 0.332, + 0.266 + ], + "angle": 0, + "content": "Sequence band from UDM10" + }, + { + "type": "image", + "bbox": [ + 0.346, + 0.207, + 0.413, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.365, + 0.259, + 0.392, + 0.266 + ], + "angle": 0, + "content": "Bicubic" + }, + { + "type": "image", + "bbox": [ + 0.421, + 0.207, + 0.488, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.427, + 0.259, + 0.481, + 0.266 + ], + "angle": 0, + "content": "Real-ESRGAN" + }, + { + "type": "image", + "bbox": [ + 0.496, + 0.207, + 0.562, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.259, + 0.545, + 0.266 + ], + "angle": 0, + "content": "StableSR" + }, + { + "type": "image", + "bbox": [ + 0.57, + 0.207, + 0.637, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.577, + 0.259, + 0.63, + 0.266 + ], + "angle": 0, + "content": "RealBasicVSR" + }, + { + "type": "image", + "bbox": [ + 0.644, + 0.207, + 0.712, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.668, + 0.259, + 0.686, + 0.266 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.718, + 0.207, + 0.786, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.746, + 0.259, + 0.759, + 0.265 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.27, + 0.338, + 0.32 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.225, + 0.321, + 0.334, + 0.328 + ], + "angle": 0, + "content": "Sequence 000002 from SPMCS" + }, + { + "type": "image", + "bbox": [ + 0.346, + 0.269, + 0.413, + 0.32 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.365, + 0.321, + 0.393, + 0.328 + ], + "angle": 0, + "content": "Bicubic" + }, + { + "type": "image", + "bbox": [ + 0.421, + 0.269, + 0.488, + 0.32 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.427, + 0.321, + 0.481, + 0.328 + ], + "angle": 0, + "content": "Real-ESRGAN" + }, + { + "type": "image", + "bbox": [ + 0.496, + 0.269, + 0.562, + 0.32 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.321, + 0.545, + 0.328 + ], + "angle": 0, + "content": "StableSR" + }, + { + "type": "image", + "bbox": [ + 0.57, + 0.269, + 0.637, + 0.32 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.577, + 0.321, + 0.631, + 0.328 + ], + "angle": 0, + "content": "RealBasicVSR" + }, + { + "type": "image", + "bbox": [ + 0.643, + 0.269, + 0.712, + 0.32 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.668, + 0.321, + 0.687, + 0.328 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.718, + 0.269, + 0.786, + 0.32 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.746, + 0.321, + 0.758, + 0.328 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.34, + 0.785, + 0.368 + ], + "angle": 0, + "content": "Fig. 3: Qualitative comparison on synthetic datasets (REDS4, UDM10, SPMCS) for \\(\\times 4\\) VSR. (Zoom-in for best view.)" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.382, + 0.785, + 0.41 + ], + "angle": 0, + "content": "Table 2: Temporal consistency comparison of competing VSR methods by using WE \\(\\downarrow\\). Bold and underline represent the best and second best scores, respectively." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.432, + 0.787, + 0.498 + ], + "angle": 0, + "content": "
DatasetsBicubicRVRTBSRGANRealESRGANStableSRDBVSRRealVSRRealBasicVSROurs
REDS48.576.5010.2511.9111.818.518.6110.0910.08
UDM103.103.064.824.994.803.042.754.334.55
SPMCS3.013.585.187.124.623.584.182.723.87
VideoLQ5.486.387.727.649.335.985.746.767.56
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.538, + 0.784, + 0.584 + ], + "angle": 0, + "content": "is asked to choose the better one from two HR videos obtained by different methods, considering both the reconstruction quality and the temporal consistency. As shown in Figure 6, our method gains more votes in the competitions." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.605, + 0.389, + 0.621 + ], + "angle": 0, + "content": "4.3 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.785, + 0.704 + ], + "angle": 0, + "content": "In this section, we conduct a series of analyses on the proposed framework, including its Motion-guided Diffusion Sampling (MDS) module and Temporal-aware Sequence Decoder (TSD) module, and employ the VideoLQ dataset for experiments. We adopt WE to measure temporal consistency and no-reference metrics NIQE, BRISQUE, MUSIQ and DOVER to measure perceptual quality." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.785, + 0.78 + ], + "angle": 0, + "content": "Different Components of MGLD-VSR. As can seen from Table 3, the introduction of MDS improves all the metrics over the baseline (i.e., without using MDS and TSD), suggesting that MDS helps not only in reducing the temporal differences (i.e., WE), but also enhancing the perceptual quality of the reconstructed video sequence." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.785, + 0.84 + ], + "angle": 0, + "content": "The incorporation of MDS aids in maintaining perceptual continuity by ensuring that the restored frames could transition smoothly from one to the next. This contributes to more natural and visually pleasing VSR results. The incorporation of TSD also yields better results across all the metrics compared to" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.367, + 0.129 + ], + "angle": 0, + "content": "X. Yang et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.169, + 0.394, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.241, + 0.275, + 0.371, + 0.285 + ], + "angle": 0, + "content": "Sequence 008 from VideoLQ" + }, + { + "type": "image", + "bbox": [ + 0.399, + 0.144, + 0.484, + 0.211 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.423, + 0.211, + 0.46, + 0.219 + ], + "angle": 0, + "content": "Bicubic" + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.144, + 0.585, + 0.211 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.519, + 0.211, + 0.565, + 0.219 + ], + "angle": 0, + "content": "BSRGAN" + }, + { + "type": "image", + "bbox": [ + 0.599, + 0.144, + 0.686, + 0.211 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.607, + 0.211, + 0.676, + 0.219 + ], + "angle": 0, + "content": "Real-ESRGAN" + }, + { + "type": "image", + "bbox": [ + 0.7, + 0.144, + 0.786, + 0.211 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.722, + 0.211, + 0.765, + 0.219 + ], + "angle": 0, + "content": "StableSR" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.338, + 0.394, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.241, + 0.418, + 0.371, + 0.429 + ], + "angle": 0, + "content": "Sequence 013 from VideoLQ" + }, + { + "type": "image", + "bbox": [ + 0.399, + 0.302, + 0.484, + 0.369 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.423, + 0.369, + 0.459, + 0.377 + ], + "angle": 0, + "content": "Bicubic" + }, + { + "type": "image", + "bbox": [ + 0.399, + 0.381, + 0.484, + 0.446 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.42, + 0.447, + 0.462, + 0.455 + ], + "angle": 0, + "content": "RealVSR" + }, + { + "type": "image", + "bbox": [ + 0.499, + 0.302, + 0.584, + 0.369 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.519, + 0.369, + 0.565, + 0.377 + ], + "angle": 0, + "content": "BSRGAN" + }, + { + "type": "image", + "bbox": [ + 0.499, + 0.381, + 0.584, + 0.446 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.523, + 0.447, + 0.561, + 0.455 + ], + "angle": 0, + "content": "DBVSR" + }, + { + "type": "image", + "bbox": [ + 0.599, + 0.302, + 0.686, + 0.369 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.607, + 0.369, + 0.676, + 0.377 + ], + "angle": 0, + "content": "Real-ESRGAN" + }, + { + "type": "image", + "bbox": [ + 0.6, + 0.381, + 0.686, + 0.446 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.609, + 0.447, + 0.675, + 0.455 + ], + "angle": 0, + "content": "RealBasicVSR" + }, + { + "type": "image", + "bbox": [ + 0.7, + 0.222, + 0.786, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.731, + 0.289, + 0.755, + 0.297 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.7, + 0.302, + 0.786, + 0.369 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.722, + 0.369, + 0.765, + 0.377 + ], + "angle": 0, + "content": "StableSR" + }, + { + "type": "image", + "bbox": [ + 0.7, + 0.381, + 0.786, + 0.446 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.731, + 0.447, + 0.755, + 0.455 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.467, + 0.789, + 0.496 + ], + "angle": 0, + "content": "Fig. 4: Qualitative comparison on the real-world video dataset (VideoLQ) for \\(\\times 4\\) video SR. (Zoom-in for best view.)" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.508, + 0.321, + 0.555 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.323, + 0.508, + 0.396, + 0.555 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.4, + 0.508, + 0.474, + 0.555 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.478, + 0.508, + 0.552, + 0.555 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.555, + 0.508, + 0.629, + 0.555 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.633, + 0.508, + 0.707, + 0.555 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.508, + 0.785, + 0.555 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.56, + 0.321, + 0.606 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.322, + 0.56, + 0.396, + 0.606 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.4, + 0.56, + 0.474, + 0.606 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.478, + 0.56, + 0.552, + 0.606 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.555, + 0.56, + 0.629, + 0.606 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.633, + 0.56, + 0.707, + 0.606 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.56, + 0.785, + 0.606 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.285, + 0.631, + 0.719, + 0.646 + ], + "angle": 0, + "content": "Fig. 5: Temporal profiles of competing real-world VSR methods." + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.659, + 0.499, + 0.758 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.659, + 0.785, + 0.758 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.769, + 0.789, + 0.799 + ], + "angle": 0, + "content": "Fig. 6: User study on 20 videos from VideoLQ evaluated by 10 subjects. MGLD-VSR gains more votes than competing methods." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.645, + 0.114, + 0.732, + 0.127 + ], + "angle": 0, + "content": "MGLD-VSR" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.784, + 0.186 + ], + "angle": 0, + "content": "Table 3: Ablation study on the different components of the proposed algorithm. \"MDS\" refers to Motion-guided Diffusion Sampling and \"TSD\" refers to Temporal-aware Sequence Decoder." + }, + { + "type": "table", + "bbox": [ + 0.328, + 0.199, + 0.671, + 0.265 + ], + "angle": 0, + "content": "
MDS TSDWE ↓NIQE ↓BRISQUE ↓MUSIQ ↑DOVER ↑
✓9.30694.109825.324951.56770.7412
8.98483.987624.084052.41430.7435
✓8.36793.628122.080652.68740.7468
✓7.54783.534621.983952.78120.7481
" + }, + { + "type": "image", + "bbox": [ + 0.232, + 0.284, + 0.415, + 0.391 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.251, + 0.393, + 0.396, + 0.404 + ], + "angle": 0, + "content": "Sequence 002 from VideoLQ" + }, + { + "type": "image", + "bbox": [ + 0.418, + 0.284, + 0.486, + 0.39 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.448, + 0.393, + 0.457, + 0.402 + ], + "angle": 0, + "content": "0" + }, + { + "type": "image", + "bbox": [ + 0.49, + 0.284, + 0.556, + 0.39 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.517, + 0.393, + 0.532, + 0.402 + ], + "angle": 0, + "content": "10" + }, + { + "type": "image", + "bbox": [ + 0.561, + 0.284, + 0.628, + 0.39 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.588, + 0.393, + 0.603, + 0.402 + ], + "angle": 0, + "content": "20" + }, + { + "type": "image", + "bbox": [ + 0.633, + 0.284, + 0.701, + 0.39 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.659, + 0.392, + 0.674, + 0.402 + ], + "angle": 0, + "content": "30" + }, + { + "type": "image", + "bbox": [ + 0.705, + 0.284, + 0.772, + 0.39 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.731, + 0.393, + 0.746, + 0.402 + ], + "angle": 0, + "content": "40" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.415, + 0.784, + 0.444 + ], + "angle": 0, + "content": "Fig. 7: Impact of guidance strength during sampling. The numbers below the cropped patches indicate the motion guidance weighting factor \\(\\eta\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.481, + 0.785, + 0.557 + ], + "angle": 0, + "content": "the baseline. This occurs because TSD is capable of extracting richer contextual information from neighboring frames. When both MDS and TSD are used, the results show further improvements. The full MGLD-VSR approach achieves the best scores on all the metrics, indicating the synergy of the two proposed components for real-world VSR." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.561, + 0.785, + 0.607 + ], + "angle": 0, + "content": "Impact of Flow Network. As shown in Table 4, the flow estimation network in motion guided sampling does not show big influence on the results. This might because the motion guided sampling is applied in low-resolution space." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.61, + 0.785, + 0.701 + ], + "angle": 0, + "content": "Impact of Guidance Strength. During guided sampling process, we introduce a weighting factor to control the guidance strength. As shown in Figure 7, a moderate guidance strength could reduce the artifacts and improve the temporal consistency of the generated results. A large guidance strength will blur the results and degrade the reconstructed details. To balance the generation power and temporal consistency, we set the weighting factor \\(\\eta = 10\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.787, + 0.841 + ], + "angle": 0, + "content": "The Design of VAE Decoder. To validate the effectiveness of the proposed architecture and loss design for the VAE decoder, we conduct a series of study and the results are shown in Table 5. As can be seen from the table, the introduction of the adaptive temporal modeling module improves the temporal consistency as well as perceptual quality (measured by no-reference metrics), and incorporating the feature from encoder to decoder during decoder fine-tuning further improves the results. By introducing two sequence based losses during fine-tuning, we improve the temporal consistency of the results while maintaining or improving the perceptual quality." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.366, + 0.129 + ], + "angle": 0, + "content": "X. Yang et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.225, + 0.145, + 0.777, + 0.16 + ], + "angle": 0, + "content": "Table 4: Ablation study on the flow network in motion guided diffusion sampling." + }, + { + "type": "table", + "bbox": [ + 0.356, + 0.171, + 0.644, + 0.214 + ], + "angle": 0, + "content": "
Flow NetworkWE ↓NIQE ↓MUSIQ ↑DOVER ↑
SpyNet [41]7.67343.554552.69440.7478
RAFT [47]7.54783.534652.78120.7481
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.228, + 0.787, + 0.27 + ], + "angle": 0, + "content": "Table 5: Ablation study on the different designs of VAE decoder. \"TM\" refers to the temporal modeling module. \"CFW\" refers to the controllable feature warping module that injects VAE encoder feature into VAE decoder." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.283, + 0.782, + 0.344 + ], + "angle": 0, + "content": "
ModelsTM CFWWE ↓ NIQE ↓ MUSIQ ↑ DOVER ↑LossesLdiffLswcWE ↓ NIQE ↓ MUSIQ ↑ DOVER ↑
(a)9.86704.067150.27070.7360(a)7.94803.535952.7894
(b)✓8.68463.845152.54080.7431(b)✓7.75793.540952.5648
(c)✓8.98423.923652.43860.7426(c)✓7.76883.537652.7456
(d)✓7.54783.534652.78120.7481(d)✓✓7.54783.534652.7812
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.375, + 0.357, + 0.388 + ], + "angle": 0, + "content": "4.4 Limitations" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.407, + 0.789, + 0.575 + ], + "angle": 0, + "content": "Our proposed MGLD-VSR method leverages the diffusion prior from pre-trained SD model, aiming to achieve perceptually realistic and temporally consistent VSR results. Due to the scale of pre-trained SD model, it does not have advantages in terms of number of parameters and inference time compared with those non-diffusion VSR methods (comparisons can be found in the supplementary materials). In addition, since MGLD-VSR will generate more visual details than those non-diffusion VSR methods, it does not have advantages in terms of temporal consistency metrics either. In the future, we can resort to model distillation [2] and more advanced diffusion sampling [1,59] techniques to improve the efficiency of MGLD-VSR, and resort to the recently developed video synthesis models [4] to improve the temporal consistency." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.605, + 0.36, + 0.62 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.789, + 0.842 + ], + "angle": 0, + "content": "We presented a motion-guided latent diffusion algorithm, namely MGLD-VSR, for high-quality and temporally consistent real-world VSR. Our method exploited the strong generative prior of pre-trained latent diffusion model to restore video sequence with realistic details. To ensure the content consistency among frames, we developed a motion-guided sampling mechanism, where the motion dynamics of input video were incorporated in the diffusion sampling process. To further improve the continuity of generated details, we introduced a temporal module into the decoder, and fine-tuned it by high quality video sequences with delicately designed sequence-based losses. The fine-tuned sequence decoder mitigated the flickering artifacts and further improved the perceptual quality. Our proposed MGLD-VSR achieved state-of-the-art visual results on the widely adopted real-world VSR benchmarks, validating the great potentials of latent diffusion priors in enhancing real-world video quality." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.646, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "MGLD-VSR" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.18, + 0.785, + 0.208 + ], + "angle": 0, + "content": "1. Lu et al., C.: Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. NeurIPS 35, 5775-5787 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.21, + 0.785, + 0.236 + ], + "angle": 0, + "content": "2. Luo et al., S.: Latent consistency models: Synthesizing high-resolution images with few-step inference. arXiv preprint arXiv:2310.04378 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.237, + 0.785, + 0.264 + ], + "angle": 0, + "content": "3. Bishop, C.M., Blake, A., Marthi, B.: Super-resolution enhancement of video. In: AISTATS (2003)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.265, + 0.785, + 0.32 + ], + "angle": 0, + "content": "4. Blattmann, A., Dockhorn, T., Kulal, S., Mendelevitch, D., Kilian, M., Lorenz, D., Levi, Y., English, Z., Voleti, V., Letts, A., et al.: Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.322, + 0.785, + 0.377 + ], + "angle": 0, + "content": "5. Blattmann, A., Rombach, R., Ling, H., Dockhorn, T., Kim, S.W., Fidler, S., Kreis, K.: Align your latents: High-resolution video synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22563-22575 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.378, + 0.785, + 0.433 + ], + "angle": 0, + "content": "6. Caballero, J., Ledig, C., Aitken, A., Acosta, A., Totz, J., Wang, Z., Shi, W.: Real-time video super-resolution with spatio-temporal networks and motion compensation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4778-4787 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.434, + 0.785, + 0.475 + ], + "angle": 0, + "content": "7. Carrillo, H., Clément, M., Bugeau, A., Simo-Serra, E.: Diffusart: Enhancing line art colorization with conditional diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3485-3489 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.476, + 0.785, + 0.517 + ], + "angle": 0, + "content": "8. Chan, K.C., Wang, X., Xu, X., Gu, J., Loy, C.C.: Glean: Generative latent bank for large-factor image super-resolution. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 14245-14254 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.518, + 0.785, + 0.572 + ], + "angle": 0, + "content": "9. Chan, K.C., Wang, X., Yu, K., Dong, C., Loy, C.C.: Basicvsr: The search for essential components in video super-resolution and beyond. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4947-4956 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.574, + 0.785, + 0.628 + ], + "angle": 0, + "content": "10. Chan, K.C., Zhou, S., Xu, X., Loy, C.C.: Basicvsr++: Improving video superresolution with enhanced propagation and alignment. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5972-5981 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.63, + 0.785, + 0.671 + ], + "angle": 0, + "content": "11. Chan, K.C., Zhou, S., Xu, X., Loy, C.C.: Investigating tradeoffs in real-world video super-resolution. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5962-5971 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.673, + 0.785, + 0.699 + ], + "angle": 0, + "content": "12. Chen, M., Laina, I., Vedaldi, A.: Training-free layout control with cross-attention guidance. arXiv preprint arXiv:2304.03373 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.701, + 0.785, + 0.741 + ], + "angle": 0, + "content": "13. Chu, M., Xie, Y., Mayer, J., Leal-Taixe, L., Thuerey, N.: Learning temporal coherence via self-supervision for gan-based video generation. ACM Transactions on Graphics (TOG) 39(4), 75-1 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.743, + 0.785, + 0.811 + ], + "angle": 0, + "content": "14. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. In: Ranzato, M., Beygelzimer, A., Dauphin, Y., Liang, P., Vaughan, J.W. (eds.) Advances in Neural Information Processing Systems. vol. 34, pp. 8780-8794. Curran Associates, Inc. (2021), https://proceedings.neurips.cc/paper_files/paper/2021/file/49ad23d1ec9fa4bd8d77d02681df5cfa-Paper.pdf" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "15. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. In: Ranzato, M., Beygelzimer, A., Dauphin, Y., Liang, P., Vaughan, J.W. (eds.) Advances" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.18, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.366, + 0.129 + ], + "angle": 0, + "content": "X. Yang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.241, + 0.148, + 0.786, + 0.189 + ], + "angle": 0, + "content": "in Neural Information Processing Systems. vol. 34, pp. 8780-8794. Curran Associates, Inc. (2021), https://proceedings.neurips.cc/paper_files/paper/2021/file/49ad23d1ec9fa4bd8d77d02681df5cfa-Paper.pdf" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.787, + 0.232 + ], + "angle": 0, + "content": "16. Ding, K., Ma, K., Wang, S., Simoncelli, E.P.: Image quality assessment: Unifying structure and texture similarity. IEEE transactions on pattern analysis and machine intelligence 44(5), 2567-2581 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.232, + 0.787, + 0.286 + ], + "angle": 0, + "content": "17. Fei, B., Lyu, Z., Pan, L., Zhang, J., Yang, W., Luo, T., Zhang, B., Dai, B.: Generative diffusion prior for unified image restoration and enhancement. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9935-9946 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.287, + 0.787, + 0.342 + ], + "angle": 0, + "content": "18. Ge, S., Nah, S., Liu, G., Poon, T., Tao, A., Catanzaro, B., Jacobs, D., Huang, J.B., Liu, M.Y., Balaji, Y.: Preserve your own correlation: A noise prior for video diffusion models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 22930-22941 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.343, + 0.787, + 0.384 + ], + "angle": 0, + "content": "19. Gu, J., Shen, Y., Zhou, B.: Image processing using multi-code gan prior. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 3012-3021 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.787, + 0.412 + ], + "angle": 0, + "content": "20. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in neural information processing systems 33, 6840-6851 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.412, + 0.787, + 0.481 + ], + "angle": 0, + "content": "21. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. In: Larochelle, H., Ranzato, M., Hadsell, R., Balcan, M., Lin, H. (eds.) Advances in Neural Information Processing Systems. vol. 33, pp. 6840-6851. Curran Associates, Inc. (2020), https://proceedings.neurips.cc/paper_files/paper/2020/file/4c5bcfec8584af0d967f1ab10179ca4b-Paper.pdf" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.481, + 0.787, + 0.535 + ], + "angle": 0, + "content": "22. Jo, Y., Oh, S.W., Kang, J., Kim, S.J.: Deep video super-resolution network using dynamic upsampling filters without explicit motion compensation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3224-3232 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.536, + 0.787, + 0.591 + ], + "angle": 0, + "content": "23. Johnson, J., Alahi, A., Fei-Fei, L.: Perceptual losses for real-time style transfer and super-resolution. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14. pp. 694-711. Springer (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.591, + 0.787, + 0.647 + ], + "angle": 0, + "content": "24. Kawar, B., Zada, S., Lang, O., Tov, O., Chang, H., Dekel, T., Mosseri, I., Irani, M.: Imagic: Text-based real image editing with diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6007-6017 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.647, + 0.787, + 0.688 + ], + "angle": 0, + "content": "25. Ke, J., Wang, Q., Wang, Y., Milanfar, P., Yang, F.: Musiq: Multi-scale image quality transformer. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5148-5157 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.689, + 0.787, + 0.716 + ], + "angle": 0, + "content": "26. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.716, + 0.787, + 0.758 + ], + "angle": 0, + "content": "27. Lai, W.S., Huang, J.B., Wang, O., Shechtman, E., Yumer, E., Yang, M.H.: Learning blind video temporal consistency. In: Proceedings of the European conference on computer vision (ECCV). pp. 170-185 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.787, + 0.813 + ], + "angle": 0, + "content": "28. Li, W., Tao, X., Guo, T., Qi, L., Lu, J., Jia, J.: Mucan: Multi-correspondence aggregation network for video super-resolution. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part X 16. pp. 335-351. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.787, + 0.841 + ], + "angle": 0, + "content": "29. Li, Z., Bampis, C., Novak, J., Aaron, A., Swanson, K., Moorthy, A., Cock, J.: Vmaf: The journey continues. Netflix Technology Blog 25(1) (2018)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.646, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "MGLD-VSR" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "30. Liang, J., Cao, J., Fan, Y., Zhang, K., Ranjan, R., Li, Y., Timofte, R., Van Gool, L.: Vrt: A video restoration transformer. arXiv preprint arXiv:2201.12288 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.177, + 0.787, + 0.232 + ], + "angle": 0, + "content": "31. Liang, J., Fan, Y., Xiang, X., Ranjan, R., Ilg, E., Green, S., Cao, J., Zhang, K., Timofte, R., Gool, L.V.: Recurrent video restoration transformer with guided deformable attention. Advances in Neural Information Processing Systems 35, 378-393 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.233, + 0.785, + 0.285 + ], + "angle": 0, + "content": "32. Lugmayr, A., Danelljan, M., Romero, A., Yu, F., Timofte, R., Van Gool, L.: Repaint: Inpainting using denoising diffusion probabilistic models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11461-11471 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.785, + 0.327 + ], + "angle": 0, + "content": "33. Meister, S., Hur, J., Roth, S.: Unflow: Unsupervised learning of optical flow with a bidirectional census loss. In: Proceedings of the AAAI conference on artificial intelligence. vol. 32 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.329, + 0.785, + 0.383 + ], + "angle": 0, + "content": "34. Menon, S., Damian, A., Hu, S., Ravi, N., Rudin, C.: Pulse: Self-supervised photoupsampling via latent space exploration of generative models. In: Proceedings of the iee/cvf conference on computer vision and pattern recognition. pp. 2437-2445 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.785, + 0.424 + ], + "angle": 0, + "content": "35. Mittal, A., Moorthy, A.K., Bovik, A.C.: No-reference image quality assessment in the spatial domain. IEEE Transactions on image processing 21(12), 4695-4708 (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.425, + 0.785, + 0.452 + ], + "angle": 0, + "content": "36. Mittal, A., Soundararajan, R., Bovik, A.C.: Making a \"completely blind\" image quality analyzer. IEEE Signal processing letters 20(3), 209-212 (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.785, + 0.508 + ], + "angle": 0, + "content": "37. Nah, S., Baik, S., Hong, S., Moon, G., Son, S., Timofte, R., Mu Lee, K.: Ntire 2019 challenge on video deblurring and super-resolution: Dataset and study. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. pp. 0-0 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.509, + 0.785, + 0.604 + ], + "angle": 0, + "content": "38. Nichol, A.Q., Dhariwal, P., Ramesh, A., Shyam, P., Mishkin, P., Mcgrew, B., Sutskever, I., Chen, M.: GLIDE: Towards photorealistic image generation and editing with text-guided diffusion models. In: Chaudhuri, K., Jegelka, S., Song, L., Szepesvari, C., Niu, G., Sabato, S. (eds.) Proceedings of the 39th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 162, pp. 16784-16804. PMLR (17-23 Jul 2022), https://proceedings.mlr.press/v162/nichol22a.html" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.605, + 0.785, + 0.646 + ], + "angle": 0, + "content": "39. Pan, J., Bai, H., Dong, J., Zhang, J., Tang, J.: Deep blind video super-resolution. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4811-4820 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.647, + 0.785, + 0.702 + ], + "angle": 0, + "content": "40. Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., et al.: Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems 32 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.702, + 0.785, + 0.743 + ], + "angle": 0, + "content": "41. Ranjan, A., Black, M.J.: Optical flow estimation using a spatial pyramid network. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4161-4170 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.744, + 0.785, + 0.771 + ], + "angle": 0, + "content": "42. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.772, + 0.785, + 0.812 + ], + "angle": 0, + "content": "43. Singer, U., Polyak, A., Hayes, T., Yin, X., An, J., Zhang, S., Hu, Q., Yang, H., Ashual, O., Gafni, O., et al.: Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "44. Song, Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score-based generative modeling through stochastic differential equations. In: Interna" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.366, + 0.129 + ], + "angle": 0, + "content": "X. Yang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.245, + 0.148, + 0.785, + 0.174 + ], + "angle": 0, + "content": "tional Conference on Learning Representations (2021), https://openreview.net/ forum?id=PxTIG12RRHS" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.176, + 0.785, + 0.217 + ], + "angle": 0, + "content": "45. Tao, X., Gao, H., Liao, R., Wang, J., Jia, J.: Detail-revealing deep video superresolution. In: Proceedings of the IEEE international conference on computer vision. pp. 4472-4480 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.218, + 0.785, + 0.259 + ], + "angle": 0, + "content": "46. Tao Yang, Peiran Ren, X.X., Zhang, L.: Gan prior embedded network for blind face restoration in the wild. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.26, + 0.785, + 0.3 + ], + "angle": 0, + "content": "47. Teed, Z., Deng, J.: Raft: Recurrent all-pairs field transforms for optical flow. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16. pp. 402-419. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.301, + 0.785, + 0.328 + ], + "angle": 0, + "content": "48. Wang, J., Yue, Z., Zhou, S., Chan, K.C., Loy, C.C.: Exploiting diffusion prior for real-world image super-resolution. arXiv preprint arXiv:2305.07015 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.329, + 0.785, + 0.383 + ], + "angle": 0, + "content": "49. Wang, X., Chan, K.C., Yu, K., Dong, C., Change Loy, C.: Edvr: Video restoration with enhanced deformable convolutional networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. pp. 0-0 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.384, + 0.785, + 0.424 + ], + "angle": 0, + "content": "50. Wang, X., Li, Y., Zhang, H., Shan, Y.: Towards real-world blind face restoration with generative facial prior. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9168-9178 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.425, + 0.785, + 0.466 + ], + "angle": 0, + "content": "51. Wang, X., Xie, L., Dong, C., Shan, Y.: Real-esrgan: Training real-world blind superresolution with pure synthetic data. In: International Conference on Computer Vision Workshops (ICCVW) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.467, + 0.785, + 0.522 + ], + "angle": 0, + "content": "52. Wu, H., Zhang, E., Liao, L., Chen, C., Hou, J., Wang, A., Sun, W., Yan, Q., Lin, W.: Exploring video quality assessment on user generated contents from aesthetic and technical perspectives. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 20144-20154 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.523, + 0.785, + 0.563 + ], + "angle": 0, + "content": "53. Xie, L., Wang, X., Shi, S., Gu, J., Dong, C., Shan, Y.: Mitigating artifacts in real-world video super-resolution models. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 2956-2964 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.565, + 0.785, + 0.605 + ], + "angle": 0, + "content": "54. Xu, H., Yang, J., Cai, J., Zhang, J., Tong, X.: High-resolution optical flow from 1d attention and correlation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10498-10507 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.606, + 0.785, + 0.66 + ], + "angle": 0, + "content": "55. Yang, X., Xiang, W., Zeng, H., Zhang, L.: Real-world video super-resolution: A benchmark dataset and a decomposition based learning scheme. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4781-4790 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.661, + 0.785, + 0.715 + ], + "angle": 0, + "content": "56. Yi, P., Wang, Z., Jiang, K., Jiang, J., Ma, J.: Progressive fusion video superresolution network via exploiting non-local spatio-temporal correlations. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 3106-3115 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.716, + 0.785, + 0.757 + ], + "angle": 0, + "content": "57. Zhang, K., Liang, J., Van Gool, L., Timofte, R.: Designing a practical degradation model for deep blind image super-resolution. In: IEEE International Conference on Computer Vision. pp. 4791-4800 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.758, + 0.785, + 0.799 + ], + "angle": 0, + "content": "58. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 586-595 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "59. Zhao, W., Bai, L., Rao, Y., Zhou, J., Lu, J.: Unipc: A unified predictor-corrector framework for fast sampling of diffusion models. Advances in Neural Information Processing Systems 36 (2024)" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.646, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "MGLD-VSR" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.176 + ], + "angle": 0, + "content": "60. Zhou, S., Chan, K.C., Li, C., Loy, C.C.: Towards robust blind face restoration with codebook lookup transformer. In: NeurIPS (2022)" + } + ] +] \ No newline at end of file diff --git a/2024/Motion-Guided Latent Diffusion for Temporally Consistent Real-world Video Super-resolution/dbea7f31-cd1b-433d-b163-70053b8d34db_origin.pdf b/2024/Motion-Guided Latent Diffusion for Temporally Consistent Real-world Video Super-resolution/dbea7f31-cd1b-433d-b163-70053b8d34db_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..92b3aef1dc88b4e819d02d1d41dd559b756f9e5b --- /dev/null +++ b/2024/Motion-Guided Latent Diffusion for Temporally Consistent Real-world Video Super-resolution/dbea7f31-cd1b-433d-b163-70053b8d34db_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:93ed17685d5315dacd797681f71928ea7a11ea376495971d995e8f675697a266 +size 6366192 diff --git a/2024/Motion-Guided Latent Diffusion for Temporally Consistent Real-world Video Super-resolution/full.md b/2024/Motion-Guided Latent Diffusion for Temporally Consistent Real-world Video Super-resolution/full.md new file mode 100644 index 0000000000000000000000000000000000000000..5e6e3c1a539960c64631473284bef0e6c8d23f21 --- /dev/null +++ b/2024/Motion-Guided Latent Diffusion for Temporally Consistent Real-world Video Super-resolution/full.md @@ -0,0 +1,424 @@ +# Motion-Guided Latent Diffusion for Temporally Consistent Real-world Video Super-resolution + +Xi Yang $^{1,2}$ , Chenhang He $^{1}$ , Jianqi Ma $^{1,2}$ , and Lei Zhang $^{1,2^{\star}}$ + +1 The Hong Kong Polytechnic University, Hong Kong SAR, China +2 OPPO Research Institute, Shenzhen, China +{xxxxi.yang, jianqi.ma}@connect.polyu.hk, chenhang.he@polyu.edu.hk, cslzhang@comp.polyu.edu.hk + +Abstract. Real-world low-resolution (LR) videos have diverse and complex degradations, imposing great challenges on video super-resolution (VSR) algorithms to reproduce their high-resolution (HR) counterparts with high quality. Recently, the diffusion models have shown compelling performance in generating realistic details for image restoration tasks. However, the diffusion process has randomness, making it hard to control the contents of restored images. This issue becomes more serious when applying diffusion models to VSR tasks because temporal consistency is crucial to the perceptual quality of videos. In this paper, we propose an effective real-world VSR algorithm by leveraging the strength of pre-trained latent diffusion models. To ensure the content consistency among adjacent frames, we exploit the temporal dynamics in LR videos to guide the diffusion process by optimizing the latent sampling path with a motion-guided loss, ensuring that the generated HR video maintains a coherent and continuous visual flow. To further mitigate the discontinuity of generated details, we insert temporal modules to the decoder and fine-tune it with an innovative sequence-oriented loss. The proposed motion-guided latent diffusion (MGLD) based VSR algorithm achieves significantly better perceptual quality than state-of-the-arts on real-world VSR benchmark datasets, validating the effectiveness of the proposed model design and training strategies. Codes and models are available at https://github.com/IanYeung/MGLD-VSR. + +Keywords: Video super-resolution $\cdot$ Motion-guided latent diffusion + +# 1 Introduction + +Video super-resolution (VSR) [3] aims at reconstructing a high-resolution (HR) video from a given low-resolution (LR) video sequence. Thanks to the rapid development of deep learning techniques, significant progress has been made for VSR in the past decade. Representative VSR methods include sliding window based methods like EDVR [49], recurrent based methods like BasicVSR [9] and + +![](images/7b6e1ee9655944cacfa565d1a1a9b298bcfe8dd926e46e0f9828df0ca4fb1d2c.jpg) +Fig. 1: The VSR results of 4 consecutive frames in the Sequence 026 from VideoLQ. The top row shows the VSR results by RealBasicVSR [11], the middle row shows the naive VSR results by running StableSR [48] on each frame of the sequence, and the bottom row shows the VSR results produced by our proposed MGLD-VSR method. Our method generates realistic details while achieving good temporal consistency. + +more recent transformer based methods [30]. However, most of the above works assume simple degradations (e.g., bicubic resizing, downsampling after Gaussian blur) between the LR and HR videos. As a result, the trained VSR models are hard to be generalized to real-world LR videos, whose degradations are unknown and much more complex. + +Recently, real-world VSR [11, 53, 55] has gained attention from researchers for its high potential value in practical applications such as camera-phone video enhancement and online video streaming. Different from traditional VSR tasks, which focus on super-resolving clean LR videos with simple synthetic degradation (e.g., bicubic downsampling), real-world VSR aims to enhance videos with complex and unknown degradations that often appear in real-world low-quality videos. The main challenge lies in how to effectively reproduce video details while suppressing the visual artifacts caused by unknown degradations. Existing real-world VSR algorithms mostly attempt to trade-off between details and artifacts using techniques such as Laplacian pyramid generative adversarial loss [55], input pre-cleaning [11] and hidden state attention [53], etc. Despite these efforts and progresses, their performance on real-world testing sets remains limited due to the insufficient training data and the limited model capacity. + +Since the seminal work of denoising diffusion probabilistic model (DDPM) [20], diffusion-based generative models have achieved great success in image gen + +eration. In particular, the latent diffusion models (LDMs) [42] have demonstrated compelling results not only in text-to-image generation, but also in downstream tasks such as image editing [24], inpainting [32], colorization [7], etc. Recently, researchers have also attempted to employ the powerful generative priors of LDMs for real-world image restoration tasks [48] and shown promising results. It is thus interesting to investigate whether the latent diffusion priors can be used to improve real-world VSR results. However, this is a non-trivial task because of the intrinsic randomness in the diffusion process. This randomness in turn gives rise to the temporal inconsistencies among video frames in the latent space. Such inconsistencies can be further amplified by the decoder of the LDM. As shown in the middle row of Figure 1, directly employing the LDM based image super-resolution algorithm [48] for VSR will lead to inconsistent details across consecutive frames, disturbing the visual perception of the reproduced video. + +To address the above-mentioned challenges, we propose a Motion-Guided Latent Diffusion (MGLD) model for real-world VSR in this paper, aiming to produce high-quality HR video sequence with good temporal consistency. We propose to incorporate the motion dynamics of input LR video into the generation of SR outputs. This can be done through a guided diffusion process where we can generate conditional samples from a guidance model based on its score function [14, 44]. In particular, we calculate the optical flows among adjacent LR frames, and use the calculated optical flows to warp the latent features of each frame to align the features of neighboring frames. The $l_{1}$ norm of the warping error is employed as the motion-guided loss, and its gradients can be incorporated in the sampling process to update the latent features. + +To overcome the inconsistent details generated by the original decoder of LDM, we insert temporal modules to the decoder and fine-tune it with the ground-truth HR video sequences. We also introduce an innovative sequence-oriented loss to guide the decoder to improve the continuity of the details. As depicted in the bottom row of Figure 1, our proposed MGLD-VSR approach could mitigate the discontinuity of generated details while maintaining natural and visually appealing texture recovery. + +The contributions of this paper are summarized as follows. First, we propose a diffusion sampling process based on motion-guided loss, allowing the temporal dynamics of the input frame to be used in generating temporally consistent latent features. Second, we propose a temporal-aware sequence decoder, along with two sequence-oriented losses, to further enhance the continuity of generated videos. Our proposed MGLD-VSR model achieves highly competitive real-world VSR results, exhibiting perceptually much more realistic details with fewer flickering artifacts than existing state-of-the-arts. + +# 2 Related Work + +Video Super-Resolution. The recent advances in VSR research largely attribute to the rapid development of deep-learning technologies [6, 9, 22, 30, 49]. Based on the paradigms, existing deep-learning based VSR algorithms could + +be roughly classified into two categories. The first category performs VSR in a multiple-input single-output (MISO) fashion. Typical algorithms include DUF [22], EDVR [49], and MuCAN [28], etc. The MISO algorithms are usually sliding-window based, which aggregate information from neighboring frames to generate the super-resolved result of the center frame. The other category of methods adopt a multiple-input multiple-output (MIMO) design, such as BasicVSR [9], BasicVSR++ [10], VRT [30] and RVRT [31]. These methods take multiple LR frames as input and generate their corresponding SR results in one go. Compared with MISO, the amortized running cost of MIMO is usually much lower as it avoids a series of redundant operations. In addition, benefiting from the lower cost, the MIMO methods are able to aggregate information from more frames to improve the performance. We adopt the MIMO design in this work. + +Real-world VSR. Most existing VSR methods are trained on synthetic data with simple degradations (e.g., bicubic resizing, downsampling after Gaussian blur). Due to the synthetic-to-real gap, models trained with such data cannot perform well on real-world videos. To tackle this problem, Yang et al. [55] collected LR-HR video pairs with iPhone cameras to better model real-world degradations. While being effective in some cases, it is relatively labor-intensive and may not generalize well to videos collected by other devices. Later works usually adopt an easier solution: synthesizing more realistic training data by mixing more types of degradations, such as blur, downsampling, noise and video compression [11, 53]. With the improved synthetic training data, the trained models could achieve better super-resolution performance on real-world videos. However, there is still much distribution gap between the synthetic LR data and real-world LR videos. Instead of working on synthesizing more realistic training data, in this work we investigate how to leverage the strong generative priors of pre-trained latent diffusion models to improve the real-world VSR performance. + +Generative Prior for Image Restoration. The highly ill-posed nature of real-world restoration problem brings forward a great challenge on real-world image restoration algorithms. To boost the performance, a series of methods have introduced generative priors into the image restoration process. Early methods usually adopt pre-trained generative adversarial networks (GANs) as the generative prior [8,19,34,46,50]. Although effective, the priors adopted in these methods are usually tailored for specific tasks (e.g., face image restoration), limiting their capability in handling more general natural image restoration problems. Considering the higher generality of diffusion models than GANs in natural image generation, recently researchers have started to investigate how to leverage generative diffusion priors for image restoration, and demonstrated interesting results [17, 48]. In this work, we investigate the possibility of applying latent diffusion priors to real-world VSR, and develop effective techniques to tackle the flickering issue and improve the temporal consistency of generated videos. + +Diffusion for Video Synthesis. Recently, diffusion based models have demonstrated its superiority in the video synthesis [4,5,18,43]. As one of the pioneer works, Make-a-Video [43] leverages a pretrained text-to-image (T2I) model and build a text-to-video (T2V) model by inserting temporal modules into the + +![](images/ce3187b7ca565a36504f90fd37e72cbb09efd9510150185681d3ea2ce3a260c7.jpg) +MGLD Sampling Pipeline + +![](images/02c68e471630b2d143c8f181d77ecc91c3ea73b81ecd4f9887593d44e420770c.jpg) +Update with Motion Guidance + +![](images/4eb625b1717bda730056a1a487b4e67015f21241eb72a978c4a2228d3126a7e7.jpg) +Temporal-aware Sequence Decoder +Fig. 2: Overview of the proposed MGLD-VSR framework for real-world VSR. We first estimate the forward and backward optical flow with a pre-trained optical flow estimation network and then employ a motion-guided loss to update the latent diffusion sampling process. The motion-guided loss is computed by summing the forward and backward masked warping error of the latent sequence at each time-step. After a number of $T$ sampling steps, we obtain the generated latent sequence and feed it into the temporal-aware sequence decoder to reconstruct the VSR sequence. + +T2I model and designing a cascaded architecture. VideoLDM [5] adopts a latent diffusion paradigm and train the T2V model on large-scale paired text-video datasets, greatly improving the generation results. Stable Video Diffusion [4] further improve the results by conducting training on more carefully constructed datasets and introducing more advanced training paradigm. In this work, we leverage the diffusion prior and build a powerful real-world VSR model, which aims to tackle the ill-posedness of real-world VSR problem with diffusion prior and obtain temporally consistent VSR results with rich details. + +# 3 Methodology + +Given an LR video sequence of $N$ frames $I = \{I_1, I_2, \ldots, I_N\}$ , our goal is to reproduce an HR video sequence $\hat{I} = \{\hat{I}_1, \hat{I}_2, \ldots, \hat{I}_N\}$ with enhanced visual quality while maintaining the content consistency across adjacent frames. Inspired by the recently developed diffusion-based image restoration methods [17, 48], we propose to harness the generative priors of the pre-trained SD model [42] to improve the real-world VSR performance. First, we propose a motion-guided diffusion sampling process, where the temporal dynamics from the LR sequence are introduced into the diffusion sampling process to enhance the consistency of latent features across frames. Then, we further improve the quality of generated details by designing a temporal-aware sequence decoder and fine-tuning + +it with a video-oriented loss. The overall framework of our algorithm, namely motion-guided latent diffusion (MGLD), is illustrated in Figure 2. + +# 3.1 Latent Diffusion Model + +Diffusion models [21] are a class of generative models which transform Gaussian noise into data samples through an iterative reverse Markovian process. This process can be described by: $q(x_{t}|x_{t - 1})\coloneqq \mathcal{N}(x_{t},(1 - \beta_{t})x_{t - 1},\beta_{t}I)$ , where the Gaussian noise is iteratively applied to the sample $x_{t - 1}$ at time-step $t - 1$ to produce $x_{t}$ , guided by a variance scheduler $\{\beta_t\}_{t = 1}^T$ along the diffusion process. The key objective of diffusion models is to train a denoising U-Net $\epsilon_{\theta}$ parameterized by $\theta$ and conditioned on the time-step $t$ and an optional signal $y$ . Given the noise target $\epsilon_{t}$ , the diffusion loss can be defined as: $L(\theta) = \mathbb{E}_{t\sim U(1,T),\epsilon_t\sim \mathcal{N}(0,I)}\| \epsilon_t - \epsilon_\theta (\mathbf{x}_t;t,y)\| _2^2$ . + +Latent Diffusion Models (LDMs) [42] have made a remarkable impact in the field of image generation. These models employ Variational Autoencoders (VAE) to map the images into a latent space, enabling them to be trained on large-scale text-image datasets. This equips LDMs with powerful prior knowledge about natural images, which can be leveraged by downstream tasks such as image restoration and super-resolution [48] to produce images with clear details and high-quality content. However, in the case of VSR, the stochastic nature of diffusion process often leads to the temporal inconsistencies in the restored video sequence. To solve this issue, we propose a MGLD model that achieves high-quality VSR by utilizing guidance during the sampling process and incorporating temporal awareness in the decoder. + +# 3.2 Motion-guided Diffusion Sampling + +From a score-based perspective [44], the sampling process of a diffusion model involves the application of Langevin Dynamics to both gradient of the data distribution $\log p(\mathbf{z})$ and a guidance model $g = G(z)$ , which can be a network [15,38] or an energy model [12]: + +$$ +\nabla_ {\mathbf {z}} \log p (\mathbf {z}, g) = \nabla_ {\mathbf {z}} \log p (\mathbf {z}) + \nabla_ {\mathbf {z}} \log p (g | \mathbf {z}). \qquad (1) +$$ + +In this work, we introduce a technique that utilizes the temporal dynamics in low-resolution input videos to create latent features that maintain temporal consistency, thereby enhancing VSR results. Specifically, we incorporate an innovative motion-guided module within the sampling process to measure the warping errors of latent features across frames. Given the LR frames, we first compute their optical flow with $\mathcal{F}$ [47] and downsample the flow map to the dimension of latent features. Given the forward flow $O_{f}$ and backward flow $O_{b}$ , we warp the latent features to their neighbouring frames and compute the accumulated warping error along both directions: + +$$ +E (z) ^ {t} = \sum_ {i = 1} ^ {N - 1} \| \left(\operatorname {W a r p} \left(z _ {i} ^ {t}, O _ {b, i}\right) - z _ {i + 1} ^ {t})\right) \| _ {1} + \sum_ {i = 2} ^ {N} \| \left(\operatorname {W a r p} \left(z _ {i} ^ {t}, O _ {f, i - 1}\right) - z _ {i - 1} ^ {t})\right) \| _ {1}. \tag {2} +$$ + +We empirically find that the presence of occlusion will pose adverse impact on flow estimation, which in turn causes distraction in the sampling process and leads to artifacts in the final VSR results. To solve this issue, we further estimate the occlusion mask $\mathcal{M}$ in each frame [33] and ignore the contributions in the occluded region. Given the variance $\sigma_t^2$ of noise scheduler at step $t$ , the motion-guided sampling process at sampling step $t$ can be written as: + +$$ +\hat {z} ^ {t} \leftarrow \operatorname {D D P M} \left(z ^ {t + 1}, t\right) - \eta \sigma_ {t} ^ {2} \nabla_ {z} \left(\mathcal {M} \circ E (z) ^ {t}\right), \tag {3} +$$ + +where the DDPM [21] scheduler is applied for diffusion reverse (denoising) process and $\eta = 10$ is a weighting factor to control the guidance strength. After $T$ iterations, the final denoised latent sequence $z^0$ is fed into the VAE decoder to obtain the final video sequence. + +# 3.3 Temporal-aware Decoder Fine-tuning + +With the proposed motion-guided diffusion sampling, the VSR model could better exploit the temporal dynamics to generate temporally more consistent HR videos. However, since the guidance is performed in the low-resolution latent space, which is 8 times smaller in both height and width than the output image space, the video reconstruction through a VAE decoder may still lead to temporal discontinuities in the generated details of VSR output. Thus, we introduce a temporal-aware sequence decoder and fine-tune it with ground-truth video sequences to improve the smoothness and accuracy of output videos. + +As shown in Figure 2, we build our sequence decoder on top of the pretrained VAE decoder by incorporating 1D convolutions along the temporal dimension. This design facilitates interleaved spatial-temporal interactions and enables our module to restore details with high continuity while incurring minimal computational cost. We further incorporate the information $e$ from the VAE encoder through a Controllable Feature Warping (CFW) module [48, 60] for better restoration and generation results. To leverage the restorative power of the pretrained VAE, we freeze their original spatial blocks and only update the parameters in the temporal convolution and CFW module. + +We first employ the $\ell_1$ loss and perceptual loss [23, 58] as the reconstruction loss $L_{\mathrm{recon}}$ for each frame, then utilize the GAN-loss and frame difference loss that tailored for video sequences. The frame difference loss $L_{\mathrm{diff}}$ calculates the difference between consecutive predicted frames $\hat{I}$ and ground truth frames $I^{gt}$ , which is defined as: + +$$ +L _ {\text {d i f f}} = \sum \| \left(\hat {I} _ {i + 1} - \hat {I} _ {i}\right) - \left(I _ {i + 1} ^ {g t} - I _ {i} ^ {g t}\right) \| _ {1}. \tag {4} +$$ + +It is noteworthy that diffusion models typically generate strong structures. To make the super-resolved video look more natural, we need to apply consistency constraints to those structured regions. We hence introduce a structure weighted consistency loss $L_{\mathrm{swc}}$ : + +$$ +\begin{array}{l} L _ {\mathrm {s w c}} = \sum \| \mathcal {M} \circ (\mathcal {W} _ {i + 1} (\operatorname {W a r p} (\hat {I} _ {i}, O _ {b, i} ^ {g t}) - \hat {I} _ {i + 1})) \| _ {1} \tag {5} \\ + \sum \| \mathcal {M} \circ (\mathcal {W} _ {i - 1} (\operatorname {W a r p} (\hat {I _ {i}}, O _ {f, i} ^ {g t}) - \hat {I} _ {i - 1})) \| _ {1}. \\ \end{array} +$$ + +where $O_{f}^{gt}$ , $O_{b}^{gt}$ denote the forward and backward flow computed on the ground-truth sequence. This loss helps the network to generate consistent details by modulating the loss through a weighted map $\mathcal{W} = (1 + wS)$ computed on the ground-truth frames, where $w = 3$ is a weighting factor and $S$ is a structure (edge) map using the Sobel operator. Occlusion masking is also adopted. Given the GAN-based loss $L_{\mathrm{GAN}}$ and reconstruction loss $L_{\mathrm{recon}}$ , the total loss for decoder fine-tuning is defined as: + +$$ +L _ {\text {v i d e o}} = L _ {\text {r e c o n}} + \alpha L _ {\text {d i f f}} + \beta L _ {\text {s w c}} + \gamma L _ {\text {G A N}}, \tag {6} +$$ + +where $\alpha, \beta$ and $\gamma$ are empirically set to 0.5, 0.5 and 0.025. + +# 4 Experiments + +# 4.1 Experimental Settings + +Implementation Details. The training of MGLD-VSR consists of two main stages. In the first stage, we fine-tune the denoising U-Net for latent space diffusion. The weight of the denoising U-Net is initialized from Stable Diffusion V2.1. We insert 1D temporal convolution into the U-Net to help temporal dynamic modeling. We then fix the weight in the denoising U-Net of SD model and train the conditioning and temporal modeling modules. For conditioning, we adopt a small time-aware encoder to encode LR condition and inject it into the denoising U-Net with spatial feature transform operation as in [48]. The training batch size, sequence length and latent image size are set to 24, 6 and $64 \times 64$ , respectively. In the second stage, we first generate the clean latent sequence with the proposed motion-guided diffusion sampling process, then fine-tune the temporal-aware sequence decoder with the LR sequence, the generated latent sequence and the HR sequence. We fix the VAE decoder and insert temporal modeling layers and CFW modules for training. The training batch size, sequence length and image size are set to 4, 5 and $512 \times 512$ , respectively. + +Our models are trained on 4 NVIDIA A100 GPUs with the PyTorch [40] framework. Adam [26] is chosen as the optimizer. During inference, due to memory constrains, we divide the LR video into multiple sequences to run. To handle output with arbitrary resolution, we apply a progressive patch aggregation sampling algorithm [48]. For each sequence, we run the sampling for 50 steps. + +Training and Testing Datasets. As in previous works [11, 49, 53], we merge the training set and validation set of REDS [37] in training and leave 4 sequences for testing (REDS4). We follow the degradation pipeline of RealBasicVSR [11] to synthesize the training sequence pairs, which involves blur, noise, downsampling and compression degradations. We evaluate our method on both synthetic and real-world datasets. To synthesize LR-HR testing sequence pairs, we apply the RealBasicVSR degradation pipeline on several widely used VSR datasets, including REDS4 [37], UDM10 [56] and SPMCS [45]. REDS4 contains 4 video sequences, each having 100 frames. UDM10 consists of 10 sequences, each + +having 32 frames. SPMCS has 30 sequences, each having 31 frames. For real-world dataset, we adopt VideoLQ [11] for testing, which contains 50 real-world sequences with complex degradations. + +Evaluation Metrics. On synthetic datasets, we could adopt full-reference metrics to evaluate different real-world VSR methods. The perceptual quality metrics LPIPS [58] and DISTS [16] are selected. Fidelity-oriented metrics PSNR and SSIM are also calculated for reference. Apart from image-based metrics, the video quality assessment metric VMAF [29] is also employed. + +On real-world datasets, as ground truths are not available, we compute the no-reference image quality metrics NIQE [36] and BRISQUE [35] for evaluation. In addition, we also include a deep learning based image-based metric MUSIQ [25] and a state-of-the-art video quality assessment metric DOVER [52]. + +# 4.2 Experimental Results + +To demonstrate the effectiveness of the our MGLD-VSR algorithm, we compare it with seven state-of-the-art methods, including three real-world image super-resolution models (BSRGAN [57], RealESRGAN [51], StableSR [48]), one classic VSR model (RVRT [31]), and three blind or real-world VSR models (DBVSR [39], RealVSR [55], RealBasicVSR [11]. + +Quantitative Comparison. We first show the quantitative comparison on the synthetic datasets and real-world video benchmarks in Table 1. As shown from Table 1, our method achieves the best results in terms of full-reference perceptual based metrics LPIPS and DISTS on all synthetic test datasets, indicating that the proposed method is able to reconstruct high-quality plausible details from sequences degraded with complex degradations. Although methods like DBVSR may achieve better performance in terms of PSNR or SSIM on some datasets, they tend to generate blurry results, as demonstrated in the LPIPS and DISTS metrics. In terms of VMAF, our model achieves the best or second best results. As for real-world VSR dataset VideoLQ, our methods achieve the best performance in terms of NIQE, BRISQUE, and DOVER and the second best results in terms of MUSIQ, which reflects its strong capability to enhance the real-world videos and generate realistic details and textures. + +Qualitative Comparison. To demonstrate the effectiveness of MGLD-VSR, we visualize the VSR results on synthetic datasets in Figure 3 and the results on real-world VideoLQ in Figure 4. One can see that MGLD-VSR can remove complex spatial-variant degradations and generate realistic details, outperforming other state-of-the-art real-world VSR algorithms. On synthetic datasets, as shown in Figure 3, MGLD-VSR can well reconstruct the structure while generating plausible details under complex degradations. On real-world videos, as shown by the example of sequence 008 in Figure 4, MGLD-VSR reconstructs clearer characters and patterns than competing methods. For the example of sequence 013, MGLD-VSR not only reconstructs the structure of the windows, and at the same time generates plausible exterior structure of the brick wall. More visual comparisons can be found in the supplementary materials. + +Table 1: Quantitative comparison of different VSR models evaluated on popular VSR benchmarking datasets. Bold and underline represent the best and second best score, respectively. For synthetic datasets (REDS4, UDM10, SPMCS), we adopt full-reference metrics to evaluate different methods. For real-world video datasets (VideoLQ), we adopt no-reference metrics to compare different methods. + +
DatasetsMetricsBicubicRVRTBSRGANRealESRGANStableSRDBVSRRealVSRRealBasicVSROurs
REDS4PSNR ↑22.1322.2922.0221.3422.0422.2921.7822.3522.46
SSIM ↑0.57300.57590.57020.56470.56830.57600.56630.58460.5723
LPIPS ↓0.62770.62210.48020.45090.45530.62210.62670.41900.3776
DISTS ↓0.36670.35540.21180.17500.18450.35560.36200.15810.1151
VMAF ↑2.12554.800526.241333.995526.69532.60752.590334.988434.3214
UDM10PSNR ↑25.7826.1825.2924.9425.5826.1425.0925.5625.99
SSIM ↑0.75590.76290.73230.74390.75920.76350.75240.74480.7548
LPIPS ↓0.44410.44130.42170.38840.36840.43980.43300.38970.3491
DISTS ↓0.27300.25750.19260.16510.14970.25850.26080.16370.1369
VMAF ↑5.464710.172431.087436.915833.82647.20486.007938.263839.3892
SPMCSPSNR ↑22.2822.5122.1221.6622.2622.5022.0322.4122.66
SSIM ↑0.56910.57530.57300.56940.58440.57640.57380.57360.5960
LPIPS ↓0.55720.54710.45090.44350.43210.54580.54870.44200.3990
DISTS ↓0.35930.34150.23070.21500.23510.34350.34790.21300.1934
VMAF ↑0.96565.353322.482827.693521.80603.75762.702427.451928.7032
VideoLQNIQE ↓9.13196.58884.21044.20364.44426.75046.33313.69153.5346
BRISQUE ↓72.740562.176825.240029.842826.309261.189757.276524.548421.9839
MUSIQ ↑21.678928.420452.221349.838047.804529.003924.218955.969152.7812
DOVER ↑0.36770.49070.71840.73090.71080.49980.52020.74010.7481
+ +Temporal Consistency. Compared with single image super-resolution, one major aspect of VSR is the temporal consistency of the generated videos. Following most previous works, in this paper, we adopt the average warping error (WE) [13,27] of the sequence to quantitatively measure the temporal consistency: $\mathrm{WE} = \frac{1}{N - 1}\sum_{i = 1}^{N - 1}\| \hat{I}_{i + 1} - \mathrm{Warp}(\hat{I}_i,O_{b,i})\| _1$ , where $\hat{I}_i$ is the predicted frame and $O_{b,i}$ is the estimated backward optical flow [54] from ground-truth video. On real-world video datasets without ground-truth, the optical flow is computed on the predicted frames. + +It is worth mentioning that, despite the wide use of WE for temporal consistency evaluation, it may not be able to faithfully reflect the real human perception on the video. For example, WE could be easily attacked by blurring the sequence, resulting much higher WE scores but lower video quality. The WE scores of competing methods are given in Table 2. One can see that our diffusion prior based method MGLD-VSR is not good at this metric because it tends to generate more details and textures, which are not favourable by WE. Nonetheless, the WE of MGLD-VSR is better than StableSR, which is also driven by the latent diffusion priors. To better showcase the effectiveness of our method, we visualize the temporal profiles of the VSR results produced by different algorithms in Figure 5. One can see that MGLD-VSR reconstructs more accurate details while maintaining good temporal consistency. + +User Study. We also conduct a user study to further examine the effectiveness of the proposed algorithm in comparison with RealBasicVSR [11] and StableSR [48]. The study is conducted in pairs: given the LR video, the subject + +![](images/1a21be49bca64884ee81ee330d398fa5ebaac7c1c601e8e5fe8736961f50206a.jpg) +Sequence 011 from REDS4 + +![](images/8a42fa41971422baccc438633c04648760f223367e810a1f60bbd666b170b7ba.jpg) +Bicubic + +![](images/b3b26876e14df09051d32da7e9b1be2577bbaa8595e4d090fb01bed9148200c5.jpg) +Real-ESRGAN + +![](images/09490f3ead1d67dfa12550c5c9f9aea7dd6bac950b5555c0a4f00199911b956d.jpg) +StableSR + +![](images/7378eabc63867bcda96fb3c85fcdabed4ad6c519f5ec49d655eb379bbd761c77.jpg) +RealBasicVSR + +![](images/159c2189dcbabd27947074e6a56869999ff598cd02d109dc6bbcb69312fd269c.jpg) +Ours + +![](images/6a8edb8c0dbd6417bf15337a10049a46e926a2bff6602e810f7d5abe14c89214.jpg) +GT + +![](images/26f5f2891c7c911908c82f8aaa13f12b673a13cc57814840ab5158e88e1df629.jpg) +Sequence band from UDM10 + +![](images/1bfc860ed73a6becfed9f2e5255c7facc155d14f32eb369ccbb13c826dd7ea47.jpg) +Bicubic + +![](images/9805048d8ad0be3e6f49ffbe4a4794d0dad08f16ea7593523fb3898f170e054b.jpg) +Real-ESRGAN + +![](images/a4d544fc5392c0a3d07024189fe4094b2a74177237830c5ce3f75aa91ae924d1.jpg) +StableSR + +![](images/e81829d39c8daae6cd1402e573c7aadc63b3dbe8e6578d69e93e8df8e7f1f16d.jpg) +RealBasicVSR + +![](images/22ea8ba186e6d2f8f77a1abd7a78e8be7dd12d563e258d7dae4768bfa585194c.jpg) +Ours + +![](images/dc9b5fcbd53908c7cc836cbb2fff032d8a3f2ae99f7e8dc560a712ee58426030.jpg) +GT + +![](images/e892820ceaf4c9765a5508bf22ee3de820fb930065cd95e1b589ef69247898b2.jpg) +Sequence 000002 from SPMCS + +![](images/d9f9eaf4732ee295736f2928923672cd5ef09e5fe4f39767958d37fcdd0816ad.jpg) +Bicubic + +![](images/2d0c996430a220b3c1625fa3804172d3aea2e0cf06b8c14460137cc8597e3e1c.jpg) +Real-ESRGAN +Fig. 3: Qualitative comparison on synthetic datasets (REDS4, UDM10, SPMCS) for $\times 4$ VSR. (Zoom-in for best view.) + +![](images/e001fbe85d63fd5ddc80b8bf5b0989087430cca1b6d75a43733b4c7bd05cb1ca.jpg) +StableSR + +![](images/9f4897da176e1b91f6ccaa78375773acaf6f1037050de6131a49339d3f6ec88d.jpg) +RealBasicVSR + +![](images/b4fbbdb03af3b88df08c54ec8f8ae09d3f61761a3abb01106b13ff72b1f4bb77.jpg) +Ours + +![](images/a9a1a880f31d483a04caf95d4e88f63e9edc6cd8380e9f32e5de11bb85587e0b.jpg) +GT + +Table 2: Temporal consistency comparison of competing VSR methods by using WE $\downarrow$ . Bold and underline represent the best and second best scores, respectively. + +
DatasetsBicubicRVRTBSRGANRealESRGANStableSRDBVSRRealVSRRealBasicVSROurs
REDS48.576.5010.2511.9111.818.518.6110.0910.08
UDM103.103.064.824.994.803.042.754.334.55
SPMCS3.013.585.187.124.623.584.182.723.87
VideoLQ5.486.387.727.649.335.985.746.767.56
+ +is asked to choose the better one from two HR videos obtained by different methods, considering both the reconstruction quality and the temporal consistency. As shown in Figure 6, our method gains more votes in the competitions. + +# 4.3 Ablation Study + +In this section, we conduct a series of analyses on the proposed framework, including its Motion-guided Diffusion Sampling (MDS) module and Temporal-aware Sequence Decoder (TSD) module, and employ the VideoLQ dataset for experiments. We adopt WE to measure temporal consistency and no-reference metrics NIQE, BRISQUE, MUSIQ and DOVER to measure perceptual quality. + +Different Components of MGLD-VSR. As can seen from Table 3, the introduction of MDS improves all the metrics over the baseline (i.e., without using MDS and TSD), suggesting that MDS helps not only in reducing the temporal differences (i.e., WE), but also enhancing the perceptual quality of the reconstructed video sequence. + +The incorporation of MDS aids in maintaining perceptual continuity by ensuring that the restored frames could transition smoothly from one to the next. This contributes to more natural and visually pleasing VSR results. The incorporation of TSD also yields better results across all the metrics compared to + +![](images/66d1aed951c44d33847292360718af7dba651bcb9d8c3e1d3693f08aeb3682b4.jpg) +Sequence 008 from VideoLQ + +![](images/a696e082a10641f9b191d58ce98491f9830d067de57894e4b0de284ebb11a8b5.jpg) +Bicubic + +![](images/ca382d079fecafd21112c6abe1bb792379ae77338d04df50b17b97742b2aa06e.jpg) +BSRGAN + +![](images/0bcebf50697bdf5a9e9916fd0f55415b6198bbb3cfd2bb5f2c4890f5ae6b2ec1.jpg) +Real-ESRGAN + +![](images/a3b03c9d59434da6ff64abd589aca844099c8b7f96b05c19bb17c73f465fb1dc.jpg) +StableSR + +![](images/5feba6f652cf468516ad5f3f1a0f0cc68bb7a72c6d7efcdeb7bd915a6def1d3a.jpg) +Sequence 013 from VideoLQ + +![](images/489d4622053853c6083303f2bc9794fa27e8414fa17759937c74b9f885f36d02.jpg) +Bicubic + +![](images/48127e6a671b247d61e8adb6788c8ca1d73a28f0628d6b83c36fb756179b2ac2.jpg) +RealVSR + +![](images/712dcec4f5b3baa9c86dec11dd36c6a3c77f6cfd3eb20942b4881645e8526579.jpg) +BSRGAN + +![](images/bc3cf4999da48aa0574ba136619ec6f812d21b4e7f1bf7d02e54ff136187f4ce.jpg) +DBVSR + +![](images/da61c9ee62dab88f5a59bff79bd7ff741c8631ae9fcf3858f10d20ab435faae2.jpg) +Real-ESRGAN + +![](images/ed52997a91f5dd302afdfa831f62fcc94bf6989c4ced72f051d355c4cd485e60.jpg) +RealBasicVSR + +![](images/c0d3c2ffa3de0e90eaff7bfa46f74083b1dc6876dd29fac78000ffe9ed24f811.jpg) +Ours + +![](images/b46a49a679e57918d39e6bd9e82df7a38e68716fbee0f3df5d93ad6a1d12d718.jpg) +StableSR + +![](images/6ace3dbd55c52f74aebd9e603790d6d9a06104909e473232527b96b4d066c890.jpg) +Ours + +![](images/76cd62080f30b42c8b4c191efbab7bdb3ade7c1707e2a7e67eb2f87edbf7b4bd.jpg) +Fig. 4: Qualitative comparison on the real-world video dataset (VideoLQ) for $\times 4$ video SR. (Zoom-in for best view.) + +![](images/66589e67a61bd62f29bbb514747cd0e51a37359d6c51a3323433769bd1bc7f4e.jpg) + +![](images/622ca7949d5e8f5666ae9765a52dc45c2f7e9e9c74c5f4bd23bbe66c68f0987a.jpg) + +![](images/7dcd77593b048342fd1c8e8d7622d1850a294beba46d7e2ff98e4a0311a9e869.jpg) + +![](images/f07208c77bcd272c498bb30726320220336bd96c8e0d2262a9e771af03eb22ad.jpg) + +![](images/d3bc162df8c76ab72a924e375dffdd8b619f1a618ce5c6a1fce49479cfa9ec00.jpg) + +![](images/26afcc10b95caa60126528e6429bbdb47e5eef4b8e06bf4829efcb64be8a2683.jpg) + +![](images/7134fc49cba0d0d657c3cb94b509b31a6d10d24d3bdff82a1fc68a82e1c0c756.jpg) + +![](images/89927672146a443faa86ceddbacb6bbbd9730af44148e53977bfb4bfeb605361.jpg) + +![](images/c76c5094e3a20fd31a9738da9bfa69bb14819e7a30d4c0a93d8d1ac1e29d6941.jpg) + +![](images/c6e99f77deaa80429e501bac068865126af5ef8663fcb467a1060c4c8637e1a9.jpg) + +![](images/efd09aca113e4c94e9574cabfc958431a893fbbcbffae3f92b9cfc394c27d1b7.jpg) + +![](images/fee8658ada73b209db1beb2c7c6f548706e5001e4849205f645fb8c996656898.jpg) + +![](images/dec93126028840fd97e1dfa59a2a5e7130cba18b32f864374f615b07f351d052.jpg) + +![](images/b883b80d49db29bf4b07b2bb73b5c39b623578280af3134b3458c93b8e302c3c.jpg) +Fig. 5: Temporal profiles of competing real-world VSR methods. +Fig. 6: User study on 20 videos from VideoLQ evaluated by 10 subjects. MGLD-VSR gains more votes than competing methods. + +![](images/2714013b4093a5f8bc2ffd51805f1956e184ee627ec42971ea1746429b91f188.jpg) + +Table 3: Ablation study on the different components of the proposed algorithm. "MDS" refers to Motion-guided Diffusion Sampling and "TSD" refers to Temporal-aware Sequence Decoder. + +
MDS TSDWE ↓NIQE ↓BRISQUE ↓MUSIQ ↑DOVER ↑
✓9.30694.109825.324951.56770.7412
8.98483.987624.084052.41430.7435
✓8.36793.628122.080652.68740.7468
✓7.54783.534621.983952.78120.7481
+ +![](images/774778f79401f7818e2b2b94abf61de0c3f753e8e0c90daa247afcebf06dd8f2.jpg) +Sequence 002 from VideoLQ + +![](images/1c67da428bce5bf1d85d6d26285eacfd3f1ad94446e5a853cdeb3029e3ce45fc.jpg) +0 +Fig. 7: Impact of guidance strength during sampling. The numbers below the cropped patches indicate the motion guidance weighting factor $\eta$ . + +![](images/6e5f2e88a23d0c8f1bec715d642579b62990a6d5df38361d7d2696f0ea29158b.jpg) +10 + +![](images/57e87a8e5d2af8edab1741764e385b45164c7b5265e43daa99fdcdd7e87ddb6c.jpg) +20 + +![](images/4d11dc341d7114a89d0f12de62381f3f6b01066fec2f3be970f4353457b6912d.jpg) +30 + +![](images/5890a77f2e8d82ee4decff05dcb01ccb6098d1bfee84b750dfaaf75858d6c340.jpg) +40 + +the baseline. This occurs because TSD is capable of extracting richer contextual information from neighboring frames. When both MDS and TSD are used, the results show further improvements. The full MGLD-VSR approach achieves the best scores on all the metrics, indicating the synergy of the two proposed components for real-world VSR. + +Impact of Flow Network. As shown in Table 4, the flow estimation network in motion guided sampling does not show big influence on the results. This might because the motion guided sampling is applied in low-resolution space. + +Impact of Guidance Strength. During guided sampling process, we introduce a weighting factor to control the guidance strength. As shown in Figure 7, a moderate guidance strength could reduce the artifacts and improve the temporal consistency of the generated results. A large guidance strength will blur the results and degrade the reconstructed details. To balance the generation power and temporal consistency, we set the weighting factor $\eta = 10$ . + +The Design of VAE Decoder. To validate the effectiveness of the proposed architecture and loss design for the VAE decoder, we conduct a series of study and the results are shown in Table 5. As can be seen from the table, the introduction of the adaptive temporal modeling module improves the temporal consistency as well as perceptual quality (measured by no-reference metrics), and incorporating the feature from encoder to decoder during decoder fine-tuning further improves the results. By introducing two sequence based losses during fine-tuning, we improve the temporal consistency of the results while maintaining or improving the perceptual quality. + +Table 4: Ablation study on the flow network in motion guided diffusion sampling. + +
Flow NetworkWE ↓NIQE ↓MUSIQ ↑DOVER ↑
SpyNet [41]7.67343.554552.69440.7478
RAFT [47]7.54783.534652.78120.7481
+ +Table 5: Ablation study on the different designs of VAE decoder. "TM" refers to the temporal modeling module. "CFW" refers to the controllable feature warping module that injects VAE encoder feature into VAE decoder. + +
ModelsTM CFWWE ↓ NIQE ↓ MUSIQ ↑ DOVER ↑LossesLdiffLswcWE ↓ NIQE ↓ MUSIQ ↑ DOVER ↑
(a)9.86704.067150.27070.7360(a)7.94803.535952.7894
(b)✓8.68463.845152.54080.7431(b)✓7.75793.540952.5648
(c)✓8.98423.923652.43860.7426(c)✓7.76883.537652.7456
(d)✓7.54783.534652.78120.7481(d)✓✓7.54783.534652.7812
+ +# 4.4 Limitations + +Our proposed MGLD-VSR method leverages the diffusion prior from pre-trained SD model, aiming to achieve perceptually realistic and temporally consistent VSR results. Due to the scale of pre-trained SD model, it does not have advantages in terms of number of parameters and inference time compared with those non-diffusion VSR methods (comparisons can be found in the supplementary materials). In addition, since MGLD-VSR will generate more visual details than those non-diffusion VSR methods, it does not have advantages in terms of temporal consistency metrics either. In the future, we can resort to model distillation [2] and more advanced diffusion sampling [1,59] techniques to improve the efficiency of MGLD-VSR, and resort to the recently developed video synthesis models [4] to improve the temporal consistency. + +# 5 Conclusion + +We presented a motion-guided latent diffusion algorithm, namely MGLD-VSR, for high-quality and temporally consistent real-world VSR. Our method exploited the strong generative prior of pre-trained latent diffusion model to restore video sequence with realistic details. To ensure the content consistency among frames, we developed a motion-guided sampling mechanism, where the motion dynamics of input video were incorporated in the diffusion sampling process. To further improve the continuity of generated details, we introduced a temporal module into the decoder, and fine-tuned it by high quality video sequences with delicately designed sequence-based losses. The fine-tuned sequence decoder mitigated the flickering artifacts and further improved the perceptual quality. Our proposed MGLD-VSR achieved state-of-the-art visual results on the widely adopted real-world VSR benchmarks, validating the great potentials of latent diffusion priors in enhancing real-world video quality. + +# References + +1. Lu et al., C.: Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. NeurIPS 35, 5775-5787 (2022) +2. Luo et al., S.: Latent consistency models: Synthesizing high-resolution images with few-step inference. arXiv preprint arXiv:2310.04378 (2023) +3. Bishop, C.M., Blake, A., Marthi, B.: Super-resolution enhancement of video. In: AISTATS (2003) +4. Blattmann, A., Dockhorn, T., Kulal, S., Mendelevitch, D., Kilian, M., Lorenz, D., Levi, Y., English, Z., Voleti, V., Letts, A., et al.: Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127 (2023) +5. Blattmann, A., Rombach, R., Ling, H., Dockhorn, T., Kim, S.W., Fidler, S., Kreis, K.: Align your latents: High-resolution video synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22563-22575 (2023) +6. Caballero, J., Ledig, C., Aitken, A., Acosta, A., Totz, J., Wang, Z., Shi, W.: Real-time video super-resolution with spatio-temporal networks and motion compensation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4778-4787 (2017) +7. Carrillo, H., Clément, M., Bugeau, A., Simo-Serra, E.: Diffusart: Enhancing line art colorization with conditional diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3485-3489 (2023) +8. Chan, K.C., Wang, X., Xu, X., Gu, J., Loy, C.C.: Glean: Generative latent bank for large-factor image super-resolution. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 14245-14254 (2021) +9. Chan, K.C., Wang, X., Yu, K., Dong, C., Loy, C.C.: Basicvsr: The search for essential components in video super-resolution and beyond. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4947-4956 (2021) +10. Chan, K.C., Zhou, S., Xu, X., Loy, C.C.: Basicvsr++: Improving video superresolution with enhanced propagation and alignment. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5972-5981 (2022) +11. Chan, K.C., Zhou, S., Xu, X., Loy, C.C.: Investigating tradeoffs in real-world video super-resolution. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5962-5971 (2022) +12. Chen, M., Laina, I., Vedaldi, A.: Training-free layout control with cross-attention guidance. arXiv preprint arXiv:2304.03373 (2023) +13. Chu, M., Xie, Y., Mayer, J., Leal-Taixe, L., Thuerey, N.: Learning temporal coherence via self-supervision for gan-based video generation. ACM Transactions on Graphics (TOG) 39(4), 75-1 (2020) +14. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. In: Ranzato, M., Beygelzimer, A., Dauphin, Y., Liang, P., Vaughan, J.W. (eds.) Advances in Neural Information Processing Systems. vol. 34, pp. 8780-8794. Curran Associates, Inc. (2021), https://proceedings.neurips.cc/paper_files/paper/2021/file/49ad23d1ec9fa4bd8d77d02681df5cfa-Paper.pdf +15. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. In: Ranzato, M., Beygelzimer, A., Dauphin, Y., Liang, P., Vaughan, J.W. (eds.) Advances + +in Neural Information Processing Systems. vol. 34, pp. 8780-8794. Curran Associates, Inc. (2021), https://proceedings.neurips.cc/paper_files/paper/2021/file/49ad23d1ec9fa4bd8d77d02681df5cfa-Paper.pdf +16. Ding, K., Ma, K., Wang, S., Simoncelli, E.P.: Image quality assessment: Unifying structure and texture similarity. IEEE transactions on pattern analysis and machine intelligence 44(5), 2567-2581 (2020) +17. Fei, B., Lyu, Z., Pan, L., Zhang, J., Yang, W., Luo, T., Zhang, B., Dai, B.: Generative diffusion prior for unified image restoration and enhancement. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9935-9946 (2023) +18. Ge, S., Nah, S., Liu, G., Poon, T., Tao, A., Catanzaro, B., Jacobs, D., Huang, J.B., Liu, M.Y., Balaji, Y.: Preserve your own correlation: A noise prior for video diffusion models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 22930-22941 (2023) +19. Gu, J., Shen, Y., Zhou, B.: Image processing using multi-code gan prior. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 3012-3021 (2020) +20. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in neural information processing systems 33, 6840-6851 (2020) +21. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. In: Larochelle, H., Ranzato, M., Hadsell, R., Balcan, M., Lin, H. (eds.) Advances in Neural Information Processing Systems. vol. 33, pp. 6840-6851. Curran Associates, Inc. (2020), https://proceedings.neurips.cc/paper_files/paper/2020/file/4c5bcfec8584af0d967f1ab10179ca4b-Paper.pdf +22. Jo, Y., Oh, S.W., Kang, J., Kim, S.J.: Deep video super-resolution network using dynamic upsampling filters without explicit motion compensation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3224-3232 (2018) +23. Johnson, J., Alahi, A., Fei-Fei, L.: Perceptual losses for real-time style transfer and super-resolution. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14. pp. 694-711. Springer (2016) +24. Kawar, B., Zada, S., Lang, O., Tov, O., Chang, H., Dekel, T., Mosseri, I., Irani, M.: Imagic: Text-based real image editing with diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6007-6017 (2023) +25. Ke, J., Wang, Q., Wang, Y., Milanfar, P., Yang, F.: Musiq: Multi-scale image quality transformer. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5148-5157 (2021) +26. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014) +27. Lai, W.S., Huang, J.B., Wang, O., Shechtman, E., Yumer, E., Yang, M.H.: Learning blind video temporal consistency. In: Proceedings of the European conference on computer vision (ECCV). pp. 170-185 (2018) +28. Li, W., Tao, X., Guo, T., Qi, L., Lu, J., Jia, J.: Mucan: Multi-correspondence aggregation network for video super-resolution. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part X 16. pp. 335-351. Springer (2020) +29. Li, Z., Bampis, C., Novak, J., Aaron, A., Swanson, K., Moorthy, A., Cock, J.: Vmaf: The journey continues. Netflix Technology Blog 25(1) (2018) + +30. Liang, J., Cao, J., Fan, Y., Zhang, K., Ranjan, R., Li, Y., Timofte, R., Van Gool, L.: Vrt: A video restoration transformer. arXiv preprint arXiv:2201.12288 (2022) +31. Liang, J., Fan, Y., Xiang, X., Ranjan, R., Ilg, E., Green, S., Cao, J., Zhang, K., Timofte, R., Gool, L.V.: Recurrent video restoration transformer with guided deformable attention. Advances in Neural Information Processing Systems 35, 378-393 (2022) +32. Lugmayr, A., Danelljan, M., Romero, A., Yu, F., Timofte, R., Van Gool, L.: Repaint: Inpainting using denoising diffusion probabilistic models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11461-11471 (2022) +33. Meister, S., Hur, J., Roth, S.: Unflow: Unsupervised learning of optical flow with a bidirectional census loss. In: Proceedings of the AAAI conference on artificial intelligence. vol. 32 (2018) +34. Menon, S., Damian, A., Hu, S., Ravi, N., Rudin, C.: Pulse: Self-supervised photoupsampling via latent space exploration of generative models. In: Proceedings of the iee/cvf conference on computer vision and pattern recognition. pp. 2437-2445 (2020) +35. Mittal, A., Moorthy, A.K., Bovik, A.C.: No-reference image quality assessment in the spatial domain. IEEE Transactions on image processing 21(12), 4695-4708 (2012) +36. Mittal, A., Soundararajan, R., Bovik, A.C.: Making a "completely blind" image quality analyzer. IEEE Signal processing letters 20(3), 209-212 (2012) +37. Nah, S., Baik, S., Hong, S., Moon, G., Son, S., Timofte, R., Mu Lee, K.: Ntire 2019 challenge on video deblurring and super-resolution: Dataset and study. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. pp. 0-0 (2019) +38. Nichol, A.Q., Dhariwal, P., Ramesh, A., Shyam, P., Mishkin, P., Mcgrew, B., Sutskever, I., Chen, M.: GLIDE: Towards photorealistic image generation and editing with text-guided diffusion models. In: Chaudhuri, K., Jegelka, S., Song, L., Szepesvari, C., Niu, G., Sabato, S. (eds.) Proceedings of the 39th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 162, pp. 16784-16804. PMLR (17-23 Jul 2022), https://proceedings.mlr.press/v162/nichol22a.html +39. Pan, J., Bai, H., Dong, J., Zhang, J., Tang, J.: Deep blind video super-resolution. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4811-4820 (2021) +40. Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., et al.: Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems 32 (2019) +41. Ranjan, A., Black, M.J.: Optical flow estimation using a spatial pyramid network. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4161-4170 (2017) +42. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models (2021) +43. Singer, U., Polyak, A., Hayes, T., Yin, X., An, J., Zhang, S., Hu, Q., Yang, H., Ashual, O., Gafni, O., et al.: Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792 (2022) +44. Song, Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score-based generative modeling through stochastic differential equations. In: Interna + +tional Conference on Learning Representations (2021), https://openreview.net/ forum?id=PxTIG12RRHS +45. Tao, X., Gao, H., Liao, R., Wang, J., Jia, J.: Detail-revealing deep video superresolution. In: Proceedings of the IEEE international conference on computer vision. pp. 4472-4480 (2017) +46. Tao Yang, Peiran Ren, X.X., Zhang, L.: Gan prior embedded network for blind face restoration in the wild. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2021) +47. Teed, Z., Deng, J.: Raft: Recurrent all-pairs field transforms for optical flow. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16. pp. 402-419. Springer (2020) +48. Wang, J., Yue, Z., Zhou, S., Chan, K.C., Loy, C.C.: Exploiting diffusion prior for real-world image super-resolution. arXiv preprint arXiv:2305.07015 (2023) +49. Wang, X., Chan, K.C., Yu, K., Dong, C., Change Loy, C.: Edvr: Video restoration with enhanced deformable convolutional networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. pp. 0-0 (2019) +50. Wang, X., Li, Y., Zhang, H., Shan, Y.: Towards real-world blind face restoration with generative facial prior. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9168-9178 (2021) +51. Wang, X., Xie, L., Dong, C., Shan, Y.: Real-esrgan: Training real-world blind superresolution with pure synthetic data. In: International Conference on Computer Vision Workshops (ICCVW) (2021) +52. Wu, H., Zhang, E., Liao, L., Chen, C., Hou, J., Wang, A., Sun, W., Yan, Q., Lin, W.: Exploring video quality assessment on user generated contents from aesthetic and technical perspectives. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 20144-20154 (2023) +53. Xie, L., Wang, X., Shi, S., Gu, J., Dong, C., Shan, Y.: Mitigating artifacts in real-world video super-resolution models. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 2956-2964 (2023) +54. Xu, H., Yang, J., Cai, J., Zhang, J., Tong, X.: High-resolution optical flow from 1d attention and correlation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10498-10507 (2021) +55. Yang, X., Xiang, W., Zeng, H., Zhang, L.: Real-world video super-resolution: A benchmark dataset and a decomposition based learning scheme. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4781-4790 (2021) +56. Yi, P., Wang, Z., Jiang, K., Jiang, J., Ma, J.: Progressive fusion video superresolution network via exploiting non-local spatio-temporal correlations. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 3106-3115 (2019) +57. Zhang, K., Liang, J., Van Gool, L., Timofte, R.: Designing a practical degradation model for deep blind image super-resolution. In: IEEE International Conference on Computer Vision. pp. 4791-4800 (2021) +58. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 586-595 (2018) +59. Zhao, W., Bai, L., Rao, Y., Zhou, J., Lu, J.: Unipc: A unified predictor-corrector framework for fast sampling of diffusion models. Advances in Neural Information Processing Systems 36 (2024) + +60. Zhou, S., Chan, K.C., Li, C., Loy, C.C.: Towards robust blind face restoration with codebook lookup transformer. In: NeurIPS (2022) \ No newline at end of file diff --git a/2024/Motion-Guided Latent Diffusion for Temporally Consistent Real-world Video Super-resolution/images.zip b/2024/Motion-Guided Latent Diffusion for Temporally Consistent Real-world Video Super-resolution/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..1554f066883519b3ba9806fbd9a0381d84da9336 --- /dev/null +++ b/2024/Motion-Guided Latent Diffusion for Temporally Consistent Real-world Video Super-resolution/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:051fd567162c42145333afa3361ca1e74d43defb532477503f59e71571cbc990 +size 726978 diff --git a/2024/Motion-Guided Latent Diffusion for Temporally Consistent Real-world Video Super-resolution/layout.json b/2024/Motion-Guided Latent Diffusion for Temporally Consistent Real-world Video Super-resolution/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..3e451dbd2d5d051302567b4443a740705c0cd6e0 --- /dev/null +++ b/2024/Motion-Guided Latent Diffusion for Temporally Consistent Real-world Video Super-resolution/layout.json @@ -0,0 +1,12057 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 140, + 111, + 474, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 111, + 474, + 148 + ], + "spans": [ + { + "bbox": [ + 140, + 111, + 474, + 148 + ], + "type": "text", + "content": "Motion-Guided Latent Diffusion for Temporally Consistent Real-world Video Super-resolution" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 174, + 173, + 438, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 174, + 173, + 438, + 187 + ], + "spans": [ + { + "bbox": [ + 174, + 173, + 438, + 187 + ], + "type": "text", + "content": "Xi Yang" + }, + { + "bbox": [ + 174, + 173, + 438, + 187 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 174, + 173, + 438, + 187 + ], + "type": "text", + "content": ", Chenhang He" + }, + { + "bbox": [ + 174, + 173, + 438, + 187 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 174, + 173, + 438, + 187 + ], + "type": "text", + "content": ", Jianqi Ma" + }, + { + "bbox": [ + 174, + 173, + 438, + 187 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 174, + 173, + 438, + 187 + ], + "type": "text", + "content": ", and Lei Zhang" + }, + { + "bbox": [ + 174, + 173, + 438, + 187 + ], + "type": "inline_equation", + "content": "^{1,2^{\\star}}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 150, + 195, + 463, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 150, + 195, + 463, + 241 + ], + "spans": [ + { + "bbox": [ + 150, + 195, + 463, + 241 + ], + "type": "text", + "content": "1 The Hong Kong Polytechnic University, Hong Kong SAR, China \n2 OPPO Research Institute, Shenzhen, China \n{xxxxi.yang, jianqi.ma}@connect.polyu.hk, chenhang.he@polyu.edu.hk, cslzhang@comp.polyu.edu.hk" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 268, + 455, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 268, + 455, + 511 + ], + "spans": [ + { + "bbox": [ + 160, + 268, + 455, + 511 + ], + "type": "text", + "content": "Abstract. Real-world low-resolution (LR) videos have diverse and complex degradations, imposing great challenges on video super-resolution (VSR) algorithms to reproduce their high-resolution (HR) counterparts with high quality. Recently, the diffusion models have shown compelling performance in generating realistic details for image restoration tasks. However, the diffusion process has randomness, making it hard to control the contents of restored images. This issue becomes more serious when applying diffusion models to VSR tasks because temporal consistency is crucial to the perceptual quality of videos. In this paper, we propose an effective real-world VSR algorithm by leveraging the strength of pre-trained latent diffusion models. To ensure the content consistency among adjacent frames, we exploit the temporal dynamics in LR videos to guide the diffusion process by optimizing the latent sampling path with a motion-guided loss, ensuring that the generated HR video maintains a coherent and continuous visual flow. To further mitigate the discontinuity of generated details, we insert temporal modules to the decoder and fine-tune it with an innovative sequence-oriented loss. The proposed motion-guided latent diffusion (MGLD) based VSR algorithm achieves significantly better perceptual quality than state-of-the-arts on real-world VSR benchmark datasets, validating the effectiveness of the proposed model design and training strategies. Codes and models are available at https://github.com/IanYeung/MGLD-VSR." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 520, + 440, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 520, + 440, + 533 + ], + "spans": [ + { + "bbox": [ + 160, + 520, + 440, + 533 + ], + "type": "text", + "content": "Keywords: Video super-resolution " + }, + { + "bbox": [ + 160, + 520, + 440, + 533 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 520, + 440, + 533 + ], + "type": "text", + "content": " Motion-guided latent diffusion" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 552, + 230, + 565 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 552, + 230, + 565 + ], + "spans": [ + { + "bbox": [ + 132, + 552, + 230, + 565 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 575, + 482, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 575, + 482, + 638 + ], + "spans": [ + { + "bbox": [ + 130, + 575, + 482, + 638 + ], + "type": "text", + "content": "Video super-resolution (VSR) [3] aims at reconstructing a high-resolution (HR) video from a given low-resolution (LR) video sequence. Thanks to the rapid development of deep learning techniques, significant progress has been made for VSR in the past decade. Representative VSR methods include sliding window based methods like EDVR [49], recurrent based methods like BasicVSR [9] and" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 642, + 481, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 642, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 642, + 481, + 666 + ], + "type": "text", + "content": "* Corresponding author. This work is supported by the Hong Kong RGC RIF grant (R5001-18) and the PolyU-OPPO Joint Innovation Lab." + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 157, + 113, + 460, + 335 + ], + "blocks": [ + { + "bbox": [ + 157, + 113, + 460, + 335 + ], + "lines": [ + { + "bbox": [ + 157, + 113, + 460, + 335 + ], + "spans": [ + { + "bbox": [ + 157, + 113, + 460, + 335 + ], + "type": "image", + "image_path": "7b6e1ee9655944cacfa565d1a1a9b298bcfe8dd926e46e0f9828df0ca4fb1d2c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 344, + 482, + 401 + ], + "lines": [ + { + "bbox": [ + 130, + 344, + 482, + 401 + ], + "spans": [ + { + "bbox": [ + 130, + 344, + 482, + 401 + ], + "type": "text", + "content": "Fig. 1: The VSR results of 4 consecutive frames in the Sequence 026 from VideoLQ. The top row shows the VSR results by RealBasicVSR [11], the middle row shows the naive VSR results by running StableSR [48] on each frame of the sequence, and the bottom row shows the VSR results produced by our proposed MGLD-VSR method. Our method generates realistic details while achieving good temporal consistency." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 424, + 482, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 424, + 482, + 484 + ], + "spans": [ + { + "bbox": [ + 130, + 424, + 482, + 484 + ], + "type": "text", + "content": "more recent transformer based methods [30]. However, most of the above works assume simple degradations (e.g., bicubic resizing, downsampling after Gaussian blur) between the LR and HR videos. As a result, the trained VSR models are hard to be generalized to real-world LR videos, whose degradations are unknown and much more complex." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 484, + 482, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 484, + 482, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 484, + 482, + 640 + ], + "type": "text", + "content": "Recently, real-world VSR [11, 53, 55] has gained attention from researchers for its high potential value in practical applications such as camera-phone video enhancement and online video streaming. Different from traditional VSR tasks, which focus on super-resolving clean LR videos with simple synthetic degradation (e.g., bicubic downsampling), real-world VSR aims to enhance videos with complex and unknown degradations that often appear in real-world low-quality videos. The main challenge lies in how to effectively reproduce video details while suppressing the visual artifacts caused by unknown degradations. Existing real-world VSR algorithms mostly attempt to trade-off between details and artifacts using techniques such as Laplacian pyramid generative adversarial loss [55], input pre-cleaning [11] and hidden state attention [53], etc. Despite these efforts and progresses, their performance on real-world testing sets remains limited due to the insufficient training data and the limited model capacity." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "content": "Since the seminal work of denoising diffusion probabilistic model (DDPM) [20], diffusion-based generative models have achieved great success in image gen" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 224, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 224, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 224, + 102 + ], + "type": "text", + "content": "X. Yang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 482, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 482, + 272 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 482, + 272 + ], + "type": "text", + "content": "eration. In particular, the latent diffusion models (LDMs) [42] have demonstrated compelling results not only in text-to-image generation, but also in downstream tasks such as image editing [24], inpainting [32], colorization [7], etc. Recently, researchers have also attempted to employ the powerful generative priors of LDMs for real-world image restoration tasks [48] and shown promising results. It is thus interesting to investigate whether the latent diffusion priors can be used to improve real-world VSR results. However, this is a non-trivial task because of the intrinsic randomness in the diffusion process. This randomness in turn gives rise to the temporal inconsistencies among video frames in the latent space. Such inconsistencies can be further amplified by the decoder of the LDM. As shown in the middle row of Figure 1, directly employing the LDM based image super-resolution algorithm [48] for VSR will lead to inconsistent details across consecutive frames, disturbing the visual perception of the reproduced video." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 272, + 482, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 272, + 482, + 403 + ], + "spans": [ + { + "bbox": [ + 130, + 272, + 482, + 403 + ], + "type": "text", + "content": "To address the above-mentioned challenges, we propose a Motion-Guided Latent Diffusion (MGLD) model for real-world VSR in this paper, aiming to produce high-quality HR video sequence with good temporal consistency. We propose to incorporate the motion dynamics of input LR video into the generation of SR outputs. This can be done through a guided diffusion process where we can generate conditional samples from a guidance model based on its score function [14, 44]. In particular, we calculate the optical flows among adjacent LR frames, and use the calculated optical flows to warp the latent features of each frame to align the features of neighboring frames. The " + }, + { + "bbox": [ + 130, + 272, + 482, + 403 + ], + "type": "inline_equation", + "content": "l_{1}" + }, + { + "bbox": [ + 130, + 272, + 482, + 403 + ], + "type": "text", + "content": " norm of the warping error is employed as the motion-guided loss, and its gradients can be incorporated in the sampling process to update the latent features." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 403, + 482, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 403, + 482, + 487 + ], + "spans": [ + { + "bbox": [ + 130, + 403, + 482, + 487 + ], + "type": "text", + "content": "To overcome the inconsistent details generated by the original decoder of LDM, we insert temporal modules to the decoder and fine-tune it with the ground-truth HR video sequences. We also introduce an innovative sequence-oriented loss to guide the decoder to improve the continuity of the details. As depicted in the bottom row of Figure 1, our proposed MGLD-VSR approach could mitigate the discontinuity of generated details while maintaining natural and visually appealing texture recovery." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 487, + 482, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 487, + 482, + 582 + ], + "spans": [ + { + "bbox": [ + 130, + 487, + 482, + 582 + ], + "type": "text", + "content": "The contributions of this paper are summarized as follows. First, we propose a diffusion sampling process based on motion-guided loss, allowing the temporal dynamics of the input frame to be used in generating temporally consistent latent features. Second, we propose a temporal-aware sequence decoder, along with two sequence-oriented losses, to further enhance the continuity of generated videos. Our proposed MGLD-VSR model achieves highly competitive real-world VSR results, exhibiting perceptually much more realistic details with fewer flickering artifacts than existing state-of-the-arts." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 602, + 237, + 616 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 602, + 237, + 616 + ], + "spans": [ + { + "bbox": [ + 132, + 602, + 237, + 616 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "content": "Video Super-Resolution. The recent advances in VSR research largely attribute to the rapid development of deep-learning technologies [6, 9, 22, 30, 49]. Based on the paradigms, existing deep-learning based VSR algorithms could" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 395, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 395, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 395, + 91, + 447, + 100 + ], + "type": "text", + "content": "MGLD-VSR" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 260 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 260 + ], + "type": "text", + "content": "be roughly classified into two categories. The first category performs VSR in a multiple-input single-output (MISO) fashion. Typical algorithms include DUF [22], EDVR [49], and MuCAN [28], etc. The MISO algorithms are usually sliding-window based, which aggregate information from neighboring frames to generate the super-resolved result of the center frame. The other category of methods adopt a multiple-input multiple-output (MIMO) design, such as BasicVSR [9], BasicVSR++ [10], VRT [30] and RVRT [31]. These methods take multiple LR frames as input and generate their corresponding SR results in one go. Compared with MISO, the amortized running cost of MIMO is usually much lower as it avoids a series of redundant operations. In addition, benefiting from the lower cost, the MIMO methods are able to aggregate information from more frames to improve the performance. We adopt the MIMO design in this work." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 263, + 482, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 263, + 482, + 443 + ], + "spans": [ + { + "bbox": [ + 130, + 263, + 482, + 443 + ], + "type": "text", + "content": "Real-world VSR. Most existing VSR methods are trained on synthetic data with simple degradations (e.g., bicubic resizing, downsampling after Gaussian blur). Due to the synthetic-to-real gap, models trained with such data cannot perform well on real-world videos. To tackle this problem, Yang et al. [55] collected LR-HR video pairs with iPhone cameras to better model real-world degradations. While being effective in some cases, it is relatively labor-intensive and may not generalize well to videos collected by other devices. Later works usually adopt an easier solution: synthesizing more realistic training data by mixing more types of degradations, such as blur, downsampling, noise and video compression [11, 53]. With the improved synthetic training data, the trained models could achieve better super-resolution performance on real-world videos. However, there is still much distribution gap between the synthetic LR data and real-world LR videos. Instead of working on synthesizing more realistic training data, in this work we investigate how to leverage the strong generative priors of pre-trained latent diffusion models to improve the real-world VSR performance." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 446, + 482, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 446, + 482, + 615 + ], + "spans": [ + { + "bbox": [ + 130, + 446, + 482, + 615 + ], + "type": "text", + "content": "Generative Prior for Image Restoration. The highly ill-posed nature of real-world restoration problem brings forward a great challenge on real-world image restoration algorithms. To boost the performance, a series of methods have introduced generative priors into the image restoration process. Early methods usually adopt pre-trained generative adversarial networks (GANs) as the generative prior [8,19,34,46,50]. Although effective, the priors adopted in these methods are usually tailored for specific tasks (e.g., face image restoration), limiting their capability in handling more general natural image restoration problems. Considering the higher generality of diffusion models than GANs in natural image generation, recently researchers have started to investigate how to leverage generative diffusion priors for image restoration, and demonstrated interesting results [17, 48]. In this work, we investigate the possibility of applying latent diffusion priors to real-world VSR, and develop effective techniques to tackle the flickering issue and improve the temporal consistency of generated videos." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Diffusion for Video Synthesis. Recently, diffusion based models have demonstrated its superiority in the video synthesis [4,5,18,43]. As one of the pioneer works, Make-a-Video [43] leverages a pretrained text-to-image (T2I) model and build a text-to-video (T2V) model by inserting temporal modules into the" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "text", + "content": "X. Yang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 129, + 352, + 277 + ], + "blocks": [ + { + "bbox": [ + 214, + 116, + 279, + 124 + ], + "lines": [ + { + "bbox": [ + 214, + 116, + 279, + 124 + ], + "spans": [ + { + "bbox": [ + 214, + 116, + 279, + 124 + ], + "type": "text", + "content": "MGLD Sampling Pipeline" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 135, + 129, + 352, + 277 + ], + "lines": [ + { + "bbox": [ + 135, + 129, + 352, + 277 + ], + "spans": [ + { + "bbox": [ + 135, + 129, + 352, + 277 + ], + "type": "image", + "image_path": "ce3187b7ca565a36504f90fd37e72cbb09efd9510150185681d3ea2ce3a260c7.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 363, + 129, + 467, + 186 + ], + "blocks": [ + { + "bbox": [ + 380, + 117, + 456, + 124 + ], + "lines": [ + { + "bbox": [ + 380, + 117, + 456, + 124 + ], + "spans": [ + { + "bbox": [ + 380, + 117, + 456, + 124 + ], + "type": "text", + "content": "Update with Motion Guidance" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 363, + 129, + 467, + 186 + ], + "lines": [ + { + "bbox": [ + 363, + 129, + 467, + 186 + ], + "spans": [ + { + "bbox": [ + 363, + 129, + 467, + 186 + ], + "type": "image", + "image_path": "02c68e471630b2d143c8f181d77ecc91c3ea73b81ecd4f9887593d44e420770c.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 364, + 198, + 471, + 278 + ], + "blocks": [ + { + "bbox": [ + 375, + 190, + 465, + 197 + ], + "lines": [ + { + "bbox": [ + 375, + 190, + 465, + 197 + ], + "spans": [ + { + "bbox": [ + 375, + 190, + 465, + 197 + ], + "type": "text", + "content": "Temporal-aware Sequence Decoder" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 364, + 198, + 471, + 278 + ], + "lines": [ + { + "bbox": [ + 364, + 198, + 471, + 278 + ], + "spans": [ + { + "bbox": [ + 364, + 198, + 471, + 278 + ], + "type": "image", + "image_path": "4eb625b1717bda730056a1a487b4e67015f21241eb72a978c4a2228d3126a7e7.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 301, + 482, + 378 + ], + "lines": [ + { + "bbox": [ + 130, + 301, + 482, + 378 + ], + "spans": [ + { + "bbox": [ + 130, + 301, + 482, + 378 + ], + "type": "text", + "content": "Fig. 2: Overview of the proposed MGLD-VSR framework for real-world VSR. We first estimate the forward and backward optical flow with a pre-trained optical flow estimation network and then employ a motion-guided loss to update the latent diffusion sampling process. The motion-guided loss is computed by summing the forward and backward masked warping error of the latent sequence at each time-step. After a number of " + }, + { + "bbox": [ + 130, + 301, + 482, + 378 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 301, + 482, + 378 + ], + "type": "text", + "content": " sampling steps, we obtain the generated latent sequence and feed it into the temporal-aware sequence decoder to reconstruct the VSR sequence." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 403, + 482, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 403, + 482, + 499 + ], + "spans": [ + { + "bbox": [ + 130, + 403, + 482, + 499 + ], + "type": "text", + "content": "T2I model and designing a cascaded architecture. VideoLDM [5] adopts a latent diffusion paradigm and train the T2V model on large-scale paired text-video datasets, greatly improving the generation results. Stable Video Diffusion [4] further improve the results by conducting training on more carefully constructed datasets and introducing more advanced training paradigm. In this work, we leverage the diffusion prior and build a powerful real-world VSR model, which aims to tackle the ill-posedness of real-world VSR problem with diffusion prior and obtain temporally consistent VSR results with rich details." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 519, + 233, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 519, + 233, + 533 + ], + "spans": [ + { + "bbox": [ + 131, + 519, + 233, + 533 + ], + "type": "text", + "content": "3 Methodology" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "Given an LR video sequence of " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": " frames " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "I = \\{I_1, I_2, \\ldots, I_N\\}" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": ", our goal is to reproduce an HR video sequence " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\hat{I} = \\{\\hat{I}_1, \\hat{I}_2, \\ldots, \\hat{I}_N\\}" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": " with enhanced visual quality while maintaining the content consistency across adjacent frames. Inspired by the recently developed diffusion-based image restoration methods [17, 48], we propose to harness the generative priors of the pre-trained SD model [42] to improve the real-world VSR performance. First, we propose a motion-guided diffusion sampling process, where the temporal dynamics from the LR sequence are introduced into the diffusion sampling process to enhance the consistency of latent features across frames. Then, we further improve the quality of generated details by designing a temporal-aware sequence decoder and fine-tuning" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 395, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 395, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 395, + 91, + 447, + 100 + ], + "type": "text", + "content": "MGLD-VSR" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "it with a video-oriented loss. The overall framework of our algorithm, namely motion-guided latent diffusion (MGLD), is illustrated in Figure 2." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 156, + 279, + 167 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 156, + 279, + 167 + ], + "spans": [ + { + "bbox": [ + 131, + 156, + 279, + 167 + ], + "type": "text", + "content": "3.1 Latent Diffusion Model" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 175, + 482, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 175, + 482, + 284 + ], + "spans": [ + { + "bbox": [ + 130, + 175, + 482, + 284 + ], + "type": "text", + "content": "Diffusion models [21] are a class of generative models which transform Gaussian noise into data samples through an iterative reverse Markovian process. This process can be described by: " + }, + { + "bbox": [ + 130, + 175, + 482, + 284 + ], + "type": "inline_equation", + "content": "q(x_{t}|x_{t - 1})\\coloneqq \\mathcal{N}(x_{t},(1 - \\beta_{t})x_{t - 1},\\beta_{t}I)" + }, + { + "bbox": [ + 130, + 175, + 482, + 284 + ], + "type": "text", + "content": ", where the Gaussian noise is iteratively applied to the sample " + }, + { + "bbox": [ + 130, + 175, + 482, + 284 + ], + "type": "inline_equation", + "content": "x_{t - 1}" + }, + { + "bbox": [ + 130, + 175, + 482, + 284 + ], + "type": "text", + "content": " at time-step " + }, + { + "bbox": [ + 130, + 175, + 482, + 284 + ], + "type": "inline_equation", + "content": "t - 1" + }, + { + "bbox": [ + 130, + 175, + 482, + 284 + ], + "type": "text", + "content": " to produce " + }, + { + "bbox": [ + 130, + 175, + 482, + 284 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 130, + 175, + 482, + 284 + ], + "type": "text", + "content": ", guided by a variance scheduler " + }, + { + "bbox": [ + 130, + 175, + 482, + 284 + ], + "type": "inline_equation", + "content": "\\{\\beta_t\\}_{t = 1}^T" + }, + { + "bbox": [ + 130, + 175, + 482, + 284 + ], + "type": "text", + "content": " along the diffusion process. The key objective of diffusion models is to train a denoising U-Net " + }, + { + "bbox": [ + 130, + 175, + 482, + 284 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 130, + 175, + 482, + 284 + ], + "type": "text", + "content": " parameterized by " + }, + { + "bbox": [ + 130, + 175, + 482, + 284 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 130, + 175, + 482, + 284 + ], + "type": "text", + "content": " and conditioned on the time-step " + }, + { + "bbox": [ + 130, + 175, + 482, + 284 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 175, + 482, + 284 + ], + "type": "text", + "content": " and an optional signal " + }, + { + "bbox": [ + 130, + 175, + 482, + 284 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 130, + 175, + 482, + 284 + ], + "type": "text", + "content": ". Given the noise target " + }, + { + "bbox": [ + 130, + 175, + 482, + 284 + ], + "type": "inline_equation", + "content": "\\epsilon_{t}" + }, + { + "bbox": [ + 130, + 175, + 482, + 284 + ], + "type": "text", + "content": ", the diffusion loss can be defined as: " + }, + { + "bbox": [ + 130, + 175, + 482, + 284 + ], + "type": "inline_equation", + "content": "L(\\theta) = \\mathbb{E}_{t\\sim U(1,T),\\epsilon_t\\sim \\mathcal{N}(0,I)}\\| \\epsilon_t - \\epsilon_\\theta (\\mathbf{x}_t;t,y)\\| _2^2" + }, + { + "bbox": [ + 130, + 175, + 482, + 284 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 285, + 482, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 285, + 482, + 415 + ], + "spans": [ + { + "bbox": [ + 130, + 285, + 482, + 415 + ], + "type": "text", + "content": "Latent Diffusion Models (LDMs) [42] have made a remarkable impact in the field of image generation. These models employ Variational Autoencoders (VAE) to map the images into a latent space, enabling them to be trained on large-scale text-image datasets. This equips LDMs with powerful prior knowledge about natural images, which can be leveraged by downstream tasks such as image restoration and super-resolution [48] to produce images with clear details and high-quality content. However, in the case of VSR, the stochastic nature of diffusion process often leads to the temporal inconsistencies in the restored video sequence. To solve this issue, we propose a MGLD model that achieves high-quality VSR by utilizing guidance during the sampling process and incorporating temporal awareness in the decoder." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 431, + 334, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 431, + 334, + 443 + ], + "spans": [ + { + "bbox": [ + 131, + 431, + 334, + 443 + ], + "type": "text", + "content": "3.2 Motion-guided Diffusion Sampling" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 449, + 482, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 482, + 498 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 482, + 498 + ], + "type": "text", + "content": "From a score-based perspective [44], the sampling process of a diffusion model involves the application of Langevin Dynamics to both gradient of the data distribution " + }, + { + "bbox": [ + 130, + 449, + 482, + 498 + ], + "type": "inline_equation", + "content": "\\log p(\\mathbf{z})" + }, + { + "bbox": [ + 130, + 449, + 482, + 498 + ], + "type": "text", + "content": " and a guidance model " + }, + { + "bbox": [ + 130, + 449, + 482, + 498 + ], + "type": "inline_equation", + "content": "g = G(z)" + }, + { + "bbox": [ + 130, + 449, + 482, + 498 + ], + "type": "text", + "content": ", which can be a network [15,38] or an energy model [12]:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 212, + 504, + 481, + 519 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 212, + 504, + 481, + 519 + ], + "spans": [ + { + "bbox": [ + 212, + 504, + 481, + 519 + ], + "type": "interline_equation", + "content": "\\nabla_ {\\mathbf {z}} \\log p (\\mathbf {z}, g) = \\nabla_ {\\mathbf {z}} \\log p (\\mathbf {z}) + \\nabla_ {\\mathbf {z}} \\log p (g | \\mathbf {z}). \\qquad (1)", + "image_path": "daf5522b2fbc0227fd9434c9523cbf9e0010cf0967080e1ae75a9255a7d9d4da.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 525, + 482, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 525, + 482, + 634 + ], + "spans": [ + { + "bbox": [ + 130, + 525, + 482, + 634 + ], + "type": "text", + "content": "In this work, we introduce a technique that utilizes the temporal dynamics in low-resolution input videos to create latent features that maintain temporal consistency, thereby enhancing VSR results. Specifically, we incorporate an innovative motion-guided module within the sampling process to measure the warping errors of latent features across frames. Given the LR frames, we first compute their optical flow with " + }, + { + "bbox": [ + 130, + 525, + 482, + 634 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 130, + 525, + 482, + 634 + ], + "type": "text", + "content": " [47] and downsample the flow map to the dimension of latent features. Given the forward flow " + }, + { + "bbox": [ + 130, + 525, + 482, + 634 + ], + "type": "inline_equation", + "content": "O_{f}" + }, + { + "bbox": [ + 130, + 525, + 482, + 634 + ], + "type": "text", + "content": " and backward flow " + }, + { + "bbox": [ + 130, + 525, + 482, + 634 + ], + "type": "inline_equation", + "content": "O_{b}" + }, + { + "bbox": [ + 130, + 525, + 482, + 634 + ], + "type": "text", + "content": ", we warp the latent features to their neighbouring frames and compute the accumulated warping error along both directions:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 638, + 481, + 666 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 638, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 638, + 481, + 666 + ], + "type": "interline_equation", + "content": "E (z) ^ {t} = \\sum_ {i = 1} ^ {N - 1} \\| \\left(\\operatorname {W a r p} \\left(z _ {i} ^ {t}, O _ {b, i}\\right) - z _ {i + 1} ^ {t})\\right) \\| _ {1} + \\sum_ {i = 2} ^ {N} \\| \\left(\\operatorname {W a r p} \\left(z _ {i} ^ {t}, O _ {f, i - 1}\\right) - z _ {i - 1} ^ {t})\\right) \\| _ {1}. \\tag {2}", + "image_path": "76af992151d60675840645a7bbd02f2796477fe674e51268f536daaca473a2ec.jpg" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "text", + "content": "X. Yang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 188 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 188 + ], + "type": "text", + "content": "We empirically find that the presence of occlusion will pose adverse impact on flow estimation, which in turn causes distraction in the sampling process and leads to artifacts in the final VSR results. To solve this issue, we further estimate the occlusion mask " + }, + { + "bbox": [ + 130, + 116, + 479, + 188 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 130, + 116, + 479, + 188 + ], + "type": "text", + "content": " in each frame [33] and ignore the contributions in the occluded region. Given the variance " + }, + { + "bbox": [ + 130, + 116, + 479, + 188 + ], + "type": "inline_equation", + "content": "\\sigma_t^2" + }, + { + "bbox": [ + 130, + 116, + 479, + 188 + ], + "type": "text", + "content": " of noise scheduler at step " + }, + { + "bbox": [ + 130, + 116, + 479, + 188 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 116, + 479, + 188 + ], + "type": "text", + "content": ", the motion-guided sampling process at sampling step " + }, + { + "bbox": [ + 130, + 116, + 479, + 188 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 116, + 479, + 188 + ], + "type": "text", + "content": " can be written as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 213, + 193, + 481, + 206 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 213, + 193, + 481, + 206 + ], + "spans": [ + { + "bbox": [ + 213, + 193, + 481, + 206 + ], + "type": "interline_equation", + "content": "\\hat {z} ^ {t} \\leftarrow \\operatorname {D D P M} \\left(z ^ {t + 1}, t\\right) - \\eta \\sigma_ {t} ^ {2} \\nabla_ {z} \\left(\\mathcal {M} \\circ E (z) ^ {t}\\right), \\tag {3}", + "image_path": "0f033416ee92a0e90bdc1937499dafe6d3d29e16b667ae4354952c9be7ff23ad.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 211, + 480, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 211, + 480, + 259 + ], + "spans": [ + { + "bbox": [ + 130, + 211, + 480, + 259 + ], + "type": "text", + "content": "where the DDPM [21] scheduler is applied for diffusion reverse (denoising) process and " + }, + { + "bbox": [ + 130, + 211, + 480, + 259 + ], + "type": "inline_equation", + "content": "\\eta = 10" + }, + { + "bbox": [ + 130, + 211, + 480, + 259 + ], + "type": "text", + "content": " is a weighting factor to control the guidance strength. After " + }, + { + "bbox": [ + 130, + 211, + 480, + 259 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 211, + 480, + 259 + ], + "type": "text", + "content": " iterations, the final denoised latent sequence " + }, + { + "bbox": [ + 130, + 211, + 480, + 259 + ], + "type": "inline_equation", + "content": "z^0" + }, + { + "bbox": [ + 130, + 211, + 480, + 259 + ], + "type": "text", + "content": " is fed into the VAE decoder to obtain the final video sequence." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 274, + 349, + 286 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 274, + 349, + 286 + ], + "spans": [ + { + "bbox": [ + 131, + 274, + 349, + 286 + ], + "type": "text", + "content": "3.3 Temporal-aware Decoder Fine-tuning" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 290, + 480, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 290, + 480, + 386 + ], + "spans": [ + { + "bbox": [ + 130, + 290, + 480, + 386 + ], + "type": "text", + "content": "With the proposed motion-guided diffusion sampling, the VSR model could better exploit the temporal dynamics to generate temporally more consistent HR videos. However, since the guidance is performed in the low-resolution latent space, which is 8 times smaller in both height and width than the output image space, the video reconstruction through a VAE decoder may still lead to temporal discontinuities in the generated details of VSR output. Thus, we introduce a temporal-aware sequence decoder and fine-tune it with ground-truth video sequences to improve the smoothness and accuracy of output videos." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 386, + 480, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 386, + 480, + 493 + ], + "spans": [ + { + "bbox": [ + 130, + 386, + 480, + 493 + ], + "type": "text", + "content": "As shown in Figure 2, we build our sequence decoder on top of the pretrained VAE decoder by incorporating 1D convolutions along the temporal dimension. This design facilitates interleaved spatial-temporal interactions and enables our module to restore details with high continuity while incurring minimal computational cost. We further incorporate the information " + }, + { + "bbox": [ + 130, + 386, + 480, + 493 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 130, + 386, + 480, + 493 + ], + "type": "text", + "content": " from the VAE encoder through a Controllable Feature Warping (CFW) module [48, 60] for better restoration and generation results. To leverage the restorative power of the pretrained VAE, we freeze their original spatial blocks and only update the parameters in the temporal convolution and CFW module." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 494, + 480, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 494, + 480, + 553 + ], + "spans": [ + { + "bbox": [ + 130, + 494, + 480, + 553 + ], + "type": "text", + "content": "We first employ the " + }, + { + "bbox": [ + 130, + 494, + 480, + 553 + ], + "type": "inline_equation", + "content": "\\ell_1" + }, + { + "bbox": [ + 130, + 494, + 480, + 553 + ], + "type": "text", + "content": " loss and perceptual loss [23, 58] as the reconstruction loss " + }, + { + "bbox": [ + 130, + 494, + 480, + 553 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{recon}}" + }, + { + "bbox": [ + 130, + 494, + 480, + 553 + ], + "type": "text", + "content": " for each frame, then utilize the GAN-loss and frame difference loss that tailored for video sequences. The frame difference loss " + }, + { + "bbox": [ + 130, + 494, + 480, + 553 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{diff}}" + }, + { + "bbox": [ + 130, + 494, + 480, + 553 + ], + "type": "text", + "content": " calculates the difference between consecutive predicted frames " + }, + { + "bbox": [ + 130, + 494, + 480, + 553 + ], + "type": "inline_equation", + "content": "\\hat{I}" + }, + { + "bbox": [ + 130, + 494, + 480, + 553 + ], + "type": "text", + "content": " and ground truth frames " + }, + { + "bbox": [ + 130, + 494, + 480, + 553 + ], + "type": "inline_equation", + "content": "I^{gt}" + }, + { + "bbox": [ + 130, + 494, + 480, + 553 + ], + "type": "text", + "content": ", which is defined as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 222, + 558, + 480, + 572 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 222, + 558, + 480, + 572 + ], + "spans": [ + { + "bbox": [ + 222, + 558, + 480, + 572 + ], + "type": "interline_equation", + "content": "L _ {\\text {d i f f}} = \\sum \\| \\left(\\hat {I} _ {i + 1} - \\hat {I} _ {i}\\right) - \\left(I _ {i + 1} ^ {g t} - I _ {i} ^ {g t}\\right) \\| _ {1}. \\tag {4}", + "image_path": "90442de1670ca930b00349babf99ae2f2a99a41765796c1e5ae6e1745ac73ed0.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 577, + 479, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 577, + 479, + 625 + ], + "spans": [ + { + "bbox": [ + 130, + 577, + 479, + 625 + ], + "type": "text", + "content": "It is noteworthy that diffusion models typically generate strong structures. To make the super-resolved video look more natural, we need to apply consistency constraints to those structured regions. We hence introduce a structure weighted consistency loss " + }, + { + "bbox": [ + 130, + 577, + 479, + 625 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{swc}}" + }, + { + "bbox": [ + 130, + 577, + 479, + 625 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 200, + 630, + 481, + 663 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 630, + 481, + 663 + ], + "spans": [ + { + "bbox": [ + 200, + 630, + 481, + 663 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} L _ {\\mathrm {s w c}} = \\sum \\| \\mathcal {M} \\circ (\\mathcal {W} _ {i + 1} (\\operatorname {W a r p} (\\hat {I} _ {i}, O _ {b, i} ^ {g t}) - \\hat {I} _ {i + 1})) \\| _ {1} \\tag {5} \\\\ + \\sum \\| \\mathcal {M} \\circ (\\mathcal {W} _ {i - 1} (\\operatorname {W a r p} (\\hat {I _ {i}}, O _ {f, i} ^ {g t}) - \\hat {I} _ {i - 1})) \\| _ {1}. \\\\ \\end{array}", + "image_path": "dca10d60c4e18817cb2891c294107124a1e43bba2b68e71052803e1d51baf458.jpg" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 394, + 90, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 394, + 90, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 394, + 90, + 447, + 101 + ], + "type": "text", + "content": "MGLD-VSR" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 482, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 482, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 482, + 200 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 115, + 482, + 200 + ], + "type": "inline_equation", + "content": "O_{f}^{gt}" + }, + { + "bbox": [ + 130, + 115, + 482, + 200 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 115, + 482, + 200 + ], + "type": "inline_equation", + "content": "O_{b}^{gt}" + }, + { + "bbox": [ + 130, + 115, + 482, + 200 + ], + "type": "text", + "content": " denote the forward and backward flow computed on the ground-truth sequence. This loss helps the network to generate consistent details by modulating the loss through a weighted map " + }, + { + "bbox": [ + 130, + 115, + 482, + 200 + ], + "type": "inline_equation", + "content": "\\mathcal{W} = (1 + wS)" + }, + { + "bbox": [ + 130, + 115, + 482, + 200 + ], + "type": "text", + "content": " computed on the ground-truth frames, where " + }, + { + "bbox": [ + 130, + 115, + 482, + 200 + ], + "type": "inline_equation", + "content": "w = 3" + }, + { + "bbox": [ + 130, + 115, + 482, + 200 + ], + "type": "text", + "content": " is a weighting factor and " + }, + { + "bbox": [ + 130, + 115, + 482, + 200 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 130, + 115, + 482, + 200 + ], + "type": "text", + "content": " is a structure (edge) map using the Sobel operator. Occlusion masking is also adopted. Given the GAN-based loss " + }, + { + "bbox": [ + 130, + 115, + 482, + 200 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{GAN}}" + }, + { + "bbox": [ + 130, + 115, + 482, + 200 + ], + "type": "text", + "content": " and reconstruction loss " + }, + { + "bbox": [ + 130, + 115, + 482, + 200 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{recon}}" + }, + { + "bbox": [ + 130, + 115, + 482, + 200 + ], + "type": "text", + "content": ", the total loss for decoder fine-tuning is defined as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 214, + 212, + 481, + 224 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 214, + 212, + 481, + 224 + ], + "spans": [ + { + "bbox": [ + 214, + 212, + 481, + 224 + ], + "type": "interline_equation", + "content": "L _ {\\text {v i d e o}} = L _ {\\text {r e c o n}} + \\alpha L _ {\\text {d i f f}} + \\beta L _ {\\text {s w c}} + \\gamma L _ {\\text {G A N}}, \\tag {6}", + "image_path": "42d6ad8eea5b66a770896fe8b3018ae5db7ca2f644d91a482aa13693646e3ae8.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 233, + 389, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 233, + 389, + 246 + ], + "spans": [ + { + "bbox": [ + 131, + 233, + 389, + 246 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 233, + 389, + 246 + ], + "type": "inline_equation", + "content": "\\alpha, \\beta" + }, + { + "bbox": [ + 131, + 233, + 389, + 246 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 233, + 389, + 246 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 131, + 233, + 389, + 246 + ], + "type": "text", + "content": " are empirically set to 0.5, 0.5 and 0.025." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 267, + 230, + 281 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 267, + 230, + 281 + ], + "spans": [ + { + "bbox": [ + 131, + 267, + 230, + 281 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 295, + 274, + 308 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 295, + 274, + 308 + ], + "spans": [ + { + "bbox": [ + 131, + 295, + 274, + 308 + ], + "type": "text", + "content": "4.1 Experimental Settings" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 317, + 482, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 317, + 482, + 496 + ], + "spans": [ + { + "bbox": [ + 130, + 317, + 482, + 496 + ], + "type": "text", + "content": "Implementation Details. The training of MGLD-VSR consists of two main stages. In the first stage, we fine-tune the denoising U-Net for latent space diffusion. The weight of the denoising U-Net is initialized from Stable Diffusion V2.1. We insert 1D temporal convolution into the U-Net to help temporal dynamic modeling. We then fix the weight in the denoising U-Net of SD model and train the conditioning and temporal modeling modules. For conditioning, we adopt a small time-aware encoder to encode LR condition and inject it into the denoising U-Net with spatial feature transform operation as in [48]. The training batch size, sequence length and latent image size are set to 24, 6 and " + }, + { + "bbox": [ + 130, + 317, + 482, + 496 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 130, + 317, + 482, + 496 + ], + "type": "text", + "content": ", respectively. In the second stage, we first generate the clean latent sequence with the proposed motion-guided diffusion sampling process, then fine-tune the temporal-aware sequence decoder with the LR sequence, the generated latent sequence and the HR sequence. We fix the VAE decoder and insert temporal modeling layers and CFW modules for training. The training batch size, sequence length and image size are set to 4, 5 and " + }, + { + "bbox": [ + 130, + 317, + 482, + 496 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 130, + 317, + 482, + 496 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 497, + 482, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 497, + 482, + 557 + ], + "spans": [ + { + "bbox": [ + 130, + 497, + 482, + 557 + ], + "type": "text", + "content": "Our models are trained on 4 NVIDIA A100 GPUs with the PyTorch [40] framework. Adam [26] is chosen as the optimizer. During inference, due to memory constrains, we divide the LR video into multiple sequences to run. To handle output with arbitrary resolution, we apply a progressive patch aggregation sampling algorithm [48]. For each sequence, we run the sampling for 50 steps." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": "Training and Testing Datasets. As in previous works [11, 49, 53], we merge the training set and validation set of REDS [37] in training and leave 4 sequences for testing (REDS4). We follow the degradation pipeline of RealBasicVSR [11] to synthesize the training sequence pairs, which involves blur, noise, downsampling and compression degradations. We evaluate our method on both synthetic and real-world datasets. To synthesize LR-HR testing sequence pairs, we apply the RealBasicVSR degradation pipeline on several widely used VSR datasets, including REDS4 [37], UDM10 [56] and SPMCS [45]. REDS4 contains 4 video sequences, each having 100 frames. UDM10 consists of 10 sequences, each" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 224, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 224, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 224, + 102 + ], + "type": "text", + "content": "X. Yang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "having 32 frames. SPMCS has 30 sequences, each having 31 frames. For real-world dataset, we adopt VideoLQ [11] for testing, which contains 50 real-world sequences with complex degradations." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 480, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 480, + 213 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 480, + 213 + ], + "type": "text", + "content": "Evaluation Metrics. On synthetic datasets, we could adopt full-reference metrics to evaluate different real-world VSR methods. The perceptual quality metrics LPIPS [58] and DISTS [16] are selected. Fidelity-oriented metrics PSNR and SSIM are also calculated for reference. Apart from image-based metrics, the video quality assessment metric VMAF [29] is also employed." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 213, + 481, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 213, + 481, + 262 + ], + "spans": [ + { + "bbox": [ + 130, + 213, + 481, + 262 + ], + "type": "text", + "content": "On real-world datasets, as ground truths are not available, we compute the no-reference image quality metrics NIQE [36] and BRISQUE [35] for evaluation. In addition, we also include a deep learning based image-based metric MUSIQ [25] and a state-of-the-art video quality assessment metric DOVER [52]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 282, + 269, + 294 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 282, + 269, + 294 + ], + "spans": [ + { + "bbox": [ + 132, + 282, + 269, + 294 + ], + "type": "text", + "content": "4.2 Experimental Results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 305, + 480, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 305, + 480, + 365 + ], + "spans": [ + { + "bbox": [ + 130, + 305, + 480, + 365 + ], + "type": "text", + "content": "To demonstrate the effectiveness of the our MGLD-VSR algorithm, we compare it with seven state-of-the-art methods, including three real-world image super-resolution models (BSRGAN [57], RealESRGAN [51], StableSR [48]), one classic VSR model (RVRT [31]), and three blind or real-world VSR models (DBVSR [39], RealVSR [55], RealBasicVSR [11]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 366, + 481, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 366, + 481, + 520 + ], + "spans": [ + { + "bbox": [ + 130, + 366, + 481, + 520 + ], + "type": "text", + "content": "Quantitative Comparison. We first show the quantitative comparison on the synthetic datasets and real-world video benchmarks in Table 1. As shown from Table 1, our method achieves the best results in terms of full-reference perceptual based metrics LPIPS and DISTS on all synthetic test datasets, indicating that the proposed method is able to reconstruct high-quality plausible details from sequences degraded with complex degradations. Although methods like DBVSR may achieve better performance in terms of PSNR or SSIM on some datasets, they tend to generate blurry results, as demonstrated in the LPIPS and DISTS metrics. In terms of VMAF, our model achieves the best or second best results. As for real-world VSR dataset VideoLQ, our methods achieve the best performance in terms of NIQE, BRISQUE, and DOVER and the second best results in terms of MUSIQ, which reflects its strong capability to enhance the real-world videos and generate realistic details and textures." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 521, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 481, + 665 + ], + "type": "text", + "content": "Qualitative Comparison. To demonstrate the effectiveness of MGLD-VSR, we visualize the VSR results on synthetic datasets in Figure 3 and the results on real-world VideoLQ in Figure 4. One can see that MGLD-VSR can remove complex spatial-variant degradations and generate realistic details, outperforming other state-of-the-art real-world VSR algorithms. On synthetic datasets, as shown in Figure 3, MGLD-VSR can well reconstruct the structure while generating plausible details under complex degradations. On real-world videos, as shown by the example of sequence 008 in Figure 4, MGLD-VSR reconstructs clearer characters and patterns than competing methods. For the example of sequence 013, MGLD-VSR not only reconstructs the structure of the windows, and at the same time generates plausible exterior structure of the brick wall. More visual comparisons can be found in the supplementary materials." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 395, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 395, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 395, + 91, + 447, + 100 + ], + "type": "text", + "content": "MGLD-VSR" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 179, + 481, + 350 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "text", + "content": "Table 1: Quantitative comparison of different VSR models evaluated on popular VSR benchmarking datasets. Bold and underline represent the best and second best score, respectively. For synthetic datasets (REDS4, UDM10, SPMCS), we adopt full-reference metrics to evaluate different methods. For real-world video datasets (VideoLQ), we adopt no-reference metrics to compare different methods." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 179, + 481, + 350 + ], + "lines": [ + { + "bbox": [ + 132, + 179, + 481, + 350 + ], + "spans": [ + { + "bbox": [ + 132, + 179, + 481, + 350 + ], + "type": "table", + "html": "
DatasetsMetricsBicubicRVRTBSRGANRealESRGANStableSRDBVSRRealVSRRealBasicVSROurs
REDS4PSNR ↑22.1322.2922.0221.3422.0422.2921.7822.3522.46
SSIM ↑0.57300.57590.57020.56470.56830.57600.56630.58460.5723
LPIPS ↓0.62770.62210.48020.45090.45530.62210.62670.41900.3776
DISTS ↓0.36670.35540.21180.17500.18450.35560.36200.15810.1151
VMAF ↑2.12554.800526.241333.995526.69532.60752.590334.988434.3214
UDM10PSNR ↑25.7826.1825.2924.9425.5826.1425.0925.5625.99
SSIM ↑0.75590.76290.73230.74390.75920.76350.75240.74480.7548
LPIPS ↓0.44410.44130.42170.38840.36840.43980.43300.38970.3491
DISTS ↓0.27300.25750.19260.16510.14970.25850.26080.16370.1369
VMAF ↑5.464710.172431.087436.915833.82647.20486.007938.263839.3892
SPMCSPSNR ↑22.2822.5122.1221.6622.2622.5022.0322.4122.66
SSIM ↑0.56910.57530.57300.56940.58440.57640.57380.57360.5960
LPIPS ↓0.55720.54710.45090.44350.43210.54580.54870.44200.3990
DISTS ↓0.35930.34150.23070.21500.23510.34350.34790.21300.1934
VMAF ↑0.96565.353322.482827.693521.80603.75762.702427.451928.7032
VideoLQNIQE ↓9.13196.58884.21044.20364.44426.75046.33313.69153.5346
BRISQUE ↓72.740562.176825.240029.842826.309261.189757.276524.548421.9839
MUSIQ ↑21.678928.420452.221349.838047.804529.003924.218955.969152.7812
DOVER ↑0.36770.49070.71840.73090.71080.49980.52020.74010.7481
", + "image_path": "55f22dc9c116bf24440aad83758f454974968c85ece5dc625766e7048ebc4643.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 385, + 482, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 385, + 482, + 482 + ], + "spans": [ + { + "bbox": [ + 130, + 385, + 482, + 482 + ], + "type": "text", + "content": "Temporal Consistency. Compared with single image super-resolution, one major aspect of VSR is the temporal consistency of the generated videos. Following most previous works, in this paper, we adopt the average warping error (WE) [13,27] of the sequence to quantitatively measure the temporal consistency: " + }, + { + "bbox": [ + 130, + 385, + 482, + 482 + ], + "type": "inline_equation", + "content": "\\mathrm{WE} = \\frac{1}{N - 1}\\sum_{i = 1}^{N - 1}\\| \\hat{I}_{i + 1} - \\mathrm{Warp}(\\hat{I}_i,O_{b,i})\\| _1" + }, + { + "bbox": [ + 130, + 385, + 482, + 482 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 385, + 482, + 482 + ], + "type": "inline_equation", + "content": "\\hat{I}_i" + }, + { + "bbox": [ + 130, + 385, + 482, + 482 + ], + "type": "text", + "content": " is the predicted frame and " + }, + { + "bbox": [ + 130, + 385, + 482, + 482 + ], + "type": "inline_equation", + "content": "O_{b,i}" + }, + { + "bbox": [ + 130, + 385, + 482, + 482 + ], + "type": "text", + "content": " is the estimated backward optical flow [54] from ground-truth video. On real-world video datasets without ground-truth, the optical flow is computed on the predicted frames." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 484, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 484, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 484, + 482, + 628 + ], + "type": "text", + "content": "It is worth mentioning that, despite the wide use of WE for temporal consistency evaluation, it may not be able to faithfully reflect the real human perception on the video. For example, WE could be easily attacked by blurring the sequence, resulting much higher WE scores but lower video quality. The WE scores of competing methods are given in Table 2. One can see that our diffusion prior based method MGLD-VSR is not good at this metric because it tends to generate more details and textures, which are not favourable by WE. Nonetheless, the WE of MGLD-VSR is better than StableSR, which is also driven by the latent diffusion priors. To better showcase the effectiveness of our method, we visualize the temporal profiles of the VSR results produced by different algorithms in Figure 5. One can see that MGLD-VSR reconstructs more accurate details while maintaining good temporal consistency." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": "User Study. We also conduct a user study to further examine the effectiveness of the proposed algorithm in comparison with RealBasicVSR [11] and StableSR [48]. The study is conducted in pairs: given the LR video, the subject" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "text", + "content": "X. Yang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 114, + 206, + 154 + ], + "blocks": [ + { + "bbox": [ + 135, + 114, + 206, + 154 + ], + "lines": [ + { + "bbox": [ + 135, + 114, + 206, + 154 + ], + "spans": [ + { + "bbox": [ + 135, + 114, + 206, + 154 + ], + "type": "image", + "image_path": "1a21be49bca64884ee81ee330d398fa5ebaac7c1c601e8e5fe8736961f50206a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 141, + 155, + 201, + 161 + ], + "lines": [ + { + "bbox": [ + 141, + 155, + 201, + 161 + ], + "spans": [ + { + "bbox": [ + 141, + 155, + 201, + 161 + ], + "type": "text", + "content": "Sequence 011 from REDS4" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 211, + 114, + 252, + 154 + ], + "blocks": [ + { + "bbox": [ + 211, + 114, + 252, + 154 + ], + "lines": [ + { + "bbox": [ + 211, + 114, + 252, + 154 + ], + "spans": [ + { + "bbox": [ + 211, + 114, + 252, + 154 + ], + "type": "image", + "image_path": "8a42fa41971422baccc438633c04648760f223367e810a1f60bbd666b170b7ba.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 223, + 155, + 240, + 160 + ], + "lines": [ + { + "bbox": [ + 223, + 155, + 240, + 160 + ], + "spans": [ + { + "bbox": [ + 223, + 155, + 240, + 160 + ], + "type": "text", + "content": "Bicubic" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 257, + 114, + 298, + 154 + ], + "blocks": [ + { + "bbox": [ + 257, + 114, + 298, + 154 + ], + "lines": [ + { + "bbox": [ + 257, + 114, + 298, + 154 + ], + "spans": [ + { + "bbox": [ + 257, + 114, + 298, + 154 + ], + "type": "image", + "image_path": "b3b26876e14df09051d32da7e9b1be2577bbaa8595e4d090fb01bed9148200c5.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 260, + 155, + 294, + 161 + ], + "lines": [ + { + "bbox": [ + 260, + 155, + 294, + 161 + ], + "spans": [ + { + "bbox": [ + 260, + 155, + 294, + 161 + ], + "type": "text", + "content": "Real-ESRGAN" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 303, + 114, + 343, + 154 + ], + "blocks": [ + { + "bbox": [ + 303, + 114, + 343, + 154 + ], + "lines": [ + { + "bbox": [ + 303, + 114, + 343, + 154 + ], + "spans": [ + { + "bbox": [ + 303, + 114, + 343, + 154 + ], + "type": "image", + "image_path": "09490f3ead1d67dfa12550c5c9f9aea7dd6bac950b5555c0a4f00199911b956d.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 155, + 333, + 161 + ], + "lines": [ + { + "bbox": [ + 313, + 155, + 333, + 161 + ], + "spans": [ + { + "bbox": [ + 313, + 155, + 333, + 161 + ], + "type": "text", + "content": "StableSR" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 348, + 114, + 389, + 154 + ], + "blocks": [ + { + "bbox": [ + 348, + 114, + 389, + 154 + ], + "lines": [ + { + "bbox": [ + 348, + 114, + 389, + 154 + ], + "spans": [ + { + "bbox": [ + 348, + 114, + 389, + 154 + ], + "type": "image", + "image_path": "7378eabc63867bcda96fb3c85fcdabed4ad6c519f5ec49d655eb379bbd761c77.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 353, + 155, + 384, + 161 + ], + "lines": [ + { + "bbox": [ + 353, + 155, + 384, + 161 + ], + "spans": [ + { + "bbox": [ + 353, + 155, + 384, + 161 + ], + "type": "text", + "content": "RealBasicVSR" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 393, + 114, + 435, + 155 + ], + "blocks": [ + { + "bbox": [ + 393, + 114, + 435, + 155 + ], + "lines": [ + { + "bbox": [ + 393, + 114, + 435, + 155 + ], + "spans": [ + { + "bbox": [ + 393, + 114, + 435, + 155 + ], + "type": "image", + "image_path": "159c2189dcbabd27947074e6a56869999ff598cd02d109dc6bbcb69312fd269c.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 408, + 156, + 419, + 161 + ], + "lines": [ + { + "bbox": [ + 408, + 156, + 419, + 161 + ], + "spans": [ + { + "bbox": [ + 408, + 156, + 419, + 161 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 439, + 114, + 481, + 155 + ], + "blocks": [ + { + "bbox": [ + 439, + 114, + 481, + 155 + ], + "lines": [ + { + "bbox": [ + 439, + 114, + 481, + 155 + ], + "spans": [ + { + "bbox": [ + 439, + 114, + 481, + 155 + ], + "type": "image", + "image_path": "6a8edb8c0dbd6417bf15337a10049a46e926a2bff6602e810f7d5abe14c89214.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 455, + 156, + 464, + 160 + ], + "lines": [ + { + "bbox": [ + 455, + 156, + 464, + 160 + ], + "spans": [ + { + "bbox": [ + 455, + 156, + 464, + 160 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 135, + 163, + 206, + 204 + ], + "blocks": [ + { + "bbox": [ + 135, + 163, + 206, + 204 + ], + "lines": [ + { + "bbox": [ + 135, + 163, + 206, + 204 + ], + "spans": [ + { + "bbox": [ + 135, + 163, + 206, + 204 + ], + "type": "image", + "image_path": "26f5f2891c7c911908c82f8aaa13f12b673a13cc57814840ab5158e88e1df629.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 139, + 205, + 203, + 210 + ], + "lines": [ + { + "bbox": [ + 139, + 205, + 203, + 210 + ], + "spans": [ + { + "bbox": [ + 139, + 205, + 203, + 210 + ], + "type": "text", + "content": "Sequence band from UDM10" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 211, + 163, + 252, + 204 + ], + "blocks": [ + { + "bbox": [ + 211, + 163, + 252, + 204 + ], + "lines": [ + { + "bbox": [ + 211, + 163, + 252, + 204 + ], + "spans": [ + { + "bbox": [ + 211, + 163, + 252, + 204 + ], + "type": "image", + "image_path": "1bfc860ed73a6becfed9f2e5255c7facc155d14f32eb369ccbb13c826dd7ea47.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 223, + 205, + 239, + 210 + ], + "lines": [ + { + "bbox": [ + 223, + 205, + 239, + 210 + ], + "spans": [ + { + "bbox": [ + 223, + 205, + 239, + 210 + ], + "type": "text", + "content": "Bicubic" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 257, + 163, + 298, + 204 + ], + "blocks": [ + { + "bbox": [ + 257, + 163, + 298, + 204 + ], + "lines": [ + { + "bbox": [ + 257, + 163, + 298, + 204 + ], + "spans": [ + { + "bbox": [ + 257, + 163, + 298, + 204 + ], + "type": "image", + "image_path": "9805048d8ad0be3e6f49ffbe4a4794d0dad08f16ea7593523fb3898f170e054b.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 261, + 205, + 294, + 210 + ], + "lines": [ + { + "bbox": [ + 261, + 205, + 294, + 210 + ], + "spans": [ + { + "bbox": [ + 261, + 205, + 294, + 210 + ], + "type": "text", + "content": "Real-ESRGAN" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 303, + 163, + 343, + 204 + ], + "blocks": [ + { + "bbox": [ + 303, + 163, + 343, + 204 + ], + "lines": [ + { + "bbox": [ + 303, + 163, + 343, + 204 + ], + "spans": [ + { + "bbox": [ + 303, + 163, + 343, + 204 + ], + "type": "image", + "image_path": "a4d544fc5392c0a3d07024189fe4094b2a74177237830c5ce3f75aa91ae924d1.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 205, + 333, + 210 + ], + "lines": [ + { + "bbox": [ + 313, + 205, + 333, + 210 + ], + "spans": [ + { + "bbox": [ + 313, + 205, + 333, + 210 + ], + "type": "text", + "content": "StableSR" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 348, + 163, + 389, + 204 + ], + "blocks": [ + { + "bbox": [ + 348, + 163, + 389, + 204 + ], + "lines": [ + { + "bbox": [ + 348, + 163, + 389, + 204 + ], + "spans": [ + { + "bbox": [ + 348, + 163, + 389, + 204 + ], + "type": "image", + "image_path": "e81829d39c8daae6cd1402e573c7aadc63b3dbe8e6578d69e93e8df8e7f1f16d.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 353, + 205, + 385, + 210 + ], + "lines": [ + { + "bbox": [ + 353, + 205, + 385, + 210 + ], + "spans": [ + { + "bbox": [ + 353, + 205, + 385, + 210 + ], + "type": "text", + "content": "RealBasicVSR" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 394, + 163, + 435, + 204 + ], + "blocks": [ + { + "bbox": [ + 394, + 163, + 435, + 204 + ], + "lines": [ + { + "bbox": [ + 394, + 163, + 435, + 204 + ], + "spans": [ + { + "bbox": [ + 394, + 163, + 435, + 204 + ], + "type": "image", + "image_path": "22ea8ba186e6d2f8f77a1abd7a78e8be7dd12d563e258d7dae4768bfa585194c.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 408, + 205, + 419, + 210 + ], + "lines": [ + { + "bbox": [ + 408, + 205, + 419, + 210 + ], + "spans": [ + { + "bbox": [ + 408, + 205, + 419, + 210 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 439, + 163, + 481, + 204 + ], + "blocks": [ + { + "bbox": [ + 439, + 163, + 481, + 204 + ], + "lines": [ + { + "bbox": [ + 439, + 163, + 481, + 204 + ], + "spans": [ + { + "bbox": [ + 439, + 163, + 481, + 204 + ], + "type": "image", + "image_path": "dc9b5fcbd53908c7cc836cbb2fff032d8a3f2ae99f7e8dc560a712ee58426030.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 456, + 205, + 464, + 209 + ], + "lines": [ + { + "bbox": [ + 456, + 205, + 464, + 209 + ], + "spans": [ + { + "bbox": [ + 456, + 205, + 464, + 209 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 135, + 213, + 206, + 253 + ], + "blocks": [ + { + "bbox": [ + 135, + 213, + 206, + 253 + ], + "lines": [ + { + "bbox": [ + 135, + 213, + 206, + 253 + ], + "spans": [ + { + "bbox": [ + 135, + 213, + 206, + 253 + ], + "type": "image", + "image_path": "e892820ceaf4c9765a5508bf22ee3de820fb930065cd95e1b589ef69247898b2.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 137, + 254, + 204, + 259 + ], + "lines": [ + { + "bbox": [ + 137, + 254, + 204, + 259 + ], + "spans": [ + { + "bbox": [ + 137, + 254, + 204, + 259 + ], + "type": "text", + "content": "Sequence 000002 from SPMCS" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 211, + 213, + 252, + 253 + ], + "blocks": [ + { + "bbox": [ + 211, + 213, + 252, + 253 + ], + "lines": [ + { + "bbox": [ + 211, + 213, + 252, + 253 + ], + "spans": [ + { + "bbox": [ + 211, + 213, + 252, + 253 + ], + "type": "image", + "image_path": "d9f9eaf4732ee295736f2928923672cd5ef09e5fe4f39767958d37fcdd0816ad.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 223, + 254, + 240, + 259 + ], + "lines": [ + { + "bbox": [ + 223, + 254, + 240, + 259 + ], + "spans": [ + { + "bbox": [ + 223, + 254, + 240, + 259 + ], + "type": "text", + "content": "Bicubic" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 257, + 213, + 298, + 253 + ], + "blocks": [ + { + "bbox": [ + 257, + 213, + 298, + 253 + ], + "lines": [ + { + "bbox": [ + 257, + 213, + 298, + 253 + ], + "spans": [ + { + "bbox": [ + 257, + 213, + 298, + 253 + ], + "type": "image", + "image_path": "2d0c996430a220b3c1625fa3804172d3aea2e0cf06b8c14460137cc8597e3e1c.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 261, + 254, + 294, + 259 + ], + "lines": [ + { + "bbox": [ + 261, + 254, + 294, + 259 + ], + "spans": [ + { + "bbox": [ + 261, + 254, + 294, + 259 + ], + "type": "text", + "content": "Real-ESRGAN" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 132, + 269, + 480, + 291 + ], + "lines": [ + { + "bbox": [ + 132, + 269, + 480, + 291 + ], + "spans": [ + { + "bbox": [ + 132, + 269, + 480, + 291 + ], + "type": "text", + "content": "Fig. 3: Qualitative comparison on synthetic datasets (REDS4, UDM10, SPMCS) for " + }, + { + "bbox": [ + 132, + 269, + 480, + 291 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 132, + 269, + 480, + 291 + ], + "type": "text", + "content": " VSR. (Zoom-in for best view.)" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_caption" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 303, + 213, + 343, + 253 + ], + "blocks": [ + { + "bbox": [ + 303, + 213, + 343, + 253 + ], + "lines": [ + { + "bbox": [ + 303, + 213, + 343, + 253 + ], + "spans": [ + { + "bbox": [ + 303, + 213, + 343, + 253 + ], + "type": "image", + "image_path": "e001fbe85d63fd5ddc80b8bf5b0989087430cca1b6d75a43733b4c7bd05cb1ca.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 254, + 333, + 259 + ], + "lines": [ + { + "bbox": [ + 313, + 254, + 333, + 259 + ], + "spans": [ + { + "bbox": [ + 313, + 254, + 333, + 259 + ], + "type": "text", + "content": "StableSR" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_caption" + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 348, + 213, + 389, + 253 + ], + "blocks": [ + { + "bbox": [ + 348, + 213, + 389, + 253 + ], + "lines": [ + { + "bbox": [ + 348, + 213, + 389, + 253 + ], + "spans": [ + { + "bbox": [ + 348, + 213, + 389, + 253 + ], + "type": "image", + "image_path": "9f4897da176e1b91f6ccaa78375773acaf6f1037050de6131a49339d3f6ec88d.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 353, + 254, + 386, + 259 + ], + "lines": [ + { + "bbox": [ + 353, + 254, + 386, + 259 + ], + "spans": [ + { + "bbox": [ + 353, + 254, + 386, + 259 + ], + "type": "text", + "content": "RealBasicVSR" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_caption" + } + ], + "index": 38 + }, + { + "type": "image", + "bbox": [ + 393, + 213, + 435, + 253 + ], + "blocks": [ + { + "bbox": [ + 393, + 213, + 435, + 253 + ], + "lines": [ + { + "bbox": [ + 393, + 213, + 435, + 253 + ], + "spans": [ + { + "bbox": [ + 393, + 213, + 435, + 253 + ], + "type": "image", + "image_path": "b4fbbdb03af3b88df08c54ec8f8ae09d3f61761a3abb01106b13ff72b1f4bb77.jpg" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 408, + 254, + 420, + 259 + ], + "lines": [ + { + "bbox": [ + 408, + 254, + 420, + 259 + ], + "spans": [ + { + "bbox": [ + 408, + 254, + 420, + 259 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_caption" + } + ], + "index": 40 + }, + { + "type": "image", + "bbox": [ + 439, + 213, + 481, + 253 + ], + "blocks": [ + { + "bbox": [ + 439, + 213, + 481, + 253 + ], + "lines": [ + { + "bbox": [ + 439, + 213, + 481, + 253 + ], + "spans": [ + { + "bbox": [ + 439, + 213, + 481, + 253 + ], + "type": "image", + "image_path": "a9a1a880f31d483a04caf95d4e88f63e9edc6cd8380e9f32e5de11bb85587e0b.jpg" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 456, + 254, + 463, + 259 + ], + "lines": [ + { + "bbox": [ + 456, + 254, + 463, + 259 + ], + "spans": [ + { + "bbox": [ + 456, + 254, + 463, + 259 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_caption" + } + ], + "index": 42 + }, + { + "type": "table", + "bbox": [ + 133, + 342, + 481, + 394 + ], + "blocks": [ + { + "bbox": [ + 131, + 302, + 480, + 324 + ], + "lines": [ + { + "bbox": [ + 131, + 302, + 480, + 324 + ], + "spans": [ + { + "bbox": [ + 131, + 302, + 480, + 324 + ], + "type": "text", + "content": "Table 2: Temporal consistency comparison of competing VSR methods by using WE " + }, + { + "bbox": [ + 131, + 302, + 480, + 324 + ], + "type": "inline_equation", + "content": "\\downarrow" + }, + { + "bbox": [ + 131, + 302, + 480, + 324 + ], + "type": "text", + "content": ". Bold and underline represent the best and second best scores, respectively." + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 342, + 481, + 394 + ], + "lines": [ + { + "bbox": [ + 133, + 342, + 481, + 394 + ], + "spans": [ + { + "bbox": [ + 133, + 342, + 481, + 394 + ], + "type": "table", + "html": "
DatasetsBicubicRVRTBSRGANRealESRGANStableSRDBVSRRealVSRRealBasicVSROurs
REDS48.576.5010.2511.9111.818.518.6110.0910.08
UDM103.103.064.824.994.803.042.754.334.55
SPMCS3.013.585.187.124.623.584.182.723.87
VideoLQ5.486.387.727.649.335.985.746.767.56
", + "image_path": "97aa431f058833bc287cd30d604226b4c806f02bdba49fcea5147bec2d149c5b.jpg" + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "table_body" + } + ], + "index": 46 + }, + { + "bbox": [ + 130, + 426, + 479, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 479, + 462 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 479, + 462 + ], + "type": "text", + "content": "is asked to choose the better one from two HR videos obtained by different methods, considering both the reconstruction quality and the temporal consistency. As shown in Figure 6, our method gains more votes in the competitions." + } + ] + } + ], + "index": 47 + }, + { + "bbox": [ + 131, + 479, + 238, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 479, + 238, + 491 + ], + "spans": [ + { + "bbox": [ + 131, + 479, + 238, + 491 + ], + "type": "text", + "content": "4.3 Ablation Study" + } + ] + } + ], + "index": 48 + }, + { + "bbox": [ + 130, + 498, + 480, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 480, + 557 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 480, + 557 + ], + "type": "text", + "content": "In this section, we conduct a series of analyses on the proposed framework, including its Motion-guided Diffusion Sampling (MDS) module and Temporal-aware Sequence Decoder (TSD) module, and employ the VideoLQ dataset for experiments. We adopt WE to measure temporal consistency and no-reference metrics NIQE, BRISQUE, MUSIQ and DOVER to measure perceptual quality." + } + ] + } + ], + "index": 49 + }, + { + "bbox": [ + 130, + 558, + 480, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 480, + 617 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 480, + 617 + ], + "type": "text", + "content": "Different Components of MGLD-VSR. As can seen from Table 3, the introduction of MDS improves all the metrics over the baseline (i.e., without using MDS and TSD), suggesting that MDS helps not only in reducing the temporal differences (i.e., WE), but also enhancing the perceptual quality of the reconstructed video sequence." + } + ] + } + ], + "index": 50 + }, + { + "bbox": [ + 130, + 617, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 480, + 665 + ], + "type": "text", + "content": "The incorporation of MDS aids in maintaining perceptual continuity by ensuring that the restored frames could transition smoothly from one to the next. This contributes to more natural and visually pleasing VSR results. The incorporation of TSD also yields better results across all the metrics compared to" + } + ] + } + ], + "index": 51 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 395, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 395, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 395, + 91, + 447, + 100 + ], + "type": "text", + "content": "MGLD-VSR" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 133, + 241, + 215 + ], + "blocks": [ + { + "bbox": [ + 133, + 133, + 241, + 215 + ], + "lines": [ + { + "bbox": [ + 133, + 133, + 241, + 215 + ], + "spans": [ + { + "bbox": [ + 133, + 133, + 241, + 215 + ], + "type": "image", + "image_path": "66d1aed951c44d33847292360718af7dba651bcb9d8c3e1d3693f08aeb3682b4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 147, + 217, + 227, + 225 + ], + "lines": [ + { + "bbox": [ + 147, + 217, + 227, + 225 + ], + "spans": [ + { + "bbox": [ + 147, + 217, + 227, + 225 + ], + "type": "text", + "content": "Sequence 008 from VideoLQ" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 244, + 114, + 296, + 167 + ], + "blocks": [ + { + "bbox": [ + 244, + 114, + 296, + 167 + ], + "lines": [ + { + "bbox": [ + 244, + 114, + 296, + 167 + ], + "spans": [ + { + "bbox": [ + 244, + 114, + 296, + 167 + ], + "type": "image", + "image_path": "a696e082a10641f9b191d58ce98491f9830d067de57894e4b0de284ebb11a8b5.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 258, + 167, + 281, + 173 + ], + "lines": [ + { + "bbox": [ + 258, + 167, + 281, + 173 + ], + "spans": [ + { + "bbox": [ + 258, + 167, + 281, + 173 + ], + "type": "text", + "content": "Bicubic" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 304, + 114, + 358, + 167 + ], + "blocks": [ + { + "bbox": [ + 304, + 114, + 358, + 167 + ], + "lines": [ + { + "bbox": [ + 304, + 114, + 358, + 167 + ], + "spans": [ + { + "bbox": [ + 304, + 114, + 358, + 167 + ], + "type": "image", + "image_path": "ca382d079fecafd21112c6abe1bb792379ae77338d04df50b17b97742b2aa06e.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 317, + 167, + 345, + 173 + ], + "lines": [ + { + "bbox": [ + 317, + 167, + 345, + 173 + ], + "spans": [ + { + "bbox": [ + 317, + 167, + 345, + 173 + ], + "type": "text", + "content": "BSRGAN" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 366, + 114, + 419, + 167 + ], + "blocks": [ + { + "bbox": [ + 366, + 114, + 419, + 167 + ], + "lines": [ + { + "bbox": [ + 366, + 114, + 419, + 167 + ], + "spans": [ + { + "bbox": [ + 366, + 114, + 419, + 167 + ], + "type": "image", + "image_path": "0bcebf50697bdf5a9e9916fd0f55415b6198bbb3cfd2bb5f2c4890f5ae6b2ec1.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 371, + 167, + 413, + 173 + ], + "lines": [ + { + "bbox": [ + 371, + 167, + 413, + 173 + ], + "spans": [ + { + "bbox": [ + 371, + 167, + 413, + 173 + ], + "type": "text", + "content": "Real-ESRGAN" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 428, + 114, + 481, + 167 + ], + "blocks": [ + { + "bbox": [ + 428, + 114, + 481, + 167 + ], + "lines": [ + { + "bbox": [ + 428, + 114, + 481, + 167 + ], + "spans": [ + { + "bbox": [ + 428, + 114, + 481, + 167 + ], + "type": "image", + "image_path": "a3b03c9d59434da6ff64abd589aca844099c8b7f96b05c19bb17c73f465fb1dc.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 441, + 167, + 468, + 173 + ], + "lines": [ + { + "bbox": [ + 441, + 167, + 468, + 173 + ], + "spans": [ + { + "bbox": [ + 441, + 167, + 468, + 173 + ], + "type": "text", + "content": "StableSR" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 133, + 267, + 241, + 329 + ], + "blocks": [ + { + "bbox": [ + 133, + 267, + 241, + 329 + ], + "lines": [ + { + "bbox": [ + 133, + 267, + 241, + 329 + ], + "spans": [ + { + "bbox": [ + 133, + 267, + 241, + 329 + ], + "type": "image", + "image_path": "5feba6f652cf468516ad5f3f1a0f0cc68bb7a72c6d7efcdeb7bd915a6def1d3a.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 147, + 331, + 227, + 339 + ], + "lines": [ + { + "bbox": [ + 147, + 331, + 227, + 339 + ], + "spans": [ + { + "bbox": [ + 147, + 331, + 227, + 339 + ], + "type": "text", + "content": "Sequence 013 from VideoLQ" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 244, + 239, + 296, + 292 + ], + "blocks": [ + { + "bbox": [ + 244, + 239, + 296, + 292 + ], + "lines": [ + { + "bbox": [ + 244, + 239, + 296, + 292 + ], + "spans": [ + { + "bbox": [ + 244, + 239, + 296, + 292 + ], + "type": "image", + "image_path": "489d4622053853c6083303f2bc9794fa27e8414fa17759937c74b9f885f36d02.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 258, + 292, + 280, + 298 + ], + "lines": [ + { + "bbox": [ + 258, + 292, + 280, + 298 + ], + "spans": [ + { + "bbox": [ + 258, + 292, + 280, + 298 + ], + "type": "text", + "content": "Bicubic" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 244, + 301, + 296, + 353 + ], + "blocks": [ + { + "bbox": [ + 244, + 301, + 296, + 353 + ], + "lines": [ + { + "bbox": [ + 244, + 301, + 296, + 353 + ], + "spans": [ + { + "bbox": [ + 244, + 301, + 296, + 353 + ], + "type": "image", + "image_path": "48127e6a671b247d61e8adb6788c8ca1d73a28f0628d6b83c36fb756179b2ac2.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 257, + 354, + 282, + 360 + ], + "lines": [ + { + "bbox": [ + 257, + 354, + 282, + 360 + ], + "spans": [ + { + "bbox": [ + 257, + 354, + 282, + 360 + ], + "type": "text", + "content": "RealVSR" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 305, + 239, + 357, + 292 + ], + "blocks": [ + { + "bbox": [ + 305, + 239, + 357, + 292 + ], + "lines": [ + { + "bbox": [ + 305, + 239, + 357, + 292 + ], + "spans": [ + { + "bbox": [ + 305, + 239, + 357, + 292 + ], + "type": "image", + "image_path": "712dcec4f5b3baa9c86dec11dd36c6a3c77f6cfd3eb20942b4881645e8526579.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 317, + 292, + 345, + 298 + ], + "lines": [ + { + "bbox": [ + 317, + 292, + 345, + 298 + ], + "spans": [ + { + "bbox": [ + 317, + 292, + 345, + 298 + ], + "type": "text", + "content": "BSRGAN" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 305, + 301, + 357, + 353 + ], + "blocks": [ + { + "bbox": [ + 305, + 301, + 357, + 353 + ], + "lines": [ + { + "bbox": [ + 305, + 301, + 357, + 353 + ], + "spans": [ + { + "bbox": [ + 305, + 301, + 357, + 353 + ], + "type": "image", + "image_path": "bc3cf4999da48aa0574ba136619ec6f812d21b4e7f1bf7d02e54ff136187f4ce.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 320, + 354, + 343, + 360 + ], + "lines": [ + { + "bbox": [ + 320, + 354, + 343, + 360 + ], + "spans": [ + { + "bbox": [ + 320, + 354, + 343, + 360 + ], + "type": "text", + "content": "DBVSR" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 366, + 239, + 419, + 292 + ], + "blocks": [ + { + "bbox": [ + 366, + 239, + 419, + 292 + ], + "lines": [ + { + "bbox": [ + 366, + 239, + 419, + 292 + ], + "spans": [ + { + "bbox": [ + 366, + 239, + 419, + 292 + ], + "type": "image", + "image_path": "da61c9ee62dab88f5a59bff79bd7ff741c8631ae9fcf3858f10d20ab435faae2.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 371, + 292, + 413, + 298 + ], + "lines": [ + { + "bbox": [ + 371, + 292, + 413, + 298 + ], + "spans": [ + { + "bbox": [ + 371, + 292, + 413, + 298 + ], + "type": "text", + "content": "Real-ESRGAN" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 367, + 301, + 419, + 353 + ], + "blocks": [ + { + "bbox": [ + 367, + 301, + 419, + 353 + ], + "lines": [ + { + "bbox": [ + 367, + 301, + 419, + 353 + ], + "spans": [ + { + "bbox": [ + 367, + 301, + 419, + 353 + ], + "type": "image", + "image_path": "ed52997a91f5dd302afdfa831f62fcc94bf6989c4ced72f051d355c4cd485e60.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 372, + 354, + 413, + 360 + ], + "lines": [ + { + "bbox": [ + 372, + 354, + 413, + 360 + ], + "spans": [ + { + "bbox": [ + 372, + 354, + 413, + 360 + ], + "type": "text", + "content": "RealBasicVSR" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 428, + 175, + 481, + 228 + ], + "blocks": [ + { + "bbox": [ + 428, + 175, + 481, + 228 + ], + "lines": [ + { + "bbox": [ + 428, + 175, + 481, + 228 + ], + "spans": [ + { + "bbox": [ + 428, + 175, + 481, + 228 + ], + "type": "image", + "image_path": "c0d3c2ffa3de0e90eaff7bfa46f74083b1dc6876dd29fac78000ffe9ed24f811.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 447, + 228, + 462, + 235 + ], + "lines": [ + { + "bbox": [ + 447, + 228, + 462, + 235 + ], + "spans": [ + { + "bbox": [ + 447, + 228, + 462, + 235 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 428, + 239, + 481, + 292 + ], + "blocks": [ + { + "bbox": [ + 428, + 239, + 481, + 292 + ], + "lines": [ + { + "bbox": [ + 428, + 239, + 481, + 292 + ], + "spans": [ + { + "bbox": [ + 428, + 239, + 481, + 292 + ], + "type": "image", + "image_path": "b46a49a679e57918d39e6bd9e82df7a38e68716fbee0f3df5d93ad6a1d12d718.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 441, + 292, + 468, + 298 + ], + "lines": [ + { + "bbox": [ + 441, + 292, + 468, + 298 + ], + "spans": [ + { + "bbox": [ + 441, + 292, + 468, + 298 + ], + "type": "text", + "content": "StableSR" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 428, + 301, + 481, + 353 + ], + "blocks": [ + { + "bbox": [ + 428, + 301, + 481, + 353 + ], + "lines": [ + { + "bbox": [ + 428, + 301, + 481, + 353 + ], + "spans": [ + { + "bbox": [ + 428, + 301, + 481, + 353 + ], + "type": "image", + "image_path": "6ace3dbd55c52f74aebd9e603790d6d9a06104909e473232527b96b4d066c890.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 447, + 354, + 462, + 360 + ], + "lines": [ + { + "bbox": [ + 447, + 354, + 462, + 360 + ], + "spans": [ + { + "bbox": [ + 447, + 354, + 462, + 360 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 133, + 402, + 196, + 439 + ], + "blocks": [ + { + "bbox": [ + 131, + 369, + 482, + 392 + ], + "lines": [ + { + "bbox": [ + 131, + 369, + 482, + 392 + ], + "spans": [ + { + "bbox": [ + 131, + 369, + 482, + 392 + ], + "type": "text", + "content": "Fig. 4: Qualitative comparison on the real-world video dataset (VideoLQ) for " + }, + { + "bbox": [ + 131, + 369, + 482, + 392 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 131, + 369, + 482, + 392 + ], + "type": "text", + "content": " video SR. (Zoom-in for best view.)" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 133, + 402, + 196, + 439 + ], + "lines": [ + { + "bbox": [ + 133, + 402, + 196, + 439 + ], + "spans": [ + { + "bbox": [ + 133, + 402, + 196, + 439 + ], + "type": "image", + "image_path": "76cd62080f30b42c8b4c191efbab7bdb3ade7c1707e2a7e67eb2f87edbf7b4bd.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 197, + 402, + 242, + 439 + ], + "blocks": [ + { + "bbox": [ + 197, + 402, + 242, + 439 + ], + "lines": [ + { + "bbox": [ + 197, + 402, + 242, + 439 + ], + "spans": [ + { + "bbox": [ + 197, + 402, + 242, + 439 + ], + "type": "image", + "image_path": "66589e67a61bd62f29bbb514747cd0e51a37359d6c51a3323433769bd1bc7f4e.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 244, + 402, + 290, + 439 + ], + "blocks": [ + { + "bbox": [ + 244, + 402, + 290, + 439 + ], + "lines": [ + { + "bbox": [ + 244, + 402, + 290, + 439 + ], + "spans": [ + { + "bbox": [ + 244, + 402, + 290, + 439 + ], + "type": "image", + "image_path": "622ca7949d5e8f5666ae9765a52dc45c2f7e9e9c74c5f4bd23bbe66c68f0987a.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 292, + 402, + 337, + 439 + ], + "blocks": [ + { + "bbox": [ + 292, + 402, + 337, + 439 + ], + "lines": [ + { + "bbox": [ + 292, + 402, + 337, + 439 + ], + "spans": [ + { + "bbox": [ + 292, + 402, + 337, + 439 + ], + "type": "image", + "image_path": "7dcd77593b048342fd1c8e8d7622d1850a294beba46d7e2ff98e4a0311a9e869.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 339, + 402, + 384, + 439 + ], + "blocks": [ + { + "bbox": [ + 339, + 402, + 384, + 439 + ], + "lines": [ + { + "bbox": [ + 339, + 402, + 384, + 439 + ], + "spans": [ + { + "bbox": [ + 339, + 402, + 384, + 439 + ], + "type": "image", + "image_path": "f07208c77bcd272c498bb30726320220336bd96c8e0d2262a9e771af03eb22ad.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 387, + 402, + 432, + 439 + ], + "blocks": [ + { + "bbox": [ + 387, + 402, + 432, + 439 + ], + "lines": [ + { + "bbox": [ + 387, + 402, + 432, + 439 + ], + "spans": [ + { + "bbox": [ + 387, + 402, + 432, + 439 + ], + "type": "image", + "image_path": "d3bc162df8c76ab72a924e375dffdd8b619f1a618ce5c6a1fce49479cfa9ec00.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + } + ], + "index": 38 + }, + { + "type": "image", + "bbox": [ + 434, + 402, + 480, + 439 + ], + "blocks": [ + { + "bbox": [ + 434, + 402, + 480, + 439 + ], + "lines": [ + { + "bbox": [ + 434, + 402, + 480, + 439 + ], + "spans": [ + { + "bbox": [ + 434, + 402, + 480, + 439 + ], + "type": "image", + "image_path": "26afcc10b95caa60126528e6429bbdb47e5eef4b8e06bf4829efcb64be8a2683.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + } + ], + "index": 39 + }, + { + "type": "image", + "bbox": [ + 133, + 443, + 196, + 479 + ], + "blocks": [ + { + "bbox": [ + 133, + 443, + 196, + 479 + ], + "lines": [ + { + "bbox": [ + 133, + 443, + 196, + 479 + ], + "spans": [ + { + "bbox": [ + 133, + 443, + 196, + 479 + ], + "type": "image", + "image_path": "7134fc49cba0d0d657c3cb94b509b31a6d10d24d3bdff82a1fc68a82e1c0c756.jpg" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_body" + } + ], + "index": 40 + }, + { + "type": "image", + "bbox": [ + 197, + 443, + 242, + 479 + ], + "blocks": [ + { + "bbox": [ + 197, + 443, + 242, + 479 + ], + "lines": [ + { + "bbox": [ + 197, + 443, + 242, + 479 + ], + "spans": [ + { + "bbox": [ + 197, + 443, + 242, + 479 + ], + "type": "image", + "image_path": "89927672146a443faa86ceddbacb6bbbd9730af44148e53977bfb4bfeb605361.jpg" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_body" + } + ], + "index": 41 + }, + { + "type": "image", + "bbox": [ + 244, + 443, + 290, + 479 + ], + "blocks": [ + { + "bbox": [ + 244, + 443, + 290, + 479 + ], + "lines": [ + { + "bbox": [ + 244, + 443, + 290, + 479 + ], + "spans": [ + { + "bbox": [ + 244, + 443, + 290, + 479 + ], + "type": "image", + "image_path": "c76c5094e3a20fd31a9738da9bfa69bb14819e7a30d4c0a93d8d1ac1e29d6941.jpg" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_body" + } + ], + "index": 42 + }, + { + "type": "image", + "bbox": [ + 292, + 443, + 337, + 479 + ], + "blocks": [ + { + "bbox": [ + 292, + 443, + 337, + 479 + ], + "lines": [ + { + "bbox": [ + 292, + 443, + 337, + 479 + ], + "spans": [ + { + "bbox": [ + 292, + 443, + 337, + 479 + ], + "type": "image", + "image_path": "c6e99f77deaa80429e501bac068865126af5ef8663fcb467a1060c4c8637e1a9.jpg" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_body" + } + ], + "index": 43 + }, + { + "type": "image", + "bbox": [ + 339, + 443, + 384, + 479 + ], + "blocks": [ + { + "bbox": [ + 339, + 443, + 384, + 479 + ], + "lines": [ + { + "bbox": [ + 339, + 443, + 384, + 479 + ], + "spans": [ + { + "bbox": [ + 339, + 443, + 384, + 479 + ], + "type": "image", + "image_path": "efd09aca113e4c94e9574cabfc958431a893fbbcbffae3f92b9cfc394c27d1b7.jpg" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_body" + } + ], + "index": 44 + }, + { + "type": "image", + "bbox": [ + 387, + 443, + 432, + 479 + ], + "blocks": [ + { + "bbox": [ + 387, + 443, + 432, + 479 + ], + "lines": [ + { + "bbox": [ + 387, + 443, + 432, + 479 + ], + "spans": [ + { + "bbox": [ + 387, + 443, + 432, + 479 + ], + "type": "image", + "image_path": "fee8658ada73b209db1beb2c7c6f548706e5001e4849205f645fb8c996656898.jpg" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_body" + } + ], + "index": 45 + }, + { + "type": "image", + "bbox": [ + 434, + 443, + 480, + 479 + ], + "blocks": [ + { + "bbox": [ + 434, + 443, + 480, + 479 + ], + "lines": [ + { + "bbox": [ + 434, + 443, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 434, + 443, + 480, + 479 + ], + "type": "image", + "image_path": "dec93126028840fd97e1dfa59a2a5e7130cba18b32f864374f615b07f351d052.jpg" + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_body" + } + ], + "index": 46 + }, + { + "type": "image", + "bbox": [ + 134, + 521, + 305, + 600 + ], + "blocks": [ + { + "bbox": [ + 174, + 499, + 440, + 511 + ], + "lines": [ + { + "bbox": [ + 174, + 499, + 440, + 511 + ], + "spans": [ + { + "bbox": [ + 174, + 499, + 440, + 511 + ], + "type": "text", + "content": "Fig. 5: Temporal profiles of competing real-world VSR methods." + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 134, + 521, + 305, + 600 + ], + "lines": [ + { + "bbox": [ + 134, + 521, + 305, + 600 + ], + "spans": [ + { + "bbox": [ + 134, + 521, + 305, + 600 + ], + "type": "image", + "image_path": "b883b80d49db29bf4b07b2bb73b5c39b623578280af3134b3458c93b8e302c3c.jpg" + } + ] + } + ], + "index": 48, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 609, + 482, + 632 + ], + "lines": [ + { + "bbox": [ + 131, + 609, + 482, + 632 + ], + "spans": [ + { + "bbox": [ + 131, + 609, + 482, + 632 + ], + "type": "text", + "content": "Fig. 6: User study on 20 videos from VideoLQ evaluated by 10 subjects. MGLD-VSR gains more votes than competing methods." + } + ] + } + ], + "index": 50, + "angle": 0, + "type": "image_caption" + } + ], + "index": 48 + }, + { + "type": "image", + "bbox": [ + 307, + 521, + 480, + 600 + ], + "blocks": [ + { + "bbox": [ + 307, + 521, + 480, + 600 + ], + "lines": [ + { + "bbox": [ + 307, + 521, + 480, + 600 + ], + "spans": [ + { + "bbox": [ + 307, + 521, + 480, + 600 + ], + "type": "image", + "image_path": "2714013b4093a5f8bc2ffd51805f1956e184ee627ec42971ea1746429b91f188.jpg" + } + ] + } + ], + "index": 49, + "angle": 0, + "type": "image_body" + } + ], + "index": 49 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 224, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 224, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 224, + 102 + ], + "type": "text", + "content": "X. Yang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 200, + 157, + 410, + 209 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 479, + 147 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 479, + 147 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 479, + 147 + ], + "type": "text", + "content": "Table 3: Ablation study on the different components of the proposed algorithm. \"MDS\" refers to Motion-guided Diffusion Sampling and \"TSD\" refers to Temporal-aware Sequence Decoder." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 200, + 157, + 410, + 209 + ], + "lines": [ + { + "bbox": [ + 200, + 157, + 410, + 209 + ], + "spans": [ + { + "bbox": [ + 200, + 157, + 410, + 209 + ], + "type": "table", + "html": "
MDS TSDWE ↓NIQE ↓BRISQUE ↓MUSIQ ↑DOVER ↑
✓9.30694.109825.324951.56770.7412
8.98483.987624.084052.41430.7435
✓8.36793.628122.080652.68740.7468
✓7.54783.534621.983952.78120.7481
", + "image_path": "b2c5116369287659ad7031dbfb8b1224c306e9e278b2d3abc6683cbb87698ca0.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 141, + 224, + 253, + 309 + ], + "blocks": [ + { + "bbox": [ + 141, + 224, + 253, + 309 + ], + "lines": [ + { + "bbox": [ + 141, + 224, + 253, + 309 + ], + "spans": [ + { + "bbox": [ + 141, + 224, + 253, + 309 + ], + "type": "image", + "image_path": "774778f79401f7818e2b2b94abf61de0c3f753e8e0c90daa247afcebf06dd8f2.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 153, + 311, + 242, + 319 + ], + "lines": [ + { + "bbox": [ + 153, + 311, + 242, + 319 + ], + "spans": [ + { + "bbox": [ + 153, + 311, + 242, + 319 + ], + "type": "text", + "content": "Sequence 002 from VideoLQ" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 255, + 224, + 297, + 308 + ], + "blocks": [ + { + "bbox": [ + 255, + 224, + 297, + 308 + ], + "lines": [ + { + "bbox": [ + 255, + 224, + 297, + 308 + ], + "spans": [ + { + "bbox": [ + 255, + 224, + 297, + 308 + ], + "type": "image", + "image_path": "1c67da428bce5bf1d85d6d26285eacfd3f1ad94446e5a853cdeb3029e3ce45fc.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 274, + 311, + 279, + 318 + ], + "lines": [ + { + "bbox": [ + 274, + 311, + 279, + 318 + ], + "spans": [ + { + "bbox": [ + 274, + 311, + 279, + 318 + ], + "type": "text", + "content": "0" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 328, + 479, + 351 + ], + "lines": [ + { + "bbox": [ + 131, + 328, + 479, + 351 + ], + "spans": [ + { + "bbox": [ + 131, + 328, + 479, + 351 + ], + "type": "text", + "content": "Fig. 7: Impact of guidance strength during sampling. The numbers below the cropped patches indicate the motion guidance weighting factor " + }, + { + "bbox": [ + 131, + 328, + 479, + 351 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 131, + 328, + 479, + 351 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 299, + 224, + 340, + 308 + ], + "blocks": [ + { + "bbox": [ + 299, + 224, + 340, + 308 + ], + "lines": [ + { + "bbox": [ + 299, + 224, + 340, + 308 + ], + "spans": [ + { + "bbox": [ + 299, + 224, + 340, + 308 + ], + "type": "image", + "image_path": "6e5f2e88a23d0c8f1bec715d642579b62990a6d5df38361d7d2696f0ea29158b.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 316, + 311, + 325, + 318 + ], + "lines": [ + { + "bbox": [ + 316, + 311, + 325, + 318 + ], + "spans": [ + { + "bbox": [ + 316, + 311, + 325, + 318 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 343, + 224, + 384, + 308 + ], + "blocks": [ + { + "bbox": [ + 343, + 224, + 384, + 308 + ], + "lines": [ + { + "bbox": [ + 343, + 224, + 384, + 308 + ], + "spans": [ + { + "bbox": [ + 343, + 224, + 384, + 308 + ], + "type": "image", + "image_path": "57e87a8e5d2af8edab1741764e385b45164c7b5265e43daa99fdcdd7e87ddb6c.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 359, + 311, + 369, + 318 + ], + "lines": [ + { + "bbox": [ + 359, + 311, + 369, + 318 + ], + "spans": [ + { + "bbox": [ + 359, + 311, + 369, + 318 + ], + "type": "text", + "content": "20" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 387, + 224, + 429, + 308 + ], + "blocks": [ + { + "bbox": [ + 387, + 224, + 429, + 308 + ], + "lines": [ + { + "bbox": [ + 387, + 224, + 429, + 308 + ], + "spans": [ + { + "bbox": [ + 387, + 224, + 429, + 308 + ], + "type": "image", + "image_path": "4d11dc341d7114a89d0f12de62381f3f6b01066fec2f3be970f4353457b6912d.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 403, + 310, + 412, + 318 + ], + "lines": [ + { + "bbox": [ + 403, + 310, + 412, + 318 + ], + "spans": [ + { + "bbox": [ + 403, + 310, + 412, + 318 + ], + "type": "text", + "content": "30" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 431, + 224, + 472, + 308 + ], + "blocks": [ + { + "bbox": [ + 431, + 224, + 472, + 308 + ], + "lines": [ + { + "bbox": [ + 431, + 224, + 472, + 308 + ], + "spans": [ + { + "bbox": [ + 431, + 224, + 472, + 308 + ], + "type": "image", + "image_path": "5890a77f2e8d82ee4decff05dcb01ccb6098d1bfee84b750dfaaf75858d6c340.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 447, + 311, + 456, + 318 + ], + "lines": [ + { + "bbox": [ + 447, + 311, + 456, + 318 + ], + "spans": [ + { + "bbox": [ + 447, + 311, + 456, + 318 + ], + "type": "text", + "content": "40" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 380, + 480, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 380, + 480, + 441 + ], + "spans": [ + { + "bbox": [ + 130, + 380, + 480, + 441 + ], + "type": "text", + "content": "the baseline. This occurs because TSD is capable of extracting richer contextual information from neighboring frames. When both MDS and TSD are used, the results show further improvements. The full MGLD-VSR approach achieves the best scores on all the metrics, indicating the synergy of the two proposed components for real-world VSR." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 444, + 480, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 444, + 480, + 480 + ], + "spans": [ + { + "bbox": [ + 130, + 444, + 480, + 480 + ], + "type": "text", + "content": "Impact of Flow Network. As shown in Table 4, the flow estimation network in motion guided sampling does not show big influence on the results. This might because the motion guided sampling is applied in low-resolution space." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 130, + 483, + 480, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 483, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 130, + 483, + 480, + 555 + ], + "type": "text", + "content": "Impact of Guidance Strength. During guided sampling process, we introduce a weighting factor to control the guidance strength. As shown in Figure 7, a moderate guidance strength could reduce the artifacts and improve the temporal consistency of the generated results. A large guidance strength will blur the results and degrade the reconstructed details. To balance the generation power and temporal consistency, we set the weighting factor " + }, + { + "bbox": [ + 130, + 483, + 480, + 555 + ], + "type": "inline_equation", + "content": "\\eta = 10" + }, + { + "bbox": [ + 130, + 483, + 480, + 555 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 130, + 558, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 481, + 666 + ], + "type": "text", + "content": "The Design of VAE Decoder. To validate the effectiveness of the proposed architecture and loss design for the VAE decoder, we conduct a series of study and the results are shown in Table 5. As can be seen from the table, the introduction of the adaptive temporal modeling module improves the temporal consistency as well as perceptual quality (measured by no-reference metrics), and incorporating the feature from encoder to decoder during decoder fine-tuning further improves the results. By introducing two sequence based losses during fine-tuning, we improve the temporal consistency of the results while maintaining or improving the perceptual quality." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 394, + 90, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 394, + 90, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 394, + 90, + 447, + 100 + ], + "type": "text", + "content": "MGLD-VSR" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 217, + 135, + 394, + 169 + ], + "blocks": [ + { + "bbox": [ + 137, + 114, + 475, + 126 + ], + "lines": [ + { + "bbox": [ + 137, + 114, + 475, + 126 + ], + "spans": [ + { + "bbox": [ + 137, + 114, + 475, + 126 + ], + "type": "text", + "content": "Table 4: Ablation study on the flow network in motion guided diffusion sampling." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 217, + 135, + 394, + 169 + ], + "lines": [ + { + "bbox": [ + 217, + 135, + 394, + 169 + ], + "spans": [ + { + "bbox": [ + 217, + 135, + 394, + 169 + ], + "type": "table", + "html": "
Flow NetworkWE ↓NIQE ↓MUSIQ ↑DOVER ↑
SpyNet [41]7.67343.554552.69440.7478
RAFT [47]7.54783.534652.78120.7481
", + "image_path": "a95ffbbc21846a1837f4365aa124e4ee593568649492d162fbcd95b396775a35.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 133, + 224, + 478, + 272 + ], + "blocks": [ + { + "bbox": [ + 132, + 180, + 481, + 213 + ], + "lines": [ + { + "bbox": [ + 132, + 180, + 481, + 213 + ], + "spans": [ + { + "bbox": [ + 132, + 180, + 481, + 213 + ], + "type": "text", + "content": "Table 5: Ablation study on the different designs of VAE decoder. \"TM\" refers to the temporal modeling module. \"CFW\" refers to the controllable feature warping module that injects VAE encoder feature into VAE decoder." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 224, + 478, + 272 + ], + "lines": [ + { + "bbox": [ + 133, + 224, + 478, + 272 + ], + "spans": [ + { + "bbox": [ + 133, + 224, + 478, + 272 + ], + "type": "table", + "html": "
ModelsTM CFWWE ↓ NIQE ↓ MUSIQ ↑ DOVER ↑LossesLdiffLswcWE ↓ NIQE ↓ MUSIQ ↑ DOVER ↑
(a)9.86704.067150.27070.7360(a)7.94803.535952.7894
(b)✓8.68463.845152.54080.7431(b)✓7.75793.540952.5648
(c)✓8.98423.923652.43860.7426(c)✓7.76883.537652.7456
(d)✓7.54783.534652.78120.7481(d)✓✓7.54783.534652.7812
", + "image_path": "74128d21b83bd577ef27e38089f38c45ba5cc239405323257c9d64d7671883dd.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 297, + 218, + 307 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 297, + 218, + 307 + ], + "spans": [ + { + "bbox": [ + 132, + 297, + 218, + 307 + ], + "type": "text", + "content": "4.4 Limitations" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 322, + 482, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 322, + 482, + 455 + ], + "spans": [ + { + "bbox": [ + 130, + 322, + 482, + 455 + ], + "type": "text", + "content": "Our proposed MGLD-VSR method leverages the diffusion prior from pre-trained SD model, aiming to achieve perceptually realistic and temporally consistent VSR results. Due to the scale of pre-trained SD model, it does not have advantages in terms of number of parameters and inference time compared with those non-diffusion VSR methods (comparisons can be found in the supplementary materials). In addition, since MGLD-VSR will generate more visual details than those non-diffusion VSR methods, it does not have advantages in terms of temporal consistency metrics either. In the future, we can resort to model distillation [2] and more advanced diffusion sampling [1,59] techniques to improve the efficiency of MGLD-VSR, and resort to the recently developed video synthesis models [4] to improve the temporal consistency." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 479, + 220, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 220, + 491 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 220, + 491 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": "We presented a motion-guided latent diffusion algorithm, namely MGLD-VSR, for high-quality and temporally consistent real-world VSR. Our method exploited the strong generative prior of pre-trained latent diffusion model to restore video sequence with realistic details. To ensure the content consistency among frames, we developed a motion-guided sampling mechanism, where the motion dynamics of input video were incorporated in the diffusion sampling process. To further improve the continuity of generated details, we introduced a temporal module into the decoder, and fine-tuned it by high quality video sequences with delicately designed sequence-based losses. The fine-tuned sequence decoder mitigated the flickering artifacts and further improved the perceptual quality. Our proposed MGLD-VSR achieved state-of-the-art visual results on the widely adopted real-world VSR benchmarks, validating the great potentials of latent diffusion priors in enhancing real-world video quality." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "text", + "content": "X. Yang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 134, + 142, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 138, + 142, + 480, + 164 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 142, + 480, + 164 + ], + "spans": [ + { + "bbox": [ + 138, + 142, + 480, + 164 + ], + "type": "text", + "content": "1. Lu et al., C.: Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. NeurIPS 35, 5775-5787 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 166, + 480, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 166, + 480, + 186 + ], + "spans": [ + { + "bbox": [ + 138, + 166, + 480, + 186 + ], + "type": "text", + "content": "2. Luo et al., S.: Latent consistency models: Synthesizing high-resolution images with few-step inference. arXiv preprint arXiv:2310.04378 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 187, + 480, + 209 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 187, + 480, + 209 + ], + "spans": [ + { + "bbox": [ + 138, + 187, + 480, + 209 + ], + "type": "text", + "content": "3. Bishop, C.M., Blake, A., Marthi, B.: Super-resolution enhancement of video. In: AISTATS (2003)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 209, + 480, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 209, + 480, + 253 + ], + "spans": [ + { + "bbox": [ + 138, + 209, + 480, + 253 + ], + "type": "text", + "content": "4. Blattmann, A., Dockhorn, T., Kulal, S., Mendelevitch, D., Kilian, M., Lorenz, D., Levi, Y., English, Z., Voleti, V., Letts, A., et al.: Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 255, + 480, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 255, + 480, + 298 + ], + "spans": [ + { + "bbox": [ + 138, + 255, + 480, + 298 + ], + "type": "text", + "content": "5. Blattmann, A., Rombach, R., Ling, H., Dockhorn, T., Kim, S.W., Fidler, S., Kreis, K.: Align your latents: High-resolution video synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22563-22575 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 299, + 480, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 299, + 480, + 342 + ], + "spans": [ + { + "bbox": [ + 138, + 299, + 480, + 342 + ], + "type": "text", + "content": "6. Caballero, J., Ledig, C., Aitken, A., Acosta, A., Totz, J., Wang, Z., Shi, W.: Real-time video super-resolution with spatio-temporal networks and motion compensation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4778-4787 (2017)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 343, + 480, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 343, + 480, + 376 + ], + "spans": [ + { + "bbox": [ + 138, + 343, + 480, + 376 + ], + "type": "text", + "content": "7. Carrillo, H., Clément, M., Bugeau, A., Simo-Serra, E.: Diffusart: Enhancing line art colorization with conditional diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3485-3489 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 376, + 480, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 376, + 480, + 409 + ], + "spans": [ + { + "bbox": [ + 138, + 376, + 480, + 409 + ], + "type": "text", + "content": "8. Chan, K.C., Wang, X., Xu, X., Gu, J., Loy, C.C.: Glean: Generative latent bank for large-factor image super-resolution. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 14245-14254 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 410, + 480, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 410, + 480, + 453 + ], + "spans": [ + { + "bbox": [ + 138, + 410, + 480, + 453 + ], + "type": "text", + "content": "9. Chan, K.C., Wang, X., Yu, K., Dong, C., Loy, C.C.: Basicvsr: The search for essential components in video super-resolution and beyond. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4947-4956 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 134, + 454, + 480, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 454, + 480, + 497 + ], + "spans": [ + { + "bbox": [ + 134, + 454, + 480, + 497 + ], + "type": "text", + "content": "10. Chan, K.C., Zhou, S., Xu, X., Loy, C.C.: Basicvsr++: Improving video superresolution with enhanced propagation and alignment. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5972-5981 (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 134, + 498, + 480, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 498, + 480, + 531 + ], + "spans": [ + { + "bbox": [ + 134, + 498, + 480, + 531 + ], + "type": "text", + "content": "11. Chan, K.C., Zhou, S., Xu, X., Loy, C.C.: Investigating tradeoffs in real-world video super-resolution. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5962-5971 (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 533, + 480, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 533, + 480, + 553 + ], + "spans": [ + { + "bbox": [ + 134, + 533, + 480, + 553 + ], + "type": "text", + "content": "12. Chen, M., Laina, I., Vedaldi, A.: Training-free layout control with cross-attention guidance. arXiv preprint arXiv:2304.03373 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 555, + 480, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 555, + 480, + 586 + ], + "spans": [ + { + "bbox": [ + 134, + 555, + 480, + 586 + ], + "type": "text", + "content": "13. Chu, M., Xie, Y., Mayer, J., Leal-Taixe, L., Thuerey, N.: Learning temporal coherence via self-supervision for gan-based video generation. ACM Transactions on Graphics (TOG) 39(4), 75-1 (2020)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 588, + 480, + 642 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 588, + 480, + 642 + ], + "spans": [ + { + "bbox": [ + 134, + 588, + 480, + 642 + ], + "type": "text", + "content": "14. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. In: Ranzato, M., Beygelzimer, A., Dauphin, Y., Liang, P., Vaughan, J.W. (eds.) Advances in Neural Information Processing Systems. vol. 34, pp. 8780-8794. Curran Associates, Inc. (2021), https://proceedings.neurips.cc/paper_files/paper/2021/file/49ad23d1ec9fa4bd8d77d02681df5cfa-Paper.pdf" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 134, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 643, + 480, + 665 + ], + "type": "text", + "content": "15. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. In: Ranzato, M., Beygelzimer, A., Dauphin, Y., Liang, P., Vaughan, J.W. (eds.) Advances" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 395, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 395, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 395, + 91, + 447, + 100 + ], + "type": "text", + "content": "MGLD-VSR" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 147, + 117, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 117, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 147, + 117, + 481, + 149 + ], + "type": "text", + "content": "in Neural Information Processing Systems. vol. 34, pp. 8780-8794. Curran Associates, Inc. (2021), https://proceedings.neurips.cc/paper_files/paper/2021/file/49ad23d1ec9fa4bd8d77d02681df5cfa-Paper.pdf" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 481, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 481, + 183 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 481, + 183 + ], + "type": "text", + "content": "16. Ding, K., Ma, K., Wang, S., Simoncelli, E.P.: Image quality assessment: Unifying structure and texture similarity. IEEE transactions on pattern analysis and machine intelligence 44(5), 2567-2581 (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 183, + 481, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 183, + 481, + 226 + ], + "spans": [ + { + "bbox": [ + 133, + 183, + 481, + 226 + ], + "type": "text", + "content": "17. Fei, B., Lyu, Z., Pan, L., Zhang, J., Yang, W., Luo, T., Zhang, B., Dai, B.: Generative diffusion prior for unified image restoration and enhancement. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9935-9946 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 227, + 481, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 227, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 133, + 227, + 481, + 270 + ], + "type": "text", + "content": "18. Ge, S., Nah, S., Liu, G., Poon, T., Tao, A., Catanzaro, B., Jacobs, D., Huang, J.B., Liu, M.Y., Balaji, Y.: Preserve your own correlation: A noise prior for video diffusion models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 22930-22941 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 271, + 481, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 271, + 481, + 304 + ], + "spans": [ + { + "bbox": [ + 133, + 271, + 481, + 304 + ], + "type": "text", + "content": "19. Gu, J., Shen, Y., Zhou, B.: Image processing using multi-code gan prior. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 3012-3021 (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 304, + 481, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 481, + 326 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 481, + 326 + ], + "type": "text", + "content": "20. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in neural information processing systems 33, 6840-6851 (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 326, + 481, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 481, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 481, + 380 + ], + "type": "text", + "content": "21. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. In: Larochelle, H., Ranzato, M., Hadsell, R., Balcan, M., Lin, H. (eds.) Advances in Neural Information Processing Systems. vol. 33, pp. 6840-6851. Curran Associates, Inc. (2020), https://proceedings.neurips.cc/paper_files/paper/2020/file/4c5bcfec8584af0d967f1ab10179ca4b-Paper.pdf" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 380, + 481, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 481, + 423 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 481, + 423 + ], + "type": "text", + "content": "22. Jo, Y., Oh, S.W., Kang, J., Kim, S.J.: Deep video super-resolution network using dynamic upsampling filters without explicit motion compensation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3224-3232 (2018)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 424, + 481, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 424, + 481, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 424, + 481, + 468 + ], + "type": "text", + "content": "23. Johnson, J., Alahi, A., Fei-Fei, L.: Perceptual losses for real-time style transfer and super-resolution. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14. pp. 694-711. Springer (2016)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 468, + 481, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 481, + 512 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 481, + 512 + ], + "type": "text", + "content": "24. Kawar, B., Zada, S., Lang, O., Tov, O., Chang, H., Dekel, T., Mosseri, I., Irani, M.: Imagic: Text-based real image editing with diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6007-6017 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 512, + 481, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 481, + 544 + ], + "type": "text", + "content": "25. Ke, J., Wang, Q., Wang, Y., Milanfar, P., Yang, F.: Musiq: Multi-scale image quality transformer. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5148-5157 (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 545, + 481, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 545, + 481, + 567 + ], + "spans": [ + { + "bbox": [ + 132, + 545, + 481, + 567 + ], + "type": "text", + "content": "26. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 567, + 481, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 481, + 600 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 481, + 600 + ], + "type": "text", + "content": "27. Lai, W.S., Huang, J.B., Wang, O., Shechtman, E., Yumer, E., Yang, M.H.: Learning blind video temporal consistency. In: Proceedings of the European conference on computer vision (ECCV). pp. 170-185 (2018)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 600, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 481, + 643 + ], + "type": "text", + "content": "28. Li, W., Tao, X., Guo, T., Qi, L., Lu, J., Jia, J.: Mucan: Multi-correspondence aggregation network for video super-resolution. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part X 16. pp. 335-351. Springer (2020)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 643, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 481, + 666 + ], + "type": "text", + "content": "29. Li, Z., Bampis, C., Novak, J., Aaron, A., Swanson, K., Moorthy, A., Cock, J.: Vmaf: The journey continues. Netflix Technology Blog 25(1) (2018)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "text", + "content": "X. Yang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 130, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 480, + 138 + ], + "type": "text", + "content": "30. Liang, J., Cao, J., Fan, Y., Zhang, K., Ranjan, R., Li, Y., Timofte, R., Van Gool, L.: Vrt: A video restoration transformer. arXiv preprint arXiv:2201.12288 (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 481, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 481, + 183 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 481, + 183 + ], + "type": "text", + "content": "31. Liang, J., Fan, Y., Xiang, X., Ranjan, R., Ilg, E., Green, S., Cao, J., Zhang, K., Timofte, R., Gool, L.V.: Recurrent video restoration transformer with guided deformable attention. Advances in Neural Information Processing Systems 35, 378-393 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 184, + 480, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 184, + 480, + 225 + ], + "spans": [ + { + "bbox": [ + 132, + 184, + 480, + 225 + ], + "type": "text", + "content": "32. Lugmayr, A., Danelljan, M., Romero, A., Yu, F., Timofte, R., Van Gool, L.: Repaint: Inpainting using denoising diffusion probabilistic models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11461-11471 (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 227, + 480, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 480, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 480, + 258 + ], + "type": "text", + "content": "33. Meister, S., Hur, J., Roth, S.: Unflow: Unsupervised learning of optical flow with a bidirectional census loss. In: Proceedings of the AAAI conference on artificial intelligence. vol. 32 (2018)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 260, + 480, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 260, + 480, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 260, + 480, + 303 + ], + "type": "text", + "content": "34. Menon, S., Damian, A., Hu, S., Ravi, N., Rudin, C.: Pulse: Self-supervised photoupsampling via latent space exploration of generative models. In: Proceedings of the iee/cvf conference on computer vision and pattern recognition. pp. 2437-2445 (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 304, + 480, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 480, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 480, + 335 + ], + "type": "text", + "content": "35. Mittal, A., Moorthy, A.K., Bovik, A.C.: No-reference image quality assessment in the spatial domain. IEEE Transactions on image processing 21(12), 4695-4708 (2012)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 336, + 480, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 480, + 357 + ], + "type": "text", + "content": "36. Mittal, A., Soundararajan, R., Bovik, A.C.: Making a \"completely blind\" image quality analyzer. IEEE Signal processing letters 20(3), 209-212 (2012)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 358, + 480, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 480, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 480, + 402 + ], + "type": "text", + "content": "37. Nah, S., Baik, S., Hong, S., Moon, G., Son, S., Timofte, R., Mu Lee, K.: Ntire 2019 challenge on video deblurring and super-resolution: Dataset and study. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. pp. 0-0 (2019)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 403, + 480, + 478 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 480, + 478 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 480, + 478 + ], + "type": "text", + "content": "38. Nichol, A.Q., Dhariwal, P., Ramesh, A., Shyam, P., Mishkin, P., Mcgrew, B., Sutskever, I., Chen, M.: GLIDE: Towards photorealistic image generation and editing with text-guided diffusion models. In: Chaudhuri, K., Jegelka, S., Song, L., Szepesvari, C., Niu, G., Sabato, S. (eds.) Proceedings of the 39th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 162, pp. 16784-16804. PMLR (17-23 Jul 2022), https://proceedings.mlr.press/v162/nichol22a.html" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 479, + 480, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 480, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 480, + 511 + ], + "type": "text", + "content": "39. Pan, J., Bai, H., Dong, J., Zhang, J., Tang, J.: Deep blind video super-resolution. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4811-4820 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 512, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 480, + 555 + ], + "type": "text", + "content": "40. Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., et al.: Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems 32 (2019)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 555, + 480, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 480, + 588 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 480, + 588 + ], + "type": "text", + "content": "41. Ranjan, A., Black, M.J.: Optical flow estimation using a spatial pyramid network. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4161-4170 (2017)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 589, + 480, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 589, + 480, + 610 + ], + "spans": [ + { + "bbox": [ + 132, + 589, + 480, + 610 + ], + "type": "text", + "content": "42. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 611, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 480, + 643 + ], + "type": "text", + "content": "43. Singer, U., Polyak, A., Hayes, T., Yin, X., An, J., Zhang, S., Hu, Q., Yang, H., Ashual, O., Gafni, O., et al.: Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792 (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "text", + "content": "44. Song, Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score-based generative modeling through stochastic differential equations. In: Interna" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 395, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 395, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 395, + 91, + 447, + 100 + ], + "type": "text", + "content": "MGLD-VSR" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 149, + 117, + 480, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 117, + 480, + 137 + ], + "spans": [ + { + "bbox": [ + 149, + 117, + 480, + 137 + ], + "type": "text", + "content": "tional Conference on Learning Representations (2021), https://openreview.net/ forum?id=PxTIG12RRHS" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 139, + 480, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 139, + 480, + 171 + ], + "spans": [ + { + "bbox": [ + 133, + 139, + 480, + 171 + ], + "type": "text", + "content": "45. Tao, X., Gao, H., Liao, R., Wang, J., Jia, J.: Detail-revealing deep video superresolution. In: Proceedings of the IEEE international conference on computer vision. pp. 4472-4480 (2017)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 172, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 172, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 133, + 172, + 480, + 205 + ], + "type": "text", + "content": "46. Tao Yang, Peiran Ren, X.X., Zhang, L.: Gan prior embedded network for blind face restoration in the wild. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 205, + 480, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 205, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 133, + 205, + 480, + 237 + ], + "type": "text", + "content": "47. Teed, Z., Deng, J.: Raft: Recurrent all-pairs field transforms for optical flow. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16. pp. 402-419. Springer (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 238, + 480, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 238, + 480, + 259 + ], + "spans": [ + { + "bbox": [ + 133, + 238, + 480, + 259 + ], + "type": "text", + "content": "48. Wang, J., Yue, Z., Zhou, S., Chan, K.C., Loy, C.C.: Exploiting diffusion prior for real-world image super-resolution. arXiv preprint arXiv:2305.07015 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 260, + 480, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 260, + 480, + 303 + ], + "spans": [ + { + "bbox": [ + 133, + 260, + 480, + 303 + ], + "type": "text", + "content": "49. Wang, X., Chan, K.C., Yu, K., Dong, C., Change Loy, C.: Edvr: Video restoration with enhanced deformable convolutional networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. pp. 0-0 (2019)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 304, + 480, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 304, + 480, + 335 + ], + "spans": [ + { + "bbox": [ + 133, + 304, + 480, + 335 + ], + "type": "text", + "content": "50. Wang, X., Li, Y., Zhang, H., Shan, Y.: Towards real-world blind face restoration with generative facial prior. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9168-9178 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 133, + 336, + 480, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 336, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 133, + 336, + 480, + 369 + ], + "type": "text", + "content": "51. Wang, X., Xie, L., Dong, C., Shan, Y.: Real-esrgan: Training real-world blind superresolution with pure synthetic data. In: International Conference on Computer Vision Workshops (ICCVW) (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 133, + 369, + 480, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 369, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 133, + 369, + 480, + 413 + ], + "type": "text", + "content": "52. Wu, H., Zhang, E., Liao, L., Chen, C., Hou, J., Wang, A., Sun, W., Yan, Q., Lin, W.: Exploring video quality assessment on user generated contents from aesthetic and technical perspectives. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 20144-20154 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 133, + 414, + 480, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 414, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 133, + 414, + 480, + 445 + ], + "type": "text", + "content": "53. Xie, L., Wang, X., Shi, S., Gu, J., Dong, C., Shan, Y.: Mitigating artifacts in real-world video super-resolution models. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 2956-2964 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 133, + 447, + 480, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 447, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 133, + 447, + 480, + 479 + ], + "type": "text", + "content": "54. Xu, H., Yang, J., Cai, J., Zhang, J., Tong, X.: High-resolution optical flow from 1d attention and correlation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10498-10507 (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 133, + 479, + 480, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 479, + 480, + 522 + ], + "spans": [ + { + "bbox": [ + 133, + 479, + 480, + 522 + ], + "type": "text", + "content": "55. Yang, X., Xiang, W., Zeng, H., Zhang, L.: Real-world video super-resolution: A benchmark dataset and a decomposition based learning scheme. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4781-4790 (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 133, + 523, + 480, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 523, + 480, + 566 + ], + "spans": [ + { + "bbox": [ + 133, + 523, + 480, + 566 + ], + "type": "text", + "content": "56. Yi, P., Wang, Z., Jiang, K., Jiang, J., Ma, J.: Progressive fusion video superresolution network via exploiting non-local spatio-temporal correlations. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 3106-3115 (2019)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 567, + 480, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 567, + 480, + 599 + ], + "spans": [ + { + "bbox": [ + 133, + 567, + 480, + 599 + ], + "type": "text", + "content": "57. Zhang, K., Liang, J., Van Gool, L., Timofte, R.: Designing a practical degradation model for deep blind image super-resolution. In: IEEE International Conference on Computer Vision. pp. 4791-4800 (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 133, + 600, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 600, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 133, + 600, + 480, + 632 + ], + "type": "text", + "content": "58. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 586-595 (2018)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 133, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 633, + 480, + 665 + ], + "type": "text", + "content": "59. Zhao, W., Bai, L., Rao, Y., Zhou, J., Lu, J.: Unipc: A unified predictor-corrector framework for fast sampling of diffusion models. Advances in Neural Information Processing Systems 36 (2024)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "text", + "content": "X. Yang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 139 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 139 + ], + "type": "text", + "content": "60. Zhou, S., Chan, K.C., Li, C., Loy, C.C.: Towards robust blind face restoration with codebook lookup transformer. In: NeurIPS (2022)" + } + ] + } + ], + "index": 2 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 395, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 395, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 395, + 91, + 447, + 100 + ], + "type": "text", + "content": "MGLD-VSR" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Motion-Oriented Compositional Neural Radiance Fields for Monocular Dynamic Human Modeling/6e1b3c80-b2f1-4690-8b23-807237bf60da_content_list.json b/2024/Motion-Oriented Compositional Neural Radiance Fields for Monocular Dynamic Human Modeling/6e1b3c80-b2f1-4690-8b23-807237bf60da_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..517ae3c22d96961206493fcb0d27f9096e9554f5 --- /dev/null +++ b/2024/Motion-Oriented Compositional Neural Radiance Fields for Monocular Dynamic Human Modeling/6e1b3c80-b2f1-4690-8b23-807237bf60da_content_list.json @@ -0,0 +1,1693 @@ +[ + { + "type": "text", + "text": "Motion-Oriented Compositional Neural Radiance Fields for Monocular Dynamic Human Modeling", + "text_level": 1, + "bbox": [ + 220, + 141, + 782, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jaehyeok $\\mathrm{Kim}^{1}$ , Dongyoon $\\mathrm{Wee}^{2}$ , and Dan $\\mathrm{Xu}^{1\\boxtimes}$", + "bbox": [ + 305, + 210, + 696, + 228 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ HKUST, Hong Kong $^{2}$ Naver Cloud Corp., South Korea jkimbf@cse.ust.hk, dongyoon.wee@navercorp.com, danxu@cse.ust.hk", + "bbox": [ + 261, + 238, + 738, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. This paper introduces Motion-oriented Compositional Neural Radiance Fields (MoCo-NeRF), a framework designed to perform free-viewpoint rendering of monocular human videos via novel non-rigid motion modeling approach. In the context of dynamic clothed humans, complex cloth dynamics generate non-rigid motions that are intrinsically distinct from skeletal articulations and critically important for the rendering quality. The conventional approach models non-rigid motions as spatial (3D) deviations in addition to skeletal transformations. However, it is either time-consuming or challenging to achieve optimal quality due to its high learning complexity without a direct supervision. To target this problem, we propose a novel approach of modeling non-rigid motions as radiance residual fields to benefit from more direct color supervision in the rendering and utilize the rigid radiance fields as a prior to reduce the complexity of the learning process. Our approach utilizes a single multiresolution hash encoding (MHE) to concurrently learn the canonical T-pose representation from rigid skeletal motions and the radiance residual field for non-rigid motions. Additionally, to further improve both training efficiency and usability, we extend MoCo-NeRF to support simultaneous training of multiple subjects within a single framework, thanks to our effective design for modeling non-rigid motions. This scalability is achieved through the integration of a global MHE and learnable identity codes in addition to multiple local MHEs. We present extensive results on ZJU-MoCap and MonoCap, clearly demonstrating state-of-the-art performance in both single- and multi-subject settings. The code and model will be made publicly available at the project page: https://stevejaehyeok.github.io/publications/moco-nerf.", + "bbox": [ + 261, + 297, + 743, + 659 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Dynamic human modeling from monocular videos $\\cdot$ Neural Radiance Fields $\\cdot$ Novel view synthesis of humans", + "bbox": [ + 261, + 671, + 740, + 700 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 742, + 377, + 758 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Novel view synthesis of monocular human video with photo-realistic quality is a challenging problem, as it involves complex articulated body motions and necessitates sophisticated modeling of non-skeletal or non-rigid motions (e.g., wrinkles", + "bbox": [ + 212, + 771, + 787, + 816 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$\\boxtimes$ Corresponding author.", + "bbox": [ + 230, + 824, + 400, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/98ab8cebc2cfd6b75e6abd8042a52f0e7f1fef5ecedab8c678c2c544464b74fb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 220, + 146, + 781, + 268 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/64c9c69f2ff2f04e7bf36bee908d712b9ff294b69e7aa7018752917f01d95370.jpg", + "image_caption": [ + "Fig. 1: a) We introduce a motion-oriented compositional NeRF for photo-realistic modeling of dynamic humans from monocular videos. Our approach innovatively employs radiance compositions to capture pose-adaptive non-rigid motions, overcoming the limitations of skeletal transformations that typically yield an average of observed deformations (mean motions). b) The proposed MoCo-NeRF achieves state-of-the-art rendering quality and noteworthy efficiency in novel view synthesis, compared to leading competitors [14,21,73]. The bold denotes the proposed training duration of each comparison model. MoCo-NeRF uniquely captures coherent non-rigid motions, like T-shirt wrinkles relative to body pose, from entirely new viewpoints. Moreover, MoCo-NeRF significantly surpasses the efficiency of another photo-realistic method, HumanNeRF [73]." + ], + "image_footnote": [], + "bbox": [ + 220, + 273, + 781, + 406 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "of clothes) from unseen viewpoints. In particular, non-rigid motion modeling is essential for achieving photo-realism and coherence in free-viewpoint renderings. A prevalent solution adopted by recent methods [14,21,73] models the non-rigid motions as pose-dependent 3D spatial offsets of canonical point samples. However, the offset-based strategies lead to either extensive training time or irrelevant modeling of non-rigid motions due to its high learning complexity, incurred by the following reasons. First, a single spatial offset spans an unbounded range and is unique for every 3D point across all input frames and body poses, thereby escalating computational complexity. Second, the learning objective of spatial offsets is not directly aligned with the conventional RGB-based supervision, further complicating the modeling process.", + "bbox": [ + 212, + 595, + 787, + 762 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To target the issues above-discussed, in this paper, we propose Motion-oriented Compositional Neural Radiance Fields (MoCo-NeRF), a new framework that effectively models non-rigid motions at a radiance-level through a unique design of residual radiance field learning for free-viewpoint renderings of monocular human videos. Specifically, our proposed framework decomposes a single", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "J. Kim et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "MHE [44] into two branches: a rigid branch that learns a canonical T-pose radiance field solely via rigid skeletal transformations, and a non-rigid branch that optimizes pose-driven radiance residuals for capturing appearance discrepancies between renderings of the rigid branch and the ground-truth images, which are typically caused by non-rigid motions. The proposed approach circumvents the complexities associated with learning unbounded spatial offsets by reducing the prediction range to the radiance space with the rigid radiance field as a prior. It further establishes a more direct supervision from the ground-truth pixel colors for optimizing the rigid and non-rigid radiance fields by embedding them together in the rendering process, which can further enhance the learning efficiency. Besides, we present a cross-attention module to query a pose conditioned implicit feature from a learnable base code. This process allows our framework to extract more discriminative implicit features tailored to each particular pose, helping our non-rigid branch to learn coherent pose-driven radiance residuals.", + "bbox": [ + 212, + 146, + 787, + 357 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In addition, thanks to our effective design for modeling non-rigid motions, the proposed MoCo-NeRF architecture can be flexibly extended to support unified training across multiple subjects while maintaining photo-realistic quality, without a significant increase in training time. For this architectural extension, we introduce multiple local MHEs and a global MHE, and also extend the aforementioned base code to a set of identity (ID) codes as a learnable codebook. The multiple local MHEs are designed for modeling multiple subjects and the global MHE targets the learning of generic and universal body representations. Furthermore, subject-specific learnable ID codes that are pivotal for generating pose-embedded implicit features enable effective sharing of the MLP decoder in the rigid branch by all subjects, enhancing the model's efficiency and scalability.", + "bbox": [ + 212, + 359, + 787, + 523 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "MoCo-NeRF achieves state-of-the-art performance on free-viewpoint rendering of monocular human videos in both single- and multi-subject settings, showing clear improvements compared to the prior works. We present our extensive results on ZJU-MoCap [52] and MonoCap [17, 18, 50] to demonstrate the effectiveness of the proposed MoCo-NeRF in learning photo-realistic representation and modeling complex non-rigid motions from monocular videos.", + "bbox": [ + 212, + 525, + 787, + 614 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In summary, the contribution of our work is threefold:", + "bbox": [ + 238, + 616, + 632, + 631 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a novel motion-oriented compositional NeRF approach, MoCo-NeRF, which concurrently learns canonical rigid radiance fields and pose-dependent radiance residual fields, effectively handling both rigid and non-rigid motions. We also introduce a pose-embedded implicit feature, computed by cross-attention between a body pose and a designed learnable base code, to implicitly enhance pose-dependent radiance residual learning.", + "- The proposed MoCo-NeRF is the first work that can be readily extended to enable unified training of multiple subjects without significantly increasing the GPU memory demands. We achieve this through our local MHEs, a global MHE, and the extended identity codebook, which enable subject-discriminative representation learning and effective module sharing.", + "- The proposed MoCo-NeRF achieves state-of-the-art performance under both single- and multi-subject training settings with a significant efficiency im" + ], + "bbox": [ + 225, + 642, + 785, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "MoCo-NeRF", + "bbox": [ + 643, + 114, + 730, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "provement compared to the best-performing methods. MoCo-NeRF jointly learns all 6 subjects of ZJU-MoCap on one RTX 3090 GPU within 2 hours.", + "bbox": [ + 241, + 146, + 787, + 176 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 200, + 388, + 217 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We review closely related work from three perspectives: deformable neural rendering, neural human rendering, and monocular photo-realistic human rendering. Deformable neural rendering. NeRF [42] leverages an MLP to learn a static 3D representation of a scene from a dense set of images from diverse viewpoints. For its application in real-life scenarios, prior works have enhanced NeRF in various directions such as improving efficiencies [4,13,19,36,44,54,59,82] or relieving strong reliance on impractical settings like dense input views [5,24,38,70,83]. In addition, there are more branches of work such as large-scale modeling [55,66, 68,77,79], improving rendering quality [2,3], reconstruction quality [46,69,81] and more. However, they are restricted to static scenes while the majority of the real-world objects are dynamic.", + "bbox": [ + 212, + 234, + 785, + 400 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Recent works including [11, 12, 35, 47, 48, 53, 67, 76] have broadened the modeling capabilities of NeRF to dynamic scenes with small and simple movements/deformations. However, they often fail in handling more dynamic or complex motions like human body articulations. In contrast, our approach enables learning of human-articulated motions from monocular videos, and performing free-viewpoint rendering of a human performer at any time frame of the videos. Neural rendering of humans. Many early-stage prior works for neural human renderings utilize traditional techniques like multi-view stereo [16, 60, 61, 75] or to build sophisticated systems for capturing human motions [8, 10, 40, 65]. However, constructing a dense set of training views and building those capturing systems are expensive, making them unsuitable for real-life situations. [1, 6, 57, 58] address 3D human reconstruction by utilizing ground truth human scans, relieving the aforementioned requirements. [9, 20, 45] further improves it by enabling human reconstruction from 2D image collections. In recent days, there have been many works learning an implicit human representation for novel pose synthesis [22, 23, 25, 27, 34, 50, 56, 64, 71, 87-89] or for a novel-view synthesis [27, 28, 31, 37, 51, 52, 64, 73, 74, 78, 85, 86]. There also have been several generalizable works [41, 43, 62] for a novel human subject with pre-training on multi-view dataset and a few-shot fine-tuning. [7, 80] focus on improving rendering efficiency to achieve real-time free-view rendering of dynamic humans. Furthermore, [14, 26, 28, 49] adopt explicit hash encoding [44] in order to boost the training efficiency for human representation modeling. In contrast to our problem setup, most of those works require multi-view videos or do not model the non-rigid motions.", + "bbox": [ + 212, + 401, + 785, + 748 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Monocular photo-realistic human rendering. HumanNeRF [73] presents a state-of-the-art performance by optimizing motion fields and canonical T-type representation of a human performer. However, its implicit modeling of pose-dependent non-rigid motions requires tens of hours to converge. Instant-NVR [14] achieves efficient training with part-based representation and 2D motion parameterization for learning non-rigid motions as 3D offsets. Note that the", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 230, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "J. Kim et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/5ad053192ebfef455342c6dfacb401ad87379dfee095188170b3a32624e90a18.jpg", + "image_caption": [ + "Fig. 2: Overview of the proposed framework MoCo-NeRF for free-view human rendering from a monocular video. Without estimating a geometrical offset of each continuous canonical point for each body pose, our framework is able to handle all deformations via its radiance-compositional approach with a single MHE [44] and achieve state-of-the-art performance. The pose-embedded implicit feature further enhances the learning of non-rigid radiance residuals by enabling pose-distinctive representation learning." + ], + "image_footnote": [], + "bbox": [ + 217, + 142, + 787, + 287 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "human parametric model (SMPL) [39], an extra requirement for training, plays a significant role in their framework. In addition, a concurrent work GauHuman [21], which also employs spatial offset approach with Gaussian Splatting (GS) [29], further enhances the efficiency and performance upon [14] by leveraging SMPL for the density control of GS. Nevertheless, [14, 21] struggle to learn complex non-rigid motions, even with an extended training duration as shown in Fig. 1. In this paper, instead of directly learning spatial offsets like these existing works, we propose a novel approach that effectively models non-rigid body motions by learning motion-oriented compositional radiance fields conditioned by pose-embedded implicit features. Our approach clearly outperforms the best-performing competitors HumanNeRF [73], Instant-NVR [14], and GauHuman [21] and can further extend to support joint training of multiple subjects.", + "bbox": [ + 212, + 412, + 787, + 595 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 The Proposed MoCo-NeRF Method", + "text_level": 1, + "bbox": [ + 215, + 622, + 602, + 640 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The proposed MoCo-NeRF aims to effectively learn high-fidelity representations for free-viewpoint rendering of dynamic human subjects in monocular videos. We introduce our novel radiance compositional NeRF approach based on different motion types: rigid and non-rigid motions, as illustrated in Fig. 2. The composition of canonical radiance and the per-sample pose-dependent radiance residuals (Sec. 3.2) facilitates photo-realistic quality of rendering with fine-grained modeling of non-rigid motions, rather than leveraging the conventional spatial 3D offset learning approach. We further improve the quality of non-rigid radiance residuals by introducing a pose-embedded implicit features (Sec. 3.2) that is jointly optimized with the non-rigid radiance branch. In Sec. 3.3, thanks to our effective design of non-rigid motion modeling, it enables MoCo-NeRF to be flexibly extended for unified training of multiple subjects with a high efficiency.", + "bbox": [ + 212, + 657, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "MoCo-NeRF", + "bbox": [ + 643, + 114, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Preliminary: Skeletal Transformations", + "text_level": 1, + "bbox": [ + 215, + 146, + 576, + 162 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The skeletal motion learns an inverse formulation of the linear blend skinning [32], which transforms canonical points to different body poses by skinning weights. Weng et al. [73] reformulate the linear blend skinning to transform a point $\\mathbf{x}$ in the observation space to a canonical point $\\mathbf{x}_c$ :", + "bbox": [ + 212, + 170, + 787, + 232 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {x} _ {c} = \\sum_ {k = 1} ^ {K} w _ {k} ^ {o} \\left(R _ {k} \\mathbf {x} + \\mathbf {t} _ {k}\\right), w _ {k} ^ {o} = \\frac {w _ {k} ^ {c} \\left(R _ {k} \\mathbf {x} + \\mathbf {t} _ {k}\\right)}{\\sum_ {j = 1} ^ {K} w _ {j} ^ {c} \\left(R _ {j} \\mathbf {x} + \\mathbf {t} _ {j}\\right)}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 321, + 243, + 785, + 284 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $w_{k}^{o}$ and $w_{k}^{c}$ respectively represent observation and canonical skinning weights of bone $k$ , and $R_{k}$ and $\\mathbf{t}_{k}$ are the rotation and translation of bone $k$ , respectively. Given Eq. 1, following [73], we optimize an explicit volume decoder to optimize $\\{w_{k}^{c}\\}_{k = 1}^{K}$ where $K$ is the total number of joints.", + "bbox": [ + 214, + 292, + 787, + 354 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Motion-Oriented Compositional Neural Radiance Fields", + "text_level": 1, + "bbox": [ + 214, + 375, + 723, + 390 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Rigid neural branch for learning canonical radiance field. The rigid branch of MoCo-NeRF is specifically designed to learn the canonical T-pose representation of the target subject using only rigid body movements. Our approach is different from but simpler than the prevalent spatial offset learning approach, which incorporates complex modeling of non-rigid motions as $\\Delta \\mathbf{x}_c$ . Our rigid branch instead directly predicts the color and density of the canonical point $\\mathbf{x}_c$ transformed by inverse linear blend skinning (Sec. 3.1), as depicted in Fig. 2. First, we query multiresolution volumetric features $f_{\\mathbf{x}_c}$ of the canonical point $\\mathbf{x}_c$ from the MHE $\\psi_{hash}$ as follows:", + "bbox": [ + 212, + 398, + 787, + 537 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nf _ {\\mathbf {x} _ {c}} = \\psi_ {\\text {h a s h}} (\\mathbf {x} _ {c}). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 442, + 547, + 784, + 564 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Then, we slice the first half ( $H$ dimensions) of $f_{\\mathbf{x}_c}$ at each level of resolution to compute $f_{\\mathbf{x}_c}^{:H}$ and predicts the radiance $c$ and density $\\sigma$ of the rigid canonical point $\\mathbf{x}_c$ with an MLP decoder $\\mathbf{M}_{rigid}$ as follows:", + "bbox": [ + 214, + 574, + 787, + 621 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nc, \\sigma = \\mathbf {M} _ {\\text {r i g i d}} \\left(f _ {\\mathbf {x} _ {c}} ^ {H}\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 429, + 631, + 784, + 648 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Non-rigid neural branch for learning radiance residual field. While the canonical representation does not explicitly model pose-dependent non-rigid motions, it still captures an average of such motions (mean motions) after training. However, the mean motions cannot effectively reflect the complex and realistic non-rigid motions under different body poses. It is only capable of rendering identical deformations across different poses, as can be observed in Fig. 1. To address this effectively, the non-rigid branch of MoCo-NeRF is designed to concurrently optimize for a radiance residual field. This models the appearance discrepancies between the renderings of the rigid branch and the ground-truth colors, typically caused by the complex non-rigid motions or deformations. We additionally condition the radiance residual predictions with the canonical radiance representation of the rigid branch as prior information. This enables more", + "bbox": [ + 212, + 657, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "J. Kim et al.", + "bbox": [ + 271, + 114, + 357, + 127 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/37ddffe77c036dd982cc06c380f82c3d3434ccd85bbd069e44416d0e9e5cb26c.jpg", + "image_caption": [ + "Fig.3: Illustration of the proposed pose-embedded implicit feature generation. We employ cross-attention to modulate the single learnable base code to pose-adaptive features." + ], + "image_footnote": [], + "bbox": [ + 222, + 143, + 460, + 292 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ddac47f517266c75f171259fdc89ab52a4be2d2a8b2542097939aaddf41fa83a.jpg", + "image_caption": [ + "Fig. 4: Illustration of extended architecture of MoCo-NeRF for the multi-subject unified training (Sec. 3.3). Major components consist of the global MHE, set of local MHEs, and the dictionary of learnable base codes as ID codes." + ], + "image_footnote": [], + "bbox": [ + 478, + 143, + 781, + 292 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "coherent modeling of non-rigid motions, as the non-rigid and rigid radiance fields are highly correlated.", + "bbox": [ + 212, + 398, + 782, + 428 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Specifically, unlike the rigid branch, the queried feature $f_{\\mathbf{x}_c}$ is not sliced, enabling the non-rigid branch to learn the radiance residuals with reference to the canonical representation. Instead, we freeze the first half of $f_{\\mathbf{x}_c}$ that represents the canonical radiance field, and only optimize the other half to learn the radiance residuals for non-rigid motion modeling. We denote this partially frozen feature as $f_{\\mathbf{x}_c}^*$ . We observe that providing frozen $f_{\\mathbf{x}_c}^{H}$ to the non-rigid radiance branch enables more effective and coherent residual learning. In addition, we also condition the radiance residual field prediction with our pose-embedded implicit feature $\\mathbf{P}_{nr}$ , which is achieved by concatenating $\\mathbf{P}_{nr}$ to the first intermediate output of the MLP decoder $\\mathbf{M}_{nr}$ , as illustrated in Fig. 2. The generation of $\\mathbf{P}_{nr}$ is explained in the last subsection of Sec. 3.2. Eventually, the color and density residuals are predicted by $\\mathbf{M}_{nr}$ for the final compositional radiance computation:", + "bbox": [ + 212, + 429, + 787, + 609 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta c, \\Delta \\sigma = \\mathbf {M} _ {n r} \\left(f _ {\\mathbf {x} _ {c}} ^ {*}; \\mathbf {P} _ {n r}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 403, + 625, + 784, + 642 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Compositional volume rendering. Following [42], we adopt stratified sampling and volume rendering to predict the RGB color of each ray. For final rendering, we query $M$ samples per ray $\\mathbf{r}$ and perform approximated integration to obtain final RGB $C_{final}$ from $(c + \\Delta c, \\sigma + \\Delta \\sigma)$ as follows:", + "bbox": [ + 212, + 646, + 782, + 708 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nC _ {f i n a l} (\\mathbf {r}) = \\sum_ {m = 1} ^ {M} T _ {m} \\left(1 - \\exp \\left(- \\left(\\sigma_ {m} + \\Delta \\sigma_ {m}\\right) \\delta_ {m}\\right)\\right) \\left(c _ {m} + \\Delta c _ {m}\\right). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 272, + 719, + 784, + 738 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\delta_{m}$ is the adjacent distance from the $m$ -th to the $m+1$ -th sample and $T_{m} = \\exp(-\\sum_{n=1}^{m-1} \\sigma_{n} \\delta_{n})$ as the transmittance. In addition, we compute $C_{rigid}$ in the same way from $(c, \\sigma)$ for an additional loss term that encourages a decomposed learning of the canonical radiance field and the radiance residuals (Sec. 3.4).", + "bbox": [ + 212, + 750, + 787, + 809 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Learnable Pose-Embedded Implicit Feature. We introduce high-frequency modulated implicit features computed from a learnable base code and a 3D body", + "bbox": [ + 212, + 809, + 787, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "MoCo-NeRF", + "bbox": [ + 643, + 114, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 125 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "pose. A learnable base code $\\mathbf{S} \\in \\mathbb{R}^{64}$ is randomly initialized and jointly optimized for a single subject with our motion-oriented compositional radiance fields. To make $\\mathbf{S}$ distinguishable for each pose, we propose to employ the 3D body pose as a guidance for computing implicit features that encourage high-fidelity and pose-dependent radiance residual learning. A cross-attention mechanism is designed to perform interactions between the learnable base code and 23 joint positions (i.e. $J = \\{j_{i}\\}_{i=1}^{23}$ ). A detailed overview of the mechanism is illustrated in Fig. 3. To allow high-frequency implicit feature learning with 3D body joints, we conduct positional encoding [42] for the input joint coordinates. We first project each joint position into a higher dimension with a sinusoidal positional encoding, $\\gamma(j_{i}) = (j_{i}, \\sin(2^{0} \\pi j_{i}), \\cos(2^{0} \\pi j_{i}), \\dots, \\sin(2^{L-1} \\pi j_{i}), \\cos(2^{L-1} \\pi j_{i}))$ , where $L = 10$ is the number of frequency bands. We then project the encoded points (i.e., $\\gamma(J) \\in \\mathbb{R}^{23 \\times 63}$ ) to a raw pose-embedded code $\\mathbf{P}_{raw}$ by a projection matrix $\\mathbf{W}_{p}$ :", + "bbox": [ + 212, + 146, + 787, + 345 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {P} _ {\\text {r a w}} = \\mathbf {W} _ {p} \\cdot \\gamma (J). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 433, + 358, + 785, + 375 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Then, we generate a query signal $\\mathbf{P}_q$ from $\\mathbf{S}$ , and key and value signals from $\\mathbf{P}_{raw}$ via linear projections as follows:", + "bbox": [ + 214, + 380, + 785, + 411 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {P} _ {q} = \\mathbf {W} _ {Q} \\cdot \\mathbf {S}, \\quad \\mathbf {P} _ {k} = \\mathbf {W} _ {K} \\cdot \\mathbf {P} _ {\\text {r a w}}, \\quad \\mathbf {P} _ {v} = \\mathbf {W} _ {V} \\cdot \\mathbf {P} _ {\\text {r a w}}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 310, + 422, + 784, + 439 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Finally, we generate the learnable pose-embedded implicit feature $\\mathbf{P}_{nr}$ for the radiance residual estimation as follows:", + "bbox": [ + 214, + 450, + 785, + 479 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {P} _ {n r} = \\operatorname {s o f t m a x} \\left(\\mathbf {P} _ {q} \\cdot \\left(\\mathbf {P} _ {k}\\right) ^ {\\top}\\right) \\cdot \\mathbf {P} _ {v}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 385, + 492, + 784, + 508 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As explained earlier, we employ $\\mathbf{P}_{nr}$ to condition $\\mathbf{M}_{nr}$ , the MLP decoder of the non-rigid radiance branch. $\\mathbf{M}_{nr}$ optimizes the implicit feature $\\mathbf{P}_{nr}$ and its generation module for estimating pose-coherent radiance residuals instead of solely relying on explicit volumetric features.", + "bbox": [ + 214, + 520, + 785, + 580 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.3 Multi-subject Unified Training", + "text_level": 1, + "bbox": [ + 214, + 603, + 516, + 619 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our effective radiance compositional approach for non-rigid motion modeling further enables a flexible extension of our single-subject MoCo-NeRF for unified training with multiple subjects, which is very beneficial for practical applications. Each monocular video exhibits distinct non-rigid motions, thereby requiring individualized modeling. However, it is impossible for the traditional spatial offset learning approach [73] that typically requires a heavy module for photo-realistic quality. In contrast, our radiance compositional approach can achieve it by employing a separate, lightweight MHE for each subject, referred to as local MHE.", + "bbox": [ + 212, + 628, + 787, + 750 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Building on this foundation, we propose three architectural modifications as illustrated in Fig. 4. To benefit from unified training of multiple subjects, we introduce a coarse global MHE $\\psi_{hash}^{G}$ shared across all subjects to learn subject-generic representations (e.g., body shapes). We also ensure learning of subject-specific details with $N_{s}$ local MHEs $\\{\\psi_{hash}^{i}\\}_{i = 1}^{N_{s}}$ , where $N_{s}$ is the number of subjects. For each canonical sample $\\mathbf{x}_c$ , we query global features from the", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "J. Kim et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "global MHE and concatenate it to the sliced feature from the corresponding local MHE for radiance predictions. We additionally propose to leverage the learnable base codes of all subjects $\\{\\mathbf{S}_i\\}_{i=1}^{N_s}$ as ID codes, for sharing MLP decoders while ensuring subject-distinctive canonical representations. The modified radiance estimations of the rigid branch for the multi-subject training are as follows:", + "bbox": [ + 212, + 146, + 787, + 223 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nc, \\sigma = M _ {r i g i d} \\left(\\psi_ {h a s h} ^ {G} \\left(\\mathbf {x} _ {c}\\right) \\oplus \\left(\\psi_ {h a s h} ^ {i} \\left(\\mathbf {x} _ {c}\\right)\\right) ^ {: H} \\oplus \\mathbf {S} _ {i}\\right) \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 334, + 229, + 785, + 250 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "3.4 Model Optimization", + "text_level": 1, + "bbox": [ + 214, + 267, + 429, + 282 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We optimize MoCo-NeRF with all $N_{i}$ image frames $\\{I_j\\}_{j=1}^{N_i}$ from the monocular video via patch-based ray sampling. Given an image frame $I_j$ at each iteration, we randomly sample $N_p$ patches denoted as $\\{P_k\\}_{k=1}^{N_p}$ to construct the LPIPS [84] loss. The LPIPS loss term can measure the perceptual distance between two image patches, and thus we also adopt it for more perceptually coherent renderings. Our framework first renders $\\{\\hat{P}_k^r, \\hat{P}_k^f\\}_{k=1}^{N_p}$ respectively from $C_{rigid}$ and $C_{final}$ , and compute two LPIPS loss terms $\\mathcal{L}_{\\mathrm{LPIPS}}^r$ and $\\mathcal{L}_{\\mathrm{LPIPS}}^f$ against the ground truth patch $P_k^{GT}$ by using the pre-trained VGGNet [63] to as $\\mathcal{L}_{\\mathrm{LPIPS}} = \\mathrm{LPIPS}(VGG(\\hat{P}_k), VGG(P_k^{GT}))$ . We respectively compute the MSE loss terms $\\mathcal{L}_{\\mathrm{MSE}}^r, \\mathcal{L}_{\\mathrm{MSE}}^f$ from each of the estimated patches $\\{\\hat{P}_k^r, \\hat{P}_k^f\\}_{k=1}^{N_p}$ . Moreover, we additionally compute an appearance matching loss with SSIM $\\mathcal{L}_{\\mathrm{SSIM}}^f$ [15] only for the final RGB $\\{\\hat{P}_k^f\\}_{k=1}^{N_p}$ , to improve the rendering quality. In summary, there are 5 loss terms in total: $\\mathcal{L}_{\\mathrm{LPIPS}}^r, \\mathcal{L}_{\\mathrm{MSE}}^r$ for the canonical radiance field of the rigid branch, and $\\mathcal{L}_{\\mathrm{LPIPS}}^f, \\mathcal{L}_{\\mathrm{MSE}}^f, \\mathcal{L}_{\\mathrm{SSIM}}^f$ for the final motion-oriented compositional renderings. To enforce the rigid branch to learn the main canonical radiance fields while the non-rigid branch to learn radiance residuals, we compute the final overall loss function $\\mathcal{L}^o$ with a hyperparameter $\\lambda = 0.2$ as:", + "bbox": [ + 212, + 287, + 787, + 564 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} ^ {o} = \\lambda \\mathcal {L} ^ {r} + (1 - \\lambda) \\mathcal {L} ^ {f} \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 423, + 573, + 784, + 588 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} ^ {r} = \\mathcal {L} _ {\\text {L P I P S}} ^ {r} + \\mathcal {L} _ {\\text {M S E}} ^ {r}, \\quad \\mathcal {L} ^ {f} = \\mathcal {L} _ {\\text {L P I P S}} ^ {f} + \\mathcal {L} _ {\\text {M S E}} ^ {f} + \\mathcal {L} _ {\\text {S S I M}} ^ {f} \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 313, + 593, + 784, + 611 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 214, + 631, + 375, + 648 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We conduct extensive experiments on ZJU-MoCap [52] and MonoCap [17, 18, 50] to verify the effectiveness of our proposed approach on both single- and multi-subject settings. The multi-subject setting indicates multiple subjects are jointly optimized. Our findings demonstrate that MoCo-NeRF can learn photo-realistic dynamic human representations by effectively handling non-rigid motions. Moreover, it learns discriminative representations of each performer in the multi-subject setting, while prior works struggle.", + "bbox": [ + 212, + 660, + 787, + 767 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Datasets", + "text_level": 1, + "bbox": [ + 214, + 787, + 333, + 800 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We leverage ZJU-MoCap and MonoCap datasets for both training and evaluation, following [14,21], and provide the details in this section. For ZJU-MoCap,", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "MoCo-NeRF", + "bbox": [ + 643, + 114, + 730, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/07cb5d98e02568dadb19d9bc4b071ebf6d91abccd9ae0b98321a361ceaab7b4d.jpg", + "table_caption": [ + "Table 1: Quantitative results of single-subject methods on ZJU-MoCap and MonoCap datasets where LPIPS* = LPIPS × 10³ with the best and second-best performance notations. MoCo-NeRF outperforms the baselines on most of the evaluation metrics in both settings. Despite pixel-wise metrics (PSNR, SSIM) tend to favor smooth renderings generated from [14,21], we achieve state-of-the-art performance while effectively modeling the photo-realistic non-rigid motions as illustrated in Fig. 5." + ], + "table_footnote": [], + "table_body": "
Single-subject methodsZJU-MoCap [52]MonoCap [50]
PSNR↑SSIM↑LPIPS*↓PSNR↑SSIM↑LPIPS*↓
HumanNeRF [73]30.180.970930.1332.920.986412.93
Instant-NVR [14]30.730.969738.8731.890.987016.07
GauHuman [21]30.980.962030.1632.870.984913.54
MoCo-NeRF (Ours)31.060.973428.8333.010.987413.02
", + "bbox": [ + 220, + 239, + 784, + 339 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/19a464fe291d78058f3c0580c246fd2007b3679c25cd5f12876f9c27dde798c5.jpg", + "image_caption": [ + "Fig.5: Qualitative comparison of novel view synthesis from single-subject training. MoCo-NeRF achieves photo-realistic rendering qualities with high-fidelity non-rigid motion modeling. Although HumanNeRF [73] presents comparable quality, however, MoCo-NeRF achieves much less training time as shown in Tab. 3." + ], + "image_footnote": [ + "HumanNeRF Instant-NVR GauHuman Ours Ground Truth HumanNeRF Instant-NVR GauHuman Ours Ground Truth" + ], + "bbox": [ + 232, + 351, + 772, + 573 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "we evaluate our MoCo-NeRF and other baselines on all 6 subjects (i.e., 377, 386, 387, 392, 393, 394). Out of 23 camera views, we use one of them for training and the other 22 views for evaluation. For MonoCap that comprises DynaCap [17] and DeepCap [18], we utilize all 4 subjects (i.e., lan, marc, olek, vlad) and use the same set of cameras as [14]. For multi-subject evaluations presented in Tab. 2, we evaluate on subjects 377, 386, and 392 of ZJU-MoCap.", + "bbox": [ + 212, + 676, + 787, + 767 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 Implementation Details", + "text_level": 1, + "bbox": [ + 214, + 786, + 457, + 801 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We optimize our model with Adam optimizer [30] with betas (0.9, 0.99), and learning rates of $5e^{-5}$ for the skeletal weights volume decoder and $5e^{-3}$ for the", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "J. Kim et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/87d34d7b5f95d9a56752b91ec2e7d0ad197ce457cdd0d9879c7a6b680468fcd0.jpg", + "table_caption": [ + "Table 2: Quantitative results of multi-subject methods on subjects #377, #386, #392 of ZJU-MoCap [52]. Despite the same evaluation settings, MoCo-NeRF significantly outperforms baseline methods under monocular settings with a shorter training time." + ], + "table_footnote": [], + "table_body": "
Multi-subject methodsPSNR↑SSIM↑LPIPS*↓Training videosTrain time
NHP [31]26.640.918540.47multi-view~86 hours
GHuNeRF [33]27.280.929140.82monocular~123 hours
MoCo-NeRF (Ours)31.970.978124.54monocular~2 hours
", + "bbox": [ + 217, + 200, + 782, + 270 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/fe8cbe06fa067da8562cf3b37049cd85ce5cb4bfd6794f43ac23813f070bd79f.jpg", + "image_caption": [ + "Fig.6: Qualitative comparison of novel view synthesis from multi-subject training. Although all subjects are seen during training, only our multi-subject MoCo-NeRF achieves photo-realistic quality comparable to the ones of single-subject models." + ], + "image_footnote": [], + "bbox": [ + 217, + 282, + 784, + 396 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "others. Each MHE has 16 levels with feature dimensions $H = 2$ , and 4 levels with $H = 2$ for the global MHE. For each iteration, six $32 \\times 32$ patches are used with 256 samples per ray. We train all models including the baselines on 1 Nvidia RTX 3090, where both baselines are trained for their suggested duration. Our model is trained for $1 \\sim 2$ hours depending on each subject's complexity for the single-subject training, and 2 hours for the multi-subject training.", + "bbox": [ + 212, + 478, + 784, + 569 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.3 State-of-the-art Comparison", + "text_level": 1, + "bbox": [ + 214, + 590, + 493, + 604 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Single-subject methods. Tab. 1 presents PSNR, SSIM [72], and LPIPS* of MoCo-NeRF and our baselines [14, 21, 73], for novel-view synthesis under the single-subject setting. MoCo-NeRF achieves state-of-the-art performance on all metrics except one comparable LPIPS* for the MonoCap dataset. As shown in Fig. 5, Instant-NVR [14] shows moderate performance, however, it often generates smoothed-out results without proper non-rigid motions due to high dependency on accurate SMPL parameters. GauHuman [21] achieves better qualitative performance than [14] but still struggles to model sophisticated non-rigid motions. Fig. 1 presents longer optimizations are surprisingly ineffective and harmful respectively for Instant-NVR [14] and GauHuman [21] in terms of non-rigid motion modeling. This implies that their approaches cannot achieve optimal non-rigid motion modeling. Moreover, we observe that PSNR and SSIM often favor those results without fine-grained non-rigid motion modeling since predicting pose-dependent deformations from novel viewpoints may not be pixel-wise accurate. Instead, the perception metric (LPIPS*) evaluates better in this case, as", + "bbox": [ + 212, + 612, + 785, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "MoCo-NeRF", + "bbox": [ + 643, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/47148ec027cdd0edf0acb096fc4a9498445118f999d5993767d80d7039735aff.jpg", + "table_caption": [ + "Table 3: Training efficiency comparison for methods with photo-realistic non-rigid motion modeling on a single RTX3090. In the table, single-sub. and multi-sub. respectively represent training for 1 and 6 subjects. MoCo-NeRF significantly outperforms HumanNeRF in terms of both GPU memory usage and training time. HumanNeRF requires sequential training to achieve comparable non-rigid motion modeling for all subjects, thus resulting in linearly-increased training time." + ], + "table_footnote": [], + "table_body": "
GPU Mem. (GB)Train time (hours)
single-sub.multi-sub.Δsingle-sub.multi-sub.Δ
HumanNeRF [73]23.723.7(+0)38.9233.3(+194.5)
MoCo-NeRF (Ours)16.220.3(+4.1)1.92.1(+0.2)
", + "bbox": [ + 220, + 239, + 784, + 311 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "HumanNeRF [73] has a better LPIPS* score and also shows qualitatively better results. Thus, we consider GauHuman's comparable PSNR might be favored by a superior representation (3D Gaussians [29]) instead of better non-rigid motion modeling. A more detailed analysis is in supplementary materials. Despite this, we still achieve state-of-the-art performance with significantly improved training efficiency compared to another photo-realistic model [73] (Tab. 3).", + "bbox": [ + 212, + 339, + 784, + 430 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Multi-subject methods. We evaluate our MoCo-NeRF against NHP [31] and GHuNeRF [33], both of which are designed to achieve generalization capabilities. This is necessitated by the absence of prior works that directly target the joint learning task for multiple subjects from monocular videos. Additionally, we observe that the single-subject baseline models fail to learn distinct representations when trained in a unified manner. This observation underscores our decision to compare against generalization models that are specifically tailored to leverage multi-subject data for training. For fairness, we use their official checkpoints and ensure all evaluated subjects were included for training.", + "bbox": [ + 212, + 430, + 785, + 566 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "As demonstrated in Tab. 2, MoCo-NeRF remarkably outperforms all the baseline methods on all metrics while also excelling in training efficiency up to 61 times faster. Even though MoCo-NeRF only utilizes a monocular video of each subject, Fig. 6 illustrates our method succeeds in learning both distinct representations and pose-dependent radiance residuals of each subject, enabling photo-realistic free-viewpoint renderings.", + "bbox": [ + 212, + 566, + 787, + 657 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.4 Model Analysis", + "text_level": 1, + "bbox": [ + 214, + 679, + 392, + 694 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "In this section, we present analyses of our contributions respectively for two major achievements: novel non-rigid motion modeling (NR modeling) and multi-subject unified training (multi-subject modeling). For both analyses, we progressively remove each component from a full model to a vanilla approach.", + "bbox": [ + 212, + 704, + 787, + 762 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "NR Modeling: Analysis of radiance compositional approach. Tab. 4 demonstrates our radiance compositional design perceptually improves the rendering results by closing the radiance discrepancies with predicted residuals. Fig. 7 shows that the representation learned solely by the rigid radiance branch (c) suffers from noisy renderings caused by incorrect correspondences due to", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "J. Kim et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/7ab68aa23ca6b2977b0ee70b98fa7d3fd2bba447e33f9324c55b2c0cf9fdc9d3.jpg", + "table_caption": [ + "Table 4: Ablation study in the single-subject task on ZJU-MoCap for b) pose-embedded implicit features and c) motion-oriented compositional NeRF approach." + ], + "table_footnote": [], + "table_body": "
Single-subjectPSNRSSIMLPIPS*
(a) MoCo-NeRF (Ours)31.060.973428.83
(b) (-) Pose features30.640.970229.65
(c) (-) Residual NeRF30.620.970229.93
", + "bbox": [ + 217, + 212, + 496, + 268 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/386a9798cc2ddcbdb54206ff4de74c2d49f0373489b3a76ddca1baf786629d9e.jpg", + "table_caption": [ + "Table 5: Ablation study of multi-subject MoCo-NeRF on ZJU-MoCap for our proposed architectural modifications." + ], + "table_footnote": [], + "table_body": "
Multi-subjectPSNRSSIMLPIPS*
(i)MoCo-NeRF (Ours)30.910.974627.69
(ii)(-) Global MHE30.830.974628.49
(iii)(-)Ns local MHEs30.700.974229.03
(iv)(-) Identity codes28.630.966638.29
", + "bbox": [ + 504, + 200, + 785, + 268 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/0ab548174b18a0c6144622ad59a44e88d1a421d2246d0e26a86202d5d3922bcb.jpg", + "image_caption": [ + "Fig.7: Qualitative results of the ablation study on single-subject MoCo-NeRF presented in Tab. 4." + ], + "image_footnote": [], + "bbox": [ + 218, + 282, + 496, + 407 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/97934d8dcb9bd581b3409de810265772b7ff5832e9695aa757622325ac14d1b4.jpg", + "image_caption": [ + "Fig. 8: Qualitative results of the ablation study in Tab. 5 where we can observe gradual improvements from (iv) to (i)." + ], + "image_footnote": [], + "bbox": [ + 508, + 284, + 785, + 406 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "non-rigid motions. With our radiance residual modeling (b), those issues are mitigated by jointly learning per-sample radiance residuals, resulting in more complete rendering and better non-rigid motion modeling as clearly shown in Fig. 7. In addition, Fig. 9 illustrates successful decomposed learning of two types of representations and validates the effectiveness of our radiance compositional approach in non-rigid motion modeling.", + "bbox": [ + 212, + 491, + 787, + 582 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "NR Modeling: Analysis of pose-embedded implicit feature. Tab. 4 further illustrates that the final version with the proposed pose-embedded implicit features (Tab. 4 (a)) significantly improves the quantitative performance on all metrics. It enables the non-rigid branch to implicitly utilize body pose information in addition to the explicit features, yielding smoother and more coherent non-rigid motion modeling as shown in Fig. 7. We can observe remarkable quality enhancements that demonstrate the effectiveness of our attention-based code generation module. We also present an ablation study on the choice of cross-attention module for the feature generation in the supplementary material.", + "bbox": [ + 212, + 582, + 787, + 718 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Multi-subject Modeling: Analysis of using learnable base codes as ID codes. In Fig. 8, the results of MoCo-NeRF without any modifications (iv) exhibit mixed-up appearances on all target subjects, similar to the ones of [73] that are presented in the supplementary material. Utilizing a set of learnable base codes as ID codes (Tab. 5 (iii)) enables a single MLP decoder $M_{rigid}$ to differentiate each subject by optimizing the ID codes as plug-in parameters. Thus, distinct representations of all subjects are successfully learned as illustrated in (iii) of Fig. 8, yielding significant quantitative improvements as well (Tab. 5).", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "MoCo-NeRF", + "bbox": [ + 643, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/f634bdf0c938ae43e180a834ea1bcd427cad6091693d578b317015426324aa83.jpg", + "image_caption": [ + "Fig. 9: Qualitative illustrations of non-rigid radiance residuals on novel-view synthesis where $\\delta \\rightarrow c' = c + \\delta \\Delta c$ , $\\sigma' = \\sigma + \\delta \\Delta \\sigma$ . The first row progressively demonstrates the capability of the non-rigid radiance branch to learn the non-rigid motions on the subject's face in addition to the cloth folds. The last two rows clearly illustrate the effects of radiance residuals on cloth folds and hems corresponding to the body pose." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 785, + 344 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Multi-subject Modeling: Analysis of $N_{s}$ local MHEs. Fig. 8 shows that the model (iii) lacks fine-grained details due to the sharing of one MHE. By utilizing separate $N_{s}$ local MHEs (Tab. 5 (ii)), the model learns subject-discriminative non-rigid motions as they now optimize separate explicit features for each subject's radiance residuals. Fig. 8 shows clear improvement of the model (ii) in non-rigid motion modeling where its quantitative gains are reflected in Tab. 5. Multi-subject Modeling: Analysis of global MHE. The addition of global MHE (i) further empowers MoCo-NeRF to learn better representations with more coherent non-rigid motions as illustrated in Fig. 8. Although the global MHE is not used for learning non-rigid radiance residuals, it encourages the rigid part of features from local MHE to concentrate on subject-specific representations. As its frozen copy is used for the non-rigid radiance branch, it facilitates better learning of radiance residuals for modeling non-rigid motions. Tab. 5 clearly shows that it can help to learn perceptually better representations.", + "bbox": [ + 212, + 454, + 787, + 667 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 688, + 359, + 704 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we presented a novel framework MoCo-NeRF for free-viewpoint dynamic human rendering that is capable of both single- and multi-subject training from monocular videos. By decomposing the neural representation of a dynamic human by motions and utilizing proposed implicit pose codes, MoCo-NeRF effectively learns radiance discrepancies induced by non-rigid motions. The proposed multi-subject MoCo-NeRF efficiently achieves comparable performance compared to the single-subject model. It has fully demonstrated its effectiveness and established new state-of-the-art performance on both problems.", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "J. Kim et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgments. This research is supported in part by the Hong Kong PhD Fellowship Scheme (HKPFS), Early Career Scheme of the Research Grants Council (RGC) of the Hong Kong SAR under grant No. 26202321, SAIL Research Project, HKUST-Zeekr Collaborative Research Fund, HKUST-WeBank Joint Lab Project, Tencent Rhino-Bird Focused Research Program, and Naver Cloud Corporation.", + "bbox": [ + 212, + 146, + 787, + 238 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 258, + 321, + 275 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Alldieck, T., Zanfir, M., Sminchisescu, C.: Photorealistic monocular 3d reconstruction of humans wearing clothing. In: CVPR (2022)", + "2. Barron, J.T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., Srinivasan, P.P.: Mip-NeRF: A multiscale representation for anti-aliasing neural radiance fields. In: ICCV (2021)", + "3. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Mip-NeRF 360: Unbounded anti-aliased neural radiance fields. In: CVPR (2022)", + "4. Chen, A., Xu, Z., Geiger, A., Yu, J., Su, H.: TensoRF: Tensorial radiance fields. In: ECCV (2022)", + "5. Chen, A., Xu, Z., Zhao, F., Zhang, X., Xiang, F., Yu, J., Su, H.: MVSNeRF: Fast generalizable radiance field reconstruction from multi-view stereo. In: ICCV (2021)", + "6. Chen, X., Jiang, T., Song, J., Yang, J., Black, M.J., Geiger, A., Hilliges, O.: gDNA: Towards generative detailed neural avatars. In: CVPR (2022)", + "7. Chen, Y., Wang, X., Chen, X., Zhang, Q., Li, X., Guo, Y., Wang, J., Wang, F.: Uv volumes for real-time rendering of editable free-view human performance. In: CVPR (2023)", + "8. Collet, A., Chuang, M., Sweeney, P., Gillett, D., Evseev, D., Calabrese, D., Hoppe, H., Kirk, A., Sullivan, S.: High-quality streamable free-viewpoint video. In: SIGGRAPH (2015)", + "9. Dong, Z., Chen, X., Yang, J., J.Black, M., Hilliges, O., Geiger, A.: AG3D: Learning to generate 3D avatars from 2D image collections. In: ICCV (2023)", + "0. Dou, M., Khamis, S., Degtyarev, Y., Davidson, P., Fanello, S.R., Kowdle, A., Escolano, S.O., Rhemann, C., Kim, D., Taylor, J., Kohli, P., Tankovich, V., Izadi, S.: Fusion4D: Real-time performance capture of challenging scenes. In: SIGGRAPH (2016)", + "1. Du, Y., Zhang, Y., Yu, H.X., Tenenbaum, J.B., Wu, J.: Neural radiance flow for 4d view synthesis and video processing. In: ICCV (2021)", + "2. Gao, C., Saraf, A., Kopf, J., Huang, J.B.: Dynamic view synthesis from dynamic monocular video. In: ICCV (2021)", + "3. Garbin, S.J., Kowalski, M., Johnson, M., Shotton, J., Valentin, J.: FastNeRF: High-fidelity neural rendering at 200fps. In: ICCV (2021)", + "4. Geng, C., Peng, S., Xu, Z., Bao, H., Zhou, X.: Learning neural volumetric representations of dynamic humans in minutes. In: CVPR (2023)", + "5. Godard, C., Mac Aodha, O., Brostow, G.J.: Unsupervised monocular depth estimation with left-right consistency. In: CVPR (2017)", + "6. Guo, K., Lincoln, P., Davidson, P., Busch, J., Yu, X., Whalen, M., Harvey, G., Orts-Escalano, S., Pandey, R., Dourgarian, J., Tang, D., Tkach, A., Kowdle, A., Cooper, E., Dou, M., Fanello, S., Fyffe, G., Rhemann, C., Taylor, J., Debevec, P., Izadi, S.: The relightables: Volumetric performance capture of humans with realistic relighting. In: SIGGRAPH Asia (2019)" + ], + "bbox": [ + 225, + 287, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "MoCo-NeRF", + "bbox": [ + 643, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "17. Habermann, M., Liu, L., Xu, W., Zollhoefer, M., Pons-Moll, G., Theobalt, C.: Real-time deep dynamic characters. In: SIGGRAPH (2021)", + "18. Habermann, M., Xu, W., Zollhoefer, M., Pons-Moll, G., Theobalt, C.: DeepCap: Monocular human performance capture using weak supervision. In: CVPR (2020)", + "19. Hedman, P., Srinivasan, P.P., Mildenhall, B., Barron, J.T., Debevec, P.: Baking neural radiance fields for real-time view synthesis. In: ICCV (2021)", + "20. Hong, F., Chen, Z., Lan, Y., Pan, L., Liu, Z.: EVA3D: Compositional 3d human generation from 2d image collections. In: ICLR (2023)", + "21. Hu, S., Liu, Z.: GauHuman: Articulated gaussian splatting from monocular human videos. In: CVPR (2024)", + "22. Hu, T., Yu, T., Zheng, Z., Zhang, H., Liu, Y., Zwicker, M.: HVTR: Hybrid volumetric-textural rendering for human avatars. In: 3DV (2022)", + "23. Huang, Y., Yi, H., Liu, W., Wang, H., Wu, B., Wang, W., Lin, B., Zhang, D., Cai, D.: One-shot implicit animatable avatars with model-based priors. In: ICCV (2023)", + "24. Jain, A., Tancik, M., Abbeel, P.: Putting nerf on a diet: Semantically consistent few-shot view synthesis. In: ICCV (2021)", + "25. Jiang, B., Hong, Y., Bao, H., Zhang, J.: SelfRecon: Self reconstruction your digital avatar from monocular video. In: CVPR (2022)", + "26. Jiang, T., Chen, X., Song, J., Hilliges, O.: Instantavatar: Learning avatars from monocular video in 60 seconds. In: CVPR (2023)", + "27. Jiang, W., Yi, K.M., Samei, G., Tuzel, O., Ranjan, A.: NeuMan: Neural human radiance field from a single video. In: ECCV (2022)", + "28. Jiang, Y., Yao, K., Su, Z., Shen, Z., Luo, H., Xu, L.: Instant-NVR: Instant neural volumetric rendering for human-object interactions from monocular rgbd stream. In: CVPR (2023)", + "29. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM TOG (July 2023)", + "30. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. In: ICLR (2015)", + "31. Kwon, Y., Kim, D., Ceylan, D., Fuchs, H.: Neural human performer: Learning generalizable radiance fields for human performance rendering. In: NeurIPS (2021)", + "32. Lewis, J.P., Cordner, M., Fong, N.: Pose space deformation: A unified approach to shape interpolation and skeleton-driven deformation. In: SIGGRAPH (2000)", + "33. Li, C., Lin, J., Lee, G.H.: Generalizable human nerf from a monocular video. In: 3DV (2024)", + "34. Li, R., Tanke, J., Vo, M., Zollhofer, M., Gall, J., Kanazawa, A., Lassner, C.: TAVA: Template-free animatable volumetric actors. In: ECCV (2022)", + "35. Li, Z., Niklaus, S., Snavely, N., Wang, O.: Neural scene flow fields for space-time view synthesis of dynamic scenes. In: CVPR (2021)", + "36. Liu, L., Gu, J., Lin, K.Z., Chua, T.S., Theobalt, C.: Neural sparse voxel fields. In: NeurIPS (2020)", + "37. Liu, L., Habermann, M., Rudnev, V., Sarkar, K., Gu, J., Theobalt, C.: Neural actor: Neural free-view synthesis of human actors with pose control. In: SIGGRAPH Asia (2021)", + "38. Liu, Y., Peng, S., Liu, L., Wang, Q., Wang, P., Christian, T., Zhou, X., Wang, W.: Neural rays for occlusion-aware image-based rendering. In: CVPR (2022)", + "39. Loper, M., Mahmood, N., Romero, J., Pons-Moll, G., Black, M.J.: SMPL: A skinned multi-person linear model. In: SIGGRAPH Asia (2015)" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "J. Kim et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "40. Martin-Brualla, R., Pandey, R., Yang, S., Pidlypenskyi, P., Taylor, J., Valentin, J., Khamis, S., Davidson, P., Tkach, A., Lincoln, P., Kowdle, A., Rhemann, C., Goldman, D.B., Keskin, C., Seitz, S., Izadi, S., Fanello, S.: LookinGood: Enhancing performance capture with real-time neural re-rendering. ACM TOG (2018)", + "41. Mihajlovic, M., Bansal, A., Zollhoefer, M., Tang, S., Saito, S.: KeypointNeRF: Generalizing image-based volumetric avatars using relative spatial encoding of keypoints. In: ECCV (2022)", + "42. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: NeRF: Representing scenes as neural radiance fields for view synthesis. In: ECCV (2020)", + "43. Mu, J., Sang, S., Vasconcelos, N., Wang, X.: ActorsNeRF: Animatable few-shot human rendering with generalizable nerfs. In: ICCV (2023)", + "44. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. In: SIGGRAPH (2022)", + "45. Noguchi, A., Sun, X., Lin, S., Harada, T.: Unsupervised learning of efficient geometry-aware neural articulated representations. In: ECCV (2022)", + "46. Oechsle, M., Peng, S., Geiger, A.: UNISURF: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In: ICCV (2021)", + "47. Park, K., Sinha, U., Barron, J.T., Bouaziz, S., Goldman, D.B., Seitz, S.M., Martin-Brualla, R.: Nerfies: Deformable neural radiance fields. In: ICCV (2021)", + "48. Park, K., Sinha, U., Hedman, P., Barron, J.T., Bouaziz, S., Goldman, D.B., Martin-Brualla, R., Seitz, S.M.: HyperNeRF: A higher-dimensional representation for topologically varying neural radiance fields. In: SIGGRAPH Asia (2021)", + "49. Peng, B., Hu, J., Zhou, J., Zhang, J.: SelfNeRF: Fast training nerf for human from monocular self-rotating video. arXiv preprint arXiv:2210.01651 (2022)", + "50. Peng, S., Dong, J., Wang, Q., Zhang, S., Shuai, Q., Zhou, X., Bao, H.: Animatable neural radiance fields for modeling dynamic human bodies. In: ICCV (2021)", + "51. Peng, S., Xu, Z., Dong, J., Wang, Q., Zhang, S., Shuai, Q., Bao, H., Zhou, X.: Animatable implicit neural representations for creating realistic avatars from videos. TPAMI (2024)", + "52. Peng, S., Zhang, Y., Xu, Y., Wang, Q., Shuai, Q., Bao, H., Zhou, X.: Neural Body: Implicit neural representations with structured latent codes for novel view synthesis of dynamic humans. In: CVPR (2021)", + "53. Pumarola, A., Corona, E., Pons-Moll, G., Moreno-Noguer, F.: D-NeRF: Neural Radiance Fields for Dynamic Scenes. In: CVPR (2020)", + "54. Reiser, C., Peng, S., Liao, Y., Geiger, A.: KiloNeRF: Speeding up neural radiance fields with thousands of tiny mlps. In: ICCV (2021)", + "55. Rematas, K., Liu, A., Srinivasan, P.P., Barron, J.T., Tagliasacchi, A., Funkhouser, T., Ferrari, V.: Urban radiance fields. In: CVPR (2022)", + "56. Remelli, E., Bagautdinov, T., Saito, S., Wu, C., Simon, T., Wei, S.E., Guo, K., Cao, Z., Prada, F., Saragih, J., et al.: Drivable volumetric avatars using texel-aligned features. In: SIGGRAPH (2022)", + "57. Saito, S., Huang, Z., Natsume, R., Morishima, S., Kanazawa, A., Li, H.: PIFu: Pixel-aligned implicit function for high-resolution clothed human digitization. In: ICCV (2019)", + "58. Saito, S., Simon, T., Saragih, J., Joo, H.: PIFuHD: Multi-level pixel-aligned implicit function for high-resolution 3d human digitization. In: CVPR (2020)", + "59. Sara Fridovich-Keil and Alex Yu, Tancik, M., Chen, Q., Recht, B., Kanazawa, A.: Plenoxels: Radiance fields without neural networks. In: CVPR (2022)", + "60. Schonberger, J.L., Zheng, E., Frahm, J.M., Pollefeys, M.: Pixelwise view selection for unstructured multi-view stereo. In: ECCV (2016)" + ], + "bbox": [ + 215, + 146, + 787, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "MoCo-NeRF", + "bbox": [ + 643, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "61. Schonberger, J.L., Frahm, J.M.: Structure-from-motion revisited. In: CVPR (2016)", + "62. Shao, R., Zhang, H., Zhang, H., Chen, M., Cao, Y., Yu, T., Liu, Y.: DoubleField: Bridging the neural surface and radiance fields for high-fidelity human reconstruction and rendering. In: CVPR (2022)", + "63. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition. In: ICLR (2015)", + "64. Su, S.Y., Yu, F., Zollhöfer, M., Rhodin, H.: A-NeRF: Articulated neural radiance fields for learning human shape, appearance, and pose. In: NeurIPS (2021)", + "65. Su, Z., Xu, L., Zheng, Z., Yu, T., Liu, Y., Fang, L.: RobustFusion: Human volumetric capture with data-driven visual cues using a rgbd camera. In: ECCV (2020)", + "66. Tancik, M., Casser, V., Yan, X., Pradhan, S., Mildenhall, B., Srinivasan, P., Barron, J.T., Kretzschmar, H.: Block-NeRF: Scalable large scene neural view synthesis. In: CVPR (2022)", + "67. Tretschk, E., Tewari, A., Golyanik, V., Zollhöfer, M., Lassner, C., Theobalt, C.: Non-rigid neural radiance fields: Reconstruction and novel view synthesis of a dynamic scene from monocular video. In: ICCV (2021)", + "68. Turki, H., Ramanan, D., Satyanarayanan, M.: Mega-NERF: Scalable construction of large-scale nerfs for virtual fly-throughs. In: CVPR (2022)", + "69. Wang, P., Liu, L., Liu, Y., Theobalt, C., Komura, T., Wang, W.: NeuS: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. In: NeurIPS (2021)", + "70. Wang, Q., Wang, Z., Genova, K., Srinivasan, P., Zhou, H., Barron, J.T., Martin-Brualla, R., Snavely, N., Funkhouser, T.: IBRNet: Learning multi-view image-based rendering. In: CVPR (2021)", + "71. Wang, S., Schwarz, K., Geiger, A., Tang, S.: ARAH: Animatable volume rendering of articulated human sdfs. In: ECCV (2022)", + "72. Wang, Z., Bovik, A., Sheikh, H., Simoncelli, E.: Image quality assessment: from error visibility to structural similarity. TIP (2004)", + "73. Weng, C.Y., Curless, B., Srinivasan, P.P., Barron, J.T., Kemelmacher-Shlizerman, I.: HumanNeRF: Free-viewpoint rendering of moving people from monocular video. In: CVPR (2022)", + "74. Weng, C.Y., Srinivasan, P.P., Curless, B., Kemelmacher-Shlizerman, I.: Person-NeRF: Personalized reconstruction from photo collections. In: CVPR (2023)", + "75. Wu, M., Wang, Y., Hu, Q., Yu, J.: Multi-view neural human rendering. In: CVPR (2020)", + "76. Xian, W., Huang, J.B., Kopf, J., Kim, C.: Space-time neural irradiance fields for free-viewpoint video. In: CVPR (2021)", + "77. Xiangli, Y., Xu, L., Pan, X., Zhao, N., Rao, A., Theobalt, C., Dai, B., Lin, D.: Bungeenerf: Progressive neural radiance field for extreme multi-scale scene rendering. In: ECCV (2022)", + "78. Xu, H., Alldieck, T., Sminchisescu, C.: H-NeRF: Neural radiance fields for rendering and temporal reconstruction of humans in motion. In: NeurIPS (2021)", + "79. Xu, L., Xiangli, Y., Peng, S., Pan, X., Zhao, N., Theobalt, C., Dai, B., Lin, D.: Grid-guided neural radiance fields for large urban scenes. In: CVPR (2023)", + "80. Xu, Z., Peng, S., Lin, H., He, G., Sun, J., Shen, Y., Bao, H., Zhou, X.: 4K4D: Real-time 4d view synthesis at 4k resolution. In: CVPR (2024)", + "81. Yariv, L., Gu, J., Kasten, Y., Lipman, Y.: Volume rendering of neural implicit surfaces. In: NeurIPS (2021)", + "82. Yu, A., Li, R., Tancik, M., Li, H., Ng, R., Kanazawa, A.: PlenOctrees for real-time rendering of neural radiance fields. In: ICCV (2021)" + ], + "bbox": [ + 215, + 147, + 784, + 839 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "J. Kim et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "83. Yu, A., Ye, V., Tancik, M., Kanazawa, A.: pixelNeRF: Neural radiance fields from one or few images. In: CVPR (2021)", + "84. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: CVPR (2018)", + "85. Zhang, R., Chen, J.: NDF: Neural deformable fields for dynamic human modelling. In: ECCV (2022)", + "86. Zhao, F., Yang, W., Zhang, J., Lin, P., Zhang, Y., Yu, J., Xu, L.: HumanNeRF: Efficiently generated human radiance field from sparse inputs. In: CVPR (2022)", + "87. Zheng, Z., Huang, H., Yu, T., Zhang, H., Guo, Y., Liu, Y.: Structured local radiance fields for human avatar modeling. In: CVPR (2022)", + "88. Zheng, Z., Zhao, X., Zhang, H., Liu, B., Liu, Y.: AvatarRex: Real-time expressive full-body avatars. In: SIGGRAPH (2023)", + "89. Zhi, Y., Qian, S., Yan, X., Gao, S.: Dual-space nerf: Learning animatable avatars and scene lighting in separate spaces. In: 3DV (2022)" + ], + "bbox": [ + 212, + 146, + 787, + 340 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "MoCo-NeRF", + "bbox": [ + 643, + 114, + 730, + 126 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 18 + } +] \ No newline at end of file diff --git a/2024/Motion-Oriented Compositional Neural Radiance Fields for Monocular Dynamic Human Modeling/6e1b3c80-b2f1-4690-8b23-807237bf60da_model.json b/2024/Motion-Oriented Compositional Neural Radiance Fields for Monocular Dynamic Human Modeling/6e1b3c80-b2f1-4690-8b23-807237bf60da_model.json new file mode 100644 index 0000000000000000000000000000000000000000..880240c52c5d401688e2e23e9a81d84f64c7063f --- /dev/null +++ b/2024/Motion-Oriented Compositional Neural Radiance Fields for Monocular Dynamic Human Modeling/6e1b3c80-b2f1-4690-8b23-807237bf60da_model.json @@ -0,0 +1,2713 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.221, + 0.142, + 0.784, + 0.187 + ], + "angle": 0, + "content": "Motion-Oriented Compositional Neural Radiance Fields for Monocular Dynamic Human Modeling" + }, + { + "type": "text", + "bbox": [ + 0.307, + 0.212, + 0.697, + 0.229 + ], + "angle": 0, + "content": "Jaehyeok \\(\\mathrm{Kim}^{1}\\), Dongyoon \\(\\mathrm{Wee}^{2}\\), and Dan \\(\\mathrm{Xu}^{1\\boxtimes}\\)" + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.239, + 0.74, + 0.269 + ], + "angle": 0, + "content": "\\(^{1}\\) HKUST, Hong Kong \\(^{2}\\) Naver Cloud Corp., South Korea jkimbf@cse.ust.hk, dongyoon.wee@navercorp.com, danxu@cse.ust.hk" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.299, + 0.744, + 0.66 + ], + "angle": 0, + "content": "Abstract. This paper introduces Motion-oriented Compositional Neural Radiance Fields (MoCo-NeRF), a framework designed to perform free-viewpoint rendering of monocular human videos via novel non-rigid motion modeling approach. In the context of dynamic clothed humans, complex cloth dynamics generate non-rigid motions that are intrinsically distinct from skeletal articulations and critically important for the rendering quality. The conventional approach models non-rigid motions as spatial (3D) deviations in addition to skeletal transformations. However, it is either time-consuming or challenging to achieve optimal quality due to its high learning complexity without a direct supervision. To target this problem, we propose a novel approach of modeling non-rigid motions as radiance residual fields to benefit from more direct color supervision in the rendering and utilize the rigid radiance fields as a prior to reduce the complexity of the learning process. Our approach utilizes a single multiresolution hash encoding (MHE) to concurrently learn the canonical T-pose representation from rigid skeletal motions and the radiance residual field for non-rigid motions. Additionally, to further improve both training efficiency and usability, we extend MoCo-NeRF to support simultaneous training of multiple subjects within a single framework, thanks to our effective design for modeling non-rigid motions. This scalability is achieved through the integration of a global MHE and learnable identity codes in addition to multiple local MHEs. We present extensive results on ZJU-MoCap and MonoCap, clearly demonstrating state-of-the-art performance in both single- and multi-subject settings. The code and model will be made publicly available at the project page: https://stevejaehyeok.github.io/publications/moco-nerf." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.672, + 0.741, + 0.701 + ], + "angle": 0, + "content": "Keywords: Dynamic human modeling from monocular videos \\(\\cdot\\) Neural Radiance Fields \\(\\cdot\\) Novel view synthesis of humans" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.743, + 0.378, + 0.759 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.772, + 0.788, + 0.818 + ], + "angle": 0, + "content": "Novel view synthesis of monocular human video with photo-realistic quality is a challenging problem, as it involves complex articulated body motions and necessitates sophisticated modeling of non-skeletal or non-rigid motions (e.g., wrinkles" + }, + { + "type": "page_footnote", + "bbox": [ + 0.232, + 0.825, + 0.401, + 0.841 + ], + "angle": 0, + "content": "\\(\\boxtimes\\) Corresponding author." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.127 + ], + "angle": 0, + "content": "J. Kim et al." + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.147, + 0.782, + 0.269 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.275, + 0.782, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.421, + 0.788, + 0.561 + ], + "angle": 0, + "content": "Fig. 1: a) We introduce a motion-oriented compositional NeRF for photo-realistic modeling of dynamic humans from monocular videos. Our approach innovatively employs radiance compositions to capture pose-adaptive non-rigid motions, overcoming the limitations of skeletal transformations that typically yield an average of observed deformations (mean motions). b) The proposed MoCo-NeRF achieves state-of-the-art rendering quality and noteworthy efficiency in novel view synthesis, compared to leading competitors [14,21,73]. The bold denotes the proposed training duration of each comparison model. MoCo-NeRF uniquely captures coherent non-rigid motions, like T-shirt wrinkles relative to body pose, from entirely new viewpoints. Moreover, MoCo-NeRF significantly surpasses the efficiency of another photo-realistic method, HumanNeRF [73]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.596, + 0.788, + 0.763 + ], + "angle": 0, + "content": "of clothes) from unseen viewpoints. In particular, non-rigid motion modeling is essential for achieving photo-realism and coherence in free-viewpoint renderings. A prevalent solution adopted by recent methods [14,21,73] models the non-rigid motions as pose-dependent 3D spatial offsets of canonical point samples. However, the offset-based strategies lead to either extensive training time or irrelevant modeling of non-rigid motions due to its high learning complexity, incurred by the following reasons. First, a single spatial offset spans an unbounded range and is unique for every 3D point across all input frames and body poses, thereby escalating computational complexity. Second, the learning objective of spatial offsets is not directly aligned with the conventional RGB-based supervision, further complicating the modeling process." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.842 + ], + "angle": 0, + "content": "To target the issues above-discussed, in this paper, we propose Motion-oriented Compositional Neural Radiance Fields (MoCo-NeRF), a new framework that effectively models non-rigid motions at a radiance-level through a unique design of residual radiance field learning for free-viewpoint renderings of monocular human videos. Specifically, our proposed framework decomposes a single" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.644, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "MoCo-NeRF" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.358 + ], + "angle": 0, + "content": "MHE [44] into two branches: a rigid branch that learns a canonical T-pose radiance field solely via rigid skeletal transformations, and a non-rigid branch that optimizes pose-driven radiance residuals for capturing appearance discrepancies between renderings of the rigid branch and the ground-truth images, which are typically caused by non-rigid motions. The proposed approach circumvents the complexities associated with learning unbounded spatial offsets by reducing the prediction range to the radiance space with the rigid radiance field as a prior. It further establishes a more direct supervision from the ground-truth pixel colors for optimizing the rigid and non-rigid radiance fields by embedding them together in the rendering process, which can further enhance the learning efficiency. Besides, we present a cross-attention module to query a pose conditioned implicit feature from a learnable base code. This process allows our framework to extract more discriminative implicit features tailored to each particular pose, helping our non-rigid branch to learn coherent pose-driven radiance residuals." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.36, + 0.788, + 0.525 + ], + "angle": 0, + "content": "In addition, thanks to our effective design for modeling non-rigid motions, the proposed MoCo-NeRF architecture can be flexibly extended to support unified training across multiple subjects while maintaining photo-realistic quality, without a significant increase in training time. For this architectural extension, we introduce multiple local MHEs and a global MHE, and also extend the aforementioned base code to a set of identity (ID) codes as a learnable codebook. The multiple local MHEs are designed for modeling multiple subjects and the global MHE targets the learning of generic and universal body representations. Furthermore, subject-specific learnable ID codes that are pivotal for generating pose-embedded implicit features enable effective sharing of the MLP decoder in the rigid branch by all subjects, enhancing the model's efficiency and scalability." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.526, + 0.788, + 0.616 + ], + "angle": 0, + "content": "MoCo-NeRF achieves state-of-the-art performance on free-viewpoint rendering of monocular human videos in both single- and multi-subject settings, showing clear improvements compared to the prior works. We present our extensive results on ZJU-MoCap [52] and MonoCap [17, 18, 50] to demonstrate the effectiveness of the proposed MoCo-NeRF in learning photo-realistic representation and modeling complex non-rigid motions from monocular videos." + }, + { + "type": "text", + "bbox": [ + 0.239, + 0.617, + 0.633, + 0.632 + ], + "angle": 0, + "content": "In summary, the contribution of our work is threefold:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.643, + 0.787, + 0.734 + ], + "angle": 0, + "content": "- We propose a novel motion-oriented compositional NeRF approach, MoCo-NeRF, which concurrently learns canonical rigid radiance fields and pose-dependent radiance residual fields, effectively handling both rigid and non-rigid motions. We also introduce a pose-embedded implicit feature, computed by cross-attention between a body pose and a designed learnable base code, to implicitly enhance pose-dependent radiance residual learning." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.735, + 0.787, + 0.809 + ], + "angle": 0, + "content": "- The proposed MoCo-NeRF is the first work that can be readily extended to enable unified training of multiple subjects without significantly increasing the GPU memory demands. We achieve this through our local MHEs, a global MHE, and the extended identity codebook, which enable subject-discriminative representation learning and effective module sharing." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "- The proposed MoCo-NeRF achieves state-of-the-art performance under both single- and multi-subject training settings with a significant efficiency im" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.643, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.232, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.127 + ], + "angle": 0, + "content": "J. Kim et al." + }, + { + "type": "text", + "bbox": [ + 0.243, + 0.147, + 0.788, + 0.178 + ], + "angle": 0, + "content": "provement compared to the best-performing methods. MoCo-NeRF jointly learns all 6 subjects of ZJU-MoCap on one RTX 3090 GPU within 2 hours." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.202, + 0.389, + 0.218 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.235, + 0.786, + 0.401 + ], + "angle": 0, + "content": "We review closely related work from three perspectives: deformable neural rendering, neural human rendering, and monocular photo-realistic human rendering. Deformable neural rendering. NeRF [42] leverages an MLP to learn a static 3D representation of a scene from a dense set of images from diverse viewpoints. For its application in real-life scenarios, prior works have enhanced NeRF in various directions such as improving efficiencies [4,13,19,36,44,54,59,82] or relieving strong reliance on impractical settings like dense input views [5,24,38,70,83]. In addition, there are more branches of work such as large-scale modeling [55,66, 68,77,79], improving rendering quality [2,3], reconstruction quality [46,69,81] and more. However, they are restricted to static scenes while the majority of the real-world objects are dynamic." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.402, + 0.787, + 0.749 + ], + "angle": 0, + "content": "Recent works including [11, 12, 35, 47, 48, 53, 67, 76] have broadened the modeling capabilities of NeRF to dynamic scenes with small and simple movements/deformations. However, they often fail in handling more dynamic or complex motions like human body articulations. In contrast, our approach enables learning of human-articulated motions from monocular videos, and performing free-viewpoint rendering of a human performer at any time frame of the videos. Neural rendering of humans. Many early-stage prior works for neural human renderings utilize traditional techniques like multi-view stereo [16, 60, 61, 75] or to build sophisticated systems for capturing human motions [8, 10, 40, 65]. However, constructing a dense set of training views and building those capturing systems are expensive, making them unsuitable for real-life situations. [1, 6, 57, 58] address 3D human reconstruction by utilizing ground truth human scans, relieving the aforementioned requirements. [9, 20, 45] further improves it by enabling human reconstruction from 2D image collections. In recent days, there have been many works learning an implicit human representation for novel pose synthesis [22, 23, 25, 27, 34, 50, 56, 64, 71, 87-89] or for a novel-view synthesis [27, 28, 31, 37, 51, 52, 64, 73, 74, 78, 85, 86]. There also have been several generalizable works [41, 43, 62] for a novel human subject with pre-training on multi-view dataset and a few-shot fine-tuning. [7, 80] focus on improving rendering efficiency to achieve real-time free-view rendering of dynamic humans. Furthermore, [14, 26, 28, 49] adopt explicit hash encoding [44] in order to boost the training efficiency for human representation modeling. In contrast to our problem setup, most of those works require multi-view videos or do not model the non-rigid motions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Monocular photo-realistic human rendering. HumanNeRF [73] presents a state-of-the-art performance by optimizing motion fields and canonical T-type representation of a human performer. However, its implicit modeling of pose-dependent non-rigid motions requires tens of hours to converge. Instant-NVR [14] achieves efficient training with part-based representation and 2D motion parameterization for learning non-rigid motions as 3D offsets. Note that the" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.645, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "MoCo-NeRF" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.143, + 0.788, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.298, + 0.788, + 0.385 + ], + "angle": 0, + "content": "Fig. 2: Overview of the proposed framework MoCo-NeRF for free-view human rendering from a monocular video. Without estimating a geometrical offset of each continuous canonical point for each body pose, our framework is able to handle all deformations via its radiance-compositional approach with a single MHE [44] and achieve state-of-the-art performance. The pose-embedded implicit feature further enhances the learning of non-rigid radiance residuals by enabling pose-distinctive representation learning." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.414, + 0.789, + 0.596 + ], + "angle": 0, + "content": "human parametric model (SMPL) [39], an extra requirement for training, plays a significant role in their framework. In addition, a concurrent work GauHuman [21], which also employs spatial offset approach with Gaussian Splatting (GS) [29], further enhances the efficiency and performance upon [14] by leveraging SMPL for the density control of GS. Nevertheless, [14, 21] struggle to learn complex non-rigid motions, even with an extended training duration as shown in Fig. 1. In this paper, instead of directly learning spatial offsets like these existing works, we propose a novel approach that effectively models non-rigid body motions by learning motion-oriented compositional radiance fields conditioned by pose-embedded implicit features. Our approach clearly outperforms the best-performing competitors HumanNeRF [73], Instant-NVR [14], and GauHuman [21] and can further extend to support joint training of multiple subjects." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.623, + 0.603, + 0.641 + ], + "angle": 0, + "content": "3 The Proposed MoCo-NeRF Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.842 + ], + "angle": 0, + "content": "The proposed MoCo-NeRF aims to effectively learn high-fidelity representations for free-viewpoint rendering of dynamic human subjects in monocular videos. We introduce our novel radiance compositional NeRF approach based on different motion types: rigid and non-rigid motions, as illustrated in Fig. 2. The composition of canonical radiance and the per-sample pose-dependent radiance residuals (Sec. 3.2) facilitates photo-realistic quality of rendering with fine-grained modeling of non-rigid motions, rather than leveraging the conventional spatial 3D offset learning approach. We further improve the quality of non-rigid radiance residuals by introducing a pose-embedded implicit features (Sec. 3.2) that is jointly optimized with the non-rigid radiance branch. In Sec. 3.3, thanks to our effective design of non-rigid motion modeling, it enables MoCo-NeRF to be flexibly extended for unified training of multiple subjects with a high efficiency." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.128 + ], + "angle": 0, + "content": "J. Kim et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.577, + 0.163 + ], + "angle": 0, + "content": "3.1 Preliminary: Skeletal Transformations" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.171, + 0.788, + 0.233 + ], + "angle": 0, + "content": "The skeletal motion learns an inverse formulation of the linear blend skinning [32], which transforms canonical points to different body poses by skinning weights. Weng et al. [73] reformulate the linear blend skinning to transform a point \\(\\mathbf{x}\\) in the observation space to a canonical point \\(\\mathbf{x}_c\\):" + }, + { + "type": "equation", + "bbox": [ + 0.323, + 0.244, + 0.786, + 0.285 + ], + "angle": 0, + "content": "\\[\n\\mathbf {x} _ {c} = \\sum_ {k = 1} ^ {K} w _ {k} ^ {o} \\left(R _ {k} \\mathbf {x} + \\mathbf {t} _ {k}\\right), w _ {k} ^ {o} = \\frac {w _ {k} ^ {c} \\left(R _ {k} \\mathbf {x} + \\mathbf {t} _ {k}\\right)}{\\sum_ {j = 1} ^ {K} w _ {j} ^ {c} \\left(R _ {j} \\mathbf {x} + \\mathbf {t} _ {j}\\right)}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.294, + 0.789, + 0.356 + ], + "angle": 0, + "content": "where \\( w_{k}^{o} \\) and \\( w_{k}^{c} \\) respectively represent observation and canonical skinning weights of bone \\( k \\), and \\( R_{k} \\) and \\( \\mathbf{t}_{k} \\) are the rotation and translation of bone \\( k \\), respectively. Given Eq. 1, following [73], we optimize an explicit volume decoder to optimize \\( \\{w_{k}^{c}\\}_{k = 1}^{K} \\) where \\( K \\) is the total number of joints." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.376, + 0.724, + 0.391 + ], + "angle": 0, + "content": "3.2 Motion-Oriented Compositional Neural Radiance Fields" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.4, + 0.789, + 0.538 + ], + "angle": 0, + "content": "Rigid neural branch for learning canonical radiance field. The rigid branch of MoCo-NeRF is specifically designed to learn the canonical T-pose representation of the target subject using only rigid body movements. Our approach is different from but simpler than the prevalent spatial offset learning approach, which incorporates complex modeling of non-rigid motions as \\(\\Delta \\mathbf{x}_c\\). Our rigid branch instead directly predicts the color and density of the canonical point \\(\\mathbf{x}_c\\) transformed by inverse linear blend skinning (Sec. 3.1), as depicted in Fig. 2. First, we query multiresolution volumetric features \\(f_{\\mathbf{x}_c}\\) of the canonical point \\(\\mathbf{x}_c\\) from the MHE \\(\\psi_{hash}\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.443, + 0.549, + 0.785, + 0.565 + ], + "angle": 0, + "content": "\\[\nf _ {\\mathbf {x} _ {c}} = \\psi_ {\\text {h a s h}} (\\mathbf {x} _ {c}). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.575, + 0.788, + 0.622 + ], + "angle": 0, + "content": "Then, we slice the first half (\\(H\\) dimensions) of \\(f_{\\mathbf{x}_c}\\) at each level of resolution to compute \\(f_{\\mathbf{x}_c}^{:H}\\) and predicts the radiance \\(c\\) and density \\(\\sigma\\) of the rigid canonical point \\(\\mathbf{x}_c\\) with an MLP decoder \\(\\mathbf{M}_{rigid}\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.431, + 0.632, + 0.785, + 0.65 + ], + "angle": 0, + "content": "\\[\nc, \\sigma = \\mathbf {M} _ {\\text {r i g i d}} \\left(f _ {\\mathbf {x} _ {c}} ^ {H}\\right). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Non-rigid neural branch for learning radiance residual field. While the canonical representation does not explicitly model pose-dependent non-rigid motions, it still captures an average of such motions (mean motions) after training. However, the mean motions cannot effectively reflect the complex and realistic non-rigid motions under different body poses. It is only capable of rendering identical deformations across different poses, as can be observed in Fig. 1. To address this effectively, the non-rigid branch of MoCo-NeRF is designed to concurrently optimize for a radiance residual field. This models the appearance discrepancies between the renderings of the rigid branch and the ground-truth colors, typically caused by the complex non-rigid motions or deformations. We additionally condition the radiance residual predictions with the canonical radiance representation of the rigid branch as prior information. This enables more" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.644, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "MoCo-NeRF" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.126 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.145, + 0.462, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.219, + 0.303, + 0.465, + 0.37 + ], + "angle": 0, + "content": "Fig.3: Illustration of the proposed pose-embedded implicit feature generation. We employ cross-attention to modulate the single learnable base code to pose-adaptive features." + }, + { + "type": "image", + "bbox": [ + 0.479, + 0.145, + 0.782, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.478, + 0.303, + 0.784, + 0.369 + ], + "angle": 0, + "content": "Fig. 4: Illustration of extended architecture of MoCo-NeRF for the multi-subject unified training (Sec. 3.3). Major components consist of the global MHE, set of local MHEs, and the dictionary of learnable base codes as ID codes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.399, + 0.784, + 0.429 + ], + "angle": 0, + "content": "coherent modeling of non-rigid motions, as the non-rigid and rigid radiance fields are highly correlated." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.43, + 0.788, + 0.611 + ], + "angle": 0, + "content": "Specifically, unlike the rigid branch, the queried feature \\( f_{\\mathbf{x}_c} \\) is not sliced, enabling the non-rigid branch to learn the radiance residuals with reference to the canonical representation. Instead, we freeze the first half of \\( f_{\\mathbf{x}_c} \\) that represents the canonical radiance field, and only optimize the other half to learn the radiance residuals for non-rigid motion modeling. We denote this partially frozen feature as \\( f_{\\mathbf{x}_c}^* \\). We observe that providing frozen \\( f_{\\mathbf{x}_c}^{H} \\) to the non-rigid radiance branch enables more effective and coherent residual learning. In addition, we also condition the radiance residual field prediction with our pose-embedded implicit feature \\( \\mathbf{P}_{nr} \\), which is achieved by concatenating \\( \\mathbf{P}_{nr} \\) to the first intermediate output of the MLP decoder \\( \\mathbf{M}_{nr} \\), as illustrated in Fig. 2. The generation of \\( \\mathbf{P}_{nr} \\) is explained in the last subsection of Sec. 3.2. Eventually, the color and density residuals are predicted by \\( \\mathbf{M}_{nr} \\) for the final compositional radiance computation:" + }, + { + "type": "equation", + "bbox": [ + 0.405, + 0.626, + 0.785, + 0.643 + ], + "angle": 0, + "content": "\\[\n\\Delta c, \\Delta \\sigma = \\mathbf {M} _ {n r} \\left(f _ {\\mathbf {x} _ {c}} ^ {*}; \\mathbf {P} _ {n r}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.647, + 0.784, + 0.709 + ], + "angle": 0, + "content": "Compositional volume rendering. Following [42], we adopt stratified sampling and volume rendering to predict the RGB color of each ray. For final rendering, we query \\(M\\) samples per ray \\(\\mathbf{r}\\) and perform approximated integration to obtain final RGB \\(C_{final}\\) from \\((c + \\Delta c, \\sigma + \\Delta \\sigma)\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.274, + 0.72, + 0.785, + 0.739 + ], + "angle": 0, + "content": "\\[\nC _ {f i n a l} (\\mathbf {r}) = \\sum_ {m = 1} ^ {M} T _ {m} \\left(1 - \\exp \\left(- \\left(\\sigma_ {m} + \\Delta \\sigma_ {m}\\right) \\delta_ {m}\\right)\\right) \\left(c _ {m} + \\Delta c _ {m}\\right). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.81 + ], + "angle": 0, + "content": "where \\(\\delta_{m}\\) is the adjacent distance from the \\(m\\)-th to the \\(m+1\\)-th sample and \\(T_{m} = \\exp(-\\sum_{n=1}^{m-1} \\sigma_{n} \\delta_{n})\\) as the transmittance. In addition, we compute \\(C_{rigid}\\) in the same way from \\((c, \\sigma)\\) for an additional loss term that encourages a decomposed learning of the canonical radiance field and the radiance residuals (Sec. 3.4)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Learnable Pose-Embedded Implicit Feature. We introduce high-frequency modulated implicit features computed from a learnable base code and a 3D body" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.127 + ], + "angle": 0, + "content": "J. Kim et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.346 + ], + "angle": 0, + "content": "pose. A learnable base code \\(\\mathbf{S} \\in \\mathbb{R}^{64}\\) is randomly initialized and jointly optimized for a single subject with our motion-oriented compositional radiance fields. To make \\(\\mathbf{S}\\) distinguishable for each pose, we propose to employ the 3D body pose as a guidance for computing implicit features that encourage high-fidelity and pose-dependent radiance residual learning. A cross-attention mechanism is designed to perform interactions between the learnable base code and 23 joint positions (i.e. \\(J = \\{j_{i}\\}_{i=1}^{23}\\)). A detailed overview of the mechanism is illustrated in Fig. 3. To allow high-frequency implicit feature learning with 3D body joints, we conduct positional encoding [42] for the input joint coordinates. We first project each joint position into a higher dimension with a sinusoidal positional encoding, \\(\\gamma(j_{i}) = (j_{i}, \\sin(2^{0} \\pi j_{i}), \\cos(2^{0} \\pi j_{i}), \\dots, \\sin(2^{L-1} \\pi j_{i}), \\cos(2^{L-1} \\pi j_{i}))\\), where \\(L = 10\\) is the number of frequency bands. We then project the encoded points (i.e., \\(\\gamma(J) \\in \\mathbb{R}^{23 \\times 63}\\)) to a raw pose-embedded code \\(\\mathbf{P}_{raw}\\) by a projection matrix \\(\\mathbf{W}_{p}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.434, + 0.359, + 0.786, + 0.375 + ], + "angle": 0, + "content": "\\[\n\\mathbf {P} _ {\\text {r a w}} = \\mathbf {W} _ {p} \\cdot \\gamma (J). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.381, + 0.786, + 0.412 + ], + "angle": 0, + "content": "Then, we generate a query signal \\(\\mathbf{P}_q\\) from \\(\\mathbf{S}\\), and key and value signals from \\(\\mathbf{P}_{raw}\\) via linear projections as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.312, + 0.424, + 0.785, + 0.44 + ], + "angle": 0, + "content": "\\[\n\\mathbf {P} _ {q} = \\mathbf {W} _ {Q} \\cdot \\mathbf {S}, \\quad \\mathbf {P} _ {k} = \\mathbf {W} _ {K} \\cdot \\mathbf {P} _ {\\text {r a w}}, \\quad \\mathbf {P} _ {v} = \\mathbf {W} _ {V} \\cdot \\mathbf {P} _ {\\text {r a w}}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.451, + 0.786, + 0.48 + ], + "angle": 0, + "content": "Finally, we generate the learnable pose-embedded implicit feature \\(\\mathbf{P}_{nr}\\) for the radiance residual estimation as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.386, + 0.493, + 0.785, + 0.51 + ], + "angle": 0, + "content": "\\[\n\\mathbf {P} _ {n r} = \\operatorname {s o f t m a x} \\left(\\mathbf {P} _ {q} \\cdot \\left(\\mathbf {P} _ {k}\\right) ^ {\\top}\\right) \\cdot \\mathbf {P} _ {v}. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.521, + 0.787, + 0.582 + ], + "angle": 0, + "content": "As explained earlier, we employ \\(\\mathbf{P}_{nr}\\) to condition \\(\\mathbf{M}_{nr}\\), the MLP decoder of the non-rigid radiance branch. \\(\\mathbf{M}_{nr}\\) optimizes the implicit feature \\(\\mathbf{P}_{nr}\\) and its generation module for estimating pose-coherent radiance residuals instead of solely relying on explicit volumetric features." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.604, + 0.517, + 0.62 + ], + "angle": 0, + "content": "3.3 Multi-subject Unified Training" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.788, + 0.75 + ], + "angle": 0, + "content": "Our effective radiance compositional approach for non-rigid motion modeling further enables a flexible extension of our single-subject MoCo-NeRF for unified training with multiple subjects, which is very beneficial for practical applications. Each monocular video exhibits distinct non-rigid motions, thereby requiring individualized modeling. However, it is impossible for the traditional spatial offset learning approach [73] that typically requires a heavy module for photo-realistic quality. In contrast, our radiance compositional approach can achieve it by employing a separate, lightweight MHE for each subject, referred to as local MHE." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Building on this foundation, we propose three architectural modifications as illustrated in Fig. 4. To benefit from unified training of multiple subjects, we introduce a coarse global MHE \\(\\psi_{hash}^{G}\\) shared across all subjects to learn subject-generic representations (e.g., body shapes). We also ensure learning of subject-specific details with \\(N_{s}\\) local MHEs \\(\\{\\psi_{hash}^{i}\\}_{i = 1}^{N_{s}}\\), where \\(N_{s}\\) is the number of subjects. For each canonical sample \\(\\mathbf{x}_c\\), we query global features from the" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.644, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "MoCo-NeRF" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.224 + ], + "angle": 0, + "content": "global MHE and concatenate it to the sliced feature from the corresponding local MHE for radiance predictions. We additionally propose to leverage the learnable base codes of all subjects \\(\\{\\mathbf{S}_i\\}_{i=1}^{N_s}\\) as ID codes, for sharing MLP decoders while ensuring subject-distinctive canonical representations. The modified radiance estimations of the rigid branch for the multi-subject training are as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.335, + 0.231, + 0.786, + 0.25 + ], + "angle": 0, + "content": "\\[\nc, \\sigma = M _ {r i g i d} \\left(\\psi_ {h a s h} ^ {G} \\left(\\mathbf {x} _ {c}\\right) \\oplus \\left(\\psi_ {h a s h} ^ {i} \\left(\\mathbf {x} _ {c}\\right)\\right) ^ {: H} \\oplus \\mathbf {S} _ {i}\\right) \\tag {9}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.268, + 0.43, + 0.284 + ], + "angle": 0, + "content": "3.4 Model Optimization" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.289, + 0.789, + 0.565 + ], + "angle": 0, + "content": "We optimize MoCo-NeRF with all \\(N_{i}\\) image frames \\(\\{I_j\\}_{j=1}^{N_i}\\) from the monocular video via patch-based ray sampling. Given an image frame \\(I_j\\) at each iteration, we randomly sample \\(N_p\\) patches denoted as \\(\\{P_k\\}_{k=1}^{N_p}\\) to construct the LPIPS [84] loss. The LPIPS loss term can measure the perceptual distance between two image patches, and thus we also adopt it for more perceptually coherent renderings. Our framework first renders \\(\\{\\hat{P}_k^r, \\hat{P}_k^f\\}_{k=1}^{N_p}\\) respectively from \\(C_{rigid}\\) and \\(C_{final}\\), and compute two LPIPS loss terms \\(\\mathcal{L}_{\\mathrm{LPIPS}}^r\\) and \\(\\mathcal{L}_{\\mathrm{LPIPS}}^f\\) against the ground truth patch \\(P_k^{GT}\\) by using the pre-trained VGGNet [63] to as \\(\\mathcal{L}_{\\mathrm{LPIPS}} = \\mathrm{LPIPS}(VGG(\\hat{P}_k), VGG(P_k^{GT}))\\). We respectively compute the MSE loss terms \\(\\mathcal{L}_{\\mathrm{MSE}}^r, \\mathcal{L}_{\\mathrm{MSE}}^f\\) from each of the estimated patches \\(\\{\\hat{P}_k^r, \\hat{P}_k^f\\}_{k=1}^{N_p}\\). Moreover, we additionally compute an appearance matching loss with SSIM \\(\\mathcal{L}_{\\mathrm{SSIM}}^f\\) [15] only for the final RGB \\(\\{\\hat{P}_k^f\\}_{k=1}^{N_p}\\), to improve the rendering quality. In summary, there are 5 loss terms in total: \\(\\mathcal{L}_{\\mathrm{LPIPS}}^r, \\mathcal{L}_{\\mathrm{MSE}}^r\\) for the canonical radiance field of the rigid branch, and \\(\\mathcal{L}_{\\mathrm{LPIPS}}^f, \\mathcal{L}_{\\mathrm{MSE}}^f, \\mathcal{L}_{\\mathrm{SSIM}}^f\\) for the final motion-oriented compositional renderings. To enforce the rigid branch to learn the main canonical radiance fields while the non-rigid branch to learn radiance residuals, we compute the final overall loss function \\(\\mathcal{L}^o\\) with a hyperparameter \\(\\lambda = 0.2\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.424, + 0.574, + 0.785, + 0.589 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} ^ {o} = \\lambda \\mathcal {L} ^ {r} + (1 - \\lambda) \\mathcal {L} ^ {f} \\tag {10}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.314, + 0.594, + 0.785, + 0.612 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} ^ {r} = \\mathcal {L} _ {\\text {L P I P S}} ^ {r} + \\mathcal {L} _ {\\text {M S E}} ^ {r}, \\quad \\mathcal {L} ^ {f} = \\mathcal {L} _ {\\text {L P I P S}} ^ {f} + \\mathcal {L} _ {\\text {M S E}} ^ {f} + \\mathcal {L} _ {\\text {S S I M}} ^ {f} \\tag {11}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.632, + 0.376, + 0.649 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.661, + 0.788, + 0.768 + ], + "angle": 0, + "content": "We conduct extensive experiments on ZJU-MoCap [52] and MonoCap [17, 18, 50] to verify the effectiveness of our proposed approach on both single- and multi-subject settings. The multi-subject setting indicates multiple subjects are jointly optimized. Our findings demonstrate that MoCo-NeRF can learn photo-realistic dynamic human representations by effectively handling non-rigid motions. Moreover, it learns discriminative representations of each performer in the multi-subject setting, while prior works struggle." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.788, + 0.334, + 0.801 + ], + "angle": 0, + "content": "4.1 Datasets" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "We leverage ZJU-MoCap and MonoCap datasets for both training and evaluation, following [14,21], and provide the details in this section. For ZJU-MoCap," + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.127 + ], + "angle": 0, + "content": "J. Kim et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.158, + 0.788, + 0.24 + ], + "angle": 0, + "content": "Table 1: Quantitative results of single-subject methods on ZJU-MoCap and MonoCap datasets where LPIPS* = LPIPS × 10³ with the best and second-best performance notations. MoCo-NeRF outperforms the baselines on most of the evaluation metrics in both settings. Despite pixel-wise metrics (PSNR, SSIM) tend to favor smooth renderings generated from [14,21], we achieve state-of-the-art performance while effectively modeling the photo-realistic non-rigid motions as illustrated in Fig. 5." + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.24, + 0.785, + 0.34 + ], + "angle": 0, + "content": "
Single-subject methodsZJU-MoCap [52]MonoCap [50]
PSNR↑SSIM↑LPIPS*↓PSNR↑SSIM↑LPIPS*↓
HumanNeRF [73]30.180.970930.1332.920.986412.93
Instant-NVR [14]30.730.969738.8731.890.987016.07
GauHuman [21]30.980.962030.1632.870.984913.54
MoCo-NeRF (Ours)31.060.973428.8333.010.987413.02
" + }, + { + "type": "image", + "bbox": [ + 0.233, + 0.352, + 0.774, + 0.574 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.233, + 0.574, + 0.773, + 0.582 + ], + "angle": 0, + "content": "HumanNeRF Instant-NVR GauHuman Ours Ground Truth HumanNeRF Instant-NVR GauHuman Ours Ground Truth" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.593, + 0.788, + 0.649 + ], + "angle": 0, + "content": "Fig.5: Qualitative comparison of novel view synthesis from single-subject training. MoCo-NeRF achieves photo-realistic rendering qualities with high-fidelity non-rigid motion modeling. Although HumanNeRF [73] presents comparable quality, however, MoCo-NeRF achieves much less training time as shown in Tab. 3." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.677, + 0.788, + 0.768 + ], + "angle": 0, + "content": "we evaluate our MoCo-NeRF and other baselines on all 6 subjects (i.e., 377, 386, 387, 392, 393, 394). Out of 23 camera views, we use one of them for training and the other 22 views for evaluation. For MonoCap that comprises DynaCap [17] and DeepCap [18], we utilize all 4 subjects (i.e., lan, marc, olek, vlad) and use the same set of cameras as [14]. For multi-subject evaluations presented in Tab. 2, we evaluate on subjects 377, 386, and 392 of ZJU-MoCap." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.787, + 0.458, + 0.803 + ], + "angle": 0, + "content": "4.2 Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "We optimize our model with Adam optimizer [30] with betas (0.9, 0.99), and learning rates of \\(5e^{-5}\\) for the skeletal weights volume decoder and \\(5e^{-3}\\) for the" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.644, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "MoCo-NeRF" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.158, + 0.784, + 0.199 + ], + "angle": 0, + "content": "Table 2: Quantitative results of multi-subject methods on subjects #377, #386, #392 of ZJU-MoCap [52]. Despite the same evaluation settings, MoCo-NeRF significantly outperforms baseline methods under monocular settings with a shorter training time." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.201, + 0.784, + 0.271 + ], + "angle": 0, + "content": "
Multi-subject methodsPSNR↑SSIM↑LPIPS*↓Training videosTrain time
NHP [31]26.640.918540.47multi-view~86 hours
GHuNeRF [33]27.280.929140.82monocular~123 hours
MoCo-NeRF (Ours)31.970.978124.54monocular~2 hours
" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.283, + 0.785, + 0.397 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.409, + 0.784, + 0.451 + ], + "angle": 0, + "content": "Fig.6: Qualitative comparison of novel view synthesis from multi-subject training. Although all subjects are seen during training, only our multi-subject MoCo-NeRF achieves photo-realistic quality comparable to the ones of single-subject models." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.479, + 0.785, + 0.57 + ], + "angle": 0, + "content": "others. Each MHE has 16 levels with feature dimensions \\( H = 2 \\), and 4 levels with \\( H = 2 \\) for the global MHE. For each iteration, six \\( 32 \\times 32 \\) patches are used with 256 samples per ray. We train all models including the baselines on 1 Nvidia RTX 3090, where both baselines are trained for their suggested duration. Our model is trained for \\( 1 \\sim 2 \\) hours depending on each subject's complexity for the single-subject training, and 2 hours for the multi-subject training." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.591, + 0.494, + 0.605 + ], + "angle": 0, + "content": "4.3 State-of-the-art Comparison" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.613, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Single-subject methods. Tab. 1 presents PSNR, SSIM [72], and LPIPS* of MoCo-NeRF and our baselines [14, 21, 73], for novel-view synthesis under the single-subject setting. MoCo-NeRF achieves state-of-the-art performance on all metrics except one comparable LPIPS* for the MonoCap dataset. As shown in Fig. 5, Instant-NVR [14] shows moderate performance, however, it often generates smoothed-out results without proper non-rigid motions due to high dependency on accurate SMPL parameters. GauHuman [21] achieves better qualitative performance than [14] but still struggles to model sophisticated non-rigid motions. Fig. 1 presents longer optimizations are surprisingly ineffective and harmful respectively for Instant-NVR [14] and GauHuman [21] in terms of non-rigid motion modeling. This implies that their approaches cannot achieve optimal non-rigid motion modeling. Moreover, we observe that PSNR and SSIM often favor those results without fine-grained non-rigid motion modeling since predicting pose-dependent deformations from novel viewpoints may not be pixel-wise accurate. Instead, the perception metric (LPIPS*) evaluates better in this case, as" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.127 + ], + "angle": 0, + "content": "J. Kim et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.158, + 0.788, + 0.24 + ], + "angle": 0, + "content": "Table 3: Training efficiency comparison for methods with photo-realistic non-rigid motion modeling on a single RTX3090. In the table, single-sub. and multi-sub. respectively represent training for 1 and 6 subjects. MoCo-NeRF significantly outperforms HumanNeRF in terms of both GPU memory usage and training time. HumanNeRF requires sequential training to achieve comparable non-rigid motion modeling for all subjects, thus resulting in linearly-increased training time." + }, + { + "type": "table", + "bbox": [ + 0.222, + 0.24, + 0.785, + 0.313 + ], + "angle": 0, + "content": "
GPU Mem. (GB)Train time (hours)
single-sub.multi-sub.Δsingle-sub.multi-sub.Δ
HumanNeRF [73]23.723.7(+0)38.9233.3(+194.5)
MoCo-NeRF (Ours)16.220.3(+4.1)1.92.1(+0.2)
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.34, + 0.785, + 0.431 + ], + "angle": 0, + "content": "HumanNeRF [73] has a better LPIPS* score and also shows qualitatively better results. Thus, we consider GauHuman's comparable PSNR might be favored by a superior representation (3D Gaussians [29]) instead of better non-rigid motion modeling. A more detailed analysis is in supplementary materials. Despite this, we still achieve state-of-the-art performance with significantly improved training efficiency compared to another photo-realistic model [73] (Tab. 3)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.431, + 0.787, + 0.567 + ], + "angle": 0, + "content": "Multi-subject methods. We evaluate our MoCo-NeRF against NHP [31] and GHuNeRF [33], both of which are designed to achieve generalization capabilities. This is necessitated by the absence of prior works that directly target the joint learning task for multiple subjects from monocular videos. Additionally, we observe that the single-subject baseline models fail to learn distinct representations when trained in a unified manner. This observation underscores our decision to compare against generalization models that are specifically tailored to leverage multi-subject data for training. For fairness, we use their official checkpoints and ensure all evaluated subjects were included for training." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.567, + 0.788, + 0.658 + ], + "angle": 0, + "content": "As demonstrated in Tab. 2, MoCo-NeRF remarkably outperforms all the baseline methods on all metrics while also excelling in training efficiency up to 61 times faster. Even though MoCo-NeRF only utilizes a monocular video of each subject, Fig. 6 illustrates our method succeeds in learning both distinct representations and pose-dependent radiance residuals of each subject, enabling photo-realistic free-viewpoint renderings." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.68, + 0.393, + 0.695 + ], + "angle": 0, + "content": "4.4 Model Analysis" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.788, + 0.763 + ], + "angle": 0, + "content": "In this section, we present analyses of our contributions respectively for two major achievements: novel non-rigid motion modeling (NR modeling) and multi-subject unified training (multi-subject modeling). For both analyses, we progressively remove each component from a full model to a vanilla approach." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.842 + ], + "angle": 0, + "content": "NR Modeling: Analysis of radiance compositional approach. Tab. 4 demonstrates our radiance compositional design perceptually improves the rendering results by closing the radiance discrepancies with predicted residuals. Fig. 7 shows that the representation learned solely by the rigid radiance branch (c) suffers from noisy renderings caused by incorrect correspondences due to" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.644, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "MoCo-NeRF" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.158, + 0.499, + 0.213 + ], + "angle": 0, + "content": "Table 4: Ablation study in the single-subject task on ZJU-MoCap for b) pose-embedded implicit features and c) motion-oriented compositional NeRF approach." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.213, + 0.498, + 0.269 + ], + "angle": 0, + "content": "
Single-subjectPSNRSSIMLPIPS*
(a) MoCo-NeRF (Ours)31.060.973428.83
(b) (-) Pose features30.640.970229.65
(c) (-) Residual NeRF30.620.970229.93
" + }, + { + "type": "table_caption", + "bbox": [ + 0.504, + 0.159, + 0.787, + 0.2 + ], + "angle": 0, + "content": "Table 5: Ablation study of multi-subject MoCo-NeRF on ZJU-MoCap for our proposed architectural modifications." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.201, + 0.787, + 0.269 + ], + "angle": 0, + "content": "
Multi-subjectPSNRSSIMLPIPS*
(i)MoCo-NeRF (Ours)30.910.974627.69
(ii)(-) Global MHE30.830.974628.49
(iii)(-)Ns local MHEs30.700.974229.03
(iv)(-) Identity codes28.630.966638.29
" + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.284, + 0.498, + 0.408 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.419, + 0.499, + 0.46 + ], + "angle": 0, + "content": "Fig.7: Qualitative results of the ablation study on single-subject MoCo-NeRF presented in Tab. 4." + }, + { + "type": "image", + "bbox": [ + 0.509, + 0.285, + 0.787, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.504, + 0.419, + 0.787, + 0.461 + ], + "angle": 0, + "content": "Fig. 8: Qualitative results of the ablation study in Tab. 5 where we can observe gradual improvements from (iv) to (i)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.492, + 0.788, + 0.583 + ], + "angle": 0, + "content": "non-rigid motions. With our radiance residual modeling (b), those issues are mitigated by jointly learning per-sample radiance residuals, resulting in more complete rendering and better non-rigid motion modeling as clearly shown in Fig. 7. In addition, Fig. 9 illustrates successful decomposed learning of two types of representations and validates the effectiveness of our radiance compositional approach in non-rigid motion modeling." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.583, + 0.788, + 0.719 + ], + "angle": 0, + "content": "NR Modeling: Analysis of pose-embedded implicit feature. Tab. 4 further illustrates that the final version with the proposed pose-embedded implicit features (Tab. 4 (a)) significantly improves the quantitative performance on all metrics. It enables the non-rigid branch to implicitly utilize body pose information in addition to the explicit features, yielding smoother and more coherent non-rigid motion modeling as shown in Fig. 7. We can observe remarkable quality enhancements that demonstrate the effectiveness of our attention-based code generation module. We also present an ablation study on the choice of cross-attention module for the feature generation in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Multi-subject Modeling: Analysis of using learnable base codes as ID codes. In Fig. 8, the results of MoCo-NeRF without any modifications (iv) exhibit mixed-up appearances on all target subjects, similar to the ones of [73] that are presented in the supplementary material. Utilizing a set of learnable base codes as ID codes (Tab. 5 (iii)) enables a single MLP decoder \\( M_{rigid} \\) to differentiate each subject by optimizing the ID codes as plug-in parameters. Thus, distinct representations of all subjects are successfully learned as illustrated in (iii) of Fig. 8, yielding significant quantitative improvements as well (Tab. 5)." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.127 + ], + "angle": 0, + "content": "J. Kim et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.145, + 0.787, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.356, + 0.788, + 0.427 + ], + "angle": 0, + "content": "Fig. 9: Qualitative illustrations of non-rigid radiance residuals on novel-view synthesis where \\(\\delta \\rightarrow c' = c + \\delta \\Delta c\\), \\(\\sigma' = \\sigma + \\delta \\Delta \\sigma\\). The first row progressively demonstrates the capability of the non-rigid radiance branch to learn the non-rigid motions on the subject's face in addition to the cloth folds. The last two rows clearly illustrate the effects of radiance residuals on cloth folds and hems corresponding to the body pose." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.455, + 0.788, + 0.668 + ], + "angle": 0, + "content": "Multi-subject Modeling: Analysis of \\( N_{s} \\) local MHEs. Fig. 8 shows that the model (iii) lacks fine-grained details due to the sharing of one MHE. By utilizing separate \\( N_{s} \\) local MHEs (Tab. 5 (ii)), the model learns subject-discriminative non-rigid motions as they now optimize separate explicit features for each subject's radiance residuals. Fig. 8 shows clear improvement of the model (ii) in non-rigid motion modeling where its quantitative gains are reflected in Tab. 5. Multi-subject Modeling: Analysis of global MHE. The addition of global MHE (i) further empowers MoCo-NeRF to learn better representations with more coherent non-rigid motions as illustrated in Fig. 8. Although the global MHE is not used for learning non-rigid radiance residuals, it encourages the rigid part of features from local MHE to concentrate on subject-specific representations. As its frozen copy is used for the non-rigid radiance branch, it facilitates better learning of radiance residuals for modeling non-rigid motions. Tab. 5 clearly shows that it can help to learn perceptually better representations." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.689, + 0.36, + 0.705 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In this paper, we presented a novel framework MoCo-NeRF for free-viewpoint dynamic human rendering that is capable of both single- and multi-subject training from monocular videos. By decomposing the neural representation of a dynamic human by motions and utilizing proposed implicit pose codes, MoCo-NeRF effectively learns radiance discrepancies induced by non-rigid motions. The proposed multi-subject MoCo-NeRF efficiently achieves comparable performance compared to the single-subject model. It has fully demonstrated its effectiveness and established new state-of-the-art performance on both problems." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.644, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "MoCo-NeRF" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.239 + ], + "angle": 0, + "content": "Acknowledgments. This research is supported in part by the Hong Kong PhD Fellowship Scheme (HKPFS), Early Career Scheme of the Research Grants Council (RGC) of the Hong Kong SAR under grant No. 26202321, SAIL Research Project, HKUST-Zeekr Collaborative Research Fund, HKUST-WeBank Joint Lab Project, Tencent Rhino-Bird Focused Research Program, and Naver Cloud Corporation." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.26, + 0.323, + 0.276 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.289, + 0.785, + 0.318 + ], + "angle": 0, + "content": "1. Alldieck, T., Zanfir, M., Sminchisescu, C.: Photorealistic monocular 3d reconstruction of humans wearing clothing. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.319, + 0.785, + 0.359 + ], + "angle": 0, + "content": "2. Barron, J.T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., Srinivasan, P.P.: Mip-NeRF: A multiscale representation for anti-aliasing neural radiance fields. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.36, + 0.785, + 0.387 + ], + "angle": 0, + "content": "3. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Mip-NeRF 360: Unbounded anti-aliased neural radiance fields. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.388, + 0.785, + 0.414 + ], + "angle": 0, + "content": "4. Chen, A., Xu, Z., Geiger, A., Yu, J., Su, H.: TensoRF: Tensorial radiance fields. In: ECCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.415, + 0.785, + 0.442 + ], + "angle": 0, + "content": "5. Chen, A., Xu, Z., Zhao, F., Zhang, X., Xiang, F., Yu, J., Su, H.: MVSNeRF: Fast generalizable radiance field reconstruction from multi-view stereo. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.443, + 0.785, + 0.469 + ], + "angle": 0, + "content": "6. Chen, X., Jiang, T., Song, J., Yang, J., Black, M.J., Geiger, A., Hilliges, O.: gDNA: Towards generative detailed neural avatars. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.47, + 0.785, + 0.51 + ], + "angle": 0, + "content": "7. Chen, Y., Wang, X., Chen, X., Zhang, Q., Li, X., Guo, Y., Wang, J., Wang, F.: Uv volumes for real-time rendering of editable free-view human performance. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.511, + 0.785, + 0.551 + ], + "angle": 0, + "content": "8. Collet, A., Chuang, M., Sweeney, P., Gillett, D., Evseev, D., Calabrese, D., Hoppe, H., Kirk, A., Sullivan, S.: High-quality streamable free-viewpoint video. In: SIGGRAPH (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.552, + 0.785, + 0.579 + ], + "angle": 0, + "content": "9. Dong, Z., Chen, X., Yang, J., J.Black, M., Hilliges, O., Geiger, A.: AG3D: Learning to generate 3D avatars from 2D image collections. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.58, + 0.785, + 0.634 + ], + "angle": 0, + "content": "0. Dou, M., Khamis, S., Degtyarev, Y., Davidson, P., Fanello, S.R., Kowdle, A., Escolano, S.O., Rhemann, C., Kim, D., Taylor, J., Kohli, P., Tankovich, V., Izadi, S.: Fusion4D: Real-time performance capture of challenging scenes. In: SIGGRAPH (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.635, + 0.785, + 0.662 + ], + "angle": 0, + "content": "1. Du, Y., Zhang, Y., Yu, H.X., Tenenbaum, J.B., Wu, J.: Neural radiance flow for 4d view synthesis and video processing. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.663, + 0.785, + 0.689 + ], + "angle": 0, + "content": "2. Gao, C., Saraf, A., Kopf, J., Huang, J.B.: Dynamic view synthesis from dynamic monocular video. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.69, + 0.785, + 0.717 + ], + "angle": 0, + "content": "3. Garbin, S.J., Kowalski, M., Johnson, M., Shotton, J., Valentin, J.: FastNeRF: High-fidelity neural rendering at 200fps. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.718, + 0.785, + 0.744 + ], + "angle": 0, + "content": "4. Geng, C., Peng, S., Xu, Z., Bao, H., Zhou, X.: Learning neural volumetric representations of dynamic humans in minutes. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.745, + 0.785, + 0.772 + ], + "angle": 0, + "content": "5. Godard, C., Mac Aodha, O., Brostow, G.J.: Unsupervised monocular depth estimation with left-right consistency. In: CVPR (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.773, + 0.785, + 0.84 + ], + "angle": 0, + "content": "6. Guo, K., Lincoln, P., Davidson, P., Busch, J., Yu, X., Whalen, M., Harvey, G., Orts-Escalano, S., Pandey, R., Dourgarian, J., Tang, D., Tkach, A., Kowdle, A., Cooper, E., Dou, M., Fanello, S., Fyffe, G., Rhemann, C., Taylor, J., Debevec, P., Izadi, S.: The relightables: Volumetric performance capture of humans with realistic relighting. In: SIGGRAPH Asia (2019)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.289, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.127 + ], + "angle": 0, + "content": "J. Kim et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "17. Habermann, M., Liu, L., Xu, W., Zollhoefer, M., Pons-Moll, G., Theobalt, C.: Real-time deep dynamic characters. In: SIGGRAPH (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.177, + 0.785, + 0.204 + ], + "angle": 0, + "content": "18. Habermann, M., Xu, W., Zollhoefer, M., Pons-Moll, G., Theobalt, C.: DeepCap: Monocular human performance capture using weak supervision. In: CVPR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.205, + 0.785, + 0.232 + ], + "angle": 0, + "content": "19. Hedman, P., Srinivasan, P.P., Mildenhall, B., Barron, J.T., Debevec, P.: Baking neural radiance fields for real-time view synthesis. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.233, + 0.785, + 0.261 + ], + "angle": 0, + "content": "20. Hong, F., Chen, Z., Lan, Y., Pan, L., Liu, Z.: EVA3D: Compositional 3d human generation from 2d image collections. In: ICLR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.262, + 0.785, + 0.289 + ], + "angle": 0, + "content": "21. Hu, S., Liu, Z.: GauHuman: Articulated gaussian splatting from monocular human videos. In: CVPR (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.29, + 0.785, + 0.317 + ], + "angle": 0, + "content": "22. Hu, T., Yu, T., Zheng, Z., Zhang, H., Liu, Y., Zwicker, M.: HVTR: Hybrid volumetric-textural rendering for human avatars. In: 3DV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.318, + 0.785, + 0.359 + ], + "angle": 0, + "content": "23. Huang, Y., Yi, H., Liu, W., Wang, H., Wu, B., Wang, W., Lin, B., Zhang, D., Cai, D.: One-shot implicit animatable avatars with model-based priors. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.36, + 0.785, + 0.387 + ], + "angle": 0, + "content": "24. Jain, A., Tancik, M., Abbeel, P.: Putting nerf on a diet: Semantically consistent few-shot view synthesis. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.388, + 0.785, + 0.416 + ], + "angle": 0, + "content": "25. Jiang, B., Hong, Y., Bao, H., Zhang, J.: SelfRecon: Self reconstruction your digital avatar from monocular video. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.417, + 0.785, + 0.444 + ], + "angle": 0, + "content": "26. Jiang, T., Chen, X., Song, J., Hilliges, O.: Instantavatar: Learning avatars from monocular video in 60 seconds. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.445, + 0.785, + 0.472 + ], + "angle": 0, + "content": "27. Jiang, W., Yi, K.M., Samei, G., Tuzel, O., Ranjan, A.: NeuMan: Neural human radiance field from a single video. In: ECCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.473, + 0.785, + 0.514 + ], + "angle": 0, + "content": "28. Jiang, Y., Yao, K., Su, Z., Shen, Z., Luo, H., Xu, L.: Instant-NVR: Instant neural volumetric rendering for human-object interactions from monocular rgbd stream. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.515, + 0.785, + 0.543 + ], + "angle": 0, + "content": "29. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM TOG (July 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.544, + 0.785, + 0.571 + ], + "angle": 0, + "content": "30. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. In: ICLR (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.572, + 0.785, + 0.6 + ], + "angle": 0, + "content": "31. Kwon, Y., Kim, D., Ceylan, D., Fuchs, H.: Neural human performer: Learning generalizable radiance fields for human performance rendering. In: NeurIPS (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.601, + 0.785, + 0.628 + ], + "angle": 0, + "content": "32. Lewis, J.P., Cordner, M., Fong, N.: Pose space deformation: A unified approach to shape interpolation and skeleton-driven deformation. In: SIGGRAPH (2000)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.629, + 0.785, + 0.656 + ], + "angle": 0, + "content": "33. Li, C., Lin, J., Lee, G.H.: Generalizable human nerf from a monocular video. In: 3DV (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.657, + 0.785, + 0.685 + ], + "angle": 0, + "content": "34. Li, R., Tanke, J., Vo, M., Zollhofer, M., Gall, J., Kanazawa, A., Lassner, C.: TAVA: Template-free animatable volumetric actors. In: ECCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.686, + 0.785, + 0.713 + ], + "angle": 0, + "content": "35. Li, Z., Niklaus, S., Snavely, N., Wang, O.: Neural scene flow fields for space-time view synthesis of dynamic scenes. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.714, + 0.785, + 0.741 + ], + "angle": 0, + "content": "36. Liu, L., Gu, J., Lin, K.Z., Chua, T.S., Theobalt, C.: Neural sparse voxel fields. In: NeurIPS (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.742, + 0.785, + 0.782 + ], + "angle": 0, + "content": "37. Liu, L., Habermann, M., Rudnev, V., Sarkar, K., Gu, J., Theobalt, C.: Neural actor: Neural free-view synthesis of human actors with pose control. In: SIGGRAPH Asia (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.784, + 0.785, + 0.812 + ], + "angle": 0, + "content": "38. Liu, Y., Peng, S., Liu, L., Wang, Q., Wang, P., Christian, T., Zhou, X., Wang, W.: Neural rays for occlusion-aware image-based rendering. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "39. Loper, M., Mahmood, N., Romero, J., Pons-Moll, G., Black, M.J.: SMPL: A skinned multi-person linear model. In: SIGGRAPH Asia (2015)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.645, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "MoCo-NeRF" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.204 + ], + "angle": 0, + "content": "40. Martin-Brualla, R., Pandey, R., Yang, S., Pidlypenskyi, P., Taylor, J., Valentin, J., Khamis, S., Davidson, P., Tkach, A., Lincoln, P., Kowdle, A., Rhemann, C., Goldman, D.B., Keskin, C., Seitz, S., Izadi, S., Fanello, S.: LookinGood: Enhancing performance capture with real-time neural re-rendering. ACM TOG (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.204, + 0.787, + 0.245 + ], + "angle": 0, + "content": "41. Mihajlovic, M., Bansal, A., Zollhoefer, M., Tang, S., Saito, S.: KeypointNeRF: Generalizing image-based volumetric avatars using relative spatial encoding of keypoints. In: ECCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.245, + 0.787, + 0.284 + ], + "angle": 0, + "content": "42. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: NeRF: Representing scenes as neural radiance fields for view synthesis. In: ECCV (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.285, + 0.787, + 0.312 + ], + "angle": 0, + "content": "43. Mu, J., Sang, S., Vasconcelos, N., Wang, X.: ActorsNeRF: Animatable few-shot human rendering with generalizable nerfs. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.313, + 0.787, + 0.339 + ], + "angle": 0, + "content": "44. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. In: SIGGRAPH (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.34, + 0.787, + 0.366 + ], + "angle": 0, + "content": "45. Noguchi, A., Sun, X., Lin, S., Harada, T.: Unsupervised learning of efficient geometry-aware neural articulated representations. In: ECCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.366, + 0.787, + 0.393 + ], + "angle": 0, + "content": "46. Oechsle, M., Peng, S., Geiger, A.: UNISURF: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.394, + 0.787, + 0.42 + ], + "angle": 0, + "content": "47. Park, K., Sinha, U., Barron, J.T., Bouaziz, S., Goldman, D.B., Seitz, S.M., Martin-Brualla, R.: Nerfies: Deformable neural radiance fields. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.421, + 0.787, + 0.461 + ], + "angle": 0, + "content": "48. Park, K., Sinha, U., Hedman, P., Barron, J.T., Bouaziz, S., Goldman, D.B., Martin-Brualla, R., Seitz, S.M.: HyperNeRF: A higher-dimensional representation for topologically varying neural radiance fields. In: SIGGRAPH Asia (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.461, + 0.787, + 0.488 + ], + "angle": 0, + "content": "49. Peng, B., Hu, J., Zhou, J., Zhang, J.: SelfNeRF: Fast training nerf for human from monocular self-rotating video. arXiv preprint arXiv:2210.01651 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.488, + 0.787, + 0.515 + ], + "angle": 0, + "content": "50. Peng, S., Dong, J., Wang, Q., Zhang, S., Shuai, Q., Zhou, X., Bao, H.: Animatable neural radiance fields for modeling dynamic human bodies. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.516, + 0.787, + 0.555 + ], + "angle": 0, + "content": "51. Peng, S., Xu, Z., Dong, J., Wang, Q., Zhang, S., Shuai, Q., Bao, H., Zhou, X.: Animatable implicit neural representations for creating realistic avatars from videos. TPAMI (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.556, + 0.787, + 0.597 + ], + "angle": 0, + "content": "52. Peng, S., Zhang, Y., Xu, Y., Wang, Q., Shuai, Q., Bao, H., Zhou, X.: Neural Body: Implicit neural representations with structured latent codes for novel view synthesis of dynamic humans. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.597, + 0.787, + 0.623 + ], + "angle": 0, + "content": "53. Pumarola, A., Corona, E., Pons-Moll, G., Moreno-Noguer, F.: D-NeRF: Neural Radiance Fields for Dynamic Scenes. In: CVPR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.624, + 0.787, + 0.651 + ], + "angle": 0, + "content": "54. Reiser, C., Peng, S., Liao, Y., Geiger, A.: KiloNeRF: Speeding up neural radiance fields with thousands of tiny mlps. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.651, + 0.787, + 0.677 + ], + "angle": 0, + "content": "55. Rematas, K., Liu, A., Srinivasan, P.P., Barron, J.T., Tagliasacchi, A., Funkhouser, T., Ferrari, V.: Urban radiance fields. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.678, + 0.787, + 0.718 + ], + "angle": 0, + "content": "56. Remelli, E., Bagautdinov, T., Saito, S., Wu, C., Simon, T., Wei, S.E., Guo, K., Cao, Z., Prada, F., Saragih, J., et al.: Drivable volumetric avatars using texel-aligned features. In: SIGGRAPH (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.719, + 0.787, + 0.759 + ], + "angle": 0, + "content": "57. Saito, S., Huang, Z., Natsume, R., Morishima, S., Kanazawa, A., Li, H.: PIFu: Pixel-aligned implicit function for high-resolution clothed human digitization. In: ICCV (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.759, + 0.787, + 0.786 + ], + "angle": 0, + "content": "58. Saito, S., Simon, T., Saragih, J., Joo, H.: PIFuHD: Multi-level pixel-aligned implicit function for high-resolution 3d human digitization. In: CVPR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.786, + 0.787, + 0.813 + ], + "angle": 0, + "content": "59. Sara Fridovich-Keil and Alex Yu, Tancik, M., Chen, Q., Recht, B., Kanazawa, A.: Plenoxels: Radiance fields without neural networks. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.814, + 0.787, + 0.84 + ], + "angle": 0, + "content": "60. Schonberger, J.L., Zheng, E., Frahm, J.M., Pollefeys, M.: Pixelwise view selection for unstructured multi-view stereo. In: ECCV (2016)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.127 + ], + "angle": 0, + "content": "J. Kim et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.161 + ], + "angle": 0, + "content": "61. Schonberger, J.L., Frahm, J.M.: Structure-from-motion revisited. In: CVPR (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.162, + 0.785, + 0.203 + ], + "angle": 0, + "content": "62. Shao, R., Zhang, H., Zhang, H., Chen, M., Cao, Y., Yu, T., Liu, Y.: DoubleField: Bridging the neural surface and radiance fields for high-fidelity human reconstruction and rendering. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.204, + 0.785, + 0.231 + ], + "angle": 0, + "content": "63. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition. In: ICLR (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.785, + 0.259 + ], + "angle": 0, + "content": "64. Su, S.Y., Yu, F., Zollhöfer, M., Rhodin, H.: A-NeRF: Articulated neural radiance fields for learning human shape, appearance, and pose. In: NeurIPS (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.26, + 0.785, + 0.286 + ], + "angle": 0, + "content": "65. Su, Z., Xu, L., Zheng, Z., Yu, T., Liu, Y., Fang, L.: RobustFusion: Human volumetric capture with data-driven visual cues using a rgbd camera. In: ECCV (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.785, + 0.327 + ], + "angle": 0, + "content": "66. Tancik, M., Casser, V., Yan, X., Pradhan, S., Mildenhall, B., Srinivasan, P., Barron, J.T., Kretzschmar, H.: Block-NeRF: Scalable large scene neural view synthesis. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.785, + 0.369 + ], + "angle": 0, + "content": "67. Tretschk, E., Tewari, A., Golyanik, V., Zollhöfer, M., Lassner, C., Theobalt, C.: Non-rigid neural radiance fields: Reconstruction and novel view synthesis of a dynamic scene from monocular video. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.37, + 0.785, + 0.397 + ], + "angle": 0, + "content": "68. Turki, H., Ramanan, D., Satyanarayanan, M.: Mega-NERF: Scalable construction of large-scale nerfs for virtual fly-throughs. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.785, + 0.438 + ], + "angle": 0, + "content": "69. Wang, P., Liu, L., Liu, Y., Theobalt, C., Komura, T., Wang, W.: NeuS: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. In: NeurIPS (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.785, + 0.481 + ], + "angle": 0, + "content": "70. Wang, Q., Wang, Z., Genova, K., Srinivasan, P., Zhou, H., Barron, J.T., Martin-Brualla, R., Snavely, N., Funkhouser, T.: IBRNet: Learning multi-view image-based rendering. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.481, + 0.785, + 0.508 + ], + "angle": 0, + "content": "71. Wang, S., Schwarz, K., Geiger, A., Tang, S.: ARAH: Animatable volume rendering of articulated human sdfs. In: ECCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.509, + 0.785, + 0.536 + ], + "angle": 0, + "content": "72. Wang, Z., Bovik, A., Sheikh, H., Simoncelli, E.: Image quality assessment: from error visibility to structural similarity. TIP (2004)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.537, + 0.785, + 0.577 + ], + "angle": 0, + "content": "73. Weng, C.Y., Curless, B., Srinivasan, P.P., Barron, J.T., Kemelmacher-Shlizerman, I.: HumanNeRF: Free-viewpoint rendering of moving people from monocular video. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.578, + 0.785, + 0.605 + ], + "angle": 0, + "content": "74. Weng, C.Y., Srinivasan, P.P., Curless, B., Kemelmacher-Shlizerman, I.: Person-NeRF: Personalized reconstruction from photo collections. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.606, + 0.785, + 0.632 + ], + "angle": 0, + "content": "75. Wu, M., Wang, Y., Hu, Q., Yu, J.: Multi-view neural human rendering. In: CVPR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.633, + 0.785, + 0.66 + ], + "angle": 0, + "content": "76. Xian, W., Huang, J.B., Kopf, J., Kim, C.: Space-time neural irradiance fields for free-viewpoint video. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.661, + 0.785, + 0.702 + ], + "angle": 0, + "content": "77. Xiangli, Y., Xu, L., Pan, X., Zhao, N., Rao, A., Theobalt, C., Dai, B., Lin, D.: Bungeenerf: Progressive neural radiance field for extreme multi-scale scene rendering. In: ECCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.702, + 0.785, + 0.73 + ], + "angle": 0, + "content": "78. Xu, H., Alldieck, T., Sminchisescu, C.: H-NeRF: Neural radiance fields for rendering and temporal reconstruction of humans in motion. In: NeurIPS (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.73, + 0.785, + 0.757 + ], + "angle": 0, + "content": "79. Xu, L., Xiangli, Y., Peng, S., Pan, X., Zhao, N., Theobalt, C., Dai, B., Lin, D.: Grid-guided neural radiance fields for large urban scenes. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.785, + 0.785 + ], + "angle": 0, + "content": "80. Xu, Z., Peng, S., Lin, H., He, G., Sun, J., Shen, Y., Bao, H., Zhou, X.: 4K4D: Real-time 4d view synthesis at 4k resolution. In: CVPR (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.786, + 0.785, + 0.812 + ], + "angle": 0, + "content": "81. Yariv, L., Gu, J., Kasten, Y., Lipman, Y.: Volume rendering of neural implicit surfaces. In: NeurIPS (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "82. Yu, A., Li, R., Tancik, M., Li, H., Ng, R., Kanazawa, A.: PlenOctrees for real-time rendering of neural radiance fields. In: ICCV (2021)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.645, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "MoCo-NeRF" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.175 + ], + "angle": 0, + "content": "83. Yu, A., Ye, V., Tancik, M., Kanazawa, A.: pixelNeRF: Neural radiance fields from one or few images. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.175, + 0.788, + 0.203 + ], + "angle": 0, + "content": "84. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: CVPR (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.203, + 0.788, + 0.231 + ], + "angle": 0, + "content": "85. Zhang, R., Chen, J.: NDF: Neural deformable fields for dynamic human modelling. In: ECCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.231, + 0.788, + 0.259 + ], + "angle": 0, + "content": "86. Zhao, F., Yang, W., Zhang, J., Lin, P., Zhang, Y., Yu, J., Xu, L.: HumanNeRF: Efficiently generated human radiance field from sparse inputs. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.259, + 0.788, + 0.286 + ], + "angle": 0, + "content": "87. Zheng, Z., Huang, H., Yu, T., Zhang, H., Guo, Y., Liu, Y.: Structured local radiance fields for human avatar modeling. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.286, + 0.788, + 0.314 + ], + "angle": 0, + "content": "88. Zheng, Z., Zhao, X., Zhang, H., Liu, B., Liu, Y.: AvatarRex: Real-time expressive full-body avatars. In: SIGGRAPH (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.314, + 0.788, + 0.342 + ], + "angle": 0, + "content": "89. Zhi, Y., Qian, S., Yan, X., Gao, S.: Dual-space nerf: Learning animatable avatars and scene lighting in separate spaces. In: 3DV (2022)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.342 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Motion-Oriented Compositional Neural Radiance Fields for Monocular Dynamic Human Modeling/6e1b3c80-b2f1-4690-8b23-807237bf60da_origin.pdf b/2024/Motion-Oriented Compositional Neural Radiance Fields for Monocular Dynamic Human Modeling/6e1b3c80-b2f1-4690-8b23-807237bf60da_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..f89b588db9261816615bd8efd05fc1612d02eaf2 --- /dev/null +++ b/2024/Motion-Oriented Compositional Neural Radiance Fields for Monocular Dynamic Human Modeling/6e1b3c80-b2f1-4690-8b23-807237bf60da_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:067e149d81e690db706289c9839365c7e858ebbe51f8f74efffd3b561e5e6296 +size 11095363 diff --git a/2024/Motion-Oriented Compositional Neural Radiance Fields for Monocular Dynamic Human Modeling/full.md b/2024/Motion-Oriented Compositional Neural Radiance Fields for Monocular Dynamic Human Modeling/full.md new file mode 100644 index 0000000000000000000000000000000000000000..61f4088c15bce2bfd3e975d5eba6a2499b584c51 --- /dev/null +++ b/2024/Motion-Oriented Compositional Neural Radiance Fields for Monocular Dynamic Human Modeling/full.md @@ -0,0 +1,325 @@ +# Motion-Oriented Compositional Neural Radiance Fields for Monocular Dynamic Human Modeling + +Jaehyeok $\mathrm{Kim}^{1}$ , Dongyoon $\mathrm{Wee}^{2}$ , and Dan $\mathrm{Xu}^{1\boxtimes}$ + +$^{1}$ HKUST, Hong Kong $^{2}$ Naver Cloud Corp., South Korea jkimbf@cse.ust.hk, dongyoon.wee@navercorp.com, danxu@cse.ust.hk + +Abstract. This paper introduces Motion-oriented Compositional Neural Radiance Fields (MoCo-NeRF), a framework designed to perform free-viewpoint rendering of monocular human videos via novel non-rigid motion modeling approach. In the context of dynamic clothed humans, complex cloth dynamics generate non-rigid motions that are intrinsically distinct from skeletal articulations and critically important for the rendering quality. The conventional approach models non-rigid motions as spatial (3D) deviations in addition to skeletal transformations. However, it is either time-consuming or challenging to achieve optimal quality due to its high learning complexity without a direct supervision. To target this problem, we propose a novel approach of modeling non-rigid motions as radiance residual fields to benefit from more direct color supervision in the rendering and utilize the rigid radiance fields as a prior to reduce the complexity of the learning process. Our approach utilizes a single multiresolution hash encoding (MHE) to concurrently learn the canonical T-pose representation from rigid skeletal motions and the radiance residual field for non-rigid motions. Additionally, to further improve both training efficiency and usability, we extend MoCo-NeRF to support simultaneous training of multiple subjects within a single framework, thanks to our effective design for modeling non-rigid motions. This scalability is achieved through the integration of a global MHE and learnable identity codes in addition to multiple local MHEs. We present extensive results on ZJU-MoCap and MonoCap, clearly demonstrating state-of-the-art performance in both single- and multi-subject settings. The code and model will be made publicly available at the project page: https://stevejaehyeok.github.io/publications/moco-nerf. + +Keywords: Dynamic human modeling from monocular videos $\cdot$ Neural Radiance Fields $\cdot$ Novel view synthesis of humans + +# 1 Introduction + +Novel view synthesis of monocular human video with photo-realistic quality is a challenging problem, as it involves complex articulated body motions and necessitates sophisticated modeling of non-skeletal or non-rigid motions (e.g., wrinkles + +![](images/98ab8cebc2cfd6b75e6abd8042a52f0e7f1fef5ecedab8c678c2c544464b74fb.jpg) + +![](images/64c9c69f2ff2f04e7bf36bee908d712b9ff294b69e7aa7018752917f01d95370.jpg) +Fig. 1: a) We introduce a motion-oriented compositional NeRF for photo-realistic modeling of dynamic humans from monocular videos. Our approach innovatively employs radiance compositions to capture pose-adaptive non-rigid motions, overcoming the limitations of skeletal transformations that typically yield an average of observed deformations (mean motions). b) The proposed MoCo-NeRF achieves state-of-the-art rendering quality and noteworthy efficiency in novel view synthesis, compared to leading competitors [14,21,73]. The bold denotes the proposed training duration of each comparison model. MoCo-NeRF uniquely captures coherent non-rigid motions, like T-shirt wrinkles relative to body pose, from entirely new viewpoints. Moreover, MoCo-NeRF significantly surpasses the efficiency of another photo-realistic method, HumanNeRF [73]. + +of clothes) from unseen viewpoints. In particular, non-rigid motion modeling is essential for achieving photo-realism and coherence in free-viewpoint renderings. A prevalent solution adopted by recent methods [14,21,73] models the non-rigid motions as pose-dependent 3D spatial offsets of canonical point samples. However, the offset-based strategies lead to either extensive training time or irrelevant modeling of non-rigid motions due to its high learning complexity, incurred by the following reasons. First, a single spatial offset spans an unbounded range and is unique for every 3D point across all input frames and body poses, thereby escalating computational complexity. Second, the learning objective of spatial offsets is not directly aligned with the conventional RGB-based supervision, further complicating the modeling process. + +To target the issues above-discussed, in this paper, we propose Motion-oriented Compositional Neural Radiance Fields (MoCo-NeRF), a new framework that effectively models non-rigid motions at a radiance-level through a unique design of residual radiance field learning for free-viewpoint renderings of monocular human videos. Specifically, our proposed framework decomposes a single + +MHE [44] into two branches: a rigid branch that learns a canonical T-pose radiance field solely via rigid skeletal transformations, and a non-rigid branch that optimizes pose-driven radiance residuals for capturing appearance discrepancies between renderings of the rigid branch and the ground-truth images, which are typically caused by non-rigid motions. The proposed approach circumvents the complexities associated with learning unbounded spatial offsets by reducing the prediction range to the radiance space with the rigid radiance field as a prior. It further establishes a more direct supervision from the ground-truth pixel colors for optimizing the rigid and non-rigid radiance fields by embedding them together in the rendering process, which can further enhance the learning efficiency. Besides, we present a cross-attention module to query a pose conditioned implicit feature from a learnable base code. This process allows our framework to extract more discriminative implicit features tailored to each particular pose, helping our non-rigid branch to learn coherent pose-driven radiance residuals. + +In addition, thanks to our effective design for modeling non-rigid motions, the proposed MoCo-NeRF architecture can be flexibly extended to support unified training across multiple subjects while maintaining photo-realistic quality, without a significant increase in training time. For this architectural extension, we introduce multiple local MHEs and a global MHE, and also extend the aforementioned base code to a set of identity (ID) codes as a learnable codebook. The multiple local MHEs are designed for modeling multiple subjects and the global MHE targets the learning of generic and universal body representations. Furthermore, subject-specific learnable ID codes that are pivotal for generating pose-embedded implicit features enable effective sharing of the MLP decoder in the rigid branch by all subjects, enhancing the model's efficiency and scalability. + +MoCo-NeRF achieves state-of-the-art performance on free-viewpoint rendering of monocular human videos in both single- and multi-subject settings, showing clear improvements compared to the prior works. We present our extensive results on ZJU-MoCap [52] and MonoCap [17, 18, 50] to demonstrate the effectiveness of the proposed MoCo-NeRF in learning photo-realistic representation and modeling complex non-rigid motions from monocular videos. + +In summary, the contribution of our work is threefold: + +- We propose a novel motion-oriented compositional NeRF approach, MoCo-NeRF, which concurrently learns canonical rigid radiance fields and pose-dependent radiance residual fields, effectively handling both rigid and non-rigid motions. We also introduce a pose-embedded implicit feature, computed by cross-attention between a body pose and a designed learnable base code, to implicitly enhance pose-dependent radiance residual learning. +- The proposed MoCo-NeRF is the first work that can be readily extended to enable unified training of multiple subjects without significantly increasing the GPU memory demands. We achieve this through our local MHEs, a global MHE, and the extended identity codebook, which enable subject-discriminative representation learning and effective module sharing. +- The proposed MoCo-NeRF achieves state-of-the-art performance under both single- and multi-subject training settings with a significant efficiency im + +provement compared to the best-performing methods. MoCo-NeRF jointly learns all 6 subjects of ZJU-MoCap on one RTX 3090 GPU within 2 hours. + +# 2 Related Work + +We review closely related work from three perspectives: deformable neural rendering, neural human rendering, and monocular photo-realistic human rendering. Deformable neural rendering. NeRF [42] leverages an MLP to learn a static 3D representation of a scene from a dense set of images from diverse viewpoints. For its application in real-life scenarios, prior works have enhanced NeRF in various directions such as improving efficiencies [4,13,19,36,44,54,59,82] or relieving strong reliance on impractical settings like dense input views [5,24,38,70,83]. In addition, there are more branches of work such as large-scale modeling [55,66, 68,77,79], improving rendering quality [2,3], reconstruction quality [46,69,81] and more. However, they are restricted to static scenes while the majority of the real-world objects are dynamic. + +Recent works including [11, 12, 35, 47, 48, 53, 67, 76] have broadened the modeling capabilities of NeRF to dynamic scenes with small and simple movements/deformations. However, they often fail in handling more dynamic or complex motions like human body articulations. In contrast, our approach enables learning of human-articulated motions from monocular videos, and performing free-viewpoint rendering of a human performer at any time frame of the videos. Neural rendering of humans. Many early-stage prior works for neural human renderings utilize traditional techniques like multi-view stereo [16, 60, 61, 75] or to build sophisticated systems for capturing human motions [8, 10, 40, 65]. However, constructing a dense set of training views and building those capturing systems are expensive, making them unsuitable for real-life situations. [1, 6, 57, 58] address 3D human reconstruction by utilizing ground truth human scans, relieving the aforementioned requirements. [9, 20, 45] further improves it by enabling human reconstruction from 2D image collections. In recent days, there have been many works learning an implicit human representation for novel pose synthesis [22, 23, 25, 27, 34, 50, 56, 64, 71, 87-89] or for a novel-view synthesis [27, 28, 31, 37, 51, 52, 64, 73, 74, 78, 85, 86]. There also have been several generalizable works [41, 43, 62] for a novel human subject with pre-training on multi-view dataset and a few-shot fine-tuning. [7, 80] focus on improving rendering efficiency to achieve real-time free-view rendering of dynamic humans. Furthermore, [14, 26, 28, 49] adopt explicit hash encoding [44] in order to boost the training efficiency for human representation modeling. In contrast to our problem setup, most of those works require multi-view videos or do not model the non-rigid motions. + +Monocular photo-realistic human rendering. HumanNeRF [73] presents a state-of-the-art performance by optimizing motion fields and canonical T-type representation of a human performer. However, its implicit modeling of pose-dependent non-rigid motions requires tens of hours to converge. Instant-NVR [14] achieves efficient training with part-based representation and 2D motion parameterization for learning non-rigid motions as 3D offsets. Note that the + +![](images/5ad053192ebfef455342c6dfacb401ad87379dfee095188170b3a32624e90a18.jpg) +Fig. 2: Overview of the proposed framework MoCo-NeRF for free-view human rendering from a monocular video. Without estimating a geometrical offset of each continuous canonical point for each body pose, our framework is able to handle all deformations via its radiance-compositional approach with a single MHE [44] and achieve state-of-the-art performance. The pose-embedded implicit feature further enhances the learning of non-rigid radiance residuals by enabling pose-distinctive representation learning. + +human parametric model (SMPL) [39], an extra requirement for training, plays a significant role in their framework. In addition, a concurrent work GauHuman [21], which also employs spatial offset approach with Gaussian Splatting (GS) [29], further enhances the efficiency and performance upon [14] by leveraging SMPL for the density control of GS. Nevertheless, [14, 21] struggle to learn complex non-rigid motions, even with an extended training duration as shown in Fig. 1. In this paper, instead of directly learning spatial offsets like these existing works, we propose a novel approach that effectively models non-rigid body motions by learning motion-oriented compositional radiance fields conditioned by pose-embedded implicit features. Our approach clearly outperforms the best-performing competitors HumanNeRF [73], Instant-NVR [14], and GauHuman [21] and can further extend to support joint training of multiple subjects. + +# 3 The Proposed MoCo-NeRF Method + +The proposed MoCo-NeRF aims to effectively learn high-fidelity representations for free-viewpoint rendering of dynamic human subjects in monocular videos. We introduce our novel radiance compositional NeRF approach based on different motion types: rigid and non-rigid motions, as illustrated in Fig. 2. The composition of canonical radiance and the per-sample pose-dependent radiance residuals (Sec. 3.2) facilitates photo-realistic quality of rendering with fine-grained modeling of non-rigid motions, rather than leveraging the conventional spatial 3D offset learning approach. We further improve the quality of non-rigid radiance residuals by introducing a pose-embedded implicit features (Sec. 3.2) that is jointly optimized with the non-rigid radiance branch. In Sec. 3.3, thanks to our effective design of non-rigid motion modeling, it enables MoCo-NeRF to be flexibly extended for unified training of multiple subjects with a high efficiency. + +# 3.1 Preliminary: Skeletal Transformations + +The skeletal motion learns an inverse formulation of the linear blend skinning [32], which transforms canonical points to different body poses by skinning weights. Weng et al. [73] reformulate the linear blend skinning to transform a point $\mathbf{x}$ in the observation space to a canonical point $\mathbf{x}_c$ : + +$$ +\mathbf {x} _ {c} = \sum_ {k = 1} ^ {K} w _ {k} ^ {o} \left(R _ {k} \mathbf {x} + \mathbf {t} _ {k}\right), w _ {k} ^ {o} = \frac {w _ {k} ^ {c} \left(R _ {k} \mathbf {x} + \mathbf {t} _ {k}\right)}{\sum_ {j = 1} ^ {K} w _ {j} ^ {c} \left(R _ {j} \mathbf {x} + \mathbf {t} _ {j}\right)}, \tag {1} +$$ + +where $w_{k}^{o}$ and $w_{k}^{c}$ respectively represent observation and canonical skinning weights of bone $k$ , and $R_{k}$ and $\mathbf{t}_{k}$ are the rotation and translation of bone $k$ , respectively. Given Eq. 1, following [73], we optimize an explicit volume decoder to optimize $\{w_{k}^{c}\}_{k = 1}^{K}$ where $K$ is the total number of joints. + +# 3.2 Motion-Oriented Compositional Neural Radiance Fields + +Rigid neural branch for learning canonical radiance field. The rigid branch of MoCo-NeRF is specifically designed to learn the canonical T-pose representation of the target subject using only rigid body movements. Our approach is different from but simpler than the prevalent spatial offset learning approach, which incorporates complex modeling of non-rigid motions as $\Delta \mathbf{x}_c$ . Our rigid branch instead directly predicts the color and density of the canonical point $\mathbf{x}_c$ transformed by inverse linear blend skinning (Sec. 3.1), as depicted in Fig. 2. First, we query multiresolution volumetric features $f_{\mathbf{x}_c}$ of the canonical point $\mathbf{x}_c$ from the MHE $\psi_{hash}$ as follows: + +$$ +f _ {\mathbf {x} _ {c}} = \psi_ {\text {h a s h}} (\mathbf {x} _ {c}). \tag {2} +$$ + +Then, we slice the first half ( $H$ dimensions) of $f_{\mathbf{x}_c}$ at each level of resolution to compute $f_{\mathbf{x}_c}^{:H}$ and predicts the radiance $c$ and density $\sigma$ of the rigid canonical point $\mathbf{x}_c$ with an MLP decoder $\mathbf{M}_{rigid}$ as follows: + +$$ +c, \sigma = \mathbf {M} _ {\text {r i g i d}} \left(f _ {\mathbf {x} _ {c}} ^ {H}\right). \tag {3} +$$ + +Non-rigid neural branch for learning radiance residual field. While the canonical representation does not explicitly model pose-dependent non-rigid motions, it still captures an average of such motions (mean motions) after training. However, the mean motions cannot effectively reflect the complex and realistic non-rigid motions under different body poses. It is only capable of rendering identical deformations across different poses, as can be observed in Fig. 1. To address this effectively, the non-rigid branch of MoCo-NeRF is designed to concurrently optimize for a radiance residual field. This models the appearance discrepancies between the renderings of the rigid branch and the ground-truth colors, typically caused by the complex non-rigid motions or deformations. We additionally condition the radiance residual predictions with the canonical radiance representation of the rigid branch as prior information. This enables more + +![](images/37ddffe77c036dd982cc06c380f82c3d3434ccd85bbd069e44416d0e9e5cb26c.jpg) +Fig.3: Illustration of the proposed pose-embedded implicit feature generation. We employ cross-attention to modulate the single learnable base code to pose-adaptive features. + +![](images/ddac47f517266c75f171259fdc89ab52a4be2d2a8b2542097939aaddf41fa83a.jpg) +Fig. 4: Illustration of extended architecture of MoCo-NeRF for the multi-subject unified training (Sec. 3.3). Major components consist of the global MHE, set of local MHEs, and the dictionary of learnable base codes as ID codes. + +coherent modeling of non-rigid motions, as the non-rigid and rigid radiance fields are highly correlated. + +Specifically, unlike the rigid branch, the queried feature $f_{\mathbf{x}_c}$ is not sliced, enabling the non-rigid branch to learn the radiance residuals with reference to the canonical representation. Instead, we freeze the first half of $f_{\mathbf{x}_c}$ that represents the canonical radiance field, and only optimize the other half to learn the radiance residuals for non-rigid motion modeling. We denote this partially frozen feature as $f_{\mathbf{x}_c}^*$ . We observe that providing frozen $f_{\mathbf{x}_c}^{H}$ to the non-rigid radiance branch enables more effective and coherent residual learning. In addition, we also condition the radiance residual field prediction with our pose-embedded implicit feature $\mathbf{P}_{nr}$ , which is achieved by concatenating $\mathbf{P}_{nr}$ to the first intermediate output of the MLP decoder $\mathbf{M}_{nr}$ , as illustrated in Fig. 2. The generation of $\mathbf{P}_{nr}$ is explained in the last subsection of Sec. 3.2. Eventually, the color and density residuals are predicted by $\mathbf{M}_{nr}$ for the final compositional radiance computation: + +$$ +\Delta c, \Delta \sigma = \mathbf {M} _ {n r} \left(f _ {\mathbf {x} _ {c}} ^ {*}; \mathbf {P} _ {n r}\right), \tag {4} +$$ + +Compositional volume rendering. Following [42], we adopt stratified sampling and volume rendering to predict the RGB color of each ray. For final rendering, we query $M$ samples per ray $\mathbf{r}$ and perform approximated integration to obtain final RGB $C_{final}$ from $(c + \Delta c, \sigma + \Delta \sigma)$ as follows: + +$$ +C _ {f i n a l} (\mathbf {r}) = \sum_ {m = 1} ^ {M} T _ {m} \left(1 - \exp \left(- \left(\sigma_ {m} + \Delta \sigma_ {m}\right) \delta_ {m}\right)\right) \left(c _ {m} + \Delta c _ {m}\right). \tag {5} +$$ + +where $\delta_{m}$ is the adjacent distance from the $m$ -th to the $m+1$ -th sample and $T_{m} = \exp(-\sum_{n=1}^{m-1} \sigma_{n} \delta_{n})$ as the transmittance. In addition, we compute $C_{rigid}$ in the same way from $(c, \sigma)$ for an additional loss term that encourages a decomposed learning of the canonical radiance field and the radiance residuals (Sec. 3.4). + +Learnable Pose-Embedded Implicit Feature. We introduce high-frequency modulated implicit features computed from a learnable base code and a 3D body + +pose. A learnable base code $\mathbf{S} \in \mathbb{R}^{64}$ is randomly initialized and jointly optimized for a single subject with our motion-oriented compositional radiance fields. To make $\mathbf{S}$ distinguishable for each pose, we propose to employ the 3D body pose as a guidance for computing implicit features that encourage high-fidelity and pose-dependent radiance residual learning. A cross-attention mechanism is designed to perform interactions between the learnable base code and 23 joint positions (i.e. $J = \{j_{i}\}_{i=1}^{23}$ ). A detailed overview of the mechanism is illustrated in Fig. 3. To allow high-frequency implicit feature learning with 3D body joints, we conduct positional encoding [42] for the input joint coordinates. We first project each joint position into a higher dimension with a sinusoidal positional encoding, $\gamma(j_{i}) = (j_{i}, \sin(2^{0} \pi j_{i}), \cos(2^{0} \pi j_{i}), \dots, \sin(2^{L-1} \pi j_{i}), \cos(2^{L-1} \pi j_{i}))$ , where $L = 10$ is the number of frequency bands. We then project the encoded points (i.e., $\gamma(J) \in \mathbb{R}^{23 \times 63}$ ) to a raw pose-embedded code $\mathbf{P}_{raw}$ by a projection matrix $\mathbf{W}_{p}$ : + +$$ +\mathbf {P} _ {\text {r a w}} = \mathbf {W} _ {p} \cdot \gamma (J). \tag {6} +$$ + +Then, we generate a query signal $\mathbf{P}_q$ from $\mathbf{S}$ , and key and value signals from $\mathbf{P}_{raw}$ via linear projections as follows: + +$$ +\mathbf {P} _ {q} = \mathbf {W} _ {Q} \cdot \mathbf {S}, \quad \mathbf {P} _ {k} = \mathbf {W} _ {K} \cdot \mathbf {P} _ {\text {r a w}}, \quad \mathbf {P} _ {v} = \mathbf {W} _ {V} \cdot \mathbf {P} _ {\text {r a w}}. \tag {7} +$$ + +Finally, we generate the learnable pose-embedded implicit feature $\mathbf{P}_{nr}$ for the radiance residual estimation as follows: + +$$ +\mathbf {P} _ {n r} = \operatorname {s o f t m a x} \left(\mathbf {P} _ {q} \cdot \left(\mathbf {P} _ {k}\right) ^ {\top}\right) \cdot \mathbf {P} _ {v}. \tag {8} +$$ + +As explained earlier, we employ $\mathbf{P}_{nr}$ to condition $\mathbf{M}_{nr}$ , the MLP decoder of the non-rigid radiance branch. $\mathbf{M}_{nr}$ optimizes the implicit feature $\mathbf{P}_{nr}$ and its generation module for estimating pose-coherent radiance residuals instead of solely relying on explicit volumetric features. + +# 3.3 Multi-subject Unified Training + +Our effective radiance compositional approach for non-rigid motion modeling further enables a flexible extension of our single-subject MoCo-NeRF for unified training with multiple subjects, which is very beneficial for practical applications. Each monocular video exhibits distinct non-rigid motions, thereby requiring individualized modeling. However, it is impossible for the traditional spatial offset learning approach [73] that typically requires a heavy module for photo-realistic quality. In contrast, our radiance compositional approach can achieve it by employing a separate, lightweight MHE for each subject, referred to as local MHE. + +Building on this foundation, we propose three architectural modifications as illustrated in Fig. 4. To benefit from unified training of multiple subjects, we introduce a coarse global MHE $\psi_{hash}^{G}$ shared across all subjects to learn subject-generic representations (e.g., body shapes). We also ensure learning of subject-specific details with $N_{s}$ local MHEs $\{\psi_{hash}^{i}\}_{i = 1}^{N_{s}}$ , where $N_{s}$ is the number of subjects. For each canonical sample $\mathbf{x}_c$ , we query global features from the + +global MHE and concatenate it to the sliced feature from the corresponding local MHE for radiance predictions. We additionally propose to leverage the learnable base codes of all subjects $\{\mathbf{S}_i\}_{i=1}^{N_s}$ as ID codes, for sharing MLP decoders while ensuring subject-distinctive canonical representations. The modified radiance estimations of the rigid branch for the multi-subject training are as follows: + +$$ +c, \sigma = M _ {r i g i d} \left(\psi_ {h a s h} ^ {G} \left(\mathbf {x} _ {c}\right) \oplus \left(\psi_ {h a s h} ^ {i} \left(\mathbf {x} _ {c}\right)\right) ^ {: H} \oplus \mathbf {S} _ {i}\right) \tag {9} +$$ + +# 3.4 Model Optimization + +We optimize MoCo-NeRF with all $N_{i}$ image frames $\{I_j\}_{j=1}^{N_i}$ from the monocular video via patch-based ray sampling. Given an image frame $I_j$ at each iteration, we randomly sample $N_p$ patches denoted as $\{P_k\}_{k=1}^{N_p}$ to construct the LPIPS [84] loss. The LPIPS loss term can measure the perceptual distance between two image patches, and thus we also adopt it for more perceptually coherent renderings. Our framework first renders $\{\hat{P}_k^r, \hat{P}_k^f\}_{k=1}^{N_p}$ respectively from $C_{rigid}$ and $C_{final}$ , and compute two LPIPS loss terms $\mathcal{L}_{\mathrm{LPIPS}}^r$ and $\mathcal{L}_{\mathrm{LPIPS}}^f$ against the ground truth patch $P_k^{GT}$ by using the pre-trained VGGNet [63] to as $\mathcal{L}_{\mathrm{LPIPS}} = \mathrm{LPIPS}(VGG(\hat{P}_k), VGG(P_k^{GT}))$ . We respectively compute the MSE loss terms $\mathcal{L}_{\mathrm{MSE}}^r, \mathcal{L}_{\mathrm{MSE}}^f$ from each of the estimated patches $\{\hat{P}_k^r, \hat{P}_k^f\}_{k=1}^{N_p}$ . Moreover, we additionally compute an appearance matching loss with SSIM $\mathcal{L}_{\mathrm{SSIM}}^f$ [15] only for the final RGB $\{\hat{P}_k^f\}_{k=1}^{N_p}$ , to improve the rendering quality. In summary, there are 5 loss terms in total: $\mathcal{L}_{\mathrm{LPIPS}}^r, \mathcal{L}_{\mathrm{MSE}}^r$ for the canonical radiance field of the rigid branch, and $\mathcal{L}_{\mathrm{LPIPS}}^f, \mathcal{L}_{\mathrm{MSE}}^f, \mathcal{L}_{\mathrm{SSIM}}^f$ for the final motion-oriented compositional renderings. To enforce the rigid branch to learn the main canonical radiance fields while the non-rigid branch to learn radiance residuals, we compute the final overall loss function $\mathcal{L}^o$ with a hyperparameter $\lambda = 0.2$ as: + +$$ +\mathcal {L} ^ {o} = \lambda \mathcal {L} ^ {r} + (1 - \lambda) \mathcal {L} ^ {f} \tag {10} +$$ + +$$ +\mathcal {L} ^ {r} = \mathcal {L} _ {\text {L P I P S}} ^ {r} + \mathcal {L} _ {\text {M S E}} ^ {r}, \quad \mathcal {L} ^ {f} = \mathcal {L} _ {\text {L P I P S}} ^ {f} + \mathcal {L} _ {\text {M S E}} ^ {f} + \mathcal {L} _ {\text {S S I M}} ^ {f} \tag {11} +$$ + +# 4 Experiments + +We conduct extensive experiments on ZJU-MoCap [52] and MonoCap [17, 18, 50] to verify the effectiveness of our proposed approach on both single- and multi-subject settings. The multi-subject setting indicates multiple subjects are jointly optimized. Our findings demonstrate that MoCo-NeRF can learn photo-realistic dynamic human representations by effectively handling non-rigid motions. Moreover, it learns discriminative representations of each performer in the multi-subject setting, while prior works struggle. + +# 4.1 Datasets + +We leverage ZJU-MoCap and MonoCap datasets for both training and evaluation, following [14,21], and provide the details in this section. For ZJU-MoCap, + +Table 1: Quantitative results of single-subject methods on ZJU-MoCap and MonoCap datasets where LPIPS* = LPIPS × 10³ with the best and second-best performance notations. MoCo-NeRF outperforms the baselines on most of the evaluation metrics in both settings. Despite pixel-wise metrics (PSNR, SSIM) tend to favor smooth renderings generated from [14,21], we achieve state-of-the-art performance while effectively modeling the photo-realistic non-rigid motions as illustrated in Fig. 5. + +
Single-subject methodsZJU-MoCap [52]MonoCap [50]
PSNR↑SSIM↑LPIPS*↓PSNR↑SSIM↑LPIPS*↓
HumanNeRF [73]30.180.970930.1332.920.986412.93
Instant-NVR [14]30.730.969738.8731.890.987016.07
GauHuman [21]30.980.962030.1632.870.984913.54
MoCo-NeRF (Ours)31.060.973428.8333.010.987413.02
+ +Fig.5: Qualitative comparison of novel view synthesis from single-subject training. MoCo-NeRF achieves photo-realistic rendering qualities with high-fidelity non-rigid motion modeling. Although HumanNeRF [73] presents comparable quality, however, MoCo-NeRF achieves much less training time as shown in Tab. 3. +![](images/19a464fe291d78058f3c0580c246fd2007b3679c25cd5f12876f9c27dde798c5.jpg) +HumanNeRF Instant-NVR GauHuman Ours Ground Truth HumanNeRF Instant-NVR GauHuman Ours Ground Truth + +we evaluate our MoCo-NeRF and other baselines on all 6 subjects (i.e., 377, 386, 387, 392, 393, 394). Out of 23 camera views, we use one of them for training and the other 22 views for evaluation. For MonoCap that comprises DynaCap [17] and DeepCap [18], we utilize all 4 subjects (i.e., lan, marc, olek, vlad) and use the same set of cameras as [14]. For multi-subject evaluations presented in Tab. 2, we evaluate on subjects 377, 386, and 392 of ZJU-MoCap. + +# 4.2 Implementation Details + +We optimize our model with Adam optimizer [30] with betas (0.9, 0.99), and learning rates of $5e^{-5}$ for the skeletal weights volume decoder and $5e^{-3}$ for the + +Table 2: Quantitative results of multi-subject methods on subjects #377, #386, #392 of ZJU-MoCap [52]. Despite the same evaluation settings, MoCo-NeRF significantly outperforms baseline methods under monocular settings with a shorter training time. + +
Multi-subject methodsPSNR↑SSIM↑LPIPS*↓Training videosTrain time
NHP [31]26.640.918540.47multi-view~86 hours
GHuNeRF [33]27.280.929140.82monocular~123 hours
MoCo-NeRF (Ours)31.970.978124.54monocular~2 hours
+ +![](images/fe8cbe06fa067da8562cf3b37049cd85ce5cb4bfd6794f43ac23813f070bd79f.jpg) +Fig.6: Qualitative comparison of novel view synthesis from multi-subject training. Although all subjects are seen during training, only our multi-subject MoCo-NeRF achieves photo-realistic quality comparable to the ones of single-subject models. + +others. Each MHE has 16 levels with feature dimensions $H = 2$ , and 4 levels with $H = 2$ for the global MHE. For each iteration, six $32 \times 32$ patches are used with 256 samples per ray. We train all models including the baselines on 1 Nvidia RTX 3090, where both baselines are trained for their suggested duration. Our model is trained for $1 \sim 2$ hours depending on each subject's complexity for the single-subject training, and 2 hours for the multi-subject training. + +# 4.3 State-of-the-art Comparison + +Single-subject methods. Tab. 1 presents PSNR, SSIM [72], and LPIPS* of MoCo-NeRF and our baselines [14, 21, 73], for novel-view synthesis under the single-subject setting. MoCo-NeRF achieves state-of-the-art performance on all metrics except one comparable LPIPS* for the MonoCap dataset. As shown in Fig. 5, Instant-NVR [14] shows moderate performance, however, it often generates smoothed-out results without proper non-rigid motions due to high dependency on accurate SMPL parameters. GauHuman [21] achieves better qualitative performance than [14] but still struggles to model sophisticated non-rigid motions. Fig. 1 presents longer optimizations are surprisingly ineffective and harmful respectively for Instant-NVR [14] and GauHuman [21] in terms of non-rigid motion modeling. This implies that their approaches cannot achieve optimal non-rigid motion modeling. Moreover, we observe that PSNR and SSIM often favor those results without fine-grained non-rigid motion modeling since predicting pose-dependent deformations from novel viewpoints may not be pixel-wise accurate. Instead, the perception metric (LPIPS*) evaluates better in this case, as + +Table 3: Training efficiency comparison for methods with photo-realistic non-rigid motion modeling on a single RTX3090. In the table, single-sub. and multi-sub. respectively represent training for 1 and 6 subjects. MoCo-NeRF significantly outperforms HumanNeRF in terms of both GPU memory usage and training time. HumanNeRF requires sequential training to achieve comparable non-rigid motion modeling for all subjects, thus resulting in linearly-increased training time. + +
GPU Mem. (GB)Train time (hours)
single-sub.multi-sub.Δsingle-sub.multi-sub.Δ
HumanNeRF [73]23.723.7(+0)38.9233.3(+194.5)
MoCo-NeRF (Ours)16.220.3(+4.1)1.92.1(+0.2)
+ +HumanNeRF [73] has a better LPIPS* score and also shows qualitatively better results. Thus, we consider GauHuman's comparable PSNR might be favored by a superior representation (3D Gaussians [29]) instead of better non-rigid motion modeling. A more detailed analysis is in supplementary materials. Despite this, we still achieve state-of-the-art performance with significantly improved training efficiency compared to another photo-realistic model [73] (Tab. 3). + +Multi-subject methods. We evaluate our MoCo-NeRF against NHP [31] and GHuNeRF [33], both of which are designed to achieve generalization capabilities. This is necessitated by the absence of prior works that directly target the joint learning task for multiple subjects from monocular videos. Additionally, we observe that the single-subject baseline models fail to learn distinct representations when trained in a unified manner. This observation underscores our decision to compare against generalization models that are specifically tailored to leverage multi-subject data for training. For fairness, we use their official checkpoints and ensure all evaluated subjects were included for training. + +As demonstrated in Tab. 2, MoCo-NeRF remarkably outperforms all the baseline methods on all metrics while also excelling in training efficiency up to 61 times faster. Even though MoCo-NeRF only utilizes a monocular video of each subject, Fig. 6 illustrates our method succeeds in learning both distinct representations and pose-dependent radiance residuals of each subject, enabling photo-realistic free-viewpoint renderings. + +# 4.4 Model Analysis + +In this section, we present analyses of our contributions respectively for two major achievements: novel non-rigid motion modeling (NR modeling) and multi-subject unified training (multi-subject modeling). For both analyses, we progressively remove each component from a full model to a vanilla approach. + +NR Modeling: Analysis of radiance compositional approach. Tab. 4 demonstrates our radiance compositional design perceptually improves the rendering results by closing the radiance discrepancies with predicted residuals. Fig. 7 shows that the representation learned solely by the rigid radiance branch (c) suffers from noisy renderings caused by incorrect correspondences due to + +Table 4: Ablation study in the single-subject task on ZJU-MoCap for b) pose-embedded implicit features and c) motion-oriented compositional NeRF approach. + +
Single-subjectPSNRSSIMLPIPS*
(a) MoCo-NeRF (Ours)31.060.973428.83
(b) (-) Pose features30.640.970229.65
(c) (-) Residual NeRF30.620.970229.93
+ +Table 5: Ablation study of multi-subject MoCo-NeRF on ZJU-MoCap for our proposed architectural modifications. + +
Multi-subjectPSNRSSIMLPIPS*
(i)MoCo-NeRF (Ours)30.910.974627.69
(ii)(-) Global MHE30.830.974628.49
(iii)(-)Ns local MHEs30.700.974229.03
(iv)(-) Identity codes28.630.966638.29
+ +![](images/0ab548174b18a0c6144622ad59a44e88d1a421d2246d0e26a86202d5d3922bcb.jpg) +Fig.7: Qualitative results of the ablation study on single-subject MoCo-NeRF presented in Tab. 4. + +![](images/97934d8dcb9bd581b3409de810265772b7ff5832e9695aa757622325ac14d1b4.jpg) +Fig. 8: Qualitative results of the ablation study in Tab. 5 where we can observe gradual improvements from (iv) to (i). + +non-rigid motions. With our radiance residual modeling (b), those issues are mitigated by jointly learning per-sample radiance residuals, resulting in more complete rendering and better non-rigid motion modeling as clearly shown in Fig. 7. In addition, Fig. 9 illustrates successful decomposed learning of two types of representations and validates the effectiveness of our radiance compositional approach in non-rigid motion modeling. + +NR Modeling: Analysis of pose-embedded implicit feature. Tab. 4 further illustrates that the final version with the proposed pose-embedded implicit features (Tab. 4 (a)) significantly improves the quantitative performance on all metrics. It enables the non-rigid branch to implicitly utilize body pose information in addition to the explicit features, yielding smoother and more coherent non-rigid motion modeling as shown in Fig. 7. We can observe remarkable quality enhancements that demonstrate the effectiveness of our attention-based code generation module. We also present an ablation study on the choice of cross-attention module for the feature generation in the supplementary material. + +Multi-subject Modeling: Analysis of using learnable base codes as ID codes. In Fig. 8, the results of MoCo-NeRF without any modifications (iv) exhibit mixed-up appearances on all target subjects, similar to the ones of [73] that are presented in the supplementary material. Utilizing a set of learnable base codes as ID codes (Tab. 5 (iii)) enables a single MLP decoder $M_{rigid}$ to differentiate each subject by optimizing the ID codes as plug-in parameters. Thus, distinct representations of all subjects are successfully learned as illustrated in (iii) of Fig. 8, yielding significant quantitative improvements as well (Tab. 5). + +![](images/f634bdf0c938ae43e180a834ea1bcd427cad6091693d578b317015426324aa83.jpg) +Fig. 9: Qualitative illustrations of non-rigid radiance residuals on novel-view synthesis where $\delta \rightarrow c' = c + \delta \Delta c$ , $\sigma' = \sigma + \delta \Delta \sigma$ . The first row progressively demonstrates the capability of the non-rigid radiance branch to learn the non-rigid motions on the subject's face in addition to the cloth folds. The last two rows clearly illustrate the effects of radiance residuals on cloth folds and hems corresponding to the body pose. + +Multi-subject Modeling: Analysis of $N_{s}$ local MHEs. Fig. 8 shows that the model (iii) lacks fine-grained details due to the sharing of one MHE. By utilizing separate $N_{s}$ local MHEs (Tab. 5 (ii)), the model learns subject-discriminative non-rigid motions as they now optimize separate explicit features for each subject's radiance residuals. Fig. 8 shows clear improvement of the model (ii) in non-rigid motion modeling where its quantitative gains are reflected in Tab. 5. Multi-subject Modeling: Analysis of global MHE. The addition of global MHE (i) further empowers MoCo-NeRF to learn better representations with more coherent non-rigid motions as illustrated in Fig. 8. Although the global MHE is not used for learning non-rigid radiance residuals, it encourages the rigid part of features from local MHE to concentrate on subject-specific representations. As its frozen copy is used for the non-rigid radiance branch, it facilitates better learning of radiance residuals for modeling non-rigid motions. Tab. 5 clearly shows that it can help to learn perceptually better representations. + +# 5 Conclusion + +In this paper, we presented a novel framework MoCo-NeRF for free-viewpoint dynamic human rendering that is capable of both single- and multi-subject training from monocular videos. By decomposing the neural representation of a dynamic human by motions and utilizing proposed implicit pose codes, MoCo-NeRF effectively learns radiance discrepancies induced by non-rigid motions. The proposed multi-subject MoCo-NeRF efficiently achieves comparable performance compared to the single-subject model. It has fully demonstrated its effectiveness and established new state-of-the-art performance on both problems. + +Acknowledgments. This research is supported in part by the Hong Kong PhD Fellowship Scheme (HKPFS), Early Career Scheme of the Research Grants Council (RGC) of the Hong Kong SAR under grant No. 26202321, SAIL Research Project, HKUST-Zeekr Collaborative Research Fund, HKUST-WeBank Joint Lab Project, Tencent Rhino-Bird Focused Research Program, and Naver Cloud Corporation. + +# References + +1. Alldieck, T., Zanfir, M., Sminchisescu, C.: Photorealistic monocular 3d reconstruction of humans wearing clothing. In: CVPR (2022) +2. Barron, J.T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., Srinivasan, P.P.: Mip-NeRF: A multiscale representation for anti-aliasing neural radiance fields. In: ICCV (2021) +3. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Mip-NeRF 360: Unbounded anti-aliased neural radiance fields. In: CVPR (2022) +4. Chen, A., Xu, Z., Geiger, A., Yu, J., Su, H.: TensoRF: Tensorial radiance fields. In: ECCV (2022) +5. Chen, A., Xu, Z., Zhao, F., Zhang, X., Xiang, F., Yu, J., Su, H.: MVSNeRF: Fast generalizable radiance field reconstruction from multi-view stereo. In: ICCV (2021) +6. Chen, X., Jiang, T., Song, J., Yang, J., Black, M.J., Geiger, A., Hilliges, O.: gDNA: Towards generative detailed neural avatars. In: CVPR (2022) +7. Chen, Y., Wang, X., Chen, X., Zhang, Q., Li, X., Guo, Y., Wang, J., Wang, F.: Uv volumes for real-time rendering of editable free-view human performance. In: CVPR (2023) +8. Collet, A., Chuang, M., Sweeney, P., Gillett, D., Evseev, D., Calabrese, D., Hoppe, H., Kirk, A., Sullivan, S.: High-quality streamable free-viewpoint video. In: SIGGRAPH (2015) +9. Dong, Z., Chen, X., Yang, J., J.Black, M., Hilliges, O., Geiger, A.: AG3D: Learning to generate 3D avatars from 2D image collections. In: ICCV (2023) +0. Dou, M., Khamis, S., Degtyarev, Y., Davidson, P., Fanello, S.R., Kowdle, A., Escolano, S.O., Rhemann, C., Kim, D., Taylor, J., Kohli, P., Tankovich, V., Izadi, S.: Fusion4D: Real-time performance capture of challenging scenes. In: SIGGRAPH (2016) +1. Du, Y., Zhang, Y., Yu, H.X., Tenenbaum, J.B., Wu, J.: Neural radiance flow for 4d view synthesis and video processing. In: ICCV (2021) +2. Gao, C., Saraf, A., Kopf, J., Huang, J.B.: Dynamic view synthesis from dynamic monocular video. In: ICCV (2021) +3. Garbin, S.J., Kowalski, M., Johnson, M., Shotton, J., Valentin, J.: FastNeRF: High-fidelity neural rendering at 200fps. In: ICCV (2021) +4. Geng, C., Peng, S., Xu, Z., Bao, H., Zhou, X.: Learning neural volumetric representations of dynamic humans in minutes. In: CVPR (2023) +5. Godard, C., Mac Aodha, O., Brostow, G.J.: Unsupervised monocular depth estimation with left-right consistency. In: CVPR (2017) +6. Guo, K., Lincoln, P., Davidson, P., Busch, J., Yu, X., Whalen, M., Harvey, G., Orts-Escalano, S., Pandey, R., Dourgarian, J., Tang, D., Tkach, A., Kowdle, A., Cooper, E., Dou, M., Fanello, S., Fyffe, G., Rhemann, C., Taylor, J., Debevec, P., Izadi, S.: The relightables: Volumetric performance capture of humans with realistic relighting. In: SIGGRAPH Asia (2019) + +17. Habermann, M., Liu, L., Xu, W., Zollhoefer, M., Pons-Moll, G., Theobalt, C.: Real-time deep dynamic characters. In: SIGGRAPH (2021) +18. Habermann, M., Xu, W., Zollhoefer, M., Pons-Moll, G., Theobalt, C.: DeepCap: Monocular human performance capture using weak supervision. In: CVPR (2020) +19. Hedman, P., Srinivasan, P.P., Mildenhall, B., Barron, J.T., Debevec, P.: Baking neural radiance fields for real-time view synthesis. In: ICCV (2021) +20. Hong, F., Chen, Z., Lan, Y., Pan, L., Liu, Z.: EVA3D: Compositional 3d human generation from 2d image collections. In: ICLR (2023) +21. Hu, S., Liu, Z.: GauHuman: Articulated gaussian splatting from monocular human videos. In: CVPR (2024) +22. Hu, T., Yu, T., Zheng, Z., Zhang, H., Liu, Y., Zwicker, M.: HVTR: Hybrid volumetric-textural rendering for human avatars. In: 3DV (2022) +23. Huang, Y., Yi, H., Liu, W., Wang, H., Wu, B., Wang, W., Lin, B., Zhang, D., Cai, D.: One-shot implicit animatable avatars with model-based priors. In: ICCV (2023) +24. Jain, A., Tancik, M., Abbeel, P.: Putting nerf on a diet: Semantically consistent few-shot view synthesis. In: ICCV (2021) +25. Jiang, B., Hong, Y., Bao, H., Zhang, J.: SelfRecon: Self reconstruction your digital avatar from monocular video. In: CVPR (2022) +26. Jiang, T., Chen, X., Song, J., Hilliges, O.: Instantavatar: Learning avatars from monocular video in 60 seconds. In: CVPR (2023) +27. Jiang, W., Yi, K.M., Samei, G., Tuzel, O., Ranjan, A.: NeuMan: Neural human radiance field from a single video. In: ECCV (2022) +28. Jiang, Y., Yao, K., Su, Z., Shen, Z., Luo, H., Xu, L.: Instant-NVR: Instant neural volumetric rendering for human-object interactions from monocular rgbd stream. In: CVPR (2023) +29. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM TOG (July 2023) +30. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. In: ICLR (2015) +31. Kwon, Y., Kim, D., Ceylan, D., Fuchs, H.: Neural human performer: Learning generalizable radiance fields for human performance rendering. In: NeurIPS (2021) +32. Lewis, J.P., Cordner, M., Fong, N.: Pose space deformation: A unified approach to shape interpolation and skeleton-driven deformation. In: SIGGRAPH (2000) +33. Li, C., Lin, J., Lee, G.H.: Generalizable human nerf from a monocular video. In: 3DV (2024) +34. Li, R., Tanke, J., Vo, M., Zollhofer, M., Gall, J., Kanazawa, A., Lassner, C.: TAVA: Template-free animatable volumetric actors. In: ECCV (2022) +35. Li, Z., Niklaus, S., Snavely, N., Wang, O.: Neural scene flow fields for space-time view synthesis of dynamic scenes. In: CVPR (2021) +36. Liu, L., Gu, J., Lin, K.Z., Chua, T.S., Theobalt, C.: Neural sparse voxel fields. In: NeurIPS (2020) +37. Liu, L., Habermann, M., Rudnev, V., Sarkar, K., Gu, J., Theobalt, C.: Neural actor: Neural free-view synthesis of human actors with pose control. In: SIGGRAPH Asia (2021) +38. Liu, Y., Peng, S., Liu, L., Wang, Q., Wang, P., Christian, T., Zhou, X., Wang, W.: Neural rays for occlusion-aware image-based rendering. In: CVPR (2022) +39. Loper, M., Mahmood, N., Romero, J., Pons-Moll, G., Black, M.J.: SMPL: A skinned multi-person linear model. In: SIGGRAPH Asia (2015) + +40. Martin-Brualla, R., Pandey, R., Yang, S., Pidlypenskyi, P., Taylor, J., Valentin, J., Khamis, S., Davidson, P., Tkach, A., Lincoln, P., Kowdle, A., Rhemann, C., Goldman, D.B., Keskin, C., Seitz, S., Izadi, S., Fanello, S.: LookinGood: Enhancing performance capture with real-time neural re-rendering. ACM TOG (2018) +41. Mihajlovic, M., Bansal, A., Zollhoefer, M., Tang, S., Saito, S.: KeypointNeRF: Generalizing image-based volumetric avatars using relative spatial encoding of keypoints. In: ECCV (2022) +42. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: NeRF: Representing scenes as neural radiance fields for view synthesis. In: ECCV (2020) +43. Mu, J., Sang, S., Vasconcelos, N., Wang, X.: ActorsNeRF: Animatable few-shot human rendering with generalizable nerfs. In: ICCV (2023) +44. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. In: SIGGRAPH (2022) +45. Noguchi, A., Sun, X., Lin, S., Harada, T.: Unsupervised learning of efficient geometry-aware neural articulated representations. In: ECCV (2022) +46. Oechsle, M., Peng, S., Geiger, A.: UNISURF: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In: ICCV (2021) +47. Park, K., Sinha, U., Barron, J.T., Bouaziz, S., Goldman, D.B., Seitz, S.M., Martin-Brualla, R.: Nerfies: Deformable neural radiance fields. In: ICCV (2021) +48. Park, K., Sinha, U., Hedman, P., Barron, J.T., Bouaziz, S., Goldman, D.B., Martin-Brualla, R., Seitz, S.M.: HyperNeRF: A higher-dimensional representation for topologically varying neural radiance fields. In: SIGGRAPH Asia (2021) +49. Peng, B., Hu, J., Zhou, J., Zhang, J.: SelfNeRF: Fast training nerf for human from monocular self-rotating video. arXiv preprint arXiv:2210.01651 (2022) +50. Peng, S., Dong, J., Wang, Q., Zhang, S., Shuai, Q., Zhou, X., Bao, H.: Animatable neural radiance fields for modeling dynamic human bodies. In: ICCV (2021) +51. Peng, S., Xu, Z., Dong, J., Wang, Q., Zhang, S., Shuai, Q., Bao, H., Zhou, X.: Animatable implicit neural representations for creating realistic avatars from videos. TPAMI (2024) +52. Peng, S., Zhang, Y., Xu, Y., Wang, Q., Shuai, Q., Bao, H., Zhou, X.: Neural Body: Implicit neural representations with structured latent codes for novel view synthesis of dynamic humans. In: CVPR (2021) +53. Pumarola, A., Corona, E., Pons-Moll, G., Moreno-Noguer, F.: D-NeRF: Neural Radiance Fields for Dynamic Scenes. In: CVPR (2020) +54. Reiser, C., Peng, S., Liao, Y., Geiger, A.: KiloNeRF: Speeding up neural radiance fields with thousands of tiny mlps. In: ICCV (2021) +55. Rematas, K., Liu, A., Srinivasan, P.P., Barron, J.T., Tagliasacchi, A., Funkhouser, T., Ferrari, V.: Urban radiance fields. In: CVPR (2022) +56. Remelli, E., Bagautdinov, T., Saito, S., Wu, C., Simon, T., Wei, S.E., Guo, K., Cao, Z., Prada, F., Saragih, J., et al.: Drivable volumetric avatars using texel-aligned features. In: SIGGRAPH (2022) +57. Saito, S., Huang, Z., Natsume, R., Morishima, S., Kanazawa, A., Li, H.: PIFu: Pixel-aligned implicit function for high-resolution clothed human digitization. In: ICCV (2019) +58. Saito, S., Simon, T., Saragih, J., Joo, H.: PIFuHD: Multi-level pixel-aligned implicit function for high-resolution 3d human digitization. In: CVPR (2020) +59. Sara Fridovich-Keil and Alex Yu, Tancik, M., Chen, Q., Recht, B., Kanazawa, A.: Plenoxels: Radiance fields without neural networks. In: CVPR (2022) +60. Schonberger, J.L., Zheng, E., Frahm, J.M., Pollefeys, M.: Pixelwise view selection for unstructured multi-view stereo. In: ECCV (2016) + +61. Schonberger, J.L., Frahm, J.M.: Structure-from-motion revisited. In: CVPR (2016) +62. Shao, R., Zhang, H., Zhang, H., Chen, M., Cao, Y., Yu, T., Liu, Y.: DoubleField: Bridging the neural surface and radiance fields for high-fidelity human reconstruction and rendering. In: CVPR (2022) +63. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition. In: ICLR (2015) +64. Su, S.Y., Yu, F., Zollhöfer, M., Rhodin, H.: A-NeRF: Articulated neural radiance fields for learning human shape, appearance, and pose. In: NeurIPS (2021) +65. Su, Z., Xu, L., Zheng, Z., Yu, T., Liu, Y., Fang, L.: RobustFusion: Human volumetric capture with data-driven visual cues using a rgbd camera. In: ECCV (2020) +66. Tancik, M., Casser, V., Yan, X., Pradhan, S., Mildenhall, B., Srinivasan, P., Barron, J.T., Kretzschmar, H.: Block-NeRF: Scalable large scene neural view synthesis. In: CVPR (2022) +67. Tretschk, E., Tewari, A., Golyanik, V., Zollhöfer, M., Lassner, C., Theobalt, C.: Non-rigid neural radiance fields: Reconstruction and novel view synthesis of a dynamic scene from monocular video. In: ICCV (2021) +68. Turki, H., Ramanan, D., Satyanarayanan, M.: Mega-NERF: Scalable construction of large-scale nerfs for virtual fly-throughs. In: CVPR (2022) +69. Wang, P., Liu, L., Liu, Y., Theobalt, C., Komura, T., Wang, W.: NeuS: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. In: NeurIPS (2021) +70. Wang, Q., Wang, Z., Genova, K., Srinivasan, P., Zhou, H., Barron, J.T., Martin-Brualla, R., Snavely, N., Funkhouser, T.: IBRNet: Learning multi-view image-based rendering. In: CVPR (2021) +71. Wang, S., Schwarz, K., Geiger, A., Tang, S.: ARAH: Animatable volume rendering of articulated human sdfs. In: ECCV (2022) +72. Wang, Z., Bovik, A., Sheikh, H., Simoncelli, E.: Image quality assessment: from error visibility to structural similarity. TIP (2004) +73. Weng, C.Y., Curless, B., Srinivasan, P.P., Barron, J.T., Kemelmacher-Shlizerman, I.: HumanNeRF: Free-viewpoint rendering of moving people from monocular video. In: CVPR (2022) +74. Weng, C.Y., Srinivasan, P.P., Curless, B., Kemelmacher-Shlizerman, I.: Person-NeRF: Personalized reconstruction from photo collections. In: CVPR (2023) +75. Wu, M., Wang, Y., Hu, Q., Yu, J.: Multi-view neural human rendering. In: CVPR (2020) +76. Xian, W., Huang, J.B., Kopf, J., Kim, C.: Space-time neural irradiance fields for free-viewpoint video. In: CVPR (2021) +77. Xiangli, Y., Xu, L., Pan, X., Zhao, N., Rao, A., Theobalt, C., Dai, B., Lin, D.: Bungeenerf: Progressive neural radiance field for extreme multi-scale scene rendering. In: ECCV (2022) +78. Xu, H., Alldieck, T., Sminchisescu, C.: H-NeRF: Neural radiance fields for rendering and temporal reconstruction of humans in motion. In: NeurIPS (2021) +79. Xu, L., Xiangli, Y., Peng, S., Pan, X., Zhao, N., Theobalt, C., Dai, B., Lin, D.: Grid-guided neural radiance fields for large urban scenes. In: CVPR (2023) +80. Xu, Z., Peng, S., Lin, H., He, G., Sun, J., Shen, Y., Bao, H., Zhou, X.: 4K4D: Real-time 4d view synthesis at 4k resolution. In: CVPR (2024) +81. Yariv, L., Gu, J., Kasten, Y., Lipman, Y.: Volume rendering of neural implicit surfaces. In: NeurIPS (2021) +82. Yu, A., Li, R., Tancik, M., Li, H., Ng, R., Kanazawa, A.: PlenOctrees for real-time rendering of neural radiance fields. In: ICCV (2021) + +83. Yu, A., Ye, V., Tancik, M., Kanazawa, A.: pixelNeRF: Neural radiance fields from one or few images. In: CVPR (2021) +84. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: CVPR (2018) +85. Zhang, R., Chen, J.: NDF: Neural deformable fields for dynamic human modelling. In: ECCV (2022) +86. Zhao, F., Yang, W., Zhang, J., Lin, P., Zhang, Y., Yu, J., Xu, L.: HumanNeRF: Efficiently generated human radiance field from sparse inputs. In: CVPR (2022) +87. Zheng, Z., Huang, H., Yu, T., Zhang, H., Guo, Y., Liu, Y.: Structured local radiance fields for human avatar modeling. In: CVPR (2022) +88. Zheng, Z., Zhao, X., Zhang, H., Liu, B., Liu, Y.: AvatarRex: Real-time expressive full-body avatars. In: SIGGRAPH (2023) +89. Zhi, Y., Qian, S., Yan, X., Gao, S.: Dual-space nerf: Learning animatable avatars and scene lighting in separate spaces. In: 3DV (2022) \ No newline at end of file diff --git a/2024/Motion-Oriented Compositional Neural Radiance Fields for Monocular Dynamic Human Modeling/images.zip b/2024/Motion-Oriented Compositional Neural Radiance Fields for Monocular Dynamic Human Modeling/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..bfed2f284b61f7bbb08483e54c2f2ba92728acb1 --- /dev/null +++ b/2024/Motion-Oriented Compositional Neural Radiance Fields for Monocular Dynamic Human Modeling/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0747d0be8c3fe17ec6adae0792ea04dffb70ff464e5decc616abc24f61cec18c +size 514269 diff --git a/2024/Motion-Oriented Compositional Neural Radiance Fields for Monocular Dynamic Human Modeling/layout.json b/2024/Motion-Oriented Compositional Neural Radiance Fields for Monocular Dynamic Human Modeling/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..d66d77efa149b70cc75fe0c180be6a2cfd8321a2 --- /dev/null +++ b/2024/Motion-Oriented Compositional Neural Radiance Fields for Monocular Dynamic Human Modeling/layout.json @@ -0,0 +1,10326 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 135, + 112, + 479, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 112, + 479, + 148 + ], + "spans": [ + { + "bbox": [ + 135, + 112, + 479, + 148 + ], + "type": "text", + "content": "Motion-Oriented Compositional Neural Radiance Fields for Monocular Dynamic Human Modeling" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 187, + 167, + 426, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 187, + 167, + 426, + 181 + ], + "spans": [ + { + "bbox": [ + 187, + 167, + 426, + 181 + ], + "type": "text", + "content": "Jaehyeok " + }, + { + "bbox": [ + 187, + 167, + 426, + 181 + ], + "type": "inline_equation", + "content": "\\mathrm{Kim}^{1}" + }, + { + "bbox": [ + 187, + 167, + 426, + 181 + ], + "type": "text", + "content": ", Dongyoon " + }, + { + "bbox": [ + 187, + 167, + 426, + 181 + ], + "type": "inline_equation", + "content": "\\mathrm{Wee}^{2}" + }, + { + "bbox": [ + 187, + 167, + 426, + 181 + ], + "type": "text", + "content": ", and Dan " + }, + { + "bbox": [ + 187, + 167, + 426, + 181 + ], + "type": "inline_equation", + "content": "\\mathrm{Xu}^{1\\boxtimes}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 160, + 189, + 452, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 189, + 452, + 213 + ], + "spans": [ + { + "bbox": [ + 160, + 189, + 452, + 213 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 160, + 189, + 452, + 213 + ], + "type": "text", + "content": " HKUST, Hong Kong " + }, + { + "bbox": [ + 160, + 189, + 452, + 213 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 160, + 189, + 452, + 213 + ], + "type": "text", + "content": " Naver Cloud Corp., South Korea jkimbf@cse.ust.hk, dongyoon.wee@navercorp.com, danxu@cse.ust.hk" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 236, + 455, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 236, + 455, + 522 + ], + "spans": [ + { + "bbox": [ + 160, + 236, + 455, + 522 + ], + "type": "text", + "content": "Abstract. This paper introduces Motion-oriented Compositional Neural Radiance Fields (MoCo-NeRF), a framework designed to perform free-viewpoint rendering of monocular human videos via novel non-rigid motion modeling approach. In the context of dynamic clothed humans, complex cloth dynamics generate non-rigid motions that are intrinsically distinct from skeletal articulations and critically important for the rendering quality. The conventional approach models non-rigid motions as spatial (3D) deviations in addition to skeletal transformations. However, it is either time-consuming or challenging to achieve optimal quality due to its high learning complexity without a direct supervision. To target this problem, we propose a novel approach of modeling non-rigid motions as radiance residual fields to benefit from more direct color supervision in the rendering and utilize the rigid radiance fields as a prior to reduce the complexity of the learning process. Our approach utilizes a single multiresolution hash encoding (MHE) to concurrently learn the canonical T-pose representation from rigid skeletal motions and the radiance residual field for non-rigid motions. Additionally, to further improve both training efficiency and usability, we extend MoCo-NeRF to support simultaneous training of multiple subjects within a single framework, thanks to our effective design for modeling non-rigid motions. This scalability is achieved through the integration of a global MHE and learnable identity codes in addition to multiple local MHEs. We present extensive results on ZJU-MoCap and MonoCap, clearly demonstrating state-of-the-art performance in both single- and multi-subject settings. The code and model will be made publicly available at the project page: https://stevejaehyeok.github.io/publications/moco-nerf." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 532, + 453, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 532, + 453, + 555 + ], + "spans": [ + { + "bbox": [ + 160, + 532, + 453, + 555 + ], + "type": "text", + "content": "Keywords: Dynamic human modeling from monocular videos " + }, + { + "bbox": [ + 160, + 532, + 453, + 555 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 532, + 453, + 555 + ], + "type": "text", + "content": " Neural Radiance Fields " + }, + { + "bbox": [ + 160, + 532, + 453, + 555 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 532, + 453, + 555 + ], + "type": "text", + "content": " Novel view synthesis of humans" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 588, + 231, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 588, + 231, + 601 + ], + "spans": [ + { + "bbox": [ + 132, + 588, + 231, + 601 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 611, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 611, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 611, + 482, + 647 + ], + "type": "text", + "content": "Novel view synthesis of monocular human video with photo-realistic quality is a challenging problem, as it involves complex articulated body motions and necessitates sophisticated modeling of non-skeletal or non-rigid motions (e.g., wrinkles" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 141, + 653, + 245, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 653, + 245, + 666 + ], + "spans": [ + { + "bbox": [ + 141, + 653, + 245, + 666 + ], + "type": "inline_equation", + "content": "\\boxtimes" + }, + { + "bbox": [ + 141, + 653, + 245, + 666 + ], + "type": "text", + "content": " Corresponding author." + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 116, + 478, + 213 + ], + "blocks": [ + { + "bbox": [ + 135, + 116, + 478, + 213 + ], + "lines": [ + { + "bbox": [ + 135, + 116, + 478, + 213 + ], + "spans": [ + { + "bbox": [ + 135, + 116, + 478, + 213 + ], + "type": "image", + "image_path": "98ab8cebc2cfd6b75e6abd8042a52f0e7f1fef5ecedab8c678c2c544464b74fb.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 135, + 217, + 478, + 322 + ], + "blocks": [ + { + "bbox": [ + 135, + 217, + 478, + 322 + ], + "lines": [ + { + "bbox": [ + 135, + 217, + 478, + 322 + ], + "spans": [ + { + "bbox": [ + 135, + 217, + 478, + 322 + ], + "type": "image", + "image_path": "64c9c69f2ff2f04e7bf36bee908d712b9ff294b69e7aa7018752917f01d95370.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 333, + 482, + 444 + ], + "lines": [ + { + "bbox": [ + 130, + 333, + 482, + 444 + ], + "spans": [ + { + "bbox": [ + 130, + 333, + 482, + 444 + ], + "type": "text", + "content": "Fig. 1: a) We introduce a motion-oriented compositional NeRF for photo-realistic modeling of dynamic humans from monocular videos. Our approach innovatively employs radiance compositions to capture pose-adaptive non-rigid motions, overcoming the limitations of skeletal transformations that typically yield an average of observed deformations (mean motions). b) The proposed MoCo-NeRF achieves state-of-the-art rendering quality and noteworthy efficiency in novel view synthesis, compared to leading competitors [14,21,73]. The bold denotes the proposed training duration of each comparison model. MoCo-NeRF uniquely captures coherent non-rigid motions, like T-shirt wrinkles relative to body pose, from entirely new viewpoints. Moreover, MoCo-NeRF significantly surpasses the efficiency of another photo-realistic method, HumanNeRF [73]." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 472, + 482, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 472, + 482, + 604 + ], + "spans": [ + { + "bbox": [ + 130, + 472, + 482, + 604 + ], + "type": "text", + "content": "of clothes) from unseen viewpoints. In particular, non-rigid motion modeling is essential for achieving photo-realism and coherence in free-viewpoint renderings. A prevalent solution adopted by recent methods [14,21,73] models the non-rigid motions as pose-dependent 3D spatial offsets of canonical point samples. However, the offset-based strategies lead to either extensive training time or irrelevant modeling of non-rigid motions due to its high learning complexity, incurred by the following reasons. First, a single spatial offset spans an unbounded range and is unique for every 3D point across all input frames and body poses, thereby escalating computational complexity. Second, the learning objective of spatial offsets is not directly aligned with the conventional RGB-based supervision, further complicating the modeling process." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "To target the issues above-discussed, in this paper, we propose Motion-oriented Compositional Neural Radiance Fields (MoCo-NeRF), a new framework that effectively models non-rigid motions at a radiance-level through a unique design of residual radiance field learning for free-viewpoint renderings of monocular human videos. Specifically, our proposed framework decomposes a single" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "J. Kim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 283 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 283 + ], + "type": "text", + "content": "MHE [44] into two branches: a rigid branch that learns a canonical T-pose radiance field solely via rigid skeletal transformations, and a non-rigid branch that optimizes pose-driven radiance residuals for capturing appearance discrepancies between renderings of the rigid branch and the ground-truth images, which are typically caused by non-rigid motions. The proposed approach circumvents the complexities associated with learning unbounded spatial offsets by reducing the prediction range to the radiance space with the rigid radiance field as a prior. It further establishes a more direct supervision from the ground-truth pixel colors for optimizing the rigid and non-rigid radiance fields by embedding them together in the rendering process, which can further enhance the learning efficiency. Besides, we present a cross-attention module to query a pose conditioned implicit feature from a learnable base code. This process allows our framework to extract more discriminative implicit features tailored to each particular pose, helping our non-rigid branch to learn coherent pose-driven radiance residuals." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 285, + 482, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 285, + 482, + 415 + ], + "spans": [ + { + "bbox": [ + 130, + 285, + 482, + 415 + ], + "type": "text", + "content": "In addition, thanks to our effective design for modeling non-rigid motions, the proposed MoCo-NeRF architecture can be flexibly extended to support unified training across multiple subjects while maintaining photo-realistic quality, without a significant increase in training time. For this architectural extension, we introduce multiple local MHEs and a global MHE, and also extend the aforementioned base code to a set of identity (ID) codes as a learnable codebook. The multiple local MHEs are designed for modeling multiple subjects and the global MHE targets the learning of generic and universal body representations. Furthermore, subject-specific learnable ID codes that are pivotal for generating pose-embedded implicit features enable effective sharing of the MLP decoder in the rigid branch by all subjects, enhancing the model's efficiency and scalability." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 416, + 482, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 416, + 482, + 487 + ], + "spans": [ + { + "bbox": [ + 130, + 416, + 482, + 487 + ], + "type": "text", + "content": "MoCo-NeRF achieves state-of-the-art performance on free-viewpoint rendering of monocular human videos in both single- and multi-subject settings, showing clear improvements compared to the prior works. We present our extensive results on ZJU-MoCap [52] and MonoCap [17, 18, 50] to demonstrate the effectiveness of the proposed MoCo-NeRF in learning photo-realistic representation and modeling complex non-rigid motions from monocular videos." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 146, + 488, + 387, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 488, + 387, + 500 + ], + "spans": [ + { + "bbox": [ + 146, + 488, + 387, + 500 + ], + "type": "text", + "content": "In summary, the contribution of our work is threefold:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 509, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 138, + 509, + 481, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 509, + 481, + 581 + ], + "spans": [ + { + "bbox": [ + 138, + 509, + 481, + 581 + ], + "type": "text", + "content": "- We propose a novel motion-oriented compositional NeRF approach, MoCo-NeRF, which concurrently learns canonical rigid radiance fields and pose-dependent radiance residual fields, effectively handling both rigid and non-rigid motions. We also introduce a pose-embedded implicit feature, computed by cross-attention between a body pose and a designed learnable base code, to implicitly enhance pose-dependent radiance residual learning." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 582, + 481, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 582, + 481, + 640 + ], + "spans": [ + { + "bbox": [ + 138, + 582, + 481, + 640 + ], + "type": "text", + "content": "- The proposed MoCo-NeRF is the first work that can be readily extended to enable unified training of multiple subjects without significantly increasing the GPU memory demands. We achieve this through our local MHEs, a global MHE, and the extended identity codebook, which enable subject-discriminative representation learning and effective module sharing." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 641, + 481, + 665 + ], + "type": "text", + "content": "- The proposed MoCo-NeRF achieves state-of-the-art performance under both single- and multi-subject training settings with a significant efficiency im" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 394, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 394, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 394, + 91, + 447, + 100 + ], + "type": "text", + "content": "MoCo-NeRF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 148, + 116, + 482, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 116, + 482, + 140 + ], + "spans": [ + { + "bbox": [ + 148, + 116, + 482, + 140 + ], + "type": "text", + "content": "provement compared to the best-performing methods. MoCo-NeRF jointly learns all 6 subjects of ZJU-MoCap on one RTX 3090 GPU within 2 hours." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 159, + 238, + 172 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 159, + 238, + 172 + ], + "spans": [ + { + "bbox": [ + 132, + 159, + 238, + 172 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 186, + 481, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 186, + 481, + 317 + ], + "spans": [ + { + "bbox": [ + 130, + 186, + 481, + 317 + ], + "type": "text", + "content": "We review closely related work from three perspectives: deformable neural rendering, neural human rendering, and monocular photo-realistic human rendering. Deformable neural rendering. NeRF [42] leverages an MLP to learn a static 3D representation of a scene from a dense set of images from diverse viewpoints. For its application in real-life scenarios, prior works have enhanced NeRF in various directions such as improving efficiencies [4,13,19,36,44,54,59,82] or relieving strong reliance on impractical settings like dense input views [5,24,38,70,83]. In addition, there are more branches of work such as large-scale modeling [55,66, 68,77,79], improving rendering quality [2,3], reconstruction quality [46,69,81] and more. However, they are restricted to static scenes while the majority of the real-world objects are dynamic." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 318, + 481, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 318, + 481, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 318, + 481, + 593 + ], + "type": "text", + "content": "Recent works including [11, 12, 35, 47, 48, 53, 67, 76] have broadened the modeling capabilities of NeRF to dynamic scenes with small and simple movements/deformations. However, they often fail in handling more dynamic or complex motions like human body articulations. In contrast, our approach enables learning of human-articulated motions from monocular videos, and performing free-viewpoint rendering of a human performer at any time frame of the videos. Neural rendering of humans. Many early-stage prior works for neural human renderings utilize traditional techniques like multi-view stereo [16, 60, 61, 75] or to build sophisticated systems for capturing human motions [8, 10, 40, 65]. However, constructing a dense set of training views and building those capturing systems are expensive, making them unsuitable for real-life situations. [1, 6, 57, 58] address 3D human reconstruction by utilizing ground truth human scans, relieving the aforementioned requirements. [9, 20, 45] further improves it by enabling human reconstruction from 2D image collections. In recent days, there have been many works learning an implicit human representation for novel pose synthesis [22, 23, 25, 27, 34, 50, 56, 64, 71, 87-89] or for a novel-view synthesis [27, 28, 31, 37, 51, 52, 64, 73, 74, 78, 85, 86]. There also have been several generalizable works [41, 43, 62] for a novel human subject with pre-training on multi-view dataset and a few-shot fine-tuning. [7, 80] focus on improving rendering efficiency to achieve real-time free-view rendering of dynamic humans. Furthermore, [14, 26, 28, 49] adopt explicit hash encoding [44] in order to boost the training efficiency for human representation modeling. In contrast to our problem setup, most of those works require multi-view videos or do not model the non-rigid motions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "Monocular photo-realistic human rendering. HumanNeRF [73] presents a state-of-the-art performance by optimizing motion fields and canonical T-type representation of a human performer. However, its implicit modeling of pose-dependent non-rigid motions requires tens of hours to converge. Instant-NVR [14] achieves efficient training with part-based representation and 2D motion parameterization for learning non-rigid motions as 3D offsets. Note that the" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "J. Kim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 113, + 482, + 228 + ], + "blocks": [ + { + "bbox": [ + 133, + 113, + 482, + 228 + ], + "lines": [ + { + "bbox": [ + 133, + 113, + 482, + 228 + ], + "spans": [ + { + "bbox": [ + 133, + 113, + 482, + 228 + ], + "type": "image", + "image_path": "5ad053192ebfef455342c6dfacb401ad87379dfee095188170b3a32624e90a18.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 236, + 482, + 304 + ], + "lines": [ + { + "bbox": [ + 130, + 236, + 482, + 304 + ], + "spans": [ + { + "bbox": [ + 130, + 236, + 482, + 304 + ], + "type": "text", + "content": "Fig. 2: Overview of the proposed framework MoCo-NeRF for free-view human rendering from a monocular video. Without estimating a geometrical offset of each continuous canonical point for each body pose, our framework is able to handle all deformations via its radiance-compositional approach with a single MHE [44] and achieve state-of-the-art performance. The pose-embedded implicit feature further enhances the learning of non-rigid radiance residuals by enabling pose-distinctive representation learning." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 327, + 482, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 327, + 482, + 472 + ], + "spans": [ + { + "bbox": [ + 130, + 327, + 482, + 472 + ], + "type": "text", + "content": "human parametric model (SMPL) [39], an extra requirement for training, plays a significant role in their framework. In addition, a concurrent work GauHuman [21], which also employs spatial offset approach with Gaussian Splatting (GS) [29], further enhances the efficiency and performance upon [14] by leveraging SMPL for the density control of GS. Nevertheless, [14, 21] struggle to learn complex non-rigid motions, even with an extended training duration as shown in Fig. 1. In this paper, instead of directly learning spatial offsets like these existing works, we propose a novel approach that effectively models non-rigid body motions by learning motion-oriented compositional radiance fields conditioned by pose-embedded implicit features. Our approach clearly outperforms the best-performing competitors HumanNeRF [73], Instant-NVR [14], and GauHuman [21] and can further extend to support joint training of multiple subjects." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 493, + 369, + 507 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 493, + 369, + 507 + ], + "spans": [ + { + "bbox": [ + 132, + 493, + 369, + 507 + ], + "type": "text", + "content": "3 The Proposed MoCo-NeRF Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": "The proposed MoCo-NeRF aims to effectively learn high-fidelity representations for free-viewpoint rendering of dynamic human subjects in monocular videos. We introduce our novel radiance compositional NeRF approach based on different motion types: rigid and non-rigid motions, as illustrated in Fig. 2. The composition of canonical radiance and the per-sample pose-dependent radiance residuals (Sec. 3.2) facilitates photo-realistic quality of rendering with fine-grained modeling of non-rigid motions, rather than leveraging the conventional spatial 3D offset learning approach. We further improve the quality of non-rigid radiance residuals by introducing a pose-embedded implicit features (Sec. 3.2) that is jointly optimized with the non-rigid radiance branch. In Sec. 3.3, thanks to our effective design of non-rigid motion modeling, it enables MoCo-NeRF to be flexibly extended for unified training of multiple subjects with a high efficiency." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 394, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 394, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 394, + 91, + 447, + 100 + ], + "type": "text", + "content": "MoCo-NeRF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 353, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 353, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 353, + 129 + ], + "type": "text", + "content": "3.1 Preliminary: Skeletal Transformations" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 135, + 482, + 184 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 135, + 482, + 184 + ], + "spans": [ + { + "bbox": [ + 130, + 135, + 482, + 184 + ], + "type": "text", + "content": "The skeletal motion learns an inverse formulation of the linear blend skinning [32], which transforms canonical points to different body poses by skinning weights. Weng et al. [73] reformulate the linear blend skinning to transform a point " + }, + { + "bbox": [ + 130, + 135, + 482, + 184 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 130, + 135, + 482, + 184 + ], + "type": "text", + "content": " in the observation space to a canonical point " + }, + { + "bbox": [ + 130, + 135, + 482, + 184 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_c" + }, + { + "bbox": [ + 130, + 135, + 482, + 184 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 197, + 193, + 481, + 225 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 193, + 481, + 225 + ], + "spans": [ + { + "bbox": [ + 197, + 193, + 481, + 225 + ], + "type": "interline_equation", + "content": "\\mathbf {x} _ {c} = \\sum_ {k = 1} ^ {K} w _ {k} ^ {o} \\left(R _ {k} \\mathbf {x} + \\mathbf {t} _ {k}\\right), w _ {k} ^ {o} = \\frac {w _ {k} ^ {c} \\left(R _ {k} \\mathbf {x} + \\mathbf {t} _ {k}\\right)}{\\sum_ {j = 1} ^ {K} w _ {j} ^ {c} \\left(R _ {j} \\mathbf {x} + \\mathbf {t} _ {j}\\right)}, \\tag {1}", + "image_path": "89d039987f3b6987daa04d03dd45da34b7842978336e70885f7ca8c3da334904.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 232, + 482, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 232, + 482, + 281 + ], + "spans": [ + { + "bbox": [ + 131, + 232, + 482, + 281 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 232, + 482, + 281 + ], + "type": "inline_equation", + "content": "w_{k}^{o}" + }, + { + "bbox": [ + 131, + 232, + 482, + 281 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 232, + 482, + 281 + ], + "type": "inline_equation", + "content": "w_{k}^{c}" + }, + { + "bbox": [ + 131, + 232, + 482, + 281 + ], + "type": "text", + "content": " respectively represent observation and canonical skinning weights of bone " + }, + { + "bbox": [ + 131, + 232, + 482, + 281 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 131, + 232, + 482, + 281 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 131, + 232, + 482, + 281 + ], + "type": "inline_equation", + "content": "R_{k}" + }, + { + "bbox": [ + 131, + 232, + 482, + 281 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 232, + 482, + 281 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_{k}" + }, + { + "bbox": [ + 131, + 232, + 482, + 281 + ], + "type": "text", + "content": " are the rotation and translation of bone " + }, + { + "bbox": [ + 131, + 232, + 482, + 281 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 131, + 232, + 482, + 281 + ], + "type": "text", + "content": ", respectively. Given Eq. 1, following [73], we optimize an explicit volume decoder to optimize " + }, + { + "bbox": [ + 131, + 232, + 482, + 281 + ], + "type": "inline_equation", + "content": "\\{w_{k}^{c}\\}_{k = 1}^{K}" + }, + { + "bbox": [ + 131, + 232, + 482, + 281 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 131, + 232, + 482, + 281 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 131, + 232, + 482, + 281 + ], + "type": "text", + "content": " is the total number of joints." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 297, + 443, + 309 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 297, + 443, + 309 + ], + "spans": [ + { + "bbox": [ + 131, + 297, + 443, + 309 + ], + "type": "text", + "content": "3.2 Motion-Oriented Compositional Neural Radiance Fields" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 316, + 482, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 316, + 482, + 426 + ], + "spans": [ + { + "bbox": [ + 130, + 316, + 482, + 426 + ], + "type": "text", + "content": "Rigid neural branch for learning canonical radiance field. The rigid branch of MoCo-NeRF is specifically designed to learn the canonical T-pose representation of the target subject using only rigid body movements. Our approach is different from but simpler than the prevalent spatial offset learning approach, which incorporates complex modeling of non-rigid motions as " + }, + { + "bbox": [ + 130, + 316, + 482, + 426 + ], + "type": "inline_equation", + "content": "\\Delta \\mathbf{x}_c" + }, + { + "bbox": [ + 130, + 316, + 482, + 426 + ], + "type": "text", + "content": ". Our rigid branch instead directly predicts the color and density of the canonical point " + }, + { + "bbox": [ + 130, + 316, + 482, + 426 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_c" + }, + { + "bbox": [ + 130, + 316, + 482, + 426 + ], + "type": "text", + "content": " transformed by inverse linear blend skinning (Sec. 3.1), as depicted in Fig. 2. First, we query multiresolution volumetric features " + }, + { + "bbox": [ + 130, + 316, + 482, + 426 + ], + "type": "inline_equation", + "content": "f_{\\mathbf{x}_c}" + }, + { + "bbox": [ + 130, + 316, + 482, + 426 + ], + "type": "text", + "content": " of the canonical point " + }, + { + "bbox": [ + 130, + 316, + 482, + 426 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_c" + }, + { + "bbox": [ + 130, + 316, + 482, + 426 + ], + "type": "text", + "content": " from the MHE " + }, + { + "bbox": [ + 130, + 316, + 482, + 426 + ], + "type": "inline_equation", + "content": "\\psi_{hash}" + }, + { + "bbox": [ + 130, + 316, + 482, + 426 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 271, + 434, + 480, + 447 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 271, + 434, + 480, + 447 + ], + "spans": [ + { + "bbox": [ + 271, + 434, + 480, + 447 + ], + "type": "interline_equation", + "content": "f _ {\\mathbf {x} _ {c}} = \\psi_ {\\text {h a s h}} (\\mathbf {x} _ {c}). \\tag {2}", + "image_path": "099d27a723122a69bdf71c812768e3d0cdab20c5810869622eabeec255fd1190.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 455, + 482, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 455, + 482, + 492 + ], + "spans": [ + { + "bbox": [ + 131, + 455, + 482, + 492 + ], + "type": "text", + "content": "Then, we slice the first half (" + }, + { + "bbox": [ + 131, + 455, + 482, + 492 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 131, + 455, + 482, + 492 + ], + "type": "text", + "content": " dimensions) of " + }, + { + "bbox": [ + 131, + 455, + 482, + 492 + ], + "type": "inline_equation", + "content": "f_{\\mathbf{x}_c}" + }, + { + "bbox": [ + 131, + 455, + 482, + 492 + ], + "type": "text", + "content": " at each level of resolution to compute " + }, + { + "bbox": [ + 131, + 455, + 482, + 492 + ], + "type": "inline_equation", + "content": "f_{\\mathbf{x}_c}^{:H}" + }, + { + "bbox": [ + 131, + 455, + 482, + 492 + ], + "type": "text", + "content": " and predicts the radiance " + }, + { + "bbox": [ + 131, + 455, + 482, + 492 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 131, + 455, + 482, + 492 + ], + "type": "text", + "content": " and density " + }, + { + "bbox": [ + 131, + 455, + 482, + 492 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 131, + 455, + 482, + 492 + ], + "type": "text", + "content": " of the rigid canonical point " + }, + { + "bbox": [ + 131, + 455, + 482, + 492 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_c" + }, + { + "bbox": [ + 131, + 455, + 482, + 492 + ], + "type": "text", + "content": " with an MLP decoder " + }, + { + "bbox": [ + 131, + 455, + 482, + 492 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{rigid}" + }, + { + "bbox": [ + 131, + 455, + 482, + 492 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 263, + 500, + 480, + 514 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 263, + 500, + 480, + 514 + ], + "spans": [ + { + "bbox": [ + 263, + 500, + 480, + 514 + ], + "type": "interline_equation", + "content": "c, \\sigma = \\mathbf {M} _ {\\text {r i g i d}} \\left(f _ {\\mathbf {x} _ {c}} ^ {H}\\right). \\tag {3}", + "image_path": "3ed05bfb63d266db840771ac46fd781e5bf999edfbf8949996a8fb37fb8cd651.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": "Non-rigid neural branch for learning radiance residual field. While the canonical representation does not explicitly model pose-dependent non-rigid motions, it still captures an average of such motions (mean motions) after training. However, the mean motions cannot effectively reflect the complex and realistic non-rigid motions under different body poses. It is only capable of rendering identical deformations across different poses, as can be observed in Fig. 1. To address this effectively, the non-rigid branch of MoCo-NeRF is designed to concurrently optimize for a radiance residual field. This models the appearance discrepancies between the renderings of the rigid branch and the ground-truth colors, typically caused by the complex non-rigid motions or deformations. We additionally condition the radiance residual predictions with the canonical radiance representation of the rigid branch as prior information. This enables more" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "text", + "content": "J. Kim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 114, + 282, + 232 + ], + "blocks": [ + { + "bbox": [ + 136, + 114, + 282, + 232 + ], + "lines": [ + { + "bbox": [ + 136, + 114, + 282, + 232 + ], + "spans": [ + { + "bbox": [ + 136, + 114, + 282, + 232 + ], + "type": "image", + "image_path": "37ddffe77c036dd982cc06c380f82c3d3434ccd85bbd069e44416d0e9e5cb26c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 134, + 239, + 284, + 293 + ], + "lines": [ + { + "bbox": [ + 134, + 239, + 284, + 293 + ], + "spans": [ + { + "bbox": [ + 134, + 239, + 284, + 293 + ], + "type": "text", + "content": "Fig.3: Illustration of the proposed pose-embedded implicit feature generation. We employ cross-attention to modulate the single learnable base code to pose-adaptive features." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 293, + 114, + 478, + 232 + ], + "blocks": [ + { + "bbox": [ + 293, + 114, + 478, + 232 + ], + "lines": [ + { + "bbox": [ + 293, + 114, + 478, + 232 + ], + "spans": [ + { + "bbox": [ + 293, + 114, + 478, + 232 + ], + "type": "image", + "image_path": "ddac47f517266c75f171259fdc89ab52a4be2d2a8b2542097939aaddf41fa83a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 292, + 239, + 479, + 292 + ], + "lines": [ + { + "bbox": [ + 292, + 239, + 479, + 292 + ], + "spans": [ + { + "bbox": [ + 292, + 239, + 479, + 292 + ], + "type": "text", + "content": "Fig. 4: Illustration of extended architecture of MoCo-NeRF for the multi-subject unified training (Sec. 3.3). Major components consist of the global MHE, set of local MHEs, and the dictionary of learnable base codes as ID codes." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 316, + 479, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 316, + 479, + 339 + ], + "spans": [ + { + "bbox": [ + 130, + 316, + 479, + 339 + ], + "type": "text", + "content": "coherent modeling of non-rigid motions, as the non-rigid and rigid radiance fields are highly correlated." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 340, + 482, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 340, + 482, + 483 + ], + "spans": [ + { + "bbox": [ + 130, + 340, + 482, + 483 + ], + "type": "text", + "content": "Specifically, unlike the rigid branch, the queried feature " + }, + { + "bbox": [ + 130, + 340, + 482, + 483 + ], + "type": "inline_equation", + "content": "f_{\\mathbf{x}_c}" + }, + { + "bbox": [ + 130, + 340, + 482, + 483 + ], + "type": "text", + "content": " is not sliced, enabling the non-rigid branch to learn the radiance residuals with reference to the canonical representation. Instead, we freeze the first half of " + }, + { + "bbox": [ + 130, + 340, + 482, + 483 + ], + "type": "inline_equation", + "content": "f_{\\mathbf{x}_c}" + }, + { + "bbox": [ + 130, + 340, + 482, + 483 + ], + "type": "text", + "content": " that represents the canonical radiance field, and only optimize the other half to learn the radiance residuals for non-rigid motion modeling. We denote this partially frozen feature as " + }, + { + "bbox": [ + 130, + 340, + 482, + 483 + ], + "type": "inline_equation", + "content": "f_{\\mathbf{x}_c}^*" + }, + { + "bbox": [ + 130, + 340, + 482, + 483 + ], + "type": "text", + "content": ". We observe that providing frozen " + }, + { + "bbox": [ + 130, + 340, + 482, + 483 + ], + "type": "inline_equation", + "content": "f_{\\mathbf{x}_c}^{H}" + }, + { + "bbox": [ + 130, + 340, + 482, + 483 + ], + "type": "text", + "content": " to the non-rigid radiance branch enables more effective and coherent residual learning. In addition, we also condition the radiance residual field prediction with our pose-embedded implicit feature " + }, + { + "bbox": [ + 130, + 340, + 482, + 483 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{nr}" + }, + { + "bbox": [ + 130, + 340, + 482, + 483 + ], + "type": "text", + "content": ", which is achieved by concatenating " + }, + { + "bbox": [ + 130, + 340, + 482, + 483 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{nr}" + }, + { + "bbox": [ + 130, + 340, + 482, + 483 + ], + "type": "text", + "content": " to the first intermediate output of the MLP decoder " + }, + { + "bbox": [ + 130, + 340, + 482, + 483 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{nr}" + }, + { + "bbox": [ + 130, + 340, + 482, + 483 + ], + "type": "text", + "content": ", as illustrated in Fig. 2. The generation of " + }, + { + "bbox": [ + 130, + 340, + 482, + 483 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{nr}" + }, + { + "bbox": [ + 130, + 340, + 482, + 483 + ], + "type": "text", + "content": " is explained in the last subsection of Sec. 3.2. Eventually, the color and density residuals are predicted by " + }, + { + "bbox": [ + 130, + 340, + 482, + 483 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{nr}" + }, + { + "bbox": [ + 130, + 340, + 482, + 483 + ], + "type": "text", + "content": " for the final compositional radiance computation:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 247, + 495, + 480, + 509 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 247, + 495, + 480, + 509 + ], + "spans": [ + { + "bbox": [ + 247, + 495, + 480, + 509 + ], + "type": "interline_equation", + "content": "\\Delta c, \\Delta \\sigma = \\mathbf {M} _ {n r} \\left(f _ {\\mathbf {x} _ {c}} ^ {*}; \\mathbf {P} _ {n r}\\right), \\tag {4}", + "image_path": "076313a2b2f49bb17620f3ba1afd779fdb20211d75708a19a336346622967685.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 512, + 479, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 512, + 479, + 561 + ], + "spans": [ + { + "bbox": [ + 130, + 512, + 479, + 561 + ], + "type": "text", + "content": "Compositional volume rendering. Following [42], we adopt stratified sampling and volume rendering to predict the RGB color of each ray. For final rendering, we query " + }, + { + "bbox": [ + 130, + 512, + 479, + 561 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 512, + 479, + 561 + ], + "type": "text", + "content": " samples per ray " + }, + { + "bbox": [ + 130, + 512, + 479, + 561 + ], + "type": "inline_equation", + "content": "\\mathbf{r}" + }, + { + "bbox": [ + 130, + 512, + 479, + 561 + ], + "type": "text", + "content": " and perform approximated integration to obtain final RGB " + }, + { + "bbox": [ + 130, + 512, + 479, + 561 + ], + "type": "inline_equation", + "content": "C_{final}" + }, + { + "bbox": [ + 130, + 512, + 479, + 561 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 130, + 512, + 479, + 561 + ], + "type": "inline_equation", + "content": "(c + \\Delta c, \\sigma + \\Delta \\sigma)" + }, + { + "bbox": [ + 130, + 512, + 479, + 561 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 167, + 570, + 480, + 585 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 167, + 570, + 480, + 585 + ], + "spans": [ + { + "bbox": [ + 167, + 570, + 480, + 585 + ], + "type": "interline_equation", + "content": "C _ {f i n a l} (\\mathbf {r}) = \\sum_ {m = 1} ^ {M} T _ {m} \\left(1 - \\exp \\left(- \\left(\\sigma_ {m} + \\Delta \\sigma_ {m}\\right) \\delta_ {m}\\right)\\right) \\left(c _ {m} + \\Delta c _ {m}\\right). \\tag {5}", + "image_path": "c54cef1697cf0b384710b88c3499bf946f85eecaf09366ac38cd0654a278fa47.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 594, + 482, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 641 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 594, + 482, + 641 + ], + "type": "inline_equation", + "content": "\\delta_{m}" + }, + { + "bbox": [ + 130, + 594, + 482, + 641 + ], + "type": "text", + "content": " is the adjacent distance from the " + }, + { + "bbox": [ + 130, + 594, + 482, + 641 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 130, + 594, + 482, + 641 + ], + "type": "text", + "content": "-th to the " + }, + { + "bbox": [ + 130, + 594, + 482, + 641 + ], + "type": "inline_equation", + "content": "m+1" + }, + { + "bbox": [ + 130, + 594, + 482, + 641 + ], + "type": "text", + "content": "-th sample and " + }, + { + "bbox": [ + 130, + 594, + 482, + 641 + ], + "type": "inline_equation", + "content": "T_{m} = \\exp(-\\sum_{n=1}^{m-1} \\sigma_{n} \\delta_{n})" + }, + { + "bbox": [ + 130, + 594, + 482, + 641 + ], + "type": "text", + "content": " as the transmittance. In addition, we compute " + }, + { + "bbox": [ + 130, + 594, + 482, + 641 + ], + "type": "inline_equation", + "content": "C_{rigid}" + }, + { + "bbox": [ + 130, + 594, + 482, + 641 + ], + "type": "text", + "content": " in the same way from " + }, + { + "bbox": [ + 130, + 594, + 482, + 641 + ], + "type": "inline_equation", + "content": "(c, \\sigma)" + }, + { + "bbox": [ + 130, + 594, + 482, + 641 + ], + "type": "text", + "content": " for an additional loss term that encourages a decomposed learning of the canonical radiance field and the radiance residuals (Sec. 3.4)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "type": "text", + "content": "Learnable Pose-Embedded Implicit Feature. We introduce high-frequency modulated implicit features computed from a learnable base code and a 3D body" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 394, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 394, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 394, + 91, + 447, + 100 + ], + "type": "text", + "content": "MoCo-NeRF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 274 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 274 + ], + "type": "text", + "content": "pose. A learnable base code " + }, + { + "bbox": [ + 130, + 116, + 482, + 274 + ], + "type": "inline_equation", + "content": "\\mathbf{S} \\in \\mathbb{R}^{64}" + }, + { + "bbox": [ + 130, + 116, + 482, + 274 + ], + "type": "text", + "content": " is randomly initialized and jointly optimized for a single subject with our motion-oriented compositional radiance fields. To make " + }, + { + "bbox": [ + 130, + 116, + 482, + 274 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 130, + 116, + 482, + 274 + ], + "type": "text", + "content": " distinguishable for each pose, we propose to employ the 3D body pose as a guidance for computing implicit features that encourage high-fidelity and pose-dependent radiance residual learning. A cross-attention mechanism is designed to perform interactions between the learnable base code and 23 joint positions (i.e. " + }, + { + "bbox": [ + 130, + 116, + 482, + 274 + ], + "type": "inline_equation", + "content": "J = \\{j_{i}\\}_{i=1}^{23}" + }, + { + "bbox": [ + 130, + 116, + 482, + 274 + ], + "type": "text", + "content": "). A detailed overview of the mechanism is illustrated in Fig. 3. To allow high-frequency implicit feature learning with 3D body joints, we conduct positional encoding [42] for the input joint coordinates. We first project each joint position into a higher dimension with a sinusoidal positional encoding, " + }, + { + "bbox": [ + 130, + 116, + 482, + 274 + ], + "type": "inline_equation", + "content": "\\gamma(j_{i}) = (j_{i}, \\sin(2^{0} \\pi j_{i}), \\cos(2^{0} \\pi j_{i}), \\dots, \\sin(2^{L-1} \\pi j_{i}), \\cos(2^{L-1} \\pi j_{i}))" + }, + { + "bbox": [ + 130, + 116, + 482, + 274 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 116, + 482, + 274 + ], + "type": "inline_equation", + "content": "L = 10" + }, + { + "bbox": [ + 130, + 116, + 482, + 274 + ], + "type": "text", + "content": " is the number of frequency bands. We then project the encoded points (i.e., " + }, + { + "bbox": [ + 130, + 116, + 482, + 274 + ], + "type": "inline_equation", + "content": "\\gamma(J) \\in \\mathbb{R}^{23 \\times 63}" + }, + { + "bbox": [ + 130, + 116, + 482, + 274 + ], + "type": "text", + "content": ") to a raw pose-embedded code " + }, + { + "bbox": [ + 130, + 116, + 482, + 274 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{raw}" + }, + { + "bbox": [ + 130, + 116, + 482, + 274 + ], + "type": "text", + "content": " by a projection matrix " + }, + { + "bbox": [ + 130, + 116, + 482, + 274 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{p}" + }, + { + "bbox": [ + 130, + 116, + 482, + 274 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 265, + 284, + 481, + 297 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 265, + 284, + 481, + 297 + ], + "spans": [ + { + "bbox": [ + 265, + 284, + 481, + 297 + ], + "type": "interline_equation", + "content": "\\mathbf {P} _ {\\text {r a w}} = \\mathbf {W} _ {p} \\cdot \\gamma (J). \\tag {6}", + "image_path": "0b848c2b88e71a6787be4b1dfbbfec85f04e0c62f5472a7c3b765ecefdd7d677.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 301, + 481, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 301, + 481, + 326 + ], + "spans": [ + { + "bbox": [ + 131, + 301, + 481, + 326 + ], + "type": "text", + "content": "Then, we generate a query signal " + }, + { + "bbox": [ + 131, + 301, + 481, + 326 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_q" + }, + { + "bbox": [ + 131, + 301, + 481, + 326 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 131, + 301, + 481, + 326 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 131, + 301, + 481, + 326 + ], + "type": "text", + "content": ", and key and value signals from " + }, + { + "bbox": [ + 131, + 301, + 481, + 326 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{raw}" + }, + { + "bbox": [ + 131, + 301, + 481, + 326 + ], + "type": "text", + "content": " via linear projections as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 190, + 335, + 480, + 348 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 190, + 335, + 480, + 348 + ], + "spans": [ + { + "bbox": [ + 190, + 335, + 480, + 348 + ], + "type": "interline_equation", + "content": "\\mathbf {P} _ {q} = \\mathbf {W} _ {Q} \\cdot \\mathbf {S}, \\quad \\mathbf {P} _ {k} = \\mathbf {W} _ {K} \\cdot \\mathbf {P} _ {\\text {r a w}}, \\quad \\mathbf {P} _ {v} = \\mathbf {W} _ {V} \\cdot \\mathbf {P} _ {\\text {r a w}}. \\tag {7}", + "image_path": "63c931bd48b4580430fba0902a4e3ad0b803d5234cfa1d62d9bf2595ef8a6f09.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 357, + 481, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 357, + 481, + 380 + ], + "spans": [ + { + "bbox": [ + 131, + 357, + 481, + 380 + ], + "type": "text", + "content": "Finally, we generate the learnable pose-embedded implicit feature " + }, + { + "bbox": [ + 131, + 357, + 481, + 380 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{nr}" + }, + { + "bbox": [ + 131, + 357, + 481, + 380 + ], + "type": "text", + "content": " for the radiance residual estimation as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 236, + 390, + 480, + 403 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 236, + 390, + 480, + 403 + ], + "spans": [ + { + "bbox": [ + 236, + 390, + 480, + 403 + ], + "type": "interline_equation", + "content": "\\mathbf {P} _ {n r} = \\operatorname {s o f t m a x} \\left(\\mathbf {P} _ {q} \\cdot \\left(\\mathbf {P} _ {k}\\right) ^ {\\top}\\right) \\cdot \\mathbf {P} _ {v}. \\tag {8}", + "image_path": "2be41111776f73132e4e53ee60380826067b30672b58d6a89f8dfc315a2aa8ec.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 412, + 481, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 412, + 481, + 460 + ], + "spans": [ + { + "bbox": [ + 131, + 412, + 481, + 460 + ], + "type": "text", + "content": "As explained earlier, we employ " + }, + { + "bbox": [ + 131, + 412, + 481, + 460 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{nr}" + }, + { + "bbox": [ + 131, + 412, + 481, + 460 + ], + "type": "text", + "content": " to condition " + }, + { + "bbox": [ + 131, + 412, + 481, + 460 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{nr}" + }, + { + "bbox": [ + 131, + 412, + 481, + 460 + ], + "type": "text", + "content": ", the MLP decoder of the non-rigid radiance branch. " + }, + { + "bbox": [ + 131, + 412, + 481, + 460 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{nr}" + }, + { + "bbox": [ + 131, + 412, + 481, + 460 + ], + "type": "text", + "content": " optimizes the implicit feature " + }, + { + "bbox": [ + 131, + 412, + 481, + 460 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{nr}" + }, + { + "bbox": [ + 131, + 412, + 481, + 460 + ], + "type": "text", + "content": " and its generation module for estimating pose-coherent radiance residuals instead of solely relying on explicit volumetric features." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 478, + 316, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 478, + 316, + 491 + ], + "spans": [ + { + "bbox": [ + 131, + 478, + 316, + 491 + ], + "type": "text", + "content": "3.3 Multi-subject Unified Training" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 498, + 482, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 594 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 594 + ], + "type": "text", + "content": "Our effective radiance compositional approach for non-rigid motion modeling further enables a flexible extension of our single-subject MoCo-NeRF for unified training with multiple subjects, which is very beneficial for practical applications. Each monocular video exhibits distinct non-rigid motions, thereby requiring individualized modeling. However, it is impossible for the traditional spatial offset learning approach [73] that typically requires a heavy module for photo-realistic quality. In contrast, our radiance compositional approach can achieve it by employing a separate, lightweight MHE for each subject, referred to as local MHE." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "Building on this foundation, we propose three architectural modifications as illustrated in Fig. 4. To benefit from unified training of multiple subjects, we introduce a coarse global MHE " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\psi_{hash}^{G}" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " shared across all subjects to learn subject-generic representations (e.g., body shapes). We also ensure learning of subject-specific details with " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "N_{s}" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " local MHEs " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\{\\psi_{hash}^{i}\\}_{i = 1}^{N_{s}}" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "N_{s}" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " is the number of subjects. For each canonical sample " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_c" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": ", we query global features from the" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "J. Kim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": "global MHE and concatenate it to the sliced feature from the corresponding local MHE for radiance predictions. We additionally propose to leverage the learnable base codes of all subjects " + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{S}_i\\}_{i=1}^{N_s}" + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": " as ID codes, for sharing MLP decoders while ensuring subject-distinctive canonical representations. The modified radiance estimations of the rigid branch for the multi-subject training are as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 205, + 182, + 481, + 198 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 182, + 481, + 198 + ], + "spans": [ + { + "bbox": [ + 205, + 182, + 481, + 198 + ], + "type": "interline_equation", + "content": "c, \\sigma = M _ {r i g i d} \\left(\\psi_ {h a s h} ^ {G} \\left(\\mathbf {x} _ {c}\\right) \\oplus \\left(\\psi_ {h a s h} ^ {i} \\left(\\mathbf {x} _ {c}\\right)\\right) ^ {: H} \\oplus \\mathbf {S} _ {i}\\right) \\tag {9}", + "image_path": "31605854c96840efc1c382b330d7e11e7723f4cab47279749ec60546c6e16189.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 212, + 263, + 224 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 212, + 263, + 224 + ], + "spans": [ + { + "bbox": [ + 131, + 212, + 263, + 224 + ], + "type": "text", + "content": "3.4 Model Optimization" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 228, + 482, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 228, + 482, + 447 + ], + "spans": [ + { + "bbox": [ + 130, + 228, + 482, + 447 + ], + "type": "text", + "content": "We optimize MoCo-NeRF with all " + }, + { + "bbox": [ + 130, + 228, + 482, + 447 + ], + "type": "inline_equation", + "content": "N_{i}" + }, + { + "bbox": [ + 130, + 228, + 482, + 447 + ], + "type": "text", + "content": " image frames " + }, + { + "bbox": [ + 130, + 228, + 482, + 447 + ], + "type": "inline_equation", + "content": "\\{I_j\\}_{j=1}^{N_i}" + }, + { + "bbox": [ + 130, + 228, + 482, + 447 + ], + "type": "text", + "content": " from the monocular video via patch-based ray sampling. Given an image frame " + }, + { + "bbox": [ + 130, + 228, + 482, + 447 + ], + "type": "inline_equation", + "content": "I_j" + }, + { + "bbox": [ + 130, + 228, + 482, + 447 + ], + "type": "text", + "content": " at each iteration, we randomly sample " + }, + { + "bbox": [ + 130, + 228, + 482, + 447 + ], + "type": "inline_equation", + "content": "N_p" + }, + { + "bbox": [ + 130, + 228, + 482, + 447 + ], + "type": "text", + "content": " patches denoted as " + }, + { + "bbox": [ + 130, + 228, + 482, + 447 + ], + "type": "inline_equation", + "content": "\\{P_k\\}_{k=1}^{N_p}" + }, + { + "bbox": [ + 130, + 228, + 482, + 447 + ], + "type": "text", + "content": " to construct the LPIPS [84] loss. The LPIPS loss term can measure the perceptual distance between two image patches, and thus we also adopt it for more perceptually coherent renderings. Our framework first renders " + }, + { + "bbox": [ + 130, + 228, + 482, + 447 + ], + "type": "inline_equation", + "content": "\\{\\hat{P}_k^r, \\hat{P}_k^f\\}_{k=1}^{N_p}" + }, + { + "bbox": [ + 130, + 228, + 482, + 447 + ], + "type": "text", + "content": " respectively from " + }, + { + "bbox": [ + 130, + 228, + 482, + 447 + ], + "type": "inline_equation", + "content": "C_{rigid}" + }, + { + "bbox": [ + 130, + 228, + 482, + 447 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 228, + 482, + 447 + ], + "type": "inline_equation", + "content": "C_{final}" + }, + { + "bbox": [ + 130, + 228, + 482, + 447 + ], + "type": "text", + "content": ", and compute two LPIPS loss terms " + }, + { + "bbox": [ + 130, + 228, + 482, + 447 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{LPIPS}}^r" + }, + { + "bbox": [ + 130, + 228, + 482, + 447 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 228, + 482, + 447 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{LPIPS}}^f" + }, + { + "bbox": [ + 130, + 228, + 482, + 447 + ], + "type": "text", + "content": " against the ground truth patch " + }, + { + "bbox": [ + 130, + 228, + 482, + 447 + ], + "type": "inline_equation", + "content": "P_k^{GT}" + }, + { + "bbox": [ + 130, + 228, + 482, + 447 + ], + "type": "text", + "content": " by using the pre-trained VGGNet [63] to as " + }, + { + "bbox": [ + 130, + 228, + 482, + 447 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{LPIPS}} = \\mathrm{LPIPS}(VGG(\\hat{P}_k), VGG(P_k^{GT}))" + }, + { + "bbox": [ + 130, + 228, + 482, + 447 + ], + "type": "text", + "content": ". We respectively compute the MSE loss terms " + }, + { + "bbox": [ + 130, + 228, + 482, + 447 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{MSE}}^r, \\mathcal{L}_{\\mathrm{MSE}}^f" + }, + { + "bbox": [ + 130, + 228, + 482, + 447 + ], + "type": "text", + "content": " from each of the estimated patches " + }, + { + "bbox": [ + 130, + 228, + 482, + 447 + ], + "type": "inline_equation", + "content": "\\{\\hat{P}_k^r, \\hat{P}_k^f\\}_{k=1}^{N_p}" + }, + { + "bbox": [ + 130, + 228, + 482, + 447 + ], + "type": "text", + "content": ". Moreover, we additionally compute an appearance matching loss with SSIM " + }, + { + "bbox": [ + 130, + 228, + 482, + 447 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{SSIM}}^f" + }, + { + "bbox": [ + 130, + 228, + 482, + 447 + ], + "type": "text", + "content": " [15] only for the final RGB " + }, + { + "bbox": [ + 130, + 228, + 482, + 447 + ], + "type": "inline_equation", + "content": "\\{\\hat{P}_k^f\\}_{k=1}^{N_p}" + }, + { + "bbox": [ + 130, + 228, + 482, + 447 + ], + "type": "text", + "content": ", to improve the rendering quality. In summary, there are 5 loss terms in total: " + }, + { + "bbox": [ + 130, + 228, + 482, + 447 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{LPIPS}}^r, \\mathcal{L}_{\\mathrm{MSE}}^r" + }, + { + "bbox": [ + 130, + 228, + 482, + 447 + ], + "type": "text", + "content": " for the canonical radiance field of the rigid branch, and " + }, + { + "bbox": [ + 130, + 228, + 482, + 447 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{LPIPS}}^f, \\mathcal{L}_{\\mathrm{MSE}}^f, \\mathcal{L}_{\\mathrm{SSIM}}^f" + }, + { + "bbox": [ + 130, + 228, + 482, + 447 + ], + "type": "text", + "content": " for the final motion-oriented compositional renderings. To enforce the rigid branch to learn the main canonical radiance fields while the non-rigid branch to learn radiance residuals, we compute the final overall loss function " + }, + { + "bbox": [ + 130, + 228, + 482, + 447 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^o" + }, + { + "bbox": [ + 130, + 228, + 482, + 447 + ], + "type": "text", + "content": " with a hyperparameter " + }, + { + "bbox": [ + 130, + 228, + 482, + 447 + ], + "type": "inline_equation", + "content": "\\lambda = 0.2" + }, + { + "bbox": [ + 130, + 228, + 482, + 447 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 259, + 454, + 480, + 466 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 259, + 454, + 480, + 466 + ], + "spans": [ + { + "bbox": [ + 259, + 454, + 480, + 466 + ], + "type": "interline_equation", + "content": "\\mathcal {L} ^ {o} = \\lambda \\mathcal {L} ^ {r} + (1 - \\lambda) \\mathcal {L} ^ {f} \\tag {10}", + "image_path": "34fd811ffa63cc74c19ad1e9694e7e397f1c3ed9ed0dc07064d7799aa9dcefd7.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 192, + 470, + 480, + 484 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 192, + 470, + 480, + 484 + ], + "spans": [ + { + "bbox": [ + 192, + 470, + 480, + 484 + ], + "type": "interline_equation", + "content": "\\mathcal {L} ^ {r} = \\mathcal {L} _ {\\text {L P I P S}} ^ {r} + \\mathcal {L} _ {\\text {M S E}} ^ {r}, \\quad \\mathcal {L} ^ {f} = \\mathcal {L} _ {\\text {L P I P S}} ^ {f} + \\mathcal {L} _ {\\text {M S E}} ^ {f} + \\mathcal {L} _ {\\text {S S I M}} ^ {f} \\tag {11}", + "image_path": "2bdd836672ac5d09a51e7d958f47ee3d528113c66a42063a0f530a1fd4aceb40.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 500, + 230, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 500, + 230, + 514 + ], + "spans": [ + { + "bbox": [ + 131, + 500, + 230, + 514 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 523, + 482, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 523, + 482, + 608 + ], + "spans": [ + { + "bbox": [ + 130, + 523, + 482, + 608 + ], + "type": "text", + "content": "We conduct extensive experiments on ZJU-MoCap [52] and MonoCap [17, 18, 50] to verify the effectiveness of our proposed approach on both single- and multi-subject settings. The multi-subject setting indicates multiple subjects are jointly optimized. Our findings demonstrate that MoCo-NeRF can learn photo-realistic dynamic human representations by effectively handling non-rigid motions. Moreover, it learns discriminative representations of each performer in the multi-subject setting, while prior works struggle." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 624, + 204, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 624, + 204, + 634 + ], + "spans": [ + { + "bbox": [ + 131, + 624, + 204, + 634 + ], + "type": "text", + "content": "4.1 Datasets" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "We leverage ZJU-MoCap and MonoCap datasets for both training and evaluation, following [14,21], and provide the details in this section. For ZJU-MoCap," + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 394, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 394, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 394, + 91, + 447, + 100 + ], + "type": "text", + "content": "MoCo-NeRF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 190, + 480, + 269 + ], + "blocks": [ + { + "bbox": [ + 130, + 125, + 482, + 190 + ], + "lines": [ + { + "bbox": [ + 130, + 125, + 482, + 190 + ], + "spans": [ + { + "bbox": [ + 130, + 125, + 482, + 190 + ], + "type": "text", + "content": "Table 1: Quantitative results of single-subject methods on ZJU-MoCap and MonoCap datasets where LPIPS* = LPIPS × 10³ with the best and second-best performance notations. MoCo-NeRF outperforms the baselines on most of the evaluation metrics in both settings. Despite pixel-wise metrics (PSNR, SSIM) tend to favor smooth renderings generated from [14,21], we achieve state-of-the-art performance while effectively modeling the photo-realistic non-rigid motions as illustrated in Fig. 5." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 190, + 480, + 269 + ], + "lines": [ + { + "bbox": [ + 135, + 190, + 480, + 269 + ], + "spans": [ + { + "bbox": [ + 135, + 190, + 480, + 269 + ], + "type": "table", + "html": "
Single-subject methodsZJU-MoCap [52]MonoCap [50]
PSNR↑SSIM↑LPIPS*↓PSNR↑SSIM↑LPIPS*↓
HumanNeRF [73]30.180.970930.1332.920.986412.93
Instant-NVR [14]30.730.969738.8731.890.987016.07
GauHuman [21]30.980.962030.1632.870.984913.54
MoCo-NeRF (Ours)31.060.973428.8333.010.987413.02
", + "image_path": "07cb5d98e02568dadb19d9bc4b071ebf6d91abccd9ae0b98321a361ceaab7b4d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 142, + 278, + 473, + 454 + ], + "blocks": [ + { + "bbox": [ + 142, + 278, + 473, + 454 + ], + "lines": [ + { + "bbox": [ + 142, + 278, + 473, + 454 + ], + "spans": [ + { + "bbox": [ + 142, + 278, + 473, + 454 + ], + "type": "image", + "image_path": "19a464fe291d78058f3c0580c246fd2007b3679c25cd5f12876f9c27dde798c5.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 142, + 454, + 473, + 460 + ], + "lines": [ + { + "bbox": [ + 142, + 454, + 473, + 460 + ], + "spans": [ + { + "bbox": [ + 142, + 454, + 473, + 460 + ], + "type": "text", + "content": "HumanNeRF Instant-NVR GauHuman Ours Ground Truth HumanNeRF Instant-NVR GauHuman Ours Ground Truth" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 130, + 469, + 482, + 514 + ], + "lines": [ + { + "bbox": [ + 130, + 469, + 482, + 514 + ], + "spans": [ + { + "bbox": [ + 130, + 469, + 482, + 514 + ], + "type": "text", + "content": "Fig.5: Qualitative comparison of novel view synthesis from single-subject training. MoCo-NeRF achieves photo-realistic rendering qualities with high-fidelity non-rigid motion modeling. Although HumanNeRF [73] presents comparable quality, however, MoCo-NeRF achieves much less training time as shown in Tab. 3." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 536, + 482, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 536, + 482, + 608 + ], + "spans": [ + { + "bbox": [ + 130, + 536, + 482, + 608 + ], + "type": "text", + "content": "we evaluate our MoCo-NeRF and other baselines on all 6 subjects (i.e., 377, 386, 387, 392, 393, 394). Out of 23 camera views, we use one of them for training and the other 22 views for evaluation. For MonoCap that comprises DynaCap [17] and DeepCap [18], we utilize all 4 subjects (i.e., lan, marc, olek, vlad) and use the same set of cameras as [14]. For multi-subject evaluations presented in Tab. 2, we evaluate on subjects 377, 386, and 392 of ZJU-MoCap." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 623, + 280, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 623, + 280, + 635 + ], + "spans": [ + { + "bbox": [ + 131, + 623, + 280, + 635 + ], + "type": "text", + "content": "4.2 Implementation Details" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "We optimize our model with Adam optimizer [30] with betas (0.9, 0.99), and learning rates of " + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "inline_equation", + "content": "5e^{-5}" + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": " for the skeletal weights volume decoder and " + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "inline_equation", + "content": "5e^{-3}" + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": " for the" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "J. Kim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 133, + 159, + 479, + 214 + ], + "blocks": [ + { + "bbox": [ + 131, + 125, + 479, + 157 + ], + "lines": [ + { + "bbox": [ + 131, + 125, + 479, + 157 + ], + "spans": [ + { + "bbox": [ + 131, + 125, + 479, + 157 + ], + "type": "text", + "content": "Table 2: Quantitative results of multi-subject methods on subjects #377, #386, #392 of ZJU-MoCap [52]. Despite the same evaluation settings, MoCo-NeRF significantly outperforms baseline methods under monocular settings with a shorter training time." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 159, + 479, + 214 + ], + "lines": [ + { + "bbox": [ + 133, + 159, + 479, + 214 + ], + "spans": [ + { + "bbox": [ + 133, + 159, + 479, + 214 + ], + "type": "table", + "html": "
Multi-subject methodsPSNR↑SSIM↑LPIPS*↓Training videosTrain time
NHP [31]26.640.918540.47multi-view~86 hours
GHuNeRF [33]27.280.929140.82monocular~123 hours
MoCo-NeRF (Ours)31.970.978124.54monocular~2 hours
", + "image_path": "87d34d7b5f95d9a56752b91ec2e7d0ad197ce457cdd0d9879c7a6b680468fcd0.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 133, + 224, + 480, + 314 + ], + "blocks": [ + { + "bbox": [ + 133, + 224, + 480, + 314 + ], + "lines": [ + { + "bbox": [ + 133, + 224, + 480, + 314 + ], + "spans": [ + { + "bbox": [ + 133, + 224, + 480, + 314 + ], + "type": "image", + "image_path": "fe8cbe06fa067da8562cf3b37049cd85ce5cb4bfd6794f43ac23813f070bd79f.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 323, + 479, + 357 + ], + "lines": [ + { + "bbox": [ + 130, + 323, + 479, + 357 + ], + "spans": [ + { + "bbox": [ + 130, + 323, + 479, + 357 + ], + "type": "text", + "content": "Fig.6: Qualitative comparison of novel view synthesis from multi-subject training. Although all subjects are seen during training, only our multi-subject MoCo-NeRF achieves photo-realistic quality comparable to the ones of single-subject models." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 379, + 480, + 451 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 379, + 480, + 451 + ], + "spans": [ + { + "bbox": [ + 130, + 379, + 480, + 451 + ], + "type": "text", + "content": "others. Each MHE has 16 levels with feature dimensions " + }, + { + "bbox": [ + 130, + 379, + 480, + 451 + ], + "type": "inline_equation", + "content": "H = 2" + }, + { + "bbox": [ + 130, + 379, + 480, + 451 + ], + "type": "text", + "content": ", and 4 levels with " + }, + { + "bbox": [ + 130, + 379, + 480, + 451 + ], + "type": "inline_equation", + "content": "H = 2" + }, + { + "bbox": [ + 130, + 379, + 480, + 451 + ], + "type": "text", + "content": " for the global MHE. For each iteration, six " + }, + { + "bbox": [ + 130, + 379, + 480, + 451 + ], + "type": "inline_equation", + "content": "32 \\times 32" + }, + { + "bbox": [ + 130, + 379, + 480, + 451 + ], + "type": "text", + "content": " patches are used with 256 samples per ray. We train all models including the baselines on 1 Nvidia RTX 3090, where both baselines are trained for their suggested duration. Our model is trained for " + }, + { + "bbox": [ + 130, + 379, + 480, + 451 + ], + "type": "inline_equation", + "content": "1 \\sim 2" + }, + { + "bbox": [ + 130, + 379, + 480, + 451 + ], + "type": "text", + "content": " hours depending on each subject's complexity for the single-subject training, and 2 hours for the multi-subject training." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 468, + 302, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 468, + 302, + 479 + ], + "spans": [ + { + "bbox": [ + 131, + 468, + 302, + 479 + ], + "type": "text", + "content": "4.3 State-of-the-art Comparison" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 485, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 485, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 485, + 481, + 666 + ], + "type": "text", + "content": "Single-subject methods. Tab. 1 presents PSNR, SSIM [72], and LPIPS* of MoCo-NeRF and our baselines [14, 21, 73], for novel-view synthesis under the single-subject setting. MoCo-NeRF achieves state-of-the-art performance on all metrics except one comparable LPIPS* for the MonoCap dataset. As shown in Fig. 5, Instant-NVR [14] shows moderate performance, however, it often generates smoothed-out results without proper non-rigid motions due to high dependency on accurate SMPL parameters. GauHuman [21] achieves better qualitative performance than [14] but still struggles to model sophisticated non-rigid motions. Fig. 1 presents longer optimizations are surprisingly ineffective and harmful respectively for Instant-NVR [14] and GauHuman [21] in terms of non-rigid motion modeling. This implies that their approaches cannot achieve optimal non-rigid motion modeling. Moreover, we observe that PSNR and SSIM often favor those results without fine-grained non-rigid motion modeling since predicting pose-dependent deformations from novel viewpoints may not be pixel-wise accurate. Instead, the perception metric (LPIPS*) evaluates better in this case, as" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 394, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 394, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 394, + 91, + 447, + 100 + ], + "type": "text", + "content": "MoCo-NeRF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 190, + 480, + 247 + ], + "blocks": [ + { + "bbox": [ + 130, + 125, + 482, + 190 + ], + "lines": [ + { + "bbox": [ + 130, + 125, + 482, + 190 + ], + "spans": [ + { + "bbox": [ + 130, + 125, + 482, + 190 + ], + "type": "text", + "content": "Table 3: Training efficiency comparison for methods with photo-realistic non-rigid motion modeling on a single RTX3090. In the table, single-sub. and multi-sub. respectively represent training for 1 and 6 subjects. MoCo-NeRF significantly outperforms HumanNeRF in terms of both GPU memory usage and training time. HumanNeRF requires sequential training to achieve comparable non-rigid motion modeling for all subjects, thus resulting in linearly-increased training time." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 190, + 480, + 247 + ], + "lines": [ + { + "bbox": [ + 135, + 190, + 480, + 247 + ], + "spans": [ + { + "bbox": [ + 135, + 190, + 480, + 247 + ], + "type": "table", + "html": "
GPU Mem. (GB)Train time (hours)
single-sub.multi-sub.Δsingle-sub.multi-sub.Δ
HumanNeRF [73]23.723.7(+0)38.9233.3(+194.5)
MoCo-NeRF (Ours)16.220.3(+4.1)1.92.1(+0.2)
", + "image_path": "47148ec027cdd0edf0acb096fc4a9498445118f999d5993767d80d7039735aff.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 269, + 480, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 269, + 480, + 341 + ], + "spans": [ + { + "bbox": [ + 130, + 269, + 480, + 341 + ], + "type": "text", + "content": "HumanNeRF [73] has a better LPIPS* score and also shows qualitatively better results. Thus, we consider GauHuman's comparable PSNR might be favored by a superior representation (3D Gaussians [29]) instead of better non-rigid motion modeling. A more detailed analysis is in supplementary materials. Despite this, we still achieve state-of-the-art performance with significantly improved training efficiency compared to another photo-realistic model [73] (Tab. 3)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 341, + 481, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 341, + 481, + 449 + ], + "spans": [ + { + "bbox": [ + 130, + 341, + 481, + 449 + ], + "type": "text", + "content": "Multi-subject methods. We evaluate our MoCo-NeRF against NHP [31] and GHuNeRF [33], both of which are designed to achieve generalization capabilities. This is necessitated by the absence of prior works that directly target the joint learning task for multiple subjects from monocular videos. Additionally, we observe that the single-subject baseline models fail to learn distinct representations when trained in a unified manner. This observation underscores our decision to compare against generalization models that are specifically tailored to leverage multi-subject data for training. For fairness, we use their official checkpoints and ensure all evaluated subjects were included for training." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 449, + 482, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 482, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 482, + 521 + ], + "type": "text", + "content": "As demonstrated in Tab. 2, MoCo-NeRF remarkably outperforms all the baseline methods on all metrics while also excelling in training efficiency up to 61 times faster. Even though MoCo-NeRF only utilizes a monocular video of each subject, Fig. 6 illustrates our method succeeds in learning both distinct representations and pose-dependent radiance residuals of each subject, enabling photo-realistic free-viewpoint renderings." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 538, + 240, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 538, + 240, + 550 + ], + "spans": [ + { + "bbox": [ + 131, + 538, + 240, + 550 + ], + "type": "text", + "content": "4.4 Model Analysis" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 558, + 482, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 604 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 604 + ], + "type": "text", + "content": "In this section, we present analyses of our contributions respectively for two major achievements: novel non-rigid motion modeling (NR modeling) and multi-subject unified training (multi-subject modeling). For both analyses, we progressively remove each component from a full model to a vanilla approach." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "NR Modeling: Analysis of radiance compositional approach. Tab. 4 demonstrates our radiance compositional design perceptually improves the rendering results by closing the radiance discrepancies with predicted residuals. Fig. 7 shows that the representation learned solely by the rigid radiance branch (c) suffers from noisy renderings caused by incorrect correspondences due to" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "J. Kim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 133, + 168, + 304, + 213 + ], + "blocks": [ + { + "bbox": [ + 132, + 125, + 305, + 168 + ], + "lines": [ + { + "bbox": [ + 132, + 125, + 305, + 168 + ], + "spans": [ + { + "bbox": [ + 132, + 125, + 305, + 168 + ], + "type": "text", + "content": "Table 4: Ablation study in the single-subject task on ZJU-MoCap for b) pose-embedded implicit features and c) motion-oriented compositional NeRF approach." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 168, + 304, + 213 + ], + "lines": [ + { + "bbox": [ + 133, + 168, + 304, + 213 + ], + "spans": [ + { + "bbox": [ + 133, + 168, + 304, + 213 + ], + "type": "table", + "html": "
Single-subjectPSNRSSIMLPIPS*
(a) MoCo-NeRF (Ours)31.060.973428.83
(b) (-) Pose features30.640.970229.65
(c) (-) Residual NeRF30.620.970229.93
", + "image_path": "7ab68aa23ca6b2977b0ee70b98fa7d3fd2bba447e33f9324c55b2c0cf9fdc9d3.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 309, + 159, + 481, + 213 + ], + "blocks": [ + { + "bbox": [ + 308, + 125, + 481, + 158 + ], + "lines": [ + { + "bbox": [ + 308, + 125, + 481, + 158 + ], + "spans": [ + { + "bbox": [ + 308, + 125, + 481, + 158 + ], + "type": "text", + "content": "Table 5: Ablation study of multi-subject MoCo-NeRF on ZJU-MoCap for our proposed architectural modifications." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 159, + 481, + 213 + ], + "lines": [ + { + "bbox": [ + 309, + 159, + 481, + 213 + ], + "spans": [ + { + "bbox": [ + 309, + 159, + 481, + 213 + ], + "type": "table", + "html": "
Multi-subjectPSNRSSIMLPIPS*
(i)MoCo-NeRF (Ours)30.910.974627.69
(ii)(-) Global MHE30.830.974628.49
(iii)(-)Ns local MHEs30.700.974229.03
(iv)(-) Identity codes28.630.966638.29
", + "image_path": "386a9798cc2ddcbdb54206ff4de74c2d49f0373489b3a76ddca1baf786629d9e.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 134, + 224, + 304, + 323 + ], + "blocks": [ + { + "bbox": [ + 134, + 224, + 304, + 323 + ], + "lines": [ + { + "bbox": [ + 134, + 224, + 304, + 323 + ], + "spans": [ + { + "bbox": [ + 134, + 224, + 304, + 323 + ], + "type": "image", + "image_path": "0ab548174b18a0c6144622ad59a44e88d1a421d2246d0e26a86202d5d3922bcb.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 331, + 305, + 364 + ], + "lines": [ + { + "bbox": [ + 131, + 331, + 305, + 364 + ], + "spans": [ + { + "bbox": [ + 131, + 331, + 305, + 364 + ], + "type": "text", + "content": "Fig.7: Qualitative results of the ablation study on single-subject MoCo-NeRF presented in Tab. 4." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 311, + 225, + 481, + 322 + ], + "blocks": [ + { + "bbox": [ + 311, + 225, + 481, + 322 + ], + "lines": [ + { + "bbox": [ + 311, + 225, + 481, + 322 + ], + "spans": [ + { + "bbox": [ + 311, + 225, + 481, + 322 + ], + "type": "image", + "image_path": "97934d8dcb9bd581b3409de810265772b7ff5832e9695aa757622325ac14d1b4.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 308, + 331, + 481, + 365 + ], + "lines": [ + { + "bbox": [ + 308, + 331, + 481, + 365 + ], + "spans": [ + { + "bbox": [ + 308, + 331, + 481, + 365 + ], + "type": "text", + "content": "Fig. 8: Qualitative results of the ablation study in Tab. 5 where we can observe gradual improvements from (iv) to (i)." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 389, + 482, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 389, + 482, + 461 + ], + "spans": [ + { + "bbox": [ + 130, + 389, + 482, + 461 + ], + "type": "text", + "content": "non-rigid motions. With our radiance residual modeling (b), those issues are mitigated by jointly learning per-sample radiance residuals, resulting in more complete rendering and better non-rigid motion modeling as clearly shown in Fig. 7. In addition, Fig. 9 illustrates successful decomposed learning of two types of representations and validates the effectiveness of our radiance compositional approach in non-rigid motion modeling." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 461, + 482, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 461, + 482, + 569 + ], + "spans": [ + { + "bbox": [ + 130, + 461, + 482, + 569 + ], + "type": "text", + "content": "NR Modeling: Analysis of pose-embedded implicit feature. Tab. 4 further illustrates that the final version with the proposed pose-embedded implicit features (Tab. 4 (a)) significantly improves the quantitative performance on all metrics. It enables the non-rigid branch to implicitly utilize body pose information in addition to the explicit features, yielding smoother and more coherent non-rigid motion modeling as shown in Fig. 7. We can observe remarkable quality enhancements that demonstrate the effectiveness of our attention-based code generation module. We also present an ablation study on the choice of cross-attention module for the feature generation in the supplementary material." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "Multi-subject Modeling: Analysis of using learnable base codes as ID codes. In Fig. 8, the results of MoCo-NeRF without any modifications (iv) exhibit mixed-up appearances on all target subjects, similar to the ones of [73] that are presented in the supplementary material. Utilizing a set of learnable base codes as ID codes (Tab. 5 (iii)) enables a single MLP decoder " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "M_{rigid}" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " to differentiate each subject by optimizing the ID codes as plug-in parameters. Thus, distinct representations of all subjects are successfully learned as illustrated in (iii) of Fig. 8, yielding significant quantitative improvements as well (Tab. 5)." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 394, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 394, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 394, + 91, + 447, + 100 + ], + "type": "text", + "content": "MoCo-NeRF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 481, + 273 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 481, + 273 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 481, + 273 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 481, + 273 + ], + "type": "image", + "image_path": "f634bdf0c938ae43e180a834ea1bcd427cad6091693d578b317015426324aa83.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 281, + 482, + 338 + ], + "lines": [ + { + "bbox": [ + 130, + 281, + 482, + 338 + ], + "spans": [ + { + "bbox": [ + 130, + 281, + 482, + 338 + ], + "type": "text", + "content": "Fig. 9: Qualitative illustrations of non-rigid radiance residuals on novel-view synthesis where " + }, + { + "bbox": [ + 130, + 281, + 482, + 338 + ], + "type": "inline_equation", + "content": "\\delta \\rightarrow c' = c + \\delta \\Delta c" + }, + { + "bbox": [ + 130, + 281, + 482, + 338 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 281, + 482, + 338 + ], + "type": "inline_equation", + "content": "\\sigma' = \\sigma + \\delta \\Delta \\sigma" + }, + { + "bbox": [ + 130, + 281, + 482, + 338 + ], + "type": "text", + "content": ". The first row progressively demonstrates the capability of the non-rigid radiance branch to learn the non-rigid motions on the subject's face in addition to the cloth folds. The last two rows clearly illustrate the effects of radiance residuals on cloth folds and hems corresponding to the body pose." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 360, + 482, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 360, + 482, + 529 + ], + "spans": [ + { + "bbox": [ + 130, + 360, + 482, + 529 + ], + "type": "text", + "content": "Multi-subject Modeling: Analysis of " + }, + { + "bbox": [ + 130, + 360, + 482, + 529 + ], + "type": "inline_equation", + "content": "N_{s}" + }, + { + "bbox": [ + 130, + 360, + 482, + 529 + ], + "type": "text", + "content": " local MHEs. Fig. 8 shows that the model (iii) lacks fine-grained details due to the sharing of one MHE. By utilizing separate " + }, + { + "bbox": [ + 130, + 360, + 482, + 529 + ], + "type": "inline_equation", + "content": "N_{s}" + }, + { + "bbox": [ + 130, + 360, + 482, + 529 + ], + "type": "text", + "content": " local MHEs (Tab. 5 (ii)), the model learns subject-discriminative non-rigid motions as they now optimize separate explicit features for each subject's radiance residuals. Fig. 8 shows clear improvement of the model (ii) in non-rigid motion modeling where its quantitative gains are reflected in Tab. 5. Multi-subject Modeling: Analysis of global MHE. The addition of global MHE (i) further empowers MoCo-NeRF to learn better representations with more coherent non-rigid motions as illustrated in Fig. 8. Although the global MHE is not used for learning non-rigid radiance residuals, it encourages the rigid part of features from local MHE to concentrate on subject-specific representations. As its frozen copy is used for the non-rigid radiance branch, it facilitates better learning of radiance residuals for modeling non-rigid motions. Tab. 5 clearly shows that it can help to learn perceptually better representations." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 545, + 220, + 558 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 545, + 220, + 558 + ], + "spans": [ + { + "bbox": [ + 132, + 545, + 220, + 558 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "In this paper, we presented a novel framework MoCo-NeRF for free-viewpoint dynamic human rendering that is capable of both single- and multi-subject training from monocular videos. By decomposing the neural representation of a dynamic human by motions and utilizing proposed implicit pose codes, MoCo-NeRF effectively learns radiance discrepancies induced by non-rigid motions. The proposed multi-subject MoCo-NeRF efficiently achieves comparable performance compared to the single-subject model. It has fully demonstrated its effectiveness and established new state-of-the-art performance on both problems." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "J. Kim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": "Acknowledgments. This research is supported in part by the Hong Kong PhD Fellowship Scheme (HKPFS), Early Career Scheme of the Research Grants Council (RGC) of the Hong Kong SAR under grant No. 26202321, SAIL Research Project, HKUST-Zeekr Collaborative Research Fund, HKUST-WeBank Joint Lab Project, Tencent Rhino-Bird Focused Research Program, and Naver Cloud Corporation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 205, + 197, + 218 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 197, + 218 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 197, + 218 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 228, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 138, + 228, + 480, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 228, + 480, + 251 + ], + "spans": [ + { + "bbox": [ + 138, + 228, + 480, + 251 + ], + "type": "text", + "content": "1. Alldieck, T., Zanfir, M., Sminchisescu, C.: Photorealistic monocular 3d reconstruction of humans wearing clothing. In: CVPR (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 252, + 480, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 252, + 480, + 284 + ], + "spans": [ + { + "bbox": [ + 138, + 252, + 480, + 284 + ], + "type": "text", + "content": "2. Barron, J.T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., Srinivasan, P.P.: Mip-NeRF: A multiscale representation for anti-aliasing neural radiance fields. In: ICCV (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 285, + 480, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 285, + 480, + 306 + ], + "spans": [ + { + "bbox": [ + 138, + 285, + 480, + 306 + ], + "type": "text", + "content": "3. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Mip-NeRF 360: Unbounded anti-aliased neural radiance fields. In: CVPR (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 307, + 480, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 307, + 480, + 327 + ], + "spans": [ + { + "bbox": [ + 138, + 307, + 480, + 327 + ], + "type": "text", + "content": "4. Chen, A., Xu, Z., Geiger, A., Yu, J., Su, H.: TensoRF: Tensorial radiance fields. In: ECCV (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 328, + 480, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 328, + 480, + 350 + ], + "spans": [ + { + "bbox": [ + 138, + 328, + 480, + 350 + ], + "type": "text", + "content": "5. Chen, A., Xu, Z., Zhao, F., Zhang, X., Xiang, F., Yu, J., Su, H.: MVSNeRF: Fast generalizable radiance field reconstruction from multi-view stereo. In: ICCV (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 350, + 480, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 350, + 480, + 371 + ], + "spans": [ + { + "bbox": [ + 138, + 350, + 480, + 371 + ], + "type": "text", + "content": "6. Chen, X., Jiang, T., Song, J., Yang, J., Black, M.J., Geiger, A., Hilliges, O.: gDNA: Towards generative detailed neural avatars. In: CVPR (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 372, + 480, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 372, + 480, + 403 + ], + "spans": [ + { + "bbox": [ + 138, + 372, + 480, + 403 + ], + "type": "text", + "content": "7. Chen, Y., Wang, X., Chen, X., Zhang, Q., Li, X., Guo, Y., Wang, J., Wang, F.: Uv volumes for real-time rendering of editable free-view human performance. In: CVPR (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 404, + 480, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 404, + 480, + 436 + ], + "spans": [ + { + "bbox": [ + 138, + 404, + 480, + 436 + ], + "type": "text", + "content": "8. Collet, A., Chuang, M., Sweeney, P., Gillett, D., Evseev, D., Calabrese, D., Hoppe, H., Kirk, A., Sullivan, S.: High-quality streamable free-viewpoint video. In: SIGGRAPH (2015)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 437, + 480, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 437, + 480, + 458 + ], + "spans": [ + { + "bbox": [ + 138, + 437, + 480, + 458 + ], + "type": "text", + "content": "9. Dong, Z., Chen, X., Yang, J., J.Black, M., Hilliges, O., Geiger, A.: AG3D: Learning to generate 3D avatars from 2D image collections. In: ICCV (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 459, + 480, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 459, + 480, + 502 + ], + "spans": [ + { + "bbox": [ + 138, + 459, + 480, + 502 + ], + "type": "text", + "content": "0. Dou, M., Khamis, S., Degtyarev, Y., Davidson, P., Fanello, S.R., Kowdle, A., Escolano, S.O., Rhemann, C., Kim, D., Taylor, J., Kohli, P., Tankovich, V., Izadi, S.: Fusion4D: Real-time performance capture of challenging scenes. In: SIGGRAPH (2016)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 502, + 480, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 502, + 480, + 524 + ], + "spans": [ + { + "bbox": [ + 138, + 502, + 480, + 524 + ], + "type": "text", + "content": "1. Du, Y., Zhang, Y., Yu, H.X., Tenenbaum, J.B., Wu, J.: Neural radiance flow for 4d view synthesis and video processing. In: ICCV (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 525, + 480, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 525, + 480, + 545 + ], + "spans": [ + { + "bbox": [ + 138, + 525, + 480, + 545 + ], + "type": "text", + "content": "2. Gao, C., Saraf, A., Kopf, J., Huang, J.B.: Dynamic view synthesis from dynamic monocular video. In: ICCV (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 546, + 480, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 546, + 480, + 567 + ], + "spans": [ + { + "bbox": [ + 138, + 546, + 480, + 567 + ], + "type": "text", + "content": "3. Garbin, S.J., Kowalski, M., Johnson, M., Shotton, J., Valentin, J.: FastNeRF: High-fidelity neural rendering at 200fps. In: ICCV (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 568, + 480, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 568, + 480, + 589 + ], + "spans": [ + { + "bbox": [ + 138, + 568, + 480, + 589 + ], + "type": "text", + "content": "4. Geng, C., Peng, S., Xu, Z., Bao, H., Zhou, X.: Learning neural volumetric representations of dynamic humans in minutes. In: CVPR (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 590, + 480, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 590, + 480, + 611 + ], + "spans": [ + { + "bbox": [ + 138, + 590, + 480, + 611 + ], + "type": "text", + "content": "5. Godard, C., Mac Aodha, O., Brostow, G.J.: Unsupervised monocular depth estimation with left-right consistency. In: CVPR (2017)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 612, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 612, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 612, + 480, + 665 + ], + "type": "text", + "content": "6. Guo, K., Lincoln, P., Davidson, P., Busch, J., Yu, X., Whalen, M., Harvey, G., Orts-Escalano, S., Pandey, R., Dourgarian, J., Tang, D., Tkach, A., Kowdle, A., Cooper, E., Dou, M., Fanello, S., Fyffe, G., Rhemann, C., Taylor, J., Debevec, P., Izadi, S.: The relightables: Volumetric performance capture of humans with realistic relighting. In: SIGGRAPH Asia (2019)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 394, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 394, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 394, + 91, + 447, + 100 + ], + "type": "text", + "content": "MoCo-NeRF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "type": "text", + "content": "17. Habermann, M., Liu, L., Xu, W., Zollhoefer, M., Pons-Moll, G., Theobalt, C.: Real-time deep dynamic characters. In: SIGGRAPH (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 480, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 480, + 161 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 480, + 161 + ], + "type": "text", + "content": "18. Habermann, M., Xu, W., Zollhoefer, M., Pons-Moll, G., Theobalt, C.: DeepCap: Monocular human performance capture using weak supervision. In: CVPR (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 162, + 480, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 162, + 480, + 183 + ], + "spans": [ + { + "bbox": [ + 133, + 162, + 480, + 183 + ], + "type": "text", + "content": "19. Hedman, P., Srinivasan, P.P., Mildenhall, B., Barron, J.T., Debevec, P.: Baking neural radiance fields for real-time view synthesis. In: ICCV (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 184, + 480, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 184, + 480, + 206 + ], + "spans": [ + { + "bbox": [ + 132, + 184, + 480, + 206 + ], + "type": "text", + "content": "20. Hong, F., Chen, Z., Lan, Y., Pan, L., Liu, Z.: EVA3D: Compositional 3d human generation from 2d image collections. In: ICLR (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 207, + 480, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 207, + 480, + 228 + ], + "spans": [ + { + "bbox": [ + 132, + 207, + 480, + 228 + ], + "type": "text", + "content": "21. Hu, S., Liu, Z.: GauHuman: Articulated gaussian splatting from monocular human videos. In: CVPR (2024)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 229, + 480, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 229, + 480, + 251 + ], + "spans": [ + { + "bbox": [ + 132, + 229, + 480, + 251 + ], + "type": "text", + "content": "22. Hu, T., Yu, T., Zheng, Z., Zhang, H., Liu, Y., Zwicker, M.: HVTR: Hybrid volumetric-textural rendering for human avatars. In: 3DV (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 251, + 480, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 251, + 480, + 284 + ], + "spans": [ + { + "bbox": [ + 132, + 251, + 480, + 284 + ], + "type": "text", + "content": "23. Huang, Y., Yi, H., Liu, W., Wang, H., Wu, B., Wang, W., Lin, B., Zhang, D., Cai, D.: One-shot implicit animatable avatars with model-based priors. In: ICCV (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 285, + 480, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 285, + 480, + 306 + ], + "spans": [ + { + "bbox": [ + 132, + 285, + 480, + 306 + ], + "type": "text", + "content": "24. Jain, A., Tancik, M., Abbeel, P.: Putting nerf on a diet: Semantically consistent few-shot view synthesis. In: ICCV (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 307, + 480, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 307, + 480, + 329 + ], + "spans": [ + { + "bbox": [ + 132, + 307, + 480, + 329 + ], + "type": "text", + "content": "25. Jiang, B., Hong, Y., Bao, H., Zhang, J.: SelfRecon: Self reconstruction your digital avatar from monocular video. In: CVPR (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 330, + 480, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 330, + 480, + 351 + ], + "spans": [ + { + "bbox": [ + 132, + 330, + 480, + 351 + ], + "type": "text", + "content": "26. Jiang, T., Chen, X., Song, J., Hilliges, O.: Instantavatar: Learning avatars from monocular video in 60 seconds. In: CVPR (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 352, + 480, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 352, + 480, + 373 + ], + "spans": [ + { + "bbox": [ + 132, + 352, + 480, + 373 + ], + "type": "text", + "content": "27. Jiang, W., Yi, K.M., Samei, G., Tuzel, O., Ranjan, A.: NeuMan: Neural human radiance field from a single video. In: ECCV (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 374, + 480, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 374, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 374, + 480, + 407 + ], + "type": "text", + "content": "28. Jiang, Y., Yao, K., Su, Z., Shen, Z., Luo, H., Xu, L.: Instant-NVR: Instant neural volumetric rendering for human-object interactions from monocular rgbd stream. In: CVPR (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 407, + 480, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 407, + 480, + 430 + ], + "spans": [ + { + "bbox": [ + 132, + 407, + 480, + 430 + ], + "type": "text", + "content": "29. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM TOG (July 2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 430, + 480, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 430, + 480, + 452 + ], + "spans": [ + { + "bbox": [ + 132, + 430, + 480, + 452 + ], + "type": "text", + "content": "30. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. In: ICLR (2015)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 453, + 480, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 453, + 480, + 475 + ], + "spans": [ + { + "bbox": [ + 132, + 453, + 480, + 475 + ], + "type": "text", + "content": "31. Kwon, Y., Kim, D., Ceylan, D., Fuchs, H.: Neural human performer: Learning generalizable radiance fields for human performance rendering. In: NeurIPS (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 475, + 480, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 475, + 480, + 497 + ], + "spans": [ + { + "bbox": [ + 132, + 475, + 480, + 497 + ], + "type": "text", + "content": "32. Lewis, J.P., Cordner, M., Fong, N.: Pose space deformation: A unified approach to shape interpolation and skeleton-driven deformation. In: SIGGRAPH (2000)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 498, + 480, + 519 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 498, + 480, + 519 + ], + "spans": [ + { + "bbox": [ + 132, + 498, + 480, + 519 + ], + "type": "text", + "content": "33. Li, C., Lin, J., Lee, G.H.: Generalizable human nerf from a monocular video. In: 3DV (2024)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 520, + 480, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 520, + 480, + 542 + ], + "spans": [ + { + "bbox": [ + 132, + 520, + 480, + 542 + ], + "type": "text", + "content": "34. Li, R., Tanke, J., Vo, M., Zollhofer, M., Gall, J., Kanazawa, A., Lassner, C.: TAVA: Template-free animatable volumetric actors. In: ECCV (2022)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 543, + 480, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 543, + 480, + 564 + ], + "spans": [ + { + "bbox": [ + 132, + 543, + 480, + 564 + ], + "type": "text", + "content": "35. Li, Z., Niklaus, S., Snavely, N., Wang, O.: Neural scene flow fields for space-time view synthesis of dynamic scenes. In: CVPR (2021)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 565, + 480, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 565, + 480, + 586 + ], + "spans": [ + { + "bbox": [ + 132, + 565, + 480, + 586 + ], + "type": "text", + "content": "36. Liu, L., Gu, J., Lin, K.Z., Chua, T.S., Theobalt, C.: Neural sparse voxel fields. In: NeurIPS (2020)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 587, + 480, + 619 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 587, + 480, + 619 + ], + "spans": [ + { + "bbox": [ + 132, + 587, + 480, + 619 + ], + "type": "text", + "content": "37. Liu, L., Habermann, M., Rudnev, V., Sarkar, K., Gu, J., Theobalt, C.: Neural actor: Neural free-view synthesis of human actors with pose control. In: SIGGRAPH Asia (2021)" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 132, + 620, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 620, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 620, + 480, + 643 + ], + "type": "text", + "content": "38. Liu, Y., Peng, S., Liu, L., Wang, Q., Wang, P., Christian, T., Zhou, X., Wang, W.: Neural rays for occlusion-aware image-based rendering. In: CVPR (2022)" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "text", + "content": "39. Loper, M., Mahmood, N., Romero, J., Pons-Moll, G., Black, M.J.: SMPL: A skinned multi-person linear model. In: SIGGRAPH Asia (2015)" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "J. Kim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 161 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 161 + ], + "type": "text", + "content": "40. Martin-Brualla, R., Pandey, R., Yang, S., Pidlypenskyi, P., Taylor, J., Valentin, J., Khamis, S., Davidson, P., Tkach, A., Lincoln, P., Kowdle, A., Rhemann, C., Goldman, D.B., Keskin, C., Seitz, S., Izadi, S., Fanello, S.: LookinGood: Enhancing performance capture with real-time neural re-rendering. ACM TOG (2018)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 161, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 161, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 161, + 481, + 194 + ], + "type": "text", + "content": "41. Mihajlovic, M., Bansal, A., Zollhoefer, M., Tang, S., Saito, S.: KeypointNeRF: Generalizing image-based volumetric avatars using relative spatial encoding of keypoints. In: ECCV (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 194, + 481, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 481, + 224 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 481, + 224 + ], + "type": "text", + "content": "42. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: NeRF: Representing scenes as neural radiance fields for view synthesis. In: ECCV (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 225, + 481, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 225, + 481, + 247 + ], + "spans": [ + { + "bbox": [ + 132, + 225, + 481, + 247 + ], + "type": "text", + "content": "43. Mu, J., Sang, S., Vasconcelos, N., Wang, X.: ActorsNeRF: Animatable few-shot human rendering with generalizable nerfs. In: ICCV (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 247, + 481, + 268 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 247, + 481, + 268 + ], + "spans": [ + { + "bbox": [ + 132, + 247, + 481, + 268 + ], + "type": "text", + "content": "44. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. In: SIGGRAPH (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 269, + 481, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 269, + 481, + 289 + ], + "spans": [ + { + "bbox": [ + 132, + 269, + 481, + 289 + ], + "type": "text", + "content": "45. Noguchi, A., Sun, X., Lin, S., Harada, T.: Unsupervised learning of efficient geometry-aware neural articulated representations. In: ECCV (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 289, + 481, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 289, + 481, + 311 + ], + "spans": [ + { + "bbox": [ + 132, + 289, + 481, + 311 + ], + "type": "text", + "content": "46. Oechsle, M., Peng, S., Geiger, A.: UNISURF: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In: ICCV (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 312, + 481, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 312, + 481, + 332 + ], + "spans": [ + { + "bbox": [ + 132, + 312, + 481, + 332 + ], + "type": "text", + "content": "47. Park, K., Sinha, U., Barron, J.T., Bouaziz, S., Goldman, D.B., Seitz, S.M., Martin-Brualla, R.: Nerfies: Deformable neural radiance fields. In: ICCV (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 333, + 481, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 333, + 481, + 365 + ], + "spans": [ + { + "bbox": [ + 132, + 333, + 481, + 365 + ], + "type": "text", + "content": "48. Park, K., Sinha, U., Hedman, P., Barron, J.T., Bouaziz, S., Goldman, D.B., Martin-Brualla, R., Seitz, S.M.: HyperNeRF: A higher-dimensional representation for topologically varying neural radiance fields. In: SIGGRAPH Asia (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 365, + 481, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 365, + 481, + 386 + ], + "spans": [ + { + "bbox": [ + 132, + 365, + 481, + 386 + ], + "type": "text", + "content": "49. Peng, B., Hu, J., Zhou, J., Zhang, J.: SelfNeRF: Fast training nerf for human from monocular self-rotating video. arXiv preprint arXiv:2210.01651 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 386, + 481, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 386, + 481, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 386, + 481, + 407 + ], + "type": "text", + "content": "50. Peng, S., Dong, J., Wang, Q., Zhang, S., Shuai, Q., Zhou, X., Bao, H.: Animatable neural radiance fields for modeling dynamic human bodies. In: ICCV (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 408, + 481, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 408, + 481, + 439 + ], + "spans": [ + { + "bbox": [ + 132, + 408, + 481, + 439 + ], + "type": "text", + "content": "51. Peng, S., Xu, Z., Dong, J., Wang, Q., Zhang, S., Shuai, Q., Bao, H., Zhou, X.: Animatable implicit neural representations for creating realistic avatars from videos. TPAMI (2024)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 440, + 481, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 440, + 481, + 472 + ], + "spans": [ + { + "bbox": [ + 132, + 440, + 481, + 472 + ], + "type": "text", + "content": "52. Peng, S., Zhang, Y., Xu, Y., Wang, Q., Shuai, Q., Bao, H., Zhou, X.: Neural Body: Implicit neural representations with structured latent codes for novel view synthesis of dynamic humans. In: CVPR (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 472, + 481, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 472, + 481, + 493 + ], + "spans": [ + { + "bbox": [ + 132, + 472, + 481, + 493 + ], + "type": "text", + "content": "53. Pumarola, A., Corona, E., Pons-Moll, G., Moreno-Noguer, F.: D-NeRF: Neural Radiance Fields for Dynamic Scenes. In: CVPR (2020)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 494, + 481, + 515 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 494, + 481, + 515 + ], + "spans": [ + { + "bbox": [ + 132, + 494, + 481, + 515 + ], + "type": "text", + "content": "54. Reiser, C., Peng, S., Liao, Y., Geiger, A.: KiloNeRF: Speeding up neural radiance fields with thousands of tiny mlps. In: ICCV (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 515, + 481, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 515, + 481, + 536 + ], + "spans": [ + { + "bbox": [ + 132, + 515, + 481, + 536 + ], + "type": "text", + "content": "55. Rematas, K., Liu, A., Srinivasan, P.P., Barron, J.T., Tagliasacchi, A., Funkhouser, T., Ferrari, V.: Urban radiance fields. In: CVPR (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 536, + 481, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 536, + 481, + 568 + ], + "spans": [ + { + "bbox": [ + 132, + 536, + 481, + 568 + ], + "type": "text", + "content": "56. Remelli, E., Bagautdinov, T., Saito, S., Wu, C., Simon, T., Wei, S.E., Guo, K., Cao, Z., Prada, F., Saragih, J., et al.: Drivable volumetric avatars using texel-aligned features. In: SIGGRAPH (2022)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 569, + 481, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 569, + 481, + 601 + ], + "spans": [ + { + "bbox": [ + 132, + 569, + 481, + 601 + ], + "type": "text", + "content": "57. Saito, S., Huang, Z., Natsume, R., Morishima, S., Kanazawa, A., Li, H.: PIFu: Pixel-aligned implicit function for high-resolution clothed human digitization. In: ICCV (2019)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 601, + 481, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 481, + 622 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 481, + 622 + ], + "type": "text", + "content": "58. Saito, S., Simon, T., Saragih, J., Joo, H.: PIFuHD: Multi-level pixel-aligned implicit function for high-resolution 3d human digitization. In: CVPR (2020)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 622, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 622, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 622, + 481, + 643 + ], + "type": "text", + "content": "59. Sara Fridovich-Keil and Alex Yu, Tancik, M., Chen, Q., Recht, B., Kanazawa, A.: Plenoxels: Radiance fields without neural networks. In: CVPR (2022)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 644, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 644, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 644, + 481, + 665 + ], + "type": "text", + "content": "60. Schonberger, J.L., Zheng, E., Frahm, J.M., Pollefeys, M.: Pixelwise view selection for unstructured multi-view stereo. In: ECCV (2016)" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 394, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 394, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 394, + 91, + 447, + 100 + ], + "type": "text", + "content": "MoCo-NeRF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 127 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 127 + ], + "type": "text", + "content": "61. Schonberger, J.L., Frahm, J.M.: Structure-from-motion revisited. In: CVPR (2016)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 128, + 480, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 128, + 480, + 160 + ], + "spans": [ + { + "bbox": [ + 133, + 128, + 480, + 160 + ], + "type": "text", + "content": "62. Shao, R., Zhang, H., Zhang, H., Chen, M., Cao, Y., Yu, T., Liu, Y.: DoubleField: Bridging the neural surface and radiance fields for high-fidelity human reconstruction and rendering. In: CVPR (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 161, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 161, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 132, + 161, + 480, + 182 + ], + "type": "text", + "content": "63. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition. In: ICLR (2015)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 183, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 480, + 205 + ], + "type": "text", + "content": "64. Su, S.Y., Yu, F., Zollhöfer, M., Rhodin, H.: A-NeRF: Articulated neural radiance fields for learning human shape, appearance, and pose. In: NeurIPS (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 205, + 480, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 480, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 480, + 226 + ], + "type": "text", + "content": "65. Su, Z., Xu, L., Zheng, Z., Yu, T., Liu, Y., Fang, L.: RobustFusion: Human volumetric capture with data-driven visual cues using a rgbd camera. In: ECCV (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 227, + 480, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 480, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 480, + 258 + ], + "type": "text", + "content": "66. Tancik, M., Casser, V., Yan, X., Pradhan, S., Mildenhall, B., Srinivasan, P., Barron, J.T., Kretzschmar, H.: Block-NeRF: Scalable large scene neural view synthesis. In: CVPR (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 259, + 480, + 292 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 480, + 292 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 480, + 292 + ], + "type": "text", + "content": "67. Tretschk, E., Tewari, A., Golyanik, V., Zollhöfer, M., Lassner, C., Theobalt, C.: Non-rigid neural radiance fields: Reconstruction and novel view synthesis of a dynamic scene from monocular video. In: ICCV (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 293, + 480, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 480, + 314 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 480, + 314 + ], + "type": "text", + "content": "68. Turki, H., Ramanan, D., Satyanarayanan, M.: Mega-NERF: Scalable construction of large-scale nerfs for virtual fly-throughs. In: CVPR (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 315, + 480, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 480, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 480, + 346 + ], + "type": "text", + "content": "69. Wang, P., Liu, L., Liu, Y., Theobalt, C., Komura, T., Wang, W.: NeuS: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. In: NeurIPS (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 347, + 480, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 480, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 480, + 380 + ], + "type": "text", + "content": "70. Wang, Q., Wang, Z., Genova, K., Srinivasan, P., Zhou, H., Barron, J.T., Martin-Brualla, R., Snavely, N., Funkhouser, T.: IBRNet: Learning multi-view image-based rendering. In: CVPR (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 380, + 480, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 480, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 480, + 402 + ], + "type": "text", + "content": "71. Wang, S., Schwarz, K., Geiger, A., Tang, S.: ARAH: Animatable volume rendering of articulated human sdfs. In: ECCV (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 403, + 480, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 480, + 424 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 480, + 424 + ], + "type": "text", + "content": "72. Wang, Z., Bovik, A., Sheikh, H., Simoncelli, E.: Image quality assessment: from error visibility to structural similarity. TIP (2004)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 425, + 480, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 425, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 425, + 480, + 456 + ], + "type": "text", + "content": "73. Weng, C.Y., Curless, B., Srinivasan, P.P., Barron, J.T., Kemelmacher-Shlizerman, I.: HumanNeRF: Free-viewpoint rendering of moving people from monocular video. In: CVPR (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 457, + 480, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 480, + 479 + ], + "type": "text", + "content": "74. Weng, C.Y., Srinivasan, P.P., Curless, B., Kemelmacher-Shlizerman, I.: Person-NeRF: Personalized reconstruction from photo collections. In: CVPR (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 479, + 480, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 480, + 500 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 480, + 500 + ], + "type": "text", + "content": "75. Wu, M., Wang, Y., Hu, Q., Yu, J.: Multi-view neural human rendering. In: CVPR (2020)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 501, + 480, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 480, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 480, + 522 + ], + "type": "text", + "content": "76. Xian, W., Huang, J.B., Kopf, J., Kim, C.: Space-time neural irradiance fields for free-viewpoint video. In: CVPR (2021)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 523, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 523, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 523, + 480, + 555 + ], + "type": "text", + "content": "77. Xiangli, Y., Xu, L., Pan, X., Zhao, N., Rao, A., Theobalt, C., Dai, B., Lin, D.: Bungeenerf: Progressive neural radiance field for extreme multi-scale scene rendering. In: ECCV (2022)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 555, + 480, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 480, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 480, + 578 + ], + "type": "text", + "content": "78. Xu, H., Alldieck, T., Sminchisescu, C.: H-NeRF: Neural radiance fields for rendering and temporal reconstruction of humans in motion. In: NeurIPS (2021)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 578, + 480, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 480, + 599 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 480, + 599 + ], + "type": "text", + "content": "79. Xu, L., Xiangli, Y., Peng, S., Pan, X., Zhao, N., Theobalt, C., Dai, B., Lin, D.: Grid-guided neural radiance fields for large urban scenes. In: CVPR (2023)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 600, + 480, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 480, + 621 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 480, + 621 + ], + "type": "text", + "content": "80. Xu, Z., Peng, S., Lin, H., He, G., Sun, J., Shen, Y., Bao, H., Zhou, X.: 4K4D: Real-time 4d view synthesis at 4k resolution. In: CVPR (2024)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 622, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 622, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 622, + 480, + 643 + ], + "type": "text", + "content": "81. Yariv, L., Gu, J., Kasten, Y., Lipman, Y.: Volume rendering of neural implicit surfaces. In: NeurIPS (2021)" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "text", + "content": "82. Yu, A., Li, R., Tancik, M., Li, H., Ng, R., Kanazawa, A.: PlenOctrees for real-time rendering of neural radiance fields. In: ICCV (2021)" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "J. Kim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 270 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "type": "text", + "content": "83. Yu, A., Ye, V., Tancik, M., Kanazawa, A.: pixelNeRF: Neural radiance fields from one or few images. In: CVPR (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 138, + 482, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 138, + 482, + 160 + ], + "spans": [ + { + "bbox": [ + 130, + 138, + 482, + 160 + ], + "type": "text", + "content": "84. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: CVPR (2018)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 160, + 482, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 160, + 482, + 182 + ], + "spans": [ + { + "bbox": [ + 130, + 160, + 482, + 182 + ], + "type": "text", + "content": "85. Zhang, R., Chen, J.: NDF: Neural deformable fields for dynamic human modelling. In: ECCV (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 182, + 482, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 182, + 482, + 205 + ], + "spans": [ + { + "bbox": [ + 130, + 182, + 482, + 205 + ], + "type": "text", + "content": "86. Zhao, F., Yang, W., Zhang, J., Lin, P., Zhang, Y., Yu, J., Xu, L.: HumanNeRF: Efficiently generated human radiance field from sparse inputs. In: CVPR (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 205, + 482, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 205, + 482, + 226 + ], + "spans": [ + { + "bbox": [ + 130, + 205, + 482, + 226 + ], + "type": "text", + "content": "87. Zheng, Z., Huang, H., Yu, T., Zhang, H., Guo, Y., Liu, Y.: Structured local radiance fields for human avatar modeling. In: CVPR (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 226, + 482, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 226, + 482, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 226, + 482, + 248 + ], + "type": "text", + "content": "88. Zheng, Z., Zhao, X., Zhang, H., Liu, B., Liu, Y.: AvatarRex: Real-time expressive full-body avatars. In: SIGGRAPH (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 248, + 482, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 248, + 482, + 270 + ], + "spans": [ + { + "bbox": [ + 130, + 248, + 482, + 270 + ], + "type": "text", + "content": "89. Zhi, Y., Qian, S., Yan, X., Gao, S.: Dual-space nerf: Learning animatable avatars and scene lighting in separate spaces. In: 3DV (2022)" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 394, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 394, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 394, + 91, + 447, + 100 + ], + "type": "text", + "content": "MoCo-NeRF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Motion-prior Contrast Maximization for Dense Continuous-Time Motion Estimation/34d0d1c3-d92d-4348-bbb4-52308320fede_content_list.json b/2024/Motion-prior Contrast Maximization for Dense Continuous-Time Motion Estimation/34d0d1c3-d92d-4348-bbb4-52308320fede_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..5ec970ae9498d9c40142c41b4589288f005edfb1 --- /dev/null +++ b/2024/Motion-prior Contrast Maximization for Dense Continuous-Time Motion Estimation/34d0d1c3-d92d-4348-bbb4-52308320fede_content_list.json @@ -0,0 +1,1812 @@ +[ + { + "type": "text", + "text": "Motion-prior Contrast Maximization for Dense Continuous-Time Motion Estimation", + "text_level": 1, + "bbox": [ + 130, + 48, + 866, + 104 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Friedhelm Hamann1, Ziyun Wang2, Ioannis Asmanis2, Kenneth Chaney2, Guillermo Gallego1,3, and Kostas Daniilidis2,4", + "bbox": [ + 133, + 139, + 864, + 182 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ TU Berlin and SCIolI Excellence Cluster, Berlin, Germany \n $^{2}$ University of Pennsylvania, Philadelphia, US", + "bbox": [ + 198, + 195, + 794, + 232 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 Einstein Center Digital Future and Robotics Institute Germany, Berlin, Germany \n4 Archimedes, Athena RC, Greece", + "bbox": [ + 84, + 232, + 910, + 266 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Current optical flow and point-tracking methods rely heavily on synthetic datasets. Event cameras are novel vision sensors with advantages in challenging visual conditions, but state-of-the-art frame-based methods cannot be easily adapted to event data due to the limitations of current event simulators. We introduce a novel self-supervised loss combining the Contrast Maximization framework with a non-linear motion prior in the form of pixel-level trajectories and propose an efficient solution to solve the high-dimensional assignment problem between non-linear trajectories and events. Their effectiveness is demonstrated in two scenarios: In dense continuous-time motion estimation, our method improves the zero-shot performance of a synthetically trained model on the real-world dataset EVIMO2 by $29\\%$ . In optical flow estimation, our method elevates a simple UNet to achieve state-of-the-art performance among self-supervised methods on the DSEC optical flow benchmark. Our code is available at https://github.com/tub-rip/MotionPriorCMax.", + "bbox": [ + 142, + 323, + 854, + 593 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 77, + 630, + 309, + 648 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Determining the motion of arbitrary projected world points on the image plane over long time intervals is a difficult low-level computer vision problem. Researchers have studied it as optical flow and lately as point-tracking, with many practical applications in robotics, computational photography, video compression, and object-level tracking [26]. The best-performing frame-based methods for optical flow estimation and point tracking use large-scale synthetic datasets. Synthetic data is diverse and has highly accurate ground truth (GT), but is unrealistic and methods trained on synthetic data show a sim-to-real gap.", + "bbox": [ + 72, + 674, + 922, + 830 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Novel vision sensors called event cameras [16,17,39] have emerged as promising alternatives to take on the problem. Inspired by the transient visual pathway, which is responsible for motion perception, they are particularly fit for capturing scene motion in the form of asynchronous pixel-wise intensity changes. This working principle endows them with advantages, such as high speed, high dynamic range (HDR) and low power consumption.", + "bbox": [ + 72, + 832, + 922, + 949 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ad075927e22d4427da496a0fe9385f4e82a447e7725d122fa1ec6d2c8fe85a49.jpg", + "image_caption": [ + "a) Contrast Maximization with non-linear Trajectories" + ], + "image_footnote": [], + "bbox": [ + 164, + 74, + 409, + 178 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/d638bfaa03eae5a1b184f8a61befaab34ee71241f48f2df2c3d0192aca61e1ea.jpg", + "image_caption": [ + "b) Overview Method" + ], + "image_footnote": [], + "bbox": [ + 515, + 69, + 903, + 183 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/cfeaf3a68064fe3946cc8d75b9e32aa7998fc0aa1ee96b69fef645e7c0d3a039.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 104, + 217, + 227, + 279 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/bf7fe060a130b599dab3cda7f4387aaecc61c3e23db6b9d5519899c052f77b54.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 230, + 217, + 355, + 279 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/0501c97cc26b68f2996af6e510a69a55123e58ba6d3ced7a145af22b0bc68f36.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 355, + 217, + 481, + 279 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/9f2bed5b336b044fd57819770283e1c5d2b689e4a3744ec0883bd4a41d948e52.jpg", + "image_caption": [ + "c) Application to Trajectory Estimation" + ], + "image_footnote": [], + "bbox": [ + 104, + 279, + 227, + 341 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/a20f3b923ecfeb1822e64d8465bd12fc9a11527a90ac13c3a7adddd36ef984d4.jpg", + "image_caption": [ + "Fig.1: Summary. a) We present an approach to combine Contrast Maximization with dense non-linear trajectories. b) We show how it can be used for self-supervised learning in a pipeline to predict dense point trajectories, and c) evaluate it on the EVIMO2 dataset, for which we generate dense point tracks. d) Additionally, our approach provides state-of-the-art performance on self-supervised optical flow prediction." + ], + "image_footnote": [], + "bbox": [ + 230, + 279, + 355, + 341 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/d53aca7b0ce0b63c52044b1ab88d10fd1de4c853c12c1383a1978fe5ef6ed691.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 355, + 279, + 481, + 341 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/b3684fcdca0d6303e409a626861c564e772277f4431be555e5fb81235c4b8979.jpg", + "image_caption": [ + "d) Application to Optical Flow" + ], + "image_footnote": [], + "bbox": [ + 535, + 206, + 881, + 347 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Event-based motion estimation methods can be categorized according to the complexity of the motion considered (i.e., number of degrees-of-freedom (DOF)) and to the solution strategy: model-based or learning-based. Low-DOF motions arise in sparse feature tracking and ego-motion estimation, while high-DOF motions describe more complex scenes, e.g., via per-pixel displacement (i.e., densely, over the whole image plane). Focusing on the latter, the event-based optical flow problem has been extensively studied on mobile robotic datasets [23, 75], where GT is calculated as the motion field (from known depth and poses). Most approaches use this GT for direct supervision of dense flow prediction [22, 24, 40, 57, 65]. Alternatively, several works use a contrast loss [18, 19, 77], which allows training in a self-supervised manner. However, the contrast loss is prone to undesired local optima, called event collapse [52, 53], where many events are warped into a few pixels or lines. Evaluation of event-based optical flow has been mostly done on the MVSEC [75] and DSEC [23] datasets, which comprise largely uniform motions and test intervals from 0.022s to 0.1s. Recently, [25] has proposed a supervised method to predict pixel displacement over a larger duration (0.5s), by leveraging synthetic data. It relies on the generation of training event data from images, where current tools [21] are not as mature as frame-based simulation, therefore it suffers from a sim-to-real performance gap.", + "bbox": [ + 75, + 518, + 917, + 887 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In short, progress has been made and the research field is moving towards predicting motion over longer time intervals on the whole image plane, i.e., taking on more complex (i.e., nonlinear) motion problems. This is precisely the", + "bbox": [ + 75, + 889, + 920, + 949 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 79, + 16, + 94, + 29 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "F. Hamann et al.", + "bbox": [ + 409, + 14, + 578, + 30 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "problem tackled in this paper (Fig. 1): long-time and dense event-based motion estimation, reducing the domain adaptation gap of previous approaches. It comes with several associated challenges, mainly overcoming the lack of large labeled datasets, and dealing with event noise and data association (events depend on motion, and for large motions, the appearance of \"corresponding\" events can be wildly different due to changes in motion direction, occlusions, etc.) while leveraging the space-time characteristics of event data.", + "bbox": [ + 75, + 55, + 920, + 191 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To this end, we propose tackling the problem in two stages, by leveraging both supervised and self-supervised strategies: first, using supervised learning on synthetic data to provide initial model weights for motion estimation, and secondly, fine-tuning the network on real data via a self-supervised loss to reduce the domain adaptation gap. Our technical contributions involve extending the contrast loss framework to regress continuous-time trajectories over long time intervals via motion priors (parametric functions that provide a good balance between motion generality and regularization [64, 66]). This includes a solution to accurately and efficiently associate events to the trajectories (Fig. 2).", + "bbox": [ + 75, + 193, + 920, + 367 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "More specifically, finding the association between events and motion trajectories (to warp corresponding events and achieve event alignment) is a high-dimensional problem (e.g., for a time window of 0.3 s one can consider about ten million events, and as many trajectories as pixels), which needs to be implemented in a differentiable and parallelizable manner. We propose two actions to cope with these technical challenges. First, we relax the problem by interpolating over a coarse spatio-temporal displacement field, which serves as a lookup table. Secondly, we use a symbolic matrix framework [15] to calculate the $K = N_{\\mathrm{traj}}$ nearest neighbor (KNN) trajectories for each grid point, thereby solving KNN in a memory-efficient and differentiable way on GPU. The warp displacement at the grid point is then set to the average of the neighboring trajectories.", + "bbox": [ + 75, + 369, + 920, + 582 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our approach is versatile, allowing for different types of networks and trajectories. Hence, we evaluate its performance on two applications: dense continuous-time motion estimation and optical flow. The results on EVIMO2 [7] show that fine-tuning with our self-supervised loss improves the zero-shot performance of a model pre-trained on synthetic data by $29\\%$ . On DSEC, our model shows state-of-the-art performance among self-supervised methods, on average improving the angular error by $19\\%$ and the percentage of inliers by $14\\%$ , while having a $5\\times$ faster inference time. In summary, our contributions are (Fig. 1):", + "bbox": [ + 75, + 583, + 920, + 739 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. We introduce motion priors (parametric functions with a good balance between generality and regularization) in the event-based contrast maximization framework for continuous-time and dense motion estimation.", + "2. We combine the self-supervised loss with Bflow [25], the current top-performing supervised model for dense continuous-time event-based motion estimation trained on synthetic data, and show that it can be used to improve over zero-shot performance on unseen real data (EVIMO2).", + "3. We show that the combination of our loss with a simple U-Net architecture (à la EV-FlowNet [76]) achieves state-of-the-art performance among the self-supervised methods on the DSEC Optical Flow benchmark." + ], + "bbox": [ + 84, + 752, + 937, + 946 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Motion-prior Contrast Maximization", + "bbox": [ + 474, + 14, + 837, + 32 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 903, + 16, + 920, + 30 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 79, + 53, + 329, + 74 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Frame-based Motion Estimation. Progress in deep learning triggered a large series of learned methods for optical flow estimation in classical, frame-based computer vision [2, 14, 32, 33, 60, 62]. The solutions rely on large-scale synthetic datasets [14, 42, 59, 73]. Similarly related is the task of point tracking, which has shown impressive progress on frame-based data [13, 31, 36, 50, 74], equally relying on simulated data [8, 12, 74]. This approach is unsatisfying, as it is prone to out-of-distribution (OOD) problems that cannot be easily overcome. As an alternative, self-supervised methods have been explored [34, 35, 48, 72] and shown to improve models pre-trained on simulated data for motion estimation tasks [58, 61].", + "bbox": [ + 79, + 136, + 917, + 312 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Event-based Optical Flow. Event cameras [39, 46] are a relatively new technology, and have found applications in various computer vision domains, like mobile robotics [27, 45, 68], scene understanding [29, 30], and computational imaging [55, 63]. Their exploration for low-level vision tasks has taken a similar trajectory in a compressed timeline as previously frame-based methods [17]. The first event-based optical flow methods were model-based [3, 4, 6, 41, 43, 54], followed by learning-based approaches [11, 24, 28, 38, 44, 76, 77].", + "bbox": [ + 79, + 320, + 917, + 456 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The scarce availability of event data and less mature simulation technology compared to frame-based cameras are major obstacles [21]. Therefore, event-based optical flow has been mostly evaluated on data acquired through egomotion, such as the robotics dataset MVSEC [75], the driving dataset DSEC [23] or the M3ED dataset [9]. However, GT is calculated using the motion field equation (with data from a synchronized depth sensor) and therefore does not provide an accurate flow at the event rate, nor spatially at occlusions and independently moving objects (IMOs). This GT is used to train supervised learning approaches [22, 24, 40, 57, 65], which inherit limitations from the ground truth.", + "bbox": [ + 79, + 463, + 917, + 639 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Self-supervised methods for event-based optical flow lessen the dependency on GT labels by leveraging an event alignment error [28,54,70,77]. Notably, variations of the contrast loss have been proposed [18], however, so far the application has been limited to either low-DOF problems (e.g., feature tracking [10, 19, 51], ego-motion [19, 20, 37]) or high-DOF but short-time optical flow problem [44, 54] (max. 0.1 s). We expand the frontier to the challenging problem of long-time and high-DOF (complex) motion estimation, by exploiting trajectory priors.", + "bbox": [ + 79, + 647, + 917, + 783 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Trajectory Prior. Using trajectories as motion priors has been a widely explored scheme in computer vision. Specifically, [1, 64, 78] propose a linear combination of basis functions for structure from motion. More recently it has been explored for dynamic novel view synthesis [66]. In the context of events, [63] uses cubic motion splines for video frame interpolation, [25] uses B-splines for supervised learning of non-linear optical flow, [67] use learned basis functions for event-based video decompression, and [10, 51] show event-based feature tracking (low-DOF) using Bezier or B-spline curves in combination with a contrast loss.", + "bbox": [ + 79, + 791, + 917, + 947 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 79, + 17, + 94, + 29 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "F. Hamann et al.", + "bbox": [ + 411, + 16, + 578, + 30 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/67d6593e4476ffb8773fdbff0d1dee02a851723bc05e26dadce9f3780c710844.jpg", + "image_caption": [ + "Fig. 2: Pipeline overview. (a) Input events in a time interval are (b) voxelized and (c) passed to an artificial neural network that predicts per-pixel coefficients for continuous-time trajectories (d). The raw events and predicted trajectories are fed to the loss module (e). Here, a dense spatio-temporal displacement map is interpolated, and events are warped according to their looked-up displacement. Lastly, an image of warped events (IWE) is built at a random reference time and its gradient magnitude acts as training loss. Note that the prediction method displayed here is specific to the used Bflow backbone [25], with additional events before the prediction start time $t_s$ as input." + ], + "image_footnote": [], + "bbox": [ + 79, + 50, + 922, + 354 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Methodology", + "text_level": 1, + "bbox": [ + 75, + 541, + 319, + 565 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Contrast Loss", + "text_level": 1, + "bbox": [ + 75, + 578, + 314, + 596 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Contrast Maximization (CM) [18, 19] is used for self-supervised training. The idea is to find the point trajectories on the image plane best aligning with the events by maximizing the sharpness of warped events. It is an iterative approach, with three steps per iteration: (i) Each event $e_k = (\\mathbf{x}_k, t_k, p_k)$ contains the pixel coordinates $\\mathbf{x}_k = (x_k, y_k)^\\top$ , timestamp $t_k$ and polarity $p_k$ of a brightness change of predefined size $C$ (contrast sensitivity). Events $\\mathcal{E} = \\{e_k\\}_{k=1}^{N_e}$ are displaced according to a candidate motion hypothesis $\\mathbf{x}_k' = \\mathbf{W}(\\mathbf{x}_k, t_k; \\boldsymbol{\\theta})$ , with parameters $\\boldsymbol{\\theta}$ , to a reference time $t_\\mathrm{ref}$ , producing a set of warped events $\\mathcal{E}_{t_\\mathrm{ref}}' = \\{e_k'\\}_{k=1}^{N_e}$ :", + "bbox": [ + 72, + 604, + 922, + 767 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ne _ {k} \\doteq \\left(\\mathbf {x} _ {k}, t _ {k}, p _ {k}\\right) \\stackrel {\\mathrm {W}} {\\mapsto} e _ {k} ^ {\\prime} \\doteq \\left(\\mathbf {x} _ {k} ^ {\\prime}, t _ {\\text {r e f}}, p _ {k}\\right). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 285, + 773, + 917, + 801 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Afterwards, $(ii)$ the events are summed into an image of warped events (IWE),", + "bbox": [ + 75, + 809, + 910, + 830 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nI (\\mathbf {x}; \\mathcal {E} _ {t _ {\\text {r e f}}} ^ {\\prime}, \\boldsymbol {\\theta}) \\doteq \\sum_ {k = 1} ^ {N _ {e}} \\mathcal {N} (\\mathbf {x}; \\mathbf {x} _ {k} ^ {\\prime}, \\sigma^ {2}), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 314, + 837, + 917, + 863 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "which essentially counts the number of warped events $e_k'$ per pixel. Lastly (iii) the contrast or sharpness of the IWE is computed (e.g., the gradient magnitude of the IWE, $G \\doteq \\int \\| \\nabla I(\\mathbf{x}) \\| d\\mathbf{x}$ ), which serves as a proxy for how well the model (motion hypothesis and $\\pmb{\\theta}$ ) fit with the events produced by the true motion.", + "bbox": [ + 75, + 869, + 922, + 949 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Motion-prior Contrast Maximization", + "bbox": [ + 476, + 14, + 837, + 32 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 903, + 16, + 917, + 29 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/6d8962e708d63c5de902dfc62b06940f37f5db416b89a1ae20f70e597fea79bd.jpg", + "table_caption": [ + "Table 1: Comparison of the contrast loss formulation in the most recent methods." + ], + "table_footnote": [], + "table_body": "
StepsShiba et. al [54]Paredes et. al [44]Ours
Warp displacement (3):Δxk=(tref−tk)v(ek)Δxk=∑i(Δti v_i)(ek)Δxk=1/Ntraj ∑n Δxn
linear trajectoryflow concatenationnon-linear trajectory
IWEI(x)=∑kN(x; x′k, σ2)T±(x)=∑ktkN(x; x′k, σ2)/∑kN(x; x′k, σ2)I(x)=∑kN(x; x′k, σ2)
Reference time tref:{t0, t0.5, t1}multi-partition p, multi-time-scale s~U(0,1)
Contrast objectiveG(t0)+2G(t0.5)+G(t1)/4G(v=0)1/S ∑s=0S-11/2s ∑p=02s-1LRCM,pG(tref)
", + "bbox": [ + 84, + 87, + 915, + 234 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Estimating Continuous-Time and Dense Motion Trajectories", + "text_level": 1, + "bbox": [ + 75, + 271, + 881, + 292 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The CM framework has been extended in several works. Table 1 compares recent extensions for optical flow estimation, where events are warped as", + "bbox": [ + 75, + 302, + 917, + 341 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {x} _ {k} ^ {\\prime} = \\mathbf {x} _ {k} + \\Delta \\mathbf {x} _ {k}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 406, + 357, + 917, + 378 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The main differences between these approaches lie in the event displacement (i.e., warp) model, the reference times used, and the loss function (i.e., event-alignment metric), with the design choices having two goals: fitting the event data and regularizing the solution (e.g., avoiding event collapse [52]). Previous methods used mostly a linear model [54,56] or partitioned the inference interval into smaller flow intervals so that its concatenation need not be linear [44]. By contrast, we introduce an explicit continuous-time non-linear trajectory model, a random reference time per iteration, and an easier contrast loss.", + "bbox": [ + 72, + 393, + 917, + 547 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Trajectory Representation. Figure 2 shows an overview of the training pipeline. Events in the time interval $[t_s, t_e]$ are fed into a neural network that makes per-pixel predictions of the parameters (coefficients) $\\pmb{\\theta} \\equiv \\pmb{\\alpha}_n = (\\alpha_{n,1}, \\dots, \\alpha_{n,N_c})^\\top \\in \\mathbb{R}^{N_c}$ of a continuous-time trajectory $\\mathbf{q}_n(t; \\pmb{\\alpha}_n) \\equiv \\mathbf{q}_n(t) = (x_n(t), y_n(t))^\\top$ . There is one trajectory per pixel (i.e., \"dense\" character), $n = 1, \\dots, N_p$ , where $N_p = h w$ is the number of pixels (image height $h$ and width $w$ ). Thus, $n$ is the spatial index of the trajectory, identifying it among all trajectories on the image plane.", + "bbox": [ + 72, + 549, + 932, + 684 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We model trajectories as weighted combinations of basis functions, $\\mathbf{q}_n(t) = \\sum_{j=1}^{N_c} g_j(t) \\mathbf{p}_{n,j}$ , where $g_j(t)$ are temporal basis shares by all trajectories, and $\\mathbf{p}_{n,j} = (\\alpha_{n,j}^x, \\alpha_{n,j}^y)^\\top$ are \"control points\" (we write $x$ and $y$ components explicitly with separate coefficients in $\\alpha_n$ ). We investigate polynomial basis $g_j(t) = t^j$ , Bézier curves with basis $g_j(t) = \\binom{N_c}{j}(1 - t)^{N_c - j}t^j$ , as well as a learned basis.", + "bbox": [ + 72, + 684, + 920, + 791 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Spatio-Temporal Event Warping. The trajectories can be used to warp events. This would require finding the trajectory that passes through the spacetime coordinates of the event and then finding the value of the trajectory at the reference time (i.e., the warped event location). However, the association between events and trajectories is unknown and needs to be estimated simultaneously with the parameters (i.e., shape) of the trajectories. We circumvent the problem by using a soft association between events and trajectories. Each event $e_k$ is associated with its $N_{\\mathrm{traj}}$ nearest neighboring trajectories $\\{\\mathbf{q}_n\\}_{n=1}^{N_{\\mathrm{traj}}}$ ,", + "bbox": [ + 72, + 791, + 920, + 951 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 79, + 16, + 94, + 30 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "F. Hamann et al.", + "bbox": [ + 409, + 14, + 581, + 30 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "and the event displacement $\\varDelta\\mathbf{x}_k$ in (3) is computed as the average of the respective trajectory displacements $\\{\\varDelta\\mathbf{x}_k^n\\}_{n=1}^{N_{\\mathrm{traj}}}$ . The trajectory displacement $\\varDelta\\mathbf{x}_k^n$ is defined as the difference of trajectory locations at the time of the event and the reference time. The event displacement is defined as the mean:", + "bbox": [ + 75, + 56, + 920, + 136 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta \\mathbf {x} _ {k} \\doteq \\frac {1}{N _ {\\mathrm {t r a j}}} \\sum_ {n = 1} ^ {N _ {\\mathrm {t r a j}}} \\Delta \\mathbf {x} _ {k} ^ {n}, \\quad \\text {w i t h} \\quad \\Delta \\mathbf {x} _ {k} ^ {n} \\doteq \\mathbf {q} _ {n} \\left(t _ {\\mathrm {r e f}}\\right) - \\mathbf {q} _ {n} \\left(t _ {k}\\right). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 196, + 154, + 917, + 208 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Loss Calculation. Once the events are warped using (3), it is straightforward to compute the IWE. We adopt the magnitude of the IWE gradient as loss function [18,56]. Moreover, we use the contrast loss in a self-supervised learning setting, choosing a different reference time $t_{\\mathrm{ref}}$ , uniformly sampled in the observation interval, for every batch during training. This simplifies the objective to a single warping operation and loss calculation (e.g., compared to the three warping operations and loss calculations in the optimization-based approach by [54,56]), while it has added regularization benefits, as mentioned below.", + "bbox": [ + 72, + 226, + 920, + 380 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Memory Effective Computation of the Displacement Field. The number of events and trajectories can be very large, and even more their combination. Calculating KNN for every event can quickly become computationally unfeasible, and traditional algorithms cannot be efficiently implemented in deep-learning frameworks. We propose calculating the per-event displacement $\\Delta \\mathbf{x}_k$ in (3) by first interpolating a dense but coarser spatio-temporal displacement field, and then looking up the per-event displacement in such space-time volume. Moreover, we relax the problem by solving the KNN search in 2D instead of in the volume.", + "bbox": [ + 72, + 382, + 917, + 538 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Figure 2 shows an overview of the interpolation process. The displacement map is a tensor of shape $[N_{\\mathrm{bins}}, h/4, w/4]$ , where $N_{\\mathrm{bins}}$ is the number of temporal bins, and $h$ and $w$ are the sensor's height and width, respectively. For each temporal bin-center $t_{\\mathrm{bins}}$ the trajectories $\\mathbf{q}_n(t = t_{\\mathrm{bins}})$ are calculated and the KNN interpolation is performed between each pixel of a channel $\\mathbf{x}[t_{\\mathrm{bins}}]$ and $\\mathbf{q}_n(t = t_{\\mathrm{bins}})$ . We implement the KNN approach using KeOps [15], a framework for symbolic matrix computation, which provides a memory-efficient and differentiable solution. We perform a KNN search for every voxel in the table, and events are warped after a lookup operation. Please note that this step is entirely independent of the voxel grid passed as input to the backbone, and all raw events are used to calculate the contrast loss. From the calculated displacement map, we can directly look up per-event displacements $\\Delta \\mathbf{x}_k$ .", + "bbox": [ + 72, + 539, + 917, + 772 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Regularization. Objectives based on event alignment are prone to undesired local minima called \"event collapse\" [52,53]. Our formulation inherently provides regularization, temporally by the smoothness of the motion prior, and spatially by interpolating the event flow from several trajectories via a soft assignment. Our framework uses two additional regularization sources. Firstly, we penalize the magnitude of the spatial gradient of the interpolated displacement field between consecutive timesteps, $R \\doteq \\| \\nabla (\\Delta \\mathbf{q}_{*}(t))\\|_{L^{1}(\\Omega)}$ . To interpolate $\\mathbf{q}_{*}$ from the sparse trajectories $\\mathbf{q}_n$ , we use the same KNN indices as for the main volume. This loss encourages spatial smoothness of the trajectories.", + "bbox": [ + 72, + 773, + 917, + 947 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Motion-prior Contrast Maximization", + "bbox": [ + 474, + 14, + 837, + 32 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 903, + 16, + 917, + 29 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/6ee93ed40e4de9b5753f70848db2571291c9b4b9be11be7ae3d5c123df5fb00f.jpg", + "table_caption": [ + "Table 2: Results on EVIMO2 [7]." + ], + "table_footnote": [], + "table_body": "
MethodTEPE ↓TAE ↓%Out ↓
Paredes et. al [44]21.6951.910.63
E-RAFT (linear) [24]19.3874.520.66
BFlow (zero-shot) [25]8.6319.940.36
BFlow (in-domain)3.3811.680.17
Ours (self-supervised)6.1416.980.25
", + "bbox": [ + 198, + 87, + 808, + 216 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Secondly, we use a multi-reference formulation by randomly sampling a reference time at every training step, while previous work used multiple and fixed timestamps (Tab. 1). This has the advantage that $(i)$ IWEs are required to be sharp at truly any time and $(ii)$ the memory requirement is lowered as it scales linearly with the number of reference times used during a training step.", + "bbox": [ + 72, + 253, + 920, + 351 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In summary, the training loss is", + "bbox": [ + 108, + 352, + 455, + 370 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = 1 / G + \\lambda R, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 411, + 386, + 917, + 408 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "with $\\lambda > 0$ as regularization weight. Further details are in the supplementary. Pre-processing and Prediction-Module Architecture. Events record brightness changes asynchronously, in the form of a sparse spatio-temporal signal. As input to the Prediction Module, events are customarily converted to voxel grids [77] for compatibility with conventional artificial neural networks. An event volume is discretized in the time dimension, and each voxel counts the number of events within it (bi-linearly voted for undistorted, rectified events).", + "bbox": [ + 72, + 422, + 932, + 559 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In principle, our self-supervised loss module can be paired with any segmentation or optical flow network architecture. We use a U-Net architecture for the experiments on DSEC (Sec. 4.2), and the architecture in the recent Bflow method [25], inspired by RAFT [62], for the experiments on EVIMO2 (Sec. 4.1).", + "bbox": [ + 72, + 560, + 920, + 639 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 75, + 669, + 309, + 689 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We test our method on two applications. First, we test the capabilities for nonlinear trajectory estimation on the real-world dataset EVIMO2 [7], with additional results on the synthetic MultiFlow dataset [25], which we use for pretraining (Sec. 4.1). Secondly, we evaluate our method on the DSEC dataset [23, 24] because this has been the previous frontier for event-based optical flow estimation and it allows for direct comparison with prior work (Sec. 4.2). Additional results, including on the MVSEC dataset [75], are given in the supplement.", + "bbox": [ + 72, + 708, + 917, + 848 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.1 Results on Non-Linear Trajectory Estimation", + "text_level": 1, + "bbox": [ + 75, + 876, + 697, + 895 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "EVIMO2 Continuous Flow Dataset (CF-EVIMO2). MultiFlow [25] was first proposed to evaluate long-term \"optical flow\" methods. The dataset consists", + "bbox": [ + 72, + 908, + 920, + 949 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 79, + 16, + 94, + 30 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "F. Hamann et al.", + "bbox": [ + 409, + 14, + 581, + 30 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b0e3aa4cd2694299a90e822a42492a054bad75bcf13382c394eb69df9c30d443.jpg", + "image_caption": [ + "Fig. 3: Visualization of predicted trajectories on EVIMO2 data. $GT$ : Ground truth. In-domain: fine-tuned on EVIMO2 using GT (supervised). Zero-shot: network trained only on synthetic data (out-of-domain prediction). Ours: Pre-trained on synthetic data, fine-tuned with self-supervised loss. Note that supervision in-domain is often impossible in practice because dense trajectory labels for real data are difficult to obtain." + ], + "image_footnote": [], + "bbox": [ + 152, + 50, + 842, + 551 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "of synthetic videos generated with Internet images as foreground and Flickr30K [71] images as background. Events are generated with a simulator [47] by rendering the synthetic scenes at 1000 frames/s. However, due to the sim-to-real gap between real and synthetic events, it is challenging to infer the actual performance of the proposed and baseline methods on real datasets. Furthermore, Multiflow generates trajectories by pasting 2D foreground objects onto background images, which lacks the challenging cases of self-occlusion due to 3D rotation.", + "bbox": [ + 72, + 694, + 937, + 830 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "To address these data limitations, we present the EVIMO2 Continuous Flow Dataset based on the full 3D data and event data provided by the meticulously designed EVIMO2 [7] dataset. The EVIMO2 GT data provides high-quality 3D scans of the objects, camera poses, object poses, and camera intrinsics, which we use to compute optical flow GT as follows. Given a point $P_{t}^{i}$ on the object represented in the camera coordinate system at time $t$ , we project it onto a 2D point", + "bbox": [ + 75, + 830, + 924, + 949 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Motion-prior Contrast Maximization", + "bbox": [ + 476, + 14, + 837, + 32 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 903, + 16, + 917, + 29 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/f8b3fb6b3d8b470457f6bb9fd0a8c874f50901f8ef3b49c2b3056ef6155a147c.jpg", + "image_caption": [ + "Fig. 4: End-point-error vs. prediction time span for three methods: in-distribution, out-of-distribution and self-supervised (Ours). Using the Bézier curve results from Tab. 3." + ], + "image_footnote": [], + "bbox": [ + 322, + 53, + 680, + 248 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/1ebf7d29a0880d3a5b8df3c60225558b9508dcaa2a447b1206aadc81ff35e6bc.jpg", + "table_caption": [ + "Table 3: Sensitivity with respect to motion prior on EVIMO2 [7]. $SL$ , $ID$ : supervised, in-distribution, $SL$ , $OOD$ : supervised, out of distribution, $SSL$ : self-supervised." + ], + "table_footnote": [], + "table_body": "
MethodTEPE ↓TAE ↓%Out ↓
SL, IDBFlow (polyn.)3.5113.260.18
BFlow (learned)3.7813.870.19
BFlow (Bézier)3.3811.680.17
SL, OODBFlow (polyn.)9.3620.880.36
BFlow (learned)8.6619.640.35
BFlow (Bézier) [25]8.6319.940.36
SSLOurs, polyn.6.7819.760.27
Ours, learned7.4619.780.28
Ours, Bézier6.1416.980.25
", + "bbox": [ + 203, + 370, + 801, + 588 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "$p_{t}^{i} = \\pi (P_{t}^{i})$ , where the $\\pi (\\cdot)$ function represents the perspective projection, including the camera intrinsics. The object poses at time $t$ and $t + 1$ are provided as $T_{o}^{c_{t}}$ , $T_{o}^{c_{t + 1}}$ in the GT data. The rigid-body transformation $T_{c_t}^{t + 1} = (T_o^{c_t + 1})^{-1}T_o^{c_t}$ maps 3D points in the camera frame at time $t$ to the camera frame at time $t + 1$ by aligning the shared object coordinate frame. The \"flow\" (or dense motion, since it may not be linear) is defined as $\\Delta p_{t\\rightarrow t + 1}^i\\doteq \\pi (T_{c_t}^{c_{t + 1}}P_t^i) - \\pi (P_t^i)$ . For each sequence, we generate GT dense motion every 10ms for 300ms. We mask out areas where GT object masks are unavailable. Experiments are carried out on the IMO subset of EVIMO2 using the official train and test splits.", + "bbox": [ + 72, + 621, + 917, + 798 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Metrics. Ground truth is provided as dense motion from $t = 0$ to several timesteps in increasing order. We consider $N_{s} = 6$ timestamps (i.e., subintervals) in a 300ms window and evaluate the quality of predictions with direct extension of the common optical flow metrics end-point-error (EPE), angular error (AE), and percentage of predicted vectors above a specific EPE threshold (\\%Out). We term the corresponding trajectory metrics TEPE and TAE, where", + "bbox": [ + 72, + 800, + 920, + 917 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {T E P E} = \\frac {1}{N _ {s}} \\sum_ {k = 1} ^ {N _ {s}} \\mathrm {E P E} (\\varDelta \\mathbf {x} _ {\\mathrm {p r e d}} (t _ {k}), \\varDelta \\mathbf {x} _ {\\mathrm {g t}} (t _ {k})). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 246, + 925, + 917, + 954 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 79, + 14, + 104, + 30 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "F. Hamann et al.", + "bbox": [ + 409, + 14, + 583, + 30 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "The number of outliers is calculated on TEPE with a threshold of 3px.", + "bbox": [ + 77, + 55, + 830, + 76 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Implementation Details. We use the BFlow network architecture and test different motion priors. Specifically, we report results with polynomial, Bezier (with $N_{c} = 10$ ) and learned basis $g_{j}(t)$ (a small dense neural network with three layers and hidden dimension 64 is trained alongside the main network).", + "bbox": [ + 77, + 78, + 917, + 156 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "The contrast loss module uses one trajectory per $4 \\times 4$ px, and similarly $4 \\times 4$ px in the displacement volume. The weight for the spatial smoothness term is $\\lambda = 0.003$ , and the number of neighbors in the KNN approach is $N_{\\mathrm{traj}} = 32$ . The loss function $G$ is the $L^1$ norm of the IWE's gradient magnitude [18,56].", + "bbox": [ + 77, + 157, + 917, + 235 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Training Schedule. The model is first pre-trained on MultiFlow ( $\\approx 12000$ samples) for 50 epochs, with a batch size of 10 with an $L^1$ loss on the GT trajectory flow. We include data augmentation by flipping (horizontal and vertical) and cropping and use the training/test split provided in [25]. These weights are what we refer to as zero-shot or out-of-distribution (OOD) for the main comparison on EVIMO2. Afterwards, the pre-trained model is fine-tuned with two different losses on EVIMO2. Training using our self-supervised loss is carried out for 15 epochs with a batch size of 6. We refer to this method as Ours. Lastly, a second version is trained directly on the GT flow of EVIMO for 50 epochs, which we refer to as in-distribution (ID). All experiments are performed on Nvidia RTX A6000 GPUs with an AdamW optimizer and a learning rate of $10^{-4}$ .", + "bbox": [ + 77, + 237, + 917, + 450 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Baselines. This is the first usage of EVIMO2 for dense non-linear flow/motion. For comparison, we provide additional baselines. We use ERAFT [24] with the provided DSEC weights to infer linear flow over the whole prediction time. The timestamps of the intermediate flow are interpolated from the linear prediction. Additionally, we provide results for the prediction of Paredes et. al. [44]. The network is trained self-supervised on DSEC and performs recurrent prediction steps in time in intervals of $10\\mathrm{ms}$ , inferring the total flow at every step by accumulation of the shorter flows. We unroll this prediction over the whole interval and take the accumulated flow after each $50\\mathrm{ms}$ step as the predicted flow.", + "bbox": [ + 77, + 453, + 917, + 629 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Results. Table 2 presents the comparison of the three different training modes. The results of our self-supervised loss can improve the zero-shot performance by nearly $30\\%$ . While direct supervision on EVIMO2 GT data delivers the best results, note that this is only possible because we are comparing on a dataset that was recorded with a motion capture setup and exact object models. Under real conditions, pixel-level annotations of motion tracks are very difficult to obtain. Therefore, our loss delivers a promising tool in combination with synthetic pretraining, substantially helping in overcoming the sim-to-real gap.", + "bbox": [ + 77, + 632, + 917, + 788 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Figure 3 shows the qualitative comparison of GT tracks and the three training schedules. It delivers insights into the failure cases that our method can overcome. Specifically, the zero-shot model shows errors like overseen object motion (Examples 2 and 3), wrong scale (Ex. 5), and prediction of non-existing motion (Ex. 6), which are improved with self-supervised domain training.", + "bbox": [ + 77, + 790, + 917, + 887 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Additionally, Fig. 4 shows the error for different prediction time spans. Intuitively, the error increases for longer intervals. The visualization shows that our loss improves zero-shot performance, especially for the long prediction times.", + "bbox": [ + 77, + 889, + 917, + 947 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Motion-prior Contrast Maximization", + "bbox": [ + 474, + 14, + 837, + 32 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 893, + 16, + 917, + 30 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/a95bbfb2baa25093f29689cf475f4a5ad1e572edb33ea343d24ade742bde5cac.jpg", + "table_caption": [ + "Table 4: Results on the DSEC optical flow benchmark [24]." + ], + "table_footnote": [], + "table_body": "
Method\\( {t}_{\\text{inf }}\\left\\lbrack \\mathrm{{ms}}\\right\\rbrack \\)Allinterlaken_00_binterlaken_01_athun_01_a
EPE \\( \\downarrow \\)AE \\( \\downarrow \\)%Out ↓FWL ↑EPE \\( \\downarrow \\)AE \\( \\downarrow \\)%Out ↓FWL ↑EPE \\( \\downarrow \\)AE \\( \\downarrow \\)%Out ↓FWL ↑EPE \\( \\downarrow \\)AE \\( \\downarrow \\)%Out ↓FWL ↑
E-RAFT (SL) [24]46.330.7910.562.681.291.396.226.121.320.906.883.911.420.659.751.871.20
IDNet (SL) [69]0.722.722.04-1.252.114.35-0.772.252.60-0.572.661.47-
Paredes et al. (SSL) [44]40.102.3310.5617.77-3.346.2225.72-2.496.8819.15-1.739.7510.39-
RTEF (MB) [5]4.88-41.952.518.59-59.842.895.94-47.332.923.01-29.702.39
Ev-FlowNet (SSL) [27]3.86-31.451.306.32-47.951.464.91-36.071.422.33-20.921.32
MultiCM (MB) [54]\\( {9.9} \\cdot {10}^{3} \\)3.4713.9830.861.375.749.1938.931.503.749.7731.371.512.1211.0617.681.24
Ours (poly, \\( k = 1 \\) )7.273.208.5315.211.463.214.8920.451.582.385.4617.401.701.396.997.361.30
", + "bbox": [ + 79, + 87, + 917, + 193 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/759680ad81dd9ec551822c0b65370591bd65ae8c4f0b9060cdd4c17cfc57c95e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Methodthun_01_bzurich_city_12_azurich_city_14_czurich_city_15_a
EPE ↓AE ↓%Out ↓FWL ↑EPE ↓AE ↓%Out ↓FWL ↑EPE ↓AE ↓%Out ↓FWL ↑EPE ↓AE ↓%Out ↓FWL ↑
E-RAFT (SL) [24]0.588.411.521.180.6123.161.061.120.7110.231.911.470.598.881.301.34
IDNet (SL) [69]0.552.071.35-0.604.561.16-0.763.742.74-0.552.551.02-
Paredes et al. (SSL) [44]1.668.419.34-2.7223.1626.65-2.6410.2323.01-1.698.889.98-
RTEF (MB) [5]3.91-34.692.483.14-34.081.424.00-45.672.673.78-37.992.82
EV-FlowNet (SSL) [77]3.04-25.411.332.62-25.801.033.36-36.341.242.97-25.531.33
MultiCM (MB) [54]2.4812.0523.561.243.8628.6143.961.142.7212.6230.531.502.3511.8220.991.41
Ours (poly, k = 1)1.546.559.691.338.3320.1622.391.131.788.7912.991.561.456.278.341.51
", + "bbox": [ + 79, + 193, + 917, + 294 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Sensitivity to motion prior. Table 3 compares our method for different motion priors. While Bézier curves show slightly improved performance over the two basis function methods, self-supervised domain training can improve performance in all cases, which proves the robustness and generality of our method.", + "bbox": [ + 75, + 331, + 920, + 411 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.2 Results on Optical Flow Estimation", + "text_level": 1, + "bbox": [ + 75, + 443, + 578, + 463 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Dataset, Metrics and Details. We compare our method on the DSEC optical flow benchmark [24]. The dataset consists of sequences from a Prophesee Gen3 event camera, with a resolution of $640 \\times 480$ px, mounted on a driving car. Ground truth is calculated as the motion field from a co-deployed depth sensor.", + "bbox": [ + 72, + 479, + 917, + 557 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We compare the common flow metrics EPE, AE, and percentage of outliers (%Out), where an outlier is a flow vector with EPE $>3\\mathrm{px}$ . Additionally, we evaluate the Flow Warp Loss (FWL) [57]. FWL $>1$ means that the IWE is sharper than pixel-wise event accumulation (i.e., IWE with zero optical flow).", + "bbox": [ + 75, + 557, + 917, + 635 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We use a simple U-Net [49] as the backbone in the linear flow experiments (similar architecture like EV-FlowNet [76]) and choose polynomial basis $g_{j}(t) = t^{j}$ of degree 1, which effectively leads to a linear motion prior. We use $N_{\\mathrm{traj}} = 32$ nearest neighbors and $N_{\\mathrm{bins}} = 15$ time bins in the displacement map. The weight for the regularizer is $\\lambda = 0.003$ . The network trains for 50 epochs with and Adam optimizer, a learning rate of $10^{-4}$ , and a total batch size of 28 on two RTX A6000. Results. Table 4 shows the DSEC benchmark results, confirming that our model can generate high-quality optical flow results without having access to any GT. It furthermore achieves the best results among all contrast-maximization-based methods on six out of seven sequences, and only methods directly supervised on GT perform better. Within the self-supervised methods, it improves the AE by $19\\%$ and the number of inliers by $14\\%$ . The only exception is the night sequence zurich_city_14. Contrast maximization is based on the brightness constancy assumption, which is violated here by flickering street lights. Hence, in these areas, the contrast loss does not provide reasonable flow predictions. While our methods still provide a better AE and %Out at night, higher outliers in the", + "bbox": [ + 72, + 635, + 920, + 949 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 79, + 14, + 104, + 30 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "F. Hamann et al.", + "bbox": [ + 409, + 14, + 583, + 30 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/eaa3d82bf0778e152eaf6607b1c9b2320ebd76e2302e39fd4a869b50b4749b38.jpg", + "image_caption": [ + "Fig. 5: Results on DSEC. Image of warped events and predicted flow by three methods." + ], + "image_footnote": [], + "bbox": [ + 84, + 52, + 912, + 328 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "flickering regions prevent the difference in the average EPE from reflecting the improved performance of our method on the other six sequences.", + "bbox": [ + 72, + 401, + 917, + 440 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Figure 5 shows qualitative results of our method. In comparison to another self-supervised optical flow method (d), the visualizations are noticeably sharper, allowing for a more precise representation of the motion in the scene. Moreover, the model performs well in delineating the contours of foreground objects, without the over-smoothing effect often observable in flow methods. The IWE (Fig. 5b) shows sharp results, aligning with the overall high FWL values reported in Tab. 4. While the FWL metric also increases in event collapse (e.g., for RTEF [5]), the IWE visualization reveals that here the model is well-regularized, confirming that the predicted motion aligns the events correctly. The third example in Fig. 5 highlights how our model has improved predictions with a single forward pass, where models relying on temporal recurrence (see (d)) fail at the beginning and need a warm-up stage resulting in higher latency.", + "bbox": [ + 72, + 440, + 920, + 674 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Inference time. Additionally, Tab. 4 provides inference times $t_{\\mathrm{inf}}$ for several methods. Note that our method is about $5 \\times$ faster than the competitive baselines. The reason is that we do not rely on any recurrence in this optical flow application, such as recurrence in time or RAFT-inspired refinement steps.", + "bbox": [ + 72, + 674, + 917, + 752 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Ablation and Sensitivity. Table 5 lists the performance under different loss settings. It confirms that the number of neighboring trajectories $N_{\\mathrm{traj}}$ in the KNN approach has a regularizing effect and is important for good performance. The decreased performance on the benchmark for a lower $N_{\\mathrm{traj}}$ aligns with the visual impression in Fig. 6: models with lower $N_{\\mathrm{traj}}$ show artifacts and a stronger susceptibility to the aperture problem, as is observable on the rock pattern.", + "bbox": [ + 72, + 752, + 920, + 869 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Table 5 confirms that the multi-reference approach is crucial for regularizing the contrast loss. While using three instead of one $t_{\\mathrm{ref}}$ improves performance, our approach using a randomized $t_{\\mathrm{ref}}$ leads to an additional performance boost. Lastly, the non-linear prior showed no performance improvement on DSEC.", + "bbox": [ + 72, + 871, + 920, + 949 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Motion-prior Contrast Maximization", + "bbox": [ + 476, + 14, + 837, + 32 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 893, + 16, + 917, + 30 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/4624f4afb0b623fcb8cb27e259b4472b59a262b94b03f51e8881bf60b59f093f.jpg", + "table_caption": [ + "Table 5: Sensitivity and Ablation Study for DSEC. Ours corresponds to \"Ours\" in Tab. 4. Configurations marked with “-” are unchanged from our main result. $\\Delta$ specifies the change with respect to the original configuration." + ], + "table_footnote": [], + "table_body": "
NtrajNrefNcMotion priorEPE ↓ΔEPEAE ↓ΔAE%Out ↓Δ%Out
Ours32~U(0,1)1polynomial3.208.5315.21
Number of neighbor trajectories1---4.581.0713.735.2027.9012.69
8---3.510.3113.204.6723.398.18
Number of reference times-3--4.461.2614.435.9028.6013.39
-1--7.264.0618.7210.1944.7529.54
Type and degree of motion prior--5learned3.220.028.590.0615.340.13
--5polynomial3.280.068.610.0815.680.47
", + "bbox": [ + 87, + 123, + 915, + 269 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/4dd03fe367e6c8c39a80585d3f16b81aedc791444beabe2da998261ca097fdc5.jpg", + "image_caption": [ + "Fig. 6: Visual effect of the number of neighbors $N_{\\mathrm{traj}}$ on the predicted flow." + ], + "image_footnote": [], + "bbox": [ + 82, + 282, + 489, + 403 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/36fcbb9a333d2945ec606ba165b3dd2a9bdd499f72d6e37366ac886719d386cd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 282, + 907, + 403 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Limitations", + "text_level": 1, + "bbox": [ + 75, + 471, + 295, + 491 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Like all CM-based methods, ours is based on the brightness constancy assumption. Therefore, it shows limitations in estimating flow from events that are not caused by motion, e.g., from flickering lights.", + "bbox": [ + 72, + 510, + 917, + 567 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Our KNN interpolation shows a good trade-off between granularity and regularization; nevertheless, our method is limited by the aperture problem inherent to optical flow.", + "bbox": [ + 72, + 569, + 917, + 626 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Lastly, during training, all raw events are passed to the loss module, increasing training time. Furthermore, gradient calculation is performed through the additional step of event warping. These steps increase memory requirements and training time of our loss compared to supervised methods.", + "bbox": [ + 72, + 627, + 920, + 705 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 75, + 734, + 285, + 754 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We introduced a new loss formulation based on the contrast maximization framework, by combining it with a non-linear trajectory prior. It is a versatile tool that works with a variety of model architectures and trajectory representations. We presented an efficient method to make the high-dimensional assignment between millions of events and thousands of trajectories feasible. Our experiments show clear advantages against supervised methods on unseen data, where under real-world circumstances no GT is available. Additionally, a U-Net trained with our loss shows state-of-the-art performance on the DSEC optical flow benchmark, while being substantially faster than the previously best methods.", + "bbox": [ + 72, + 773, + 920, + 949 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 79, + 16, + 104, + 29 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "F. Hamann et al.", + "bbox": [ + 409, + 14, + 581, + 30 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 77, + 53, + 336, + 76 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "We thank Dr. Cornelia Fermüller and the NeuroPAC network for fostering collaborations within the event-based community (NSF OISE 2020624). Funded by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) under Germany's Excellence Strategy - EXC 2002/1 \"Science of Intelligence\" - project number 390523135. We furthermore gratefully acknowledge the support by the following grants: NSF FRR 2220868, NSF IIS-RI 2212433, NSF TRIPODS 1934960, ONR N00014-22-1-2677.", + "bbox": [ + 75, + 95, + 922, + 232 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 77, + 263, + 237, + 282 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Akhter, I., Sheikh, Y., Khan, S., Kanade, T.: Trajectory space: A dual representation for nonrigid structure from motion. IEEE Trans. Pattern Anal. Mach. Intell. 33(7), 1442-1456 (jul 2011). https://doi.org/10.1109/TPAMI.2010.201", + "2. Bailer, C., Varanasi, K., Stricker, D.: CNN-based patch matching for optical flow with thresholded hinge embedding loss. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR). pp. 2710-2719 (2017). https://doi.org/10.1109/CVPR.2017.290", + "3. Benosman, R., Clercq, C., Lagorce, X., Ieng, S.H., Bartolozzi, C.: Event-based visual flow. IEEE Trans. Neural Netw. Learn. Syst. 25(2), 407-417 (2014). https://doi.org/10.1109/TNNLS.2013.2273537", + "4. Benosman, R., Ieng, S.H., Clercq, C., Bartolozzi, C., Srinivasan, M.: Asynchronous frameless event-based optical flow. Neural Netw. 27, 32-37 (2012). https://doi.org/10.1016/j.neunet.2011.11.001", + "5. Brebion, V., Moreau, J., Davoine, F.: Real-time optical flow for vehicular perception with low- and high-resolution event cameras. IEEE Trans. Intell. Transport. Syst. pp. 1-13 (2021). https://doi.org/10.1109/TITS.2021.3136358", + "6. Brosch, T., Tschechne, S., Neumann, H.: On event-based optical flow detection. Front. Neurosci. 9(137) (Apr 2015). https://doi.org/10.3389/fnins.2015.00137", + "7. Burner, L., Mitrokhin, A., Fermüller, C., Aloimonos, Y.: EVIMO2: An event camera dataset for motion segmentation, optical flow, structure from motion, and visual inertial odometry in indoor scenes with monocular or stereo algorithms. arXiv e-prints (May 2022). https://doi.org/10.48550/arXiv.2205.03467", + "8. Butler, D.J., Wulff, J., Stanley, G.B., Black, M.J.: A naturalistic open source movie for optical flow evaluation. In: Eur. Conf. Comput. Vis. (ECCV). pp. 611-625 (2012). https://doi.org/10.1007/978-3-642-33783-3_44", + "9. Chaney, K., Cladera Ojeda, F., Wang, Z., Bisulco, A., Hsieh, M.A., Korpela, C., Kumar, V., Taylor, C.J., Daniilidis, K.: M3ED: Multi-robot, multi-sensor, multi-environment event dataset. In: IEEE Conf. Comput. Vis. Pattern Recog. Workshops (CVPRW). pp. 4016-4023 (2023). https://doi.org/10.1109/CVPRW59228.2023.00419", + "10. Chui, J., Klenk, S., Cremers, D.: Event-based feature tracking in continuous time with sliding window optimization. In: arXiv e-prints (2021). https://doi.org/10.48550/arXiv.2107.04536", + "11. Ding, Z., Zhao, R., Zhang, J., Gao, T., Xiong, R., Yu, Z., Huang, T.: Spatiotemporal recurrent networks for event-based optical flow estimation. In: AAAI Conf. Artificial Intell. vol. 36, pp. 525-533 (2022). https://doi.org/10.1609/aaai.v36i1.19931" + ], + "bbox": [ + 79, + 302, + 917, + 947 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Motion-prior Contrast Maximization", + "bbox": [ + 474, + 14, + 837, + 32 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 893, + 16, + 917, + 30 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "12. Doersch, C., Gupta, A., Markeeva, L., Recasens, A., Smaira, L., Aytar, Y., Carreira, J., Zisserman, A., Yang, Y.: Tap-vid: A benchmark for tracking any point in a video. Adv. Neural Inf. Process. Syst. (NeurIPS) 35, 13610-13626 (2022)", + "13. Doersch, C., Yang, Y., Vecerik, M., Gokay, D., Gupta, A., Aytar, Y., Carreira, J., Zisserman, A.: Tapir: Tracking any point with per-frame initialization and temporal refinement. arXiv preprint arXiv:2306.08637 (2023)", + "14. Dosovitskiy, A., Fischer, P., Ilg, E., Häusser, P., Hazirbaş, C., Golkov, V., van der Smagt, P., Cremers, D., Brox, T.: FlowNet: Learning optical flow with convolutional networks. In: Int. Conf. Comput. Vis. (ICCV). pp. 2758-2766 (2015). https://doi.org/10.1109/ICCV.2015.316", + "15. Feydy, J., Glaunès, A., Charlier, B., Bronstein, M.: Fast geometric learning with symbolic matrices. Adv. Neural Inf. Process. Syst. (NeurIPS) 33, 14448-14462 (2020)", + "16. Finateu, T., Niwa, A., Matolin, D., Tsuchimoto, K., Mascheroni, A., Reynaud, E., Mostafalu, P., Brady, F., Chotard, L., LeGoff, F., Takahashi, H., Wakabayashi, H., Oike, Y., Posch, C.: A 1280x720 back-illuminated stacked temporal contrast event-based vision sensor with $4.86\\mu \\mathrm{m}$ pixels, 1.066Geps readout, programmable event-rate controller and compressive data-formating pipeline. In: IEEE Int. Solid-State Circuits Conf. (ISSCC). pp. 112-114 (2020). https://doi.org/10.1109/ISSCC19947.2020.9063149", + "17. Gallego, G., Delbruck, T., Orchard, G., Bartolozzi, C., Taba, B., Censi, A., Leutenegger, S., Davison, A., Conradt, J., Daniilidis, K., Scaramuzza, D.: Event-based vision: A survey. IEEE Trans. Pattern Anal. Mach. Intell. 44(1), 154-180 (2022). https://doi.org/10.1109/TPAMI.2020.3008413", + "18. Gallego, G., Gehrig, M., Scaramuzza, D.: Focus is all you need: Loss functions for event-based vision. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR). pp. 12272-12281 (2019). https://doi.org/10.1109/CVPR.2019.01256", + "19. Gallego, G., Rebecq, H., Scaramuzza, D.: A unifying contrast maximization framework for event cameras, with applications to motion, depth, and optical flow estimation. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR). pp. 3867-3876 (2018). https://doi.org/10.1109/CVPR.2018.00407", + "20. Gallego, G., Scaramuzza, D.: Accurate angular velocity estimation with an event camera. IEEE Robot. Autom. Lett. 2(2), 632-639 (2017). https://doi.org/10.1109/LRA.2016.2647639", + "21. Gehrig, D., Gehrig, M., Hidalgo-Carrió, J., Scaramuzza, D.: Video to Events: Recycl- cling video datasets for event cameras. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR). pp. 3583-3592 (2020). https://doi.org/10.1109/CVPR42600.2020.00364", + "22. Gehrig, D., Loquercio, A., Derpanis, K.G., Scaramuzza, D.: End-to-end learning of representations for asynchronous event-based data. In: Int. Conf. Comput. Vis. (ICCV). pp. 5632-5642 (2019). https://doi.org/10.1109/ICCV.2019.00573", + "23. Gehrig, M., Aarents, W., Gehrig, D., Scaramuzza, D.: DSEC: A stereo event camera dataset for driving scenarios. IEEE Robot. Autom. Lett. 6(3), 4947-4954 (2021). https://doi.org/10.1109/LRA.2021.3068942", + "24. Gehrig, M., Millhäsler, M., Gehrig, D., Scaramuzza, D.: E-RAFT: Dense optical flow from event cameras. In: Int. Conf. 3D Vision (3DV). pp. 197-206 (2021). https://doi.org/10.1109/3DV53792.2021.00030", + "25. Gehrig, M., Muglikar, M., Scaramuzza, D.: Dense continuous-time optical flow from event cameras. IEEE Trans. Pattern Anal. Mach. Intell. pp. 1-12 (2024). https://doi.org/10.1109/TPAMI.2024.3361671" + ], + "bbox": [ + 79, + 56, + 917, + 949 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 79, + 14, + 104, + 30 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "F. Hamann et al.", + "bbox": [ + 409, + 14, + 583, + 30 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "26. Guney, F., Sevilla-Lara, L., Sun, D., Wulff, J.: \"what is optical flow for?\": Workshop results and summary. In: Eur. Conf. Comput. Vis. Workshops (ECCVW). pp. 0-0 (2018)", + "27. Guo, S., Gallego, G.: CMax-SLAM: Event-based rotational-motion bundle adjustment and SLAM system using contrast maximization. IEEE Trans. Robot. 40, 2442-2461 (2024). https://doi.org/10.1109/TRO.2024.3378443", + "28. Hagenaars, J., Paredes-Vallés, F., De Croon, G.: Self-supervised learning of event-based optical flow with spiking neural networks. Adv. Neural Inf. Process. Syst. (NeurIPS) 34, 7167-7179 (2021)", + "29. Hamann, F., Gallego, G.: Stereo co-capture system for recording and tracking fish with frame- and event cameras. In: 26th Int. Conf. on Pattern Recognition (ICPR), Visual observation and analysis of Vertebrate And Insect Behavior (VAIB) Workshop (2022). https://doi.org/10.48550/ARXIV.2207.07332", + "30. Hamann, F., Ghosh, S., Juarez-Martínez, I., Hart, T., Kacelnik, A., Gallego, G.: Low-power, continuous remote behavioral localization with event cameras. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR) (2024)", + "31. Harley, A.W., Fang, Z., Fragkiadaki, K.: Particle video revisited: Tracking through occlusions using point trajectories. In: Eur. Conf. Comput. Vis. (ECCV). pp. 59-75 (2022)", + "32. Huang, Z., Shi, X., Zhang, C., Wang, Q., Cheung, K.C., Qin, H., Dai, J., Li, H.: Flowformer: A transformer architecture for optical flow. In: Eur. Conf. Comput. Vis. (ECCV). pp. 668-685 (2022)", + "33. Ilg, E., Mayer, N., Saikia, T., Keuper, M., Dosovitskiy, A., Brox, T.: FlowNet 2.0: Evolution of optical flow estimation with deep networks. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR). pp. 1647-1655 (2017). https://doi.org/10.1109/cvpr.2017.179", + "34. Janai, J., Guney, F., Ranjan, A., Black, M., Geiger, A.: Unsupervised learning of multi-frame optical flow with occlusions. In: Eur. Conf. Comput. Vis. (ECCV). pp. 690-706 (2018)", + "35. Jonschkowski, R., Stone, A., Barron, J.T., Gordon, A., Konolige, K., Angelova, A.: What matters in unsupervised optical flow. In: Eur. Conf. Comput. Vis. (ECCV). pp. 557-572 (2020)", + "36. Karaev, N., Rocco, I., Graham, B., Neverova, N., Vedaldi, A., Rupprecht, C.: Cotracker: It is better to track together. arXiv preprint arXiv:2307.07635 (2023)", + "37. Kim, H., Kim, H.J.: Real-time rotational motion estimation with contrast maximization over globally aligned events. IEEE Robot. Autom. Lett. 6(3), 6016-6023 (2021). https://doi.org/10.1109/LRA.2021.3088793", + "38. Lee, C., Kosta, A., Zhu, A.Z., Chaney, K., Daniilidis, K., Roy, K.: Spike-flownet: Event-based optical flow estimation with energy-efficient hybrid neural networks. In: Eur. Conf. Comput. Vis. (ECCV). pp. 366-382 (2020)", + "39. Lichtsteiner, P., Posch, C., Delbruck, T.: A $128 \\times 128$ 120 dB $15~\\mu \\mathrm{s}$ latency asynchronous temporal contrast vision sensor. IEEE J. Solid-State Circuits 43(2), 566-576 (2008). https://doi.org/10.1109/JSSC.2007.914337", + "40. Liu, H., Chen, G., Qu, S., Zhang, Y., Li, Z., Knoll, A., Jiang, C.: TMA: Temporal motion aggregation for event-based optical flow. In: Int. Conf. Comput. Vis. (ICCV). pp. 9651-9660 (Oct 2023). https://doi.org/10.1109/ICCV51070.2023.00888", + "41. Liu, M., Delbruck, T.: Adaptive time-slice block-matching optical flow algorithm for dynamic vision sensors. In: British Mach. Vis. Conf. (BMVC). pp. 1-12 (2018)" + ], + "bbox": [ + 77, + 56, + 920, + 949 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Motion-prior Contrast Maximization", + "bbox": [ + 474, + 14, + 837, + 30 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 893, + 16, + 917, + 30 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "42. Mayer, N., Ilg, E., Hausser, P., Fischer, P., Cremers, D., Dosovitskiy, A., Brox, T.: A large dataset to train convolutional networks for disparity, optical flow, and scene flow estimation. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR). pp. 4040-4048 (2016)", + "43. Orchard, G., Benosman, R., Etienne-Cummings, R., Thakor, N.V.: A spiking neural network architecture for visual motion estimation. In: IEEE Biomed. Circuits Syst. Conf. (BioCAS). pp. 298-301 (2013). https://doi.org/10.1109/biocas.2013.6679698", + "44. Paredes-Valles, F., Scheper, K.Y., De Wagter, C., de Croon, G.C.: Taming contrast maximization for learning sequential, low-latency, event-based optical flow. In: Int. Conf. Comput. Vis. (ICCV). pp. 9661-9671 (Oct 2023). https://doi.org/10.1109/ICCV51070.2023.00889", + "45. Paredes-Valles, F., Hagenaars, J.J., Dupeyroux, J., Stroobants, S., Xu, Y., de Croon, G.C.H.E.: Fully neuromorphic vision and control for autonomous drone flight. Science Robotics 9(90), eadi0591 (2024). https://doi.org/10.1126/scirobotics.adio591", + "46. Posch, C., Serrano-Gotarredona, T., Linares-Barranco, B., Delbruck, T.: Retinomorphic event-based vision sensors: Bioinspired cameras with spiking output. Proc. IEEE 102(10), 1470-1484 (Oct 2014). https://doi.org/10.1109/jproc.2014.2346153", + "47. Rebecq, H., Gehrig, D., Scaramuzza, D.: ESIM: an open event camera simulator. In: Conf. on Robotics Learning (CoRL). Proc. Machine Learning Research, vol. 87, pp. 969-982. PMLR (2018)", + "48. Ren, Z., Yan, J., Ni, B., Liu, B., Yang, X., Zha, H.: Unsupervised deep learning for optical flow estimation. In: AAAI Conf. Artificial Intell. vol. 31 (2017)", + "49. Ronneberger, O., Fischer, P., Brox, T.: U-Net: Convolutional networks for biomedical image segmentation. In: Int. Conf. Medical Image Computing and Computer-Assisted Intervention (MICCAI). pp. 234-241 (2015)", + "50. Sand, P., Teller, S.: Particle video: Long-range motion estimation using point trajectories. Int. J. Comput. Vis. 80, 72-91 (2008)", + "51. Seok, H., Lim, J.: Robust feature tracking in dvs event stream using Bezier mapping. In: IEEE Winter Conf. Appl. Comput. Vis. (WACV). pp. 1647-1656 (2020). https://doi.org/10.1109/WACV45572.2020.9093607", + "52. Shiba, S., Aoki, Y., Gallego, G.: Event collapse in contrast maximization frameworks. Sensors 22(14), 1-20 (2022). https://doi.org/10.3390/s22145190", + "53. Shiba, S., Aoki, Y., Gallego, G.: A fast geometric regularizer to mitigate event collapse in the contrast maximization framework. Adv. Intell. Syst. p. 2200251 (2022). https://doi.org/10.1002/aisy.202200251", + "54. Shiba, S., Aoki, Y., Gallego, G.: Secrets of event-based optical flow. In: Eur. Conf. Comput. Vis. (ECCV). pp. 628-645 (2022). https://doi.org/10.1007/978-3-031-19797-0 36", + "55. Shiba, S., Hamann, F., Aoki, Y., Gallego, G.: Event-based background oriented schlieren. IEEE Trans. Pattern Anal. Mach. Intell. 46(4), 2011-2026 (2024). https://doi.org/10.1109/TPAMI.2023.3328188", + "56. Shiba, S., Klose, Y., Aoki, Y., Gallego, G.: Secrets of event-based optical flow, depth, and ego-motion by contrast maximization. IEEE Trans. Pattern Anal. Mach. Intell. pp. 1-18 (2024). https://doi.org/10.1109/TPAMI.2024.3396116", + "57. Stoffregen, T., Scheerlinck, C., Scaramuzza, D., Drummond, T., Barnes, N., Kleeman, L., Mahony, R.: Reducing the sim-to-real gap for event cameras. In: Eur. Conf. Comput. Vis. (ECCV). pp. 534-549 (2020). https://doi.org/https://doi.org/10.1007/978-3-030-58583-9_32" + ], + "bbox": [ + 79, + 56, + 917, + 949 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 79, + 16, + 104, + 29 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "F. Hamann et al.", + "bbox": [ + 409, + 14, + 581, + 30 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "58. Stone, A., Maurer, D., Ayvaci, A., Angelova, A., Jonschkowski, R.: Smurf: Self-teaching multi-frame unsupervised raft with full-image warping. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR). pp. 3887-3896 (2021)", + "59. Sun, D., Vlasic, D., Herrmann, C., Jampani, V., Krainin, M., Chang, H., Zabih, R., Freeman, W.T., Liu, C.: Autoflow: Learning a better training set for optical flow. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR). pp. 10093-10102 (2021)", + "60. Sun, D., Yang, X., Liu, M.Y., Kautz, J.: PWC-Net: CNNs for optical flow using pyramid, warping, and cost volume. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR) (2018)", + "61. Sun, X., Harley, A.W., Guibas, L.J.: Refining pre-trained motion models. arXiv preprint arXiv:2401.00850 (2024)", + "62. Teed, Z., Deng, J.: RAFT: Recurrent all pairs field transforms for optical flow. In: Eur. Conf. Comput. Vis. (ECCV). pp. 402-419 (2020). https://doi.org/10.1007/978-3-030-58536-5_24", + "63. Tulyakov, S., Bochicchio, A., Gehrig, D., Georgoulis, S., Li, Y., Scaramuzza, D.: Time lens++: Event-based frame interpolation with parametric non-linear flow and multi-scale fusion. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR). pp. 17755-17764 (Jun 2022)", + "64. Valmadre, J., Lucey, S.: General trajectory prior for non-rigid reconstruction. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR). pp. 1394-1401 (2012)", + "65. Wan, Z., Dai, Y., Mao, Y.: Learning dense and continuous optical flow from an event camera. IEEE Trans. Image Process. 31, 7237-7251 (2022)", + "66. Wang, C., Eckart, B., Lucey, S., Gallo, O.: Neural trajectory fields for dynamic novel view synthesis. arXiv preprint arXiv:2105.05994 (2021)", + "67. Wang, Z., Hamann, F., Chaney, K., Jiang, W., Gallego, G., Daniilidis, K.: Event-based continuous color video decompression from single frames. arXiv preprint arXiv:2312.00113 (2023)", + "68. Weikersdorfer, D., Hoffmann, R., Conradt, J.: Simultaneous localization and mapping for event-based vision systems. In: Int. Conf. Comput. Vis. Syst. (ICVS). pp. 133-142 (2013). https://doi.org/10.1007/978-3-642-39402-7_14", + "69. Wu, Y., Paredes-Valles, F., de Croon, G.C.: Lightweight event-based optical flow estimation via iterative deblurring. arXiv preprint arXiv:2211.13726 (2022)", + "70. Ye, C., Mitrokhin, A., Parameshwara, C., Fermüller, C., Yorke, J.A., Aloimonos, Y.: Unsupervised learning of dense optical flow, depth and egomotion with event-based sensors. In: IEEE/RSJ Int. Conf. Intell. Robot. Syst. (IROS). pp. 5831-5838 (2020). https://doi.org/10.1109/IROS45743.2020.9341224", + "71. Young, P., Lai, A., Hodosh, M., Hockenmaier, J.: From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. Trans. Assoc. Computational Linguistics 2, 67-78 (2014)", + "72. Yu, J.J., Harley, A.W., Derpanis, K.G.: Back to basics: Unsupervised learning of optical flow via brightness constancy and motion smoothness. In: Eur. Conf. Comput. Vis. Workshops (ECCVW). pp. 3-10 (2016)", + "73. Zen, G., Ricci, E., Sebe, N.: Exploiting sparse representations for robust analysis of noisy complex video scenes. In: Eur. Conf. Comput. Vis. (ECCV). pp. 199-213 (2012)", + "74. Zheng, Y., Harley, A.W., Shen, B., Wetzstein, G., Guibas, L.J.: PointOdyssey: A large-scale synthetic dataset for long-term point tracking. In: Int. Conf. Comput. Vis. (ICCV). pp. 19855-19865 (2023)", + "75. Zhu, A.Z., Thakur, D., Ozaslan, T., Pfrommer, B., Kumar, V., Daniilidis, K.: The multivehicle stereo event camera dataset: An event camera dataset for 3D" + ], + "bbox": [ + 79, + 56, + 917, + 946 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "Motion-prior Contrast Maximization", + "bbox": [ + 474, + 14, + 837, + 30 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 893, + 16, + 917, + 30 + ], + "page_idx": 18 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "perception. IEEE Robot. Autom. Lett. 3(3), 2032-2039 (Jul 2018). https://doi.org/10.1109/lra.2018.2800793", + "76. Zhu, A.Z., Yuan, L., Chaney, K., Daniilidis, K.: EV-FlowNet: Self-supervised optical flow estimation for event-based cameras. In: Robotics: Science and Systems (RSS). pp. 1-9 (2018). https://doi.org/10.15607/RSS.2018.XIV.062", + "77. Zhu, A.Z., Yuan, L., Chaney, K., Daniilidis, K.: Unsupervised event-based learning of optical flow, depth, and egomotion. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR). pp. 989-997 (2019). https://doi.org/10.1109/CVPR.2019.00108", + "78. Zhu, Y., Lucey, S.: Convolutional sparse coding for trajectory reconstruction. IEEE Trans. Pattern Anal. Mach. Intell. 37(3), 529-540 (2013)" + ], + "bbox": [ + 77, + 56, + 920, + 235 + ], + "page_idx": 19 + }, + { + "type": "page_number", + "text": "20", + "bbox": [ + 77, + 16, + 104, + 30 + ], + "page_idx": 19 + }, + { + "type": "header", + "text": "F. Hamann et al.", + "bbox": [ + 409, + 14, + 583, + 30 + ], + "page_idx": 19 + } +] \ No newline at end of file diff --git a/2024/Motion-prior Contrast Maximization for Dense Continuous-Time Motion Estimation/34d0d1c3-d92d-4348-bbb4-52308320fede_model.json b/2024/Motion-prior Contrast Maximization for Dense Continuous-Time Motion Estimation/34d0d1c3-d92d-4348-bbb4-52308320fede_model.json new file mode 100644 index 0000000000000000000000000000000000000000..f1f73a2032db243017f2ae2b914ddf616b054012 --- /dev/null +++ b/2024/Motion-prior Contrast Maximization for Dense Continuous-Time Motion Estimation/34d0d1c3-d92d-4348-bbb4-52308320fede_model.json @@ -0,0 +1,2726 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.132, + 0.049, + 0.868, + 0.105 + ], + "angle": 0, + "content": "Motion-prior Contrast Maximization for Dense Continuous-Time Motion Estimation" + }, + { + "type": "text", + "bbox": [ + 0.135, + 0.141, + 0.865, + 0.183 + ], + "angle": 0, + "content": "Friedhelm Hamann1, Ziyun Wang2, Ioannis Asmanis2, Kenneth Chaney2, Guillermo Gallego1,3, and Kostas Daniilidis2,4" + }, + { + "type": "text", + "bbox": [ + 0.2, + 0.196, + 0.796, + 0.233 + ], + "angle": 0, + "content": "\\(^{1}\\) TU Berlin and SCIolI Excellence Cluster, Berlin, Germany \n\\(^{2}\\) University of Pennsylvania, Philadelphia, US" + }, + { + "type": "text", + "bbox": [ + 0.086, + 0.233, + 0.911, + 0.268 + ], + "angle": 0, + "content": "3 Einstein Center Digital Future and Robotics Institute Germany, Berlin, Germany \n4 Archimedes, Athena RC, Greece" + }, + { + "type": "text", + "bbox": [ + 0.143, + 0.325, + 0.857, + 0.594 + ], + "angle": 0, + "content": "Abstract. Current optical flow and point-tracking methods rely heavily on synthetic datasets. Event cameras are novel vision sensors with advantages in challenging visual conditions, but state-of-the-art frame-based methods cannot be easily adapted to event data due to the limitations of current event simulators. We introduce a novel self-supervised loss combining the Contrast Maximization framework with a non-linear motion prior in the form of pixel-level trajectories and propose an efficient solution to solve the high-dimensional assignment problem between non-linear trajectories and events. Their effectiveness is demonstrated in two scenarios: In dense continuous-time motion estimation, our method improves the zero-shot performance of a synthetically trained model on the real-world dataset EVIMO2 by \\(29\\%\\). In optical flow estimation, our method elevates a simple UNet to achieve state-of-the-art performance among self-supervised methods on the DSEC optical flow benchmark. Our code is available at https://github.com/tub-rip/MotionPriorCMax." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.631, + 0.311, + 0.65 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.676, + 0.923, + 0.832 + ], + "angle": 0, + "content": "Determining the motion of arbitrary projected world points on the image plane over long time intervals is a difficult low-level computer vision problem. Researchers have studied it as optical flow and lately as point-tracking, with many practical applications in robotics, computational photography, video compression, and object-level tracking [26]. The best-performing frame-based methods for optical flow estimation and point tracking use large-scale synthetic datasets. Synthetic data is diverse and has highly accurate ground truth (GT), but is unrealistic and methods trained on synthetic data show a sim-to-real gap." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.833, + 0.924, + 0.95 + ], + "angle": 0, + "content": "Novel vision sensors called event cameras [16,17,39] have emerged as promising alternatives to take on the problem. Inspired by the transient visual pathway, which is responsible for motion perception, they are particularly fit for capturing scene motion in the form of asynchronous pixel-wise intensity changes. This working principle endows them with advantages, such as high speed, high dynamic range (HDR) and low power consumption." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.08, + 0.017, + 0.095, + 0.03 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.411, + 0.016, + 0.581, + 0.032 + ], + "angle": 0, + "content": "F. Hamann et al." + }, + { + "type": "image", + "bbox": [ + 0.167, + 0.075, + 0.411, + 0.18 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.121, + 0.186, + 0.456, + 0.199 + ], + "angle": 0, + "content": "a) Contrast Maximization with non-linear Trajectories" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.07, + 0.904, + 0.185 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.535, + 0.186, + 0.666, + 0.198 + ], + "angle": 0, + "content": "b) Overview Method" + }, + { + "type": "image", + "bbox": [ + 0.105, + 0.218, + 0.229, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.232, + 0.218, + 0.356, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.358, + 0.218, + 0.482, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.105, + 0.281, + 0.229, + 0.343 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.232, + 0.281, + 0.356, + 0.342 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.358, + 0.281, + 0.482, + 0.342 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.121, + 0.349, + 0.366, + 0.363 + ], + "angle": 0, + "content": "c) Application to Trajectory Estimation" + }, + { + "type": "image", + "bbox": [ + 0.537, + 0.207, + 0.882, + 0.349 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.535, + 0.35, + 0.726, + 0.362 + ], + "angle": 0, + "content": "d) Application to Optical Flow" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.384, + 0.921, + 0.474 + ], + "angle": 0, + "content": "Fig.1: Summary. a) We present an approach to combine Contrast Maximization with dense non-linear trajectories. b) We show how it can be used for self-supervised learning in a pipeline to predict dense point trajectories, and c) evaluate it on the EVIMO2 dataset, for which we generate dense point tracks. d) Additionally, our approach provides state-of-the-art performance on self-supervised optical flow prediction." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.519, + 0.92, + 0.889 + ], + "angle": 0, + "content": "Event-based motion estimation methods can be categorized according to the complexity of the motion considered (i.e., number of degrees-of-freedom (DOF)) and to the solution strategy: model-based or learning-based. Low-DOF motions arise in sparse feature tracking and ego-motion estimation, while high-DOF motions describe more complex scenes, e.g., via per-pixel displacement (i.e., densely, over the whole image plane). Focusing on the latter, the event-based optical flow problem has been extensively studied on mobile robotic datasets [23, 75], where GT is calculated as the motion field (from known depth and poses). Most approaches use this GT for direct supervision of dense flow prediction [22, 24, 40, 57, 65]. Alternatively, several works use a contrast loss [18, 19, 77], which allows training in a self-supervised manner. However, the contrast loss is prone to undesired local optima, called event collapse [52, 53], where many events are warped into a few pixels or lines. Evaluation of event-based optical flow has been mostly done on the MVSEC [75] and DSEC [23] datasets, which comprise largely uniform motions and test intervals from 0.022s to 0.1s. Recently, [25] has proposed a supervised method to predict pixel displacement over a larger duration (0.5s), by leveraging synthetic data. It relies on the generation of training event data from images, where current tools [21] are not as mature as frame-based simulation, therefore it suffers from a sim-to-real performance gap." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.891, + 0.922, + 0.951 + ], + "angle": 0, + "content": "In short, progress has been made and the research field is moving towards predicting motion over longer time intervals on the whole image plane, i.e., taking on more complex (i.e., nonlinear) motion problems. This is precisely the" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.476, + 0.016, + 0.839, + 0.034 + ], + "angle": 0, + "content": "Motion-prior Contrast Maximization" + }, + { + "type": "page_number", + "bbox": [ + 0.904, + 0.017, + 0.921, + 0.031 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.056, + 0.921, + 0.193 + ], + "angle": 0, + "content": "problem tackled in this paper (Fig. 1): long-time and dense event-based motion estimation, reducing the domain adaptation gap of previous approaches. It comes with several associated challenges, mainly overcoming the lack of large labeled datasets, and dealing with event noise and data association (events depend on motion, and for large motions, the appearance of \"corresponding\" events can be wildly different due to changes in motion direction, occlusions, etc.) while leveraging the space-time characteristics of event data." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.194, + 0.921, + 0.369 + ], + "angle": 0, + "content": "To this end, we propose tackling the problem in two stages, by leveraging both supervised and self-supervised strategies: first, using supervised learning on synthetic data to provide initial model weights for motion estimation, and secondly, fine-tuning the network on real data via a self-supervised loss to reduce the domain adaptation gap. Our technical contributions involve extending the contrast loss framework to regress continuous-time trajectories over long time intervals via motion priors (parametric functions that provide a good balance between motion generality and regularization [64, 66]). This includes a solution to accurately and efficiently associate events to the trajectories (Fig. 2)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.37, + 0.921, + 0.583 + ], + "angle": 0, + "content": "More specifically, finding the association between events and motion trajectories (to warp corresponding events and achieve event alignment) is a high-dimensional problem (e.g., for a time window of 0.3 s one can consider about ten million events, and as many trajectories as pixels), which needs to be implemented in a differentiable and parallelizable manner. We propose two actions to cope with these technical challenges. First, we relax the problem by interpolating over a coarse spatio-temporal displacement field, which serves as a lookup table. Secondly, we use a symbolic matrix framework [15] to calculate the \\( K = N_{\\mathrm{traj}} \\) nearest neighbor (KNN) trajectories for each grid point, thereby solving KNN in a memory-efficient and differentiable way on GPU. The warp displacement at the grid point is then set to the average of the neighboring trajectories." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.584, + 0.922, + 0.741 + ], + "angle": 0, + "content": "Our approach is versatile, allowing for different types of networks and trajectories. Hence, we evaluate its performance on two applications: dense continuous-time motion estimation and optical flow. The results on EVIMO2 [7] show that fine-tuning with our self-supervised loss improves the zero-shot performance of a model pre-trained on synthetic data by \\(29\\%\\). On DSEC, our model shows state-of-the-art performance among self-supervised methods, on average improving the angular error by \\(19\\%\\) and the percentage of inliers by \\(14\\%\\), while having a \\(5\\times\\) faster inference time. In summary, our contributions are (Fig. 1):" + }, + { + "type": "text", + "bbox": [ + 0.087, + 0.754, + 0.92, + 0.81 + ], + "angle": 0, + "content": "1. We introduce motion priors (parametric functions with a good balance between generality and regularization) in the event-based contrast maximization framework for continuous-time and dense motion estimation." + }, + { + "type": "text", + "bbox": [ + 0.085, + 0.812, + 0.938, + 0.889 + ], + "angle": 0, + "content": "2. We combine the self-supervised loss with Bflow [25], the current top-performing supervised model for dense continuous-time event-based motion estimation trained on synthetic data, and show that it can be used to improve over zero-shot performance on unseen real data (EVIMO2)." + }, + { + "type": "text", + "bbox": [ + 0.086, + 0.891, + 0.922, + 0.947 + ], + "angle": 0, + "content": "3. We show that the combination of our loss with a simple U-Net architecture (à la EV-FlowNet [76]) achieves state-of-the-art performance among the self-supervised methods on the DSEC Optical Flow benchmark." + }, + { + "type": "list", + "bbox": [ + 0.085, + 0.754, + 0.938, + 0.947 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.08, + 0.018, + 0.095, + 0.03 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.412, + 0.017, + 0.581, + 0.032 + ], + "angle": 0, + "content": "F. Hamann et al." + }, + { + "type": "title", + "bbox": [ + 0.081, + 0.054, + 0.331, + 0.075 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.137, + 0.92, + 0.313 + ], + "angle": 0, + "content": "Frame-based Motion Estimation. Progress in deep learning triggered a large series of learned methods for optical flow estimation in classical, frame-based computer vision [2, 14, 32, 33, 60, 62]. The solutions rely on large-scale synthetic datasets [14, 42, 59, 73]. Similarly related is the task of point tracking, which has shown impressive progress on frame-based data [13, 31, 36, 50, 74], equally relying on simulated data [8, 12, 74]. This approach is unsatisfying, as it is prone to out-of-distribution (OOD) problems that cannot be easily overcome. As an alternative, self-supervised methods have been explored [34, 35, 48, 72] and shown to improve models pre-trained on simulated data for motion estimation tasks [58, 61]." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.321, + 0.92, + 0.458 + ], + "angle": 0, + "content": "Event-based Optical Flow. Event cameras [39, 46] are a relatively new technology, and have found applications in various computer vision domains, like mobile robotics [27, 45, 68], scene understanding [29, 30], and computational imaging [55, 63]. Their exploration for low-level vision tasks has taken a similar trajectory in a compressed timeline as previously frame-based methods [17]. The first event-based optical flow methods were model-based [3, 4, 6, 41, 43, 54], followed by learning-based approaches [11, 24, 28, 38, 44, 76, 77]." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.465, + 0.92, + 0.64 + ], + "angle": 0, + "content": "The scarce availability of event data and less mature simulation technology compared to frame-based cameras are major obstacles [21]. Therefore, event-based optical flow has been mostly evaluated on data acquired through egomotion, such as the robotics dataset MVSEC [75], the driving dataset DSEC [23] or the M3ED dataset [9]. However, GT is calculated using the motion field equation (with data from a synchronized depth sensor) and therefore does not provide an accurate flow at the event rate, nor spatially at occlusions and independently moving objects (IMOs). This GT is used to train supervised learning approaches [22, 24, 40, 57, 65], which inherit limitations from the ground truth." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.648, + 0.92, + 0.785 + ], + "angle": 0, + "content": "Self-supervised methods for event-based optical flow lessen the dependency on GT labels by leveraging an event alignment error [28,54,70,77]. Notably, variations of the contrast loss have been proposed [18], however, so far the application has been limited to either low-DOF problems (e.g., feature tracking [10, 19, 51], ego-motion [19, 20, 37]) or high-DOF but short-time optical flow problem [44, 54] (max. 0.1 s). We expand the frontier to the challenging problem of long-time and high-DOF (complex) motion estimation, by exploiting trajectory priors." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.793, + 0.92, + 0.948 + ], + "angle": 0, + "content": "Trajectory Prior. Using trajectories as motion priors has been a widely explored scheme in computer vision. Specifically, [1, 64, 78] propose a linear combination of basis functions for structure from motion. More recently it has been explored for dynamic novel view synthesis [66]. In the context of events, [63] uses cubic motion splines for video frame interpolation, [25] uses B-splines for supervised learning of non-linear optical flow, [67] use learned basis functions for event-based video decompression, and [10, 51] show event-based feature tracking (low-DOF) using Bezier or B-spline curves in combination with a contrast loss." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.477, + 0.016, + 0.839, + 0.034 + ], + "angle": 0, + "content": "Motion-prior Contrast Maximization" + }, + { + "type": "page_number", + "bbox": [ + 0.904, + 0.017, + 0.919, + 0.03 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.052, + 0.923, + 0.355 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.367, + 0.923, + 0.513 + ], + "angle": 0, + "content": "Fig. 2: Pipeline overview. (a) Input events in a time interval are (b) voxelized and (c) passed to an artificial neural network that predicts per-pixel coefficients for continuous-time trajectories (d). The raw events and predicted trajectories are fed to the loss module (e). Here, a dense spatio-temporal displacement map is interpolated, and events are warped according to their looked-up displacement. Lastly, an image of warped events (IWE) is built at a random reference time and its gradient magnitude acts as training loss. Note that the prediction method displayed here is specific to the used Bflow backbone [25], with additional events before the prediction start time \\( t_s \\) as input." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.543, + 0.32, + 0.566 + ], + "angle": 0, + "content": "3 Methodology" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.579, + 0.317, + 0.597 + ], + "angle": 0, + "content": "3.1 Contrast Loss" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.606, + 0.923, + 0.768 + ], + "angle": 0, + "content": "Contrast Maximization (CM) [18, 19] is used for self-supervised training. The idea is to find the point trajectories on the image plane best aligning with the events by maximizing the sharpness of warped events. It is an iterative approach, with three steps per iteration: (i) Each event \\( e_k = (\\mathbf{x}_k, t_k, p_k) \\) contains the pixel coordinates \\( \\mathbf{x}_k = (x_k, y_k)^\\top \\), timestamp \\( t_k \\) and polarity \\( p_k \\) of a brightness change of predefined size \\( C \\) (contrast sensitivity). Events \\( \\mathcal{E} = \\{e_k\\}_{k=1}^{N_e} \\) are displaced according to a candidate motion hypothesis \\( \\mathbf{x}_k' = \\mathbf{W}(\\mathbf{x}_k, t_k; \\boldsymbol{\\theta}) \\), with parameters \\( \\boldsymbol{\\theta} \\), to a reference time \\( t_\\mathrm{ref} \\), producing a set of warped events \\( \\mathcal{E}_{t_\\mathrm{ref}}' = \\{e_k'\\}_{k=1}^{N_e} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.288, + 0.775, + 0.918, + 0.803 + ], + "angle": 0, + "content": "\\[\ne _ {k} \\doteq \\left(\\mathbf {x} _ {k}, t _ {k}, p _ {k}\\right) \\stackrel {\\mathrm {W}} {\\mapsto} e _ {k} ^ {\\prime} \\doteq \\left(\\mathbf {x} _ {k} ^ {\\prime}, t _ {\\text {r e f}}, p _ {k}\\right). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.81, + 0.912, + 0.831 + ], + "angle": 0, + "content": "Afterwards, \\((ii)\\) the events are summed into an image of warped events (IWE)," + }, + { + "type": "equation", + "bbox": [ + 0.315, + 0.838, + 0.918, + 0.865 + ], + "angle": 0, + "content": "\\[\nI (\\mathbf {x}; \\mathcal {E} _ {t _ {\\text {r e f}}} ^ {\\prime}, \\boldsymbol {\\theta}) \\doteq \\sum_ {k = 1} ^ {N _ {e}} \\mathcal {N} (\\mathbf {x}; \\mathbf {x} _ {k} ^ {\\prime}, \\sigma^ {2}), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.924, + 0.951 + ], + "angle": 0, + "content": "which essentially counts the number of warped events \\( e_k' \\) per pixel. Lastly (iii) the contrast or sharpness of the IWE is computed (e.g., the gradient magnitude of the IWE, \\( G \\doteq \\int \\| \\nabla I(\\mathbf{x}) \\| d\\mathbf{x} \\)), which serves as a proxy for how well the model (motion hypothesis and \\( \\pmb{\\theta} \\)) fit with the events produced by the true motion." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.08, + 0.017, + 0.095, + 0.031 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.41, + 0.016, + 0.582, + 0.032 + ], + "angle": 0, + "content": "F. Hamann et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.054, + 0.904, + 0.073 + ], + "angle": 0, + "content": "Table 1: Comparison of the contrast loss formulation in the most recent methods." + }, + { + "type": "table", + "bbox": [ + 0.085, + 0.089, + 0.916, + 0.235 + ], + "angle": 0, + "content": "
StepsShiba et. al [54]Paredes et. al [44]Ours
Warp displacement (3):Δxk=(tref−tk)v(ek)Δxk=∑i(Δti v_i)(ek)Δxk=1/Ntraj ∑n Δxn
linear trajectoryflow concatenationnon-linear trajectory
IWEI(x)=∑kN(x; x′k, σ2)T±(x)=∑ktkN(x; x′k, σ2)/∑kN(x; x′k, σ2)I(x)=∑kN(x; x′k, σ2)
Reference time tref:{t0, t0.5, t1}multi-partition p, multi-time-scale s~U(0,1)
Contrast objectiveG(t0)+2G(t0.5)+G(t1)/4G(v=0)1/S ∑s=0S-11/2s ∑p=02s-1LRCM,pG(tref)
" + }, + { + "type": "title", + "bbox": [ + 0.076, + 0.272, + 0.882, + 0.293 + ], + "angle": 0, + "content": "3.2 Estimating Continuous-Time and Dense Motion Trajectories" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.304, + 0.918, + 0.343 + ], + "angle": 0, + "content": "The CM framework has been extended in several works. Table 1 compares recent extensions for optical flow estimation, where events are warped as" + }, + { + "type": "equation", + "bbox": [ + 0.409, + 0.358, + 0.918, + 0.38 + ], + "angle": 0, + "content": "\\[\n\\mathbf {x} _ {k} ^ {\\prime} = \\mathbf {x} _ {k} + \\Delta \\mathbf {x} _ {k}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.394, + 0.92, + 0.548 + ], + "angle": 0, + "content": "The main differences between these approaches lie in the event displacement (i.e., warp) model, the reference times used, and the loss function (i.e., event-alignment metric), with the design choices having two goals: fitting the event data and regularizing the solution (e.g., avoiding event collapse [52]). Previous methods used mostly a linear model [54,56] or partitioned the inference interval into smaller flow intervals so that its concatenation need not be linear [44]. By contrast, we introduce an explicit continuous-time non-linear trajectory model, a random reference time per iteration, and an easier contrast loss." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.55, + 0.933, + 0.685 + ], + "angle": 0, + "content": "Trajectory Representation. Figure 2 shows an overview of the training pipeline. Events in the time interval \\([t_s, t_e]\\) are fed into a neural network that makes per-pixel predictions of the parameters (coefficients) \\(\\pmb{\\theta} \\equiv \\pmb{\\alpha}_n = (\\alpha_{n,1}, \\dots, \\alpha_{n,N_c})^\\top \\in \\mathbb{R}^{N_c}\\) of a continuous-time trajectory \\(\\mathbf{q}_n(t; \\pmb{\\alpha}_n) \\equiv \\mathbf{q}_n(t) = (x_n(t), y_n(t))^\\top\\). There is one trajectory per pixel (i.e., \"dense\" character), \\(n = 1, \\dots, N_p\\), where \\(N_p = h w\\) is the number of pixels (image height \\(h\\) and width \\(w\\)). Thus, \\(n\\) is the spatial index of the trajectory, identifying it among all trajectories on the image plane." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.686, + 0.922, + 0.793 + ], + "angle": 0, + "content": "We model trajectories as weighted combinations of basis functions, \\(\\mathbf{q}_n(t) = \\sum_{j=1}^{N_c} g_j(t) \\mathbf{p}_{n,j}\\), where \\(g_j(t)\\) are temporal basis shares by all trajectories, and \\(\\mathbf{p}_{n,j} = (\\alpha_{n,j}^x, \\alpha_{n,j}^y)^\\top\\) are \"control points\" (we write \\(x\\) and \\(y\\) components explicitly with separate coefficients in \\(\\alpha_n\\)). We investigate polynomial basis \\(g_j(t) = t^j\\), Bézier curves with basis \\(g_j(t) = \\binom{N_c}{j}(1 - t)^{N_c - j}t^j\\), as well as a learned basis." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.793, + 0.922, + 0.952 + ], + "angle": 0, + "content": "Spatio-Temporal Event Warping. The trajectories can be used to warp events. This would require finding the trajectory that passes through the spacetime coordinates of the event and then finding the value of the trajectory at the reference time (i.e., the warped event location). However, the association between events and trajectories is unknown and needs to be estimated simultaneously with the parameters (i.e., shape) of the trajectories. We circumvent the problem by using a soft association between events and trajectories. Each event \\( e_k \\) is associated with its \\( N_{\\mathrm{traj}} \\) nearest neighboring trajectories \\( \\{\\mathbf{q}_n\\}_{n=1}^{N_{\\mathrm{traj}}} \\)," + } + ], + [ + { + "type": "header", + "bbox": [ + 0.476, + 0.016, + 0.839, + 0.034 + ], + "angle": 0, + "content": "Motion-prior Contrast Maximization" + }, + { + "type": "page_number", + "bbox": [ + 0.904, + 0.017, + 0.92, + 0.03 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.057, + 0.921, + 0.138 + ], + "angle": 0, + "content": "and the event displacement \\(\\varDelta\\mathbf{x}_k\\) in (3) is computed as the average of the respective trajectory displacements \\(\\{\\varDelta\\mathbf{x}_k^n\\}_{n=1}^{N_{\\mathrm{traj}}}\\). The trajectory displacement \\(\\varDelta\\mathbf{x}_k^n\\) is defined as the difference of trajectory locations at the time of the event and the reference time. The event displacement is defined as the mean:" + }, + { + "type": "equation", + "bbox": [ + 0.198, + 0.156, + 0.919, + 0.209 + ], + "angle": 0, + "content": "\\[\n\\Delta \\mathbf {x} _ {k} \\doteq \\frac {1}{N _ {\\mathrm {t r a j}}} \\sum_ {n = 1} ^ {N _ {\\mathrm {t r a j}}} \\Delta \\mathbf {x} _ {k} ^ {n}, \\quad \\text {w i t h} \\quad \\Delta \\mathbf {x} _ {k} ^ {n} \\doteq \\mathbf {q} _ {n} \\left(t _ {\\mathrm {r e f}}\\right) - \\mathbf {q} _ {n} \\left(t _ {k}\\right). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.227, + 0.921, + 0.382 + ], + "angle": 0, + "content": "Loss Calculation. Once the events are warped using (3), it is straightforward to compute the IWE. We adopt the magnitude of the IWE gradient as loss function [18,56]. Moreover, we use the contrast loss in a self-supervised learning setting, choosing a different reference time \\( t_{\\mathrm{ref}} \\), uniformly sampled in the observation interval, for every batch during training. This simplifies the objective to a single warping operation and loss calculation (e.g., compared to the three warping operations and loss calculations in the optimization-based approach by [54,56]), while it has added regularization benefits, as mentioned below." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.383, + 0.92, + 0.539 + ], + "angle": 0, + "content": "Memory Effective Computation of the Displacement Field. The number of events and trajectories can be very large, and even more their combination. Calculating KNN for every event can quickly become computationally unfeasible, and traditional algorithms cannot be efficiently implemented in deep-learning frameworks. We propose calculating the per-event displacement \\(\\Delta \\mathbf{x}_k\\) in (3) by first interpolating a dense but coarser spatio-temporal displacement field, and then looking up the per-event displacement in such space-time volume. Moreover, we relax the problem by solving the KNN search in 2D instead of in the volume." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.54, + 0.92, + 0.773 + ], + "angle": 0, + "content": "Figure 2 shows an overview of the interpolation process. The displacement map is a tensor of shape \\([N_{\\mathrm{bins}}, h/4, w/4]\\), where \\(N_{\\mathrm{bins}}\\) is the number of temporal bins, and \\(h\\) and \\(w\\) are the sensor's height and width, respectively. For each temporal bin-center \\(t_{\\mathrm{bins}}\\) the trajectories \\(\\mathbf{q}_n(t = t_{\\mathrm{bins}})\\) are calculated and the KNN interpolation is performed between each pixel of a channel \\(\\mathbf{x}[t_{\\mathrm{bins}}]\\) and \\(\\mathbf{q}_n(t = t_{\\mathrm{bins}})\\). We implement the KNN approach using KeOps [15], a framework for symbolic matrix computation, which provides a memory-efficient and differentiable solution. We perform a KNN search for every voxel in the table, and events are warped after a lookup operation. Please note that this step is entirely independent of the voxel grid passed as input to the backbone, and all raw events are used to calculate the contrast loss. From the calculated displacement map, we can directly look up per-event displacements \\(\\Delta \\mathbf{x}_k\\)." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.774, + 0.92, + 0.949 + ], + "angle": 0, + "content": "Regularization. Objectives based on event alignment are prone to undesired local minima called \"event collapse\" [52,53]. Our formulation inherently provides regularization, temporally by the smoothness of the motion prior, and spatially by interpolating the event flow from several trajectories via a soft assignment. Our framework uses two additional regularization sources. Firstly, we penalize the magnitude of the spatial gradient of the interpolated displacement field between consecutive timesteps, \\( R \\doteq \\| \\nabla (\\Delta \\mathbf{q}_{*}(t))\\|_{L^{1}(\\Omega)} \\). To interpolate \\( \\mathbf{q}_{*} \\) from the sparse trajectories \\( \\mathbf{q}_n \\), we use the same KNN indices as for the main volume. This loss encourages spatial smoothness of the trajectories." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.08, + 0.017, + 0.095, + 0.031 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.41, + 0.015, + 0.582, + 0.032 + ], + "angle": 0, + "content": "F. Hamann et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.329, + 0.053, + 0.668, + 0.073 + ], + "angle": 0, + "content": "Table 2: Results on EVIMO2 [7]." + }, + { + "type": "table", + "bbox": [ + 0.199, + 0.088, + 0.809, + 0.217 + ], + "angle": 0, + "content": "
MethodTEPE ↓TAE ↓%Out ↓
Paredes et. al [44]21.6951.910.63
E-RAFT (linear) [24]19.3874.520.66
BFlow (zero-shot) [25]8.6319.940.36
BFlow (in-domain)3.3811.680.17
Ours (self-supervised)6.1416.980.25
" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.254, + 0.921, + 0.352 + ], + "angle": 0, + "content": "Secondly, we use a multi-reference formulation by randomly sampling a reference time at every training step, while previous work used multiple and fixed timestamps (Tab. 1). This has the advantage that \\((i)\\) IWEs are required to be sharp at truly any time and \\((ii)\\) the memory requirement is lowered as it scales linearly with the number of reference times used during a training step." + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.353, + 0.457, + 0.371 + ], + "angle": 0, + "content": "In summary, the training loss is" + }, + { + "type": "equation", + "bbox": [ + 0.412, + 0.387, + 0.919, + 0.409 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = 1 / G + \\lambda R, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.424, + 0.933, + 0.56 + ], + "angle": 0, + "content": "with \\(\\lambda > 0\\) as regularization weight. Further details are in the supplementary. Pre-processing and Prediction-Module Architecture. Events record brightness changes asynchronously, in the form of a sparse spatio-temporal signal. As input to the Prediction Module, events are customarily converted to voxel grids [77] for compatibility with conventional artificial neural networks. An event volume is discretized in the time dimension, and each voxel counts the number of events within it (bi-linearly voted for undistorted, rectified events)." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.561, + 0.922, + 0.64 + ], + "angle": 0, + "content": "In principle, our self-supervised loss module can be paired with any segmentation or optical flow network architecture. We use a U-Net architecture for the experiments on DSEC (Sec. 4.2), and the architecture in the recent Bflow method [25], inspired by RAFT [62], for the experiments on EVIMO2 (Sec. 4.1)." + }, + { + "type": "title", + "bbox": [ + 0.076, + 0.67, + 0.312, + 0.691 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.71, + 0.92, + 0.85 + ], + "angle": 0, + "content": "We test our method on two applications. First, we test the capabilities for nonlinear trajectory estimation on the real-world dataset EVIMO2 [7], with additional results on the synthetic MultiFlow dataset [25], which we use for pretraining (Sec. 4.1). Secondly, we evaluate our method on the DSEC dataset [23, 24] because this has been the previous frontier for event-based optical flow estimation and it allows for direct comparison with prior work (Sec. 4.2). Additional results, including on the MVSEC dataset [75], are given in the supplement." + }, + { + "type": "title", + "bbox": [ + 0.076, + 0.877, + 0.699, + 0.896 + ], + "angle": 0, + "content": "4.1 Results on Non-Linear Trajectory Estimation" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.909, + 0.922, + 0.951 + ], + "angle": 0, + "content": "EVIMO2 Continuous Flow Dataset (CF-EVIMO2). MultiFlow [25] was first proposed to evaluate long-term \"optical flow\" methods. The dataset consists" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.477, + 0.016, + 0.839, + 0.034 + ], + "angle": 0, + "content": "Motion-prior Contrast Maximization" + }, + { + "type": "page_number", + "bbox": [ + 0.904, + 0.017, + 0.92, + 0.03 + ], + "angle": 0, + "content": "9" + }, + { + "type": "image", + "bbox": [ + 0.153, + 0.052, + 0.845, + 0.552 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.564, + 0.922, + 0.655 + ], + "angle": 0, + "content": "Fig. 3: Visualization of predicted trajectories on EVIMO2 data. \\(GT\\): Ground truth. In-domain: fine-tuned on EVIMO2 using GT (supervised). Zero-shot: network trained only on synthetic data (out-of-domain prediction). Ours: Pre-trained on synthetic data, fine-tuned with self-supervised loss. Note that supervision in-domain is often impossible in practice because dense trajectory labels for real data are difficult to obtain." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.695, + 0.938, + 0.831 + ], + "angle": 0, + "content": "of synthetic videos generated with Internet images as foreground and Flickr30K [71] images as background. Events are generated with a simulator [47] by rendering the synthetic scenes at 1000 frames/s. However, due to the sim-to-real gap between real and synthetic events, it is challenging to infer the actual performance of the proposed and baseline methods on real datasets. Furthermore, Multiflow generates trajectories by pasting 2D foreground objects onto background images, which lacks the challenging cases of self-occlusion due to 3D rotation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.832, + 0.925, + 0.95 + ], + "angle": 0, + "content": "To address these data limitations, we present the EVIMO2 Continuous Flow Dataset based on the full 3D data and event data provided by the meticulously designed EVIMO2 [7] dataset. The EVIMO2 GT data provides high-quality 3D scans of the objects, camera poses, object poses, and camera intrinsics, which we use to compute optical flow GT as follows. Given a point \\( P_{t}^{i} \\) on the object represented in the camera coordinate system at time \\( t \\), we project it onto a 2D point" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.08, + 0.016, + 0.105, + 0.032 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.411, + 0.015, + 0.584, + 0.032 + ], + "angle": 0, + "content": "F. Hamann et al." + }, + { + "type": "image", + "bbox": [ + 0.324, + 0.054, + 0.681, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.265, + 0.921, + 0.304 + ], + "angle": 0, + "content": "Fig. 4: End-point-error vs. prediction time span for three methods: in-distribution, out-of-distribution and self-supervised (Ours). Using the Bézier curve results from Tab. 3." + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.319, + 0.921, + 0.356 + ], + "angle": 0, + "content": "Table 3: Sensitivity with respect to motion prior on EVIMO2 [7]. \\(SL\\), \\(ID\\): supervised, in-distribution, \\(SL\\), \\(OOD\\): supervised, out of distribution, \\(SSL\\): self-supervised." + }, + { + "type": "table", + "bbox": [ + 0.204, + 0.371, + 0.803, + 0.589 + ], + "angle": 0, + "content": "
MethodTEPE ↓TAE ↓%Out ↓
SL, IDBFlow (polyn.)3.5113.260.18
BFlow (learned)3.7813.870.19
BFlow (Bézier)3.3811.680.17
SL, OODBFlow (polyn.)9.3620.880.36
BFlow (learned)8.6619.640.35
BFlow (Bézier) [25]8.6319.940.36
SSLOurs, polyn.6.7819.760.27
Ours, learned7.4619.780.28
Ours, Bézier6.1416.980.25
" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.622, + 0.92, + 0.799 + ], + "angle": 0, + "content": "\\(p_{t}^{i} = \\pi (P_{t}^{i})\\), where the \\(\\pi (\\cdot)\\) function represents the perspective projection, including the camera intrinsics. The object poses at time \\(t\\) and \\(t + 1\\) are provided as \\(T_{o}^{c_{t}}\\), \\(T_{o}^{c_{t + 1}}\\) in the GT data. The rigid-body transformation \\(T_{c_t}^{t + 1} = (T_o^{c_t + 1})^{-1}T_o^{c_t}\\) maps 3D points in the camera frame at time \\(t\\) to the camera frame at time \\(t + 1\\) by aligning the shared object coordinate frame. The \"flow\" (or dense motion, since it may not be linear) is defined as \\(\\Delta p_{t\\rightarrow t + 1}^i\\doteq \\pi (T_{c_t}^{c_{t + 1}}P_t^i) - \\pi (P_t^i)\\). For each sequence, we generate GT dense motion every 10ms for 300ms. We mask out areas where GT object masks are unavailable. Experiments are carried out on the IMO subset of EVIMO2 using the official train and test splits." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.8, + 0.922, + 0.918 + ], + "angle": 0, + "content": "Metrics. Ground truth is provided as dense motion from \\( t = 0 \\) to several timesteps in increasing order. We consider \\( N_{s} = 6 \\) timestamps (i.e., subintervals) in a 300ms window and evaluate the quality of predictions with direct extension of the common optical flow metrics end-point-error (EPE), angular error (AE), and percentage of predicted vectors above a specific EPE threshold (\\%Out). We term the corresponding trajectory metrics TEPE and TAE, where" + }, + { + "type": "equation", + "bbox": [ + 0.249, + 0.926, + 0.919, + 0.955 + ], + "angle": 0, + "content": "\\[\n\\mathrm {T E P E} = \\frac {1}{N _ {s}} \\sum_ {k = 1} ^ {N _ {s}} \\mathrm {E P E} (\\varDelta \\mathbf {x} _ {\\mathrm {p r e d}} (t _ {k}), \\varDelta \\mathbf {x} _ {\\mathrm {g t}} (t _ {k})). \\tag {6}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.476, + 0.016, + 0.839, + 0.034 + ], + "angle": 0, + "content": "Motion-prior Contrast Maximization" + }, + { + "type": "page_number", + "bbox": [ + 0.894, + 0.017, + 0.918, + 0.031 + ], + "angle": 0, + "content": "11" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.056, + 0.832, + 0.077 + ], + "angle": 0, + "content": "The number of outliers is calculated on TEPE with a threshold of 3px." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.079, + 0.92, + 0.157 + ], + "angle": 0, + "content": "Implementation Details. We use the BFlow network architecture and test different motion priors. Specifically, we report results with polynomial, Bezier (with \\( N_{c} = 10 \\)) and learned basis \\( g_{j}(t) \\) (a small dense neural network with three layers and hidden dimension 64 is trained alongside the main network)." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.159, + 0.92, + 0.237 + ], + "angle": 0, + "content": "The contrast loss module uses one trajectory per \\(4 \\times 4\\) px, and similarly \\(4 \\times 4\\) px in the displacement volume. The weight for the spatial smoothness term is \\(\\lambda = 0.003\\), and the number of neighbors in the KNN approach is \\(N_{\\mathrm{traj}} = 32\\). The loss function \\(G\\) is the \\(L^1\\) norm of the IWE's gradient magnitude [18,56]." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.239, + 0.92, + 0.452 + ], + "angle": 0, + "content": "Training Schedule. The model is first pre-trained on MultiFlow (\\(\\approx 12000\\) samples) for 50 epochs, with a batch size of 10 with an \\(L^1\\) loss on the GT trajectory flow. We include data augmentation by flipping (horizontal and vertical) and cropping and use the training/test split provided in [25]. These weights are what we refer to as zero-shot or out-of-distribution (OOD) for the main comparison on EVIMO2. Afterwards, the pre-trained model is fine-tuned with two different losses on EVIMO2. Training using our self-supervised loss is carried out for 15 epochs with a batch size of 6. We refer to this method as Ours. Lastly, a second version is trained directly on the GT flow of EVIMO for 50 epochs, which we refer to as in-distribution (ID). All experiments are performed on Nvidia RTX A6000 GPUs with an AdamW optimizer and a learning rate of \\(10^{-4}\\)." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.455, + 0.92, + 0.63 + ], + "angle": 0, + "content": "Baselines. This is the first usage of EVIMO2 for dense non-linear flow/motion. For comparison, we provide additional baselines. We use ERAFT [24] with the provided DSEC weights to infer linear flow over the whole prediction time. The timestamps of the intermediate flow are interpolated from the linear prediction. Additionally, we provide results for the prediction of Paredes et. al. [44]. The network is trained self-supervised on DSEC and performs recurrent prediction steps in time in intervals of \\(10\\mathrm{ms}\\), inferring the total flow at every step by accumulation of the shorter flows. We unroll this prediction over the whole interval and take the accumulated flow after each \\(50\\mathrm{ms}\\) step as the predicted flow." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.633, + 0.92, + 0.789 + ], + "angle": 0, + "content": "Results. Table 2 presents the comparison of the three different training modes. The results of our self-supervised loss can improve the zero-shot performance by nearly \\(30\\%\\). While direct supervision on EVIMO2 GT data delivers the best results, note that this is only possible because we are comparing on a dataset that was recorded with a motion capture setup and exact object models. Under real conditions, pixel-level annotations of motion tracks are very difficult to obtain. Therefore, our loss delivers a promising tool in combination with synthetic pretraining, substantially helping in overcoming the sim-to-real gap." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.791, + 0.92, + 0.888 + ], + "angle": 0, + "content": "Figure 3 shows the qualitative comparison of GT tracks and the three training schedules. It delivers insights into the failure cases that our method can overcome. Specifically, the zero-shot model shows errors like overseen object motion (Examples 2 and 3), wrong scale (Ex. 5), and prediction of non-existing motion (Ex. 6), which are improved with self-supervised domain training." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.891, + 0.92, + 0.949 + ], + "angle": 0, + "content": "Additionally, Fig. 4 shows the error for different prediction time spans. Intuitively, the error increases for longer intervals. The visualization shows that our loss improves zero-shot performance, especially for the long prediction times." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.08, + 0.016, + 0.105, + 0.032 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.41, + 0.016, + 0.584, + 0.032 + ], + "angle": 0, + "content": "F. Hamann et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.202, + 0.054, + 0.795, + 0.073 + ], + "angle": 0, + "content": "Table 4: Results on the DSEC optical flow benchmark [24]." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.089, + 0.919, + 0.194 + ], + "angle": 0, + "content": "
Method\\( {t}_{\\text{inf }}\\left\\lbrack \\mathrm{{ms}}\\right\\rbrack \\)Allinterlaken_00_binterlaken_01_athun_01_a
EPE \\( \\downarrow \\)AE \\( \\downarrow \\)%Out ↓FWL ↑EPE \\( \\downarrow \\)AE \\( \\downarrow \\)%Out ↓FWL ↑EPE \\( \\downarrow \\)AE \\( \\downarrow \\)%Out ↓FWL ↑EPE \\( \\downarrow \\)AE \\( \\downarrow \\)%Out ↓FWL ↑
E-RAFT (SL) [24]46.330.7910.562.681.291.396.226.121.320.906.883.911.420.659.751.871.20
IDNet (SL) [69]0.722.722.04-1.252.114.35-0.772.252.60-0.572.661.47-
Paredes et al. (SSL) [44]40.102.3310.5617.77-3.346.2225.72-2.496.8819.15-1.739.7510.39-
RTEF (MB) [5]4.88-41.952.518.59-59.842.895.94-47.332.923.01-29.702.39
Ev-FlowNet (SSL) [27]3.86-31.451.306.32-47.951.464.91-36.071.422.33-20.921.32
MultiCM (MB) [54]\\( {9.9} \\cdot {10}^{3} \\)3.4713.9830.861.375.749.1938.931.503.749.7731.371.512.1211.0617.681.24
Ours (poly, \\( k = 1 \\) )7.273.208.5315.211.463.214.8920.451.582.385.4617.401.701.396.997.361.30
" + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.195, + 0.918, + 0.295 + ], + "angle": 0, + "content": "
Methodthun_01_bzurich_city_12_azurich_city_14_czurich_city_15_a
EPE ↓AE ↓%Out ↓FWL ↑EPE ↓AE ↓%Out ↓FWL ↑EPE ↓AE ↓%Out ↓FWL ↑EPE ↓AE ↓%Out ↓FWL ↑
E-RAFT (SL) [24]0.588.411.521.180.6123.161.061.120.7110.231.911.470.598.881.301.34
IDNet (SL) [69]0.552.071.35-0.604.561.16-0.763.742.74-0.552.551.02-
Paredes et al. (SSL) [44]1.668.419.34-2.7223.1626.65-2.6410.2323.01-1.698.889.98-
RTEF (MB) [5]3.91-34.692.483.14-34.081.424.00-45.672.673.78-37.992.82
EV-FlowNet (SSL) [77]3.04-25.411.332.62-25.801.033.36-36.341.242.97-25.531.33
MultiCM (MB) [54]2.4812.0523.561.243.8628.6143.961.142.7212.6230.531.502.3511.8220.991.41
Ours (poly, k = 1)1.546.559.691.338.3320.1622.391.131.788.7912.991.561.456.278.341.51
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.333, + 0.921, + 0.412 + ], + "angle": 0, + "content": "Sensitivity to motion prior. Table 3 compares our method for different motion priors. While Bézier curves show slightly improved performance over the two basis function methods, self-supervised domain training can improve performance in all cases, which proves the robustness and generality of our method." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.444, + 0.58, + 0.464 + ], + "angle": 0, + "content": "4.2 Results on Optical Flow Estimation" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.48, + 0.92, + 0.558 + ], + "angle": 0, + "content": "Dataset, Metrics and Details. We compare our method on the DSEC optical flow benchmark [24]. The dataset consists of sequences from a Prophesee Gen3 event camera, with a resolution of \\(640 \\times 480\\) px, mounted on a driving car. Ground truth is calculated as the motion field from a co-deployed depth sensor." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.559, + 0.92, + 0.636 + ], + "angle": 0, + "content": "We compare the common flow metrics EPE, AE, and percentage of outliers (%Out), where an outlier is a flow vector with EPE \\(>3\\mathrm{px}\\). Additionally, we evaluate the Flow Warp Loss (FWL) [57]. FWL \\(>1\\) means that the IWE is sharper than pixel-wise event accumulation (i.e., IWE with zero optical flow)." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.637, + 0.922, + 0.95 + ], + "angle": 0, + "content": "We use a simple U-Net [49] as the backbone in the linear flow experiments (similar architecture like EV-FlowNet [76]) and choose polynomial basis \\( g_{j}(t) = t^{j} \\) of degree 1, which effectively leads to a linear motion prior. We use \\( N_{\\mathrm{traj}} = 32 \\) nearest neighbors and \\( N_{\\mathrm{bins}} = 15 \\) time bins in the displacement map. The weight for the regularizer is \\( \\lambda = 0.003 \\). The network trains for 50 epochs with and Adam optimizer, a learning rate of \\( 10^{-4} \\), and a total batch size of 28 on two RTX A6000. Results. Table 4 shows the DSEC benchmark results, confirming that our model can generate high-quality optical flow results without having access to any GT. It furthermore achieves the best results among all contrast-maximization-based methods on six out of seven sequences, and only methods directly supervised on GT perform better. Within the self-supervised methods, it improves the AE by \\( 19\\% \\) and the number of inliers by \\( 14\\% \\). The only exception is the night sequence zurich_city_14. Contrast maximization is based on the brightness constancy assumption, which is violated here by flickering street lights. Hence, in these areas, the contrast loss does not provide reasonable flow predictions. While our methods still provide a better AE and %Out at night, higher outliers in the" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.477, + 0.016, + 0.839, + 0.034 + ], + "angle": 0, + "content": "Motion-prior Contrast Maximization" + }, + { + "type": "page_number", + "bbox": [ + 0.894, + 0.017, + 0.92, + 0.031 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.087, + 0.053, + 0.913, + 0.33 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.343, + 0.918, + 0.362 + ], + "angle": 0, + "content": "Fig. 5: Results on DSEC. Image of warped events and predicted flow by three methods." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.402, + 0.92, + 0.441 + ], + "angle": 0, + "content": "flickering regions prevent the difference in the average EPE from reflecting the improved performance of our method on the other six sequences." + }, + { + "type": "text", + "bbox": [ + 0.074, + 0.442, + 0.922, + 0.675 + ], + "angle": 0, + "content": "Figure 5 shows qualitative results of our method. In comparison to another self-supervised optical flow method (d), the visualizations are noticeably sharper, allowing for a more precise representation of the motion in the scene. Moreover, the model performs well in delineating the contours of foreground objects, without the over-smoothing effect often observable in flow methods. The IWE (Fig. 5b) shows sharp results, aligning with the overall high FWL values reported in Tab. 4. While the FWL metric also increases in event collapse (e.g., for RTEF [5]), the IWE visualization reveals that here the model is well-regularized, confirming that the predicted motion aligns the events correctly. The third example in Fig. 5 highlights how our model has improved predictions with a single forward pass, where models relying on temporal recurrence (see (d)) fail at the beginning and need a warm-up stage resulting in higher latency." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.676, + 0.92, + 0.754 + ], + "angle": 0, + "content": "Inference time. Additionally, Tab. 4 provides inference times \\( t_{\\mathrm{inf}} \\) for several methods. Note that our method is about \\( 5 \\times \\) faster than the competitive baselines. The reason is that we do not rely on any recurrence in this optical flow application, such as recurrence in time or RAFT-inspired refinement steps." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.754, + 0.922, + 0.87 + ], + "angle": 0, + "content": "Ablation and Sensitivity. Table 5 lists the performance under different loss settings. It confirms that the number of neighboring trajectories \\( N_{\\mathrm{traj}} \\) in the KNN approach has a regularizing effect and is important for good performance. The decreased performance on the benchmark for a lower \\( N_{\\mathrm{traj}} \\) aligns with the visual impression in Fig. 6: models with lower \\( N_{\\mathrm{traj}} \\) show artifacts and a stronger susceptibility to the aperture problem, as is observable on the rock pattern." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.872, + 0.922, + 0.95 + ], + "angle": 0, + "content": "Table 5 confirms that the multi-reference approach is crucial for regularizing the contrast loss. While using three instead of one \\( t_{\\mathrm{ref}} \\) improves performance, our approach using a randomized \\( t_{\\mathrm{ref}} \\) leads to an additional performance boost. Lastly, the non-linear prior showed no performance improvement on DSEC." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.08, + 0.017, + 0.105, + 0.03 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.41, + 0.016, + 0.582, + 0.032 + ], + "angle": 0, + "content": "F. Hamann et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.054, + 0.921, + 0.109 + ], + "angle": 0, + "content": "Table 5: Sensitivity and Ablation Study for DSEC. Ours corresponds to \"Ours\" in Tab. 4. Configurations marked with “-” are unchanged from our main result. \\(\\Delta\\) specifies the change with respect to the original configuration." + }, + { + "type": "table", + "bbox": [ + 0.088, + 0.124, + 0.917, + 0.27 + ], + "angle": 0, + "content": "
NtrajNrefNcMotion priorEPE ↓ΔEPEAE ↓ΔAE%Out ↓Δ%Out
Ours32~U(0,1)1polynomial3.208.5315.21
Number of neighbor trajectories1---4.581.0713.735.2027.9012.69
8---3.510.3113.204.6723.398.18
Number of reference times-3--4.461.2614.435.9028.6013.39
-1--7.264.0618.7210.1944.7529.54
Type and degree of motion prior--5learned3.220.028.590.0615.340.13
--5polynomial3.280.068.610.0815.680.47
" + }, + { + "type": "image", + "bbox": [ + 0.083, + 0.283, + 0.49, + 0.404 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.283, + 0.91, + 0.404 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.123, + 0.418, + 0.871, + 0.438 + ], + "angle": 0, + "content": "Fig. 6: Visual effect of the number of neighbors \\( N_{\\mathrm{traj}} \\) on the predicted flow." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.472, + 0.296, + 0.492 + ], + "angle": 0, + "content": "5 Limitations" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.511, + 0.918, + 0.569 + ], + "angle": 0, + "content": "Like all CM-based methods, ours is based on the brightness constancy assumption. Therefore, it shows limitations in estimating flow from events that are not caused by motion, e.g., from flickering lights." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.57, + 0.92, + 0.627 + ], + "angle": 0, + "content": "Our KNN interpolation shows a good trade-off between granularity and regularization; nevertheless, our method is limited by the aperture problem inherent to optical flow." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.628, + 0.922, + 0.706 + ], + "angle": 0, + "content": "Lastly, during training, all raw events are passed to the loss module, increasing training time. Furthermore, gradient calculation is performed through the additional step of event warping. These steps increase memory requirements and training time of our loss compared to supervised methods." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.735, + 0.288, + 0.755 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.774, + 0.922, + 0.95 + ], + "angle": 0, + "content": "We introduced a new loss formulation based on the contrast maximization framework, by combining it with a non-linear trajectory prior. It is a versatile tool that works with a variety of model architectures and trajectory representations. We presented an efficient method to make the high-dimensional assignment between millions of events and thousands of trajectories feasible. Our experiments show clear advantages against supervised methods on unseen data, where under real-world circumstances no GT is available. Additionally, a U-Net trained with our loss shows state-of-the-art performance on the DSEC optical flow benchmark, while being substantially faster than the previously best methods." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.476, + 0.016, + 0.839, + 0.034 + ], + "angle": 0, + "content": "Motion-prior Contrast Maximization" + }, + { + "type": "page_number", + "bbox": [ + 0.894, + 0.017, + 0.919, + 0.031 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.055, + 0.338, + 0.077 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.096, + 0.924, + 0.234 + ], + "angle": 0, + "content": "We thank Dr. Cornelia Fermüller and the NeuroPAC network for fostering collaborations within the event-based community (NSF OISE 2020624). Funded by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) under Germany's Excellence Strategy - EXC 2002/1 \"Science of Intelligence\" - project number 390523135. We furthermore gratefully acknowledge the support by the following grants: NSF FRR 2220868, NSF IIS-RI 2212433, NSF TRIPODS 1934960, ONR N00014-22-1-2677." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.264, + 0.238, + 0.284 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.304, + 0.918, + 0.359 + ], + "angle": 0, + "content": "1. Akhter, I., Sheikh, Y., Khan, S., Kanade, T.: Trajectory space: A dual representation for nonrigid structure from motion. IEEE Trans. Pattern Anal. Mach. Intell. 33(7), 1442-1456 (jul 2011). https://doi.org/10.1109/TPAMI.2010.201" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.36, + 0.92, + 0.413 + ], + "angle": 0, + "content": "2. Bailer, C., Varanasi, K., Stricker, D.: CNN-based patch matching for optical flow with thresholded hinge embedding loss. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR). pp. 2710-2719 (2017). https://doi.org/10.1109/CVPR.2017.290" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.414, + 0.92, + 0.467 + ], + "angle": 0, + "content": "3. Benosman, R., Clercq, C., Lagorce, X., Ieng, S.H., Bartolozzi, C.: Event-based visual flow. IEEE Trans. Neural Netw. Learn. Syst. 25(2), 407-417 (2014). https://doi.org/10.1109/TNNLS.2013.2273537" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.468, + 0.92, + 0.52 + ], + "angle": 0, + "content": "4. Benosman, R., Ieng, S.H., Clercq, C., Bartolozzi, C., Srinivasan, M.: Asynchronous frameless event-based optical flow. Neural Netw. 27, 32-37 (2012). https://doi.org/10.1016/j.neunet.2011.11.001" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.521, + 0.918, + 0.573 + ], + "angle": 0, + "content": "5. Brebion, V., Moreau, J., Davoine, F.: Real-time optical flow for vehicular perception with low- and high-resolution event cameras. IEEE Trans. Intell. Transport. Syst. pp. 1-13 (2021). https://doi.org/10.1109/TITS.2021.3136358" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.575, + 0.917, + 0.609 + ], + "angle": 0, + "content": "6. Brosch, T., Tschechne, S., Neumann, H.: On event-based optical flow detection. Front. Neurosci. 9(137) (Apr 2015). https://doi.org/10.3389/fnins.2015.00137" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.611, + 0.92, + 0.681 + ], + "angle": 0, + "content": "7. Burner, L., Mitrokhin, A., Fermüller, C., Aloimonos, Y.: EVIMO2: An event camera dataset for motion segmentation, optical flow, structure from motion, and visual inertial odometry in indoor scenes with monocular or stereo algorithms. arXiv e-prints (May 2022). https://doi.org/10.48550/arXiv.2205.03467" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.682, + 0.918, + 0.735 + ], + "angle": 0, + "content": "8. Butler, D.J., Wulff, J., Stanley, G.B., Black, M.J.: A naturalistic open source movie for optical flow evaluation. In: Eur. Conf. Comput. Vis. (ECCV). pp. 611-625 (2012). https://doi.org/10.1007/978-3-642-33783-3_44" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.736, + 0.918, + 0.823 + ], + "angle": 0, + "content": "9. Chaney, K., Cladera Ojeda, F., Wang, Z., Bisulco, A., Hsieh, M.A., Korpela, C., Kumar, V., Taylor, C.J., Daniilidis, K.: M3ED: Multi-robot, multi-sensor, multi-environment event dataset. In: IEEE Conf. Comput. Vis. Pattern Recog. Workshops (CVPRW). pp. 4016-4023 (2023). https://doi.org/10.1109/CVPRW59228.2023.00419" + }, + { + "type": "ref_text", + "bbox": [ + 0.082, + 0.824, + 0.918, + 0.877 + ], + "angle": 0, + "content": "10. Chui, J., Klenk, S., Cremers, D.: Event-based feature tracking in continuous time with sliding window optimization. In: arXiv e-prints (2021). https://doi.org/10.48550/arXiv.2107.04536" + }, + { + "type": "ref_text", + "bbox": [ + 0.082, + 0.878, + 0.918, + 0.948 + ], + "angle": 0, + "content": "11. Ding, Z., Zhao, R., Zhang, J., Gao, T., Xiong, R., Yu, Z., Huang, T.: Spatiotemporal recurrent networks for event-based optical flow estimation. In: AAAI Conf. Artificial Intell. vol. 36, pp. 525-533 (2022). https://doi.org/10.1609/aaai.v36i1.19931" + }, + { + "type": "list", + "bbox": [ + 0.082, + 0.304, + 0.92, + 0.948 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.08, + 0.016, + 0.105, + 0.032 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.41, + 0.016, + 0.584, + 0.032 + ], + "angle": 0, + "content": "F. Hamann et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.057, + 0.92, + 0.111 + ], + "angle": 0, + "content": "12. Doersch, C., Gupta, A., Markeeva, L., Recasens, A., Smaira, L., Aytar, Y., Carreira, J., Zisserman, A., Yang, Y.: Tap-vid: A benchmark for tracking any point in a video. Adv. Neural Inf. Process. Syst. (NeurIPS) 35, 13610-13626 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.113, + 0.92, + 0.167 + ], + "angle": 0, + "content": "13. Doersch, C., Yang, Y., Vecerik, M., Gokay, D., Gupta, A., Aytar, Y., Carreira, J., Zisserman, A.: Tapir: Tracking any point with per-frame initialization and temporal refinement. arXiv preprint arXiv:2306.08637 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.168, + 0.92, + 0.239 + ], + "angle": 0, + "content": "14. Dosovitskiy, A., Fischer, P., Ilg, E., Häusser, P., Hazirbaş, C., Golkov, V., van der Smagt, P., Cremers, D., Brox, T.: FlowNet: Learning optical flow with convolutional networks. In: Int. Conf. Comput. Vis. (ICCV). pp. 2758-2766 (2015). https://doi.org/10.1109/ICCV.2015.316" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.241, + 0.92, + 0.293 + ], + "angle": 0, + "content": "15. Feydy, J., Glaunès, A., Charlier, B., Bronstein, M.: Fast geometric learning with symbolic matrices. Adv. Neural Inf. Process. Syst. (NeurIPS) 33, 14448-14462 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.295, + 0.92, + 0.418 + ], + "angle": 0, + "content": "16. Finateu, T., Niwa, A., Matolin, D., Tsuchimoto, K., Mascheroni, A., Reynaud, E., Mostafalu, P., Brady, F., Chotard, L., LeGoff, F., Takahashi, H., Wakabayashi, H., Oike, Y., Posch, C.: A 1280x720 back-illuminated stacked temporal contrast event-based vision sensor with \\(4.86\\mu \\mathrm{m}\\) pixels, 1.066Geps readout, programmable event-rate controller and compressive data-formating pipeline. In: IEEE Int. Solid-State Circuits Conf. (ISSCC). pp. 112-114 (2020). https://doi.org/10.1109/ISSCC19947.2020.9063149" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.421, + 0.92, + 0.493 + ], + "angle": 0, + "content": "17. Gallego, G., Delbruck, T., Orchard, G., Bartolozzi, C., Taba, B., Censi, A., Leutenegger, S., Davison, A., Conradt, J., Daniilidis, K., Scaramuzza, D.: Event-based vision: A survey. IEEE Trans. Pattern Anal. Mach. Intell. 44(1), 154-180 (2022). https://doi.org/10.1109/TPAMI.2020.3008413" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.494, + 0.919, + 0.547 + ], + "angle": 0, + "content": "18. Gallego, G., Gehrig, M., Scaramuzza, D.: Focus is all you need: Loss functions for event-based vision. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR). pp. 12272-12281 (2019). https://doi.org/10.1109/CVPR.2019.01256" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.549, + 0.92, + 0.62 + ], + "angle": 0, + "content": "19. Gallego, G., Rebecq, H., Scaramuzza, D.: A unifying contrast maximization framework for event cameras, with applications to motion, depth, and optical flow estimation. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR). pp. 3867-3876 (2018). https://doi.org/10.1109/CVPR.2018.00407" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.622, + 0.92, + 0.673 + ], + "angle": 0, + "content": "20. Gallego, G., Scaramuzza, D.: Accurate angular velocity estimation with an event camera. IEEE Robot. Autom. Lett. 2(2), 632-639 (2017). https://doi.org/10.1109/LRA.2016.2647639" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.676, + 0.92, + 0.731 + ], + "angle": 0, + "content": "21. Gehrig, D., Gehrig, M., Hidalgo-Carrió, J., Scaramuzza, D.: Video to Events: Recycl- cling video datasets for event cameras. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR). pp. 3583-3592 (2020). https://doi.org/10.1109/CVPR42600.2020.00364" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.731, + 0.92, + 0.785 + ], + "angle": 0, + "content": "22. Gehrig, D., Loquercio, A., Derpanis, K.G., Scaramuzza, D.: End-to-end learning of representations for asynchronous event-based data. In: Int. Conf. Comput. Vis. (ICCV). pp. 5632-5642 (2019). https://doi.org/10.1109/ICCV.2019.00573" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.786, + 0.92, + 0.84 + ], + "angle": 0, + "content": "23. Gehrig, M., Aarents, W., Gehrig, D., Scaramuzza, D.: DSEC: A stereo event camera dataset for driving scenarios. IEEE Robot. Autom. Lett. 6(3), 4947-4954 (2021). https://doi.org/10.1109/LRA.2021.3068942" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.841, + 0.92, + 0.895 + ], + "angle": 0, + "content": "24. Gehrig, M., Millhäsler, M., Gehrig, D., Scaramuzza, D.: E-RAFT: Dense optical flow from event cameras. In: Int. Conf. 3D Vision (3DV). pp. 197-206 (2021). https://doi.org/10.1109/3DV53792.2021.00030" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.896, + 0.92, + 0.95 + ], + "angle": 0, + "content": "25. Gehrig, M., Muglikar, M., Scaramuzza, D.: Dense continuous-time optical flow from event cameras. IEEE Trans. Pattern Anal. Mach. Intell. pp. 1-12 (2024). https://doi.org/10.1109/TPAMI.2024.3361671" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.057, + 0.92, + 0.95 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.476, + 0.016, + 0.84, + 0.032 + ], + "angle": 0, + "content": "Motion-prior Contrast Maximization" + }, + { + "type": "page_number", + "bbox": [ + 0.894, + 0.017, + 0.92, + 0.031 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.057, + 0.92, + 0.111 + ], + "angle": 0, + "content": "26. Guney, F., Sevilla-Lara, L., Sun, D., Wulff, J.: \"what is optical flow for?\": Workshop results and summary. In: Eur. Conf. Comput. Vis. Workshops (ECCVW). pp. 0-0 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.113, + 0.92, + 0.166 + ], + "angle": 0, + "content": "27. Guo, S., Gallego, G.: CMax-SLAM: Event-based rotational-motion bundle adjustment and SLAM system using contrast maximization. IEEE Trans. Robot. 40, 2442-2461 (2024). https://doi.org/10.1109/TRO.2024.3378443" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.168, + 0.918, + 0.222 + ], + "angle": 0, + "content": "28. Hagenaars, J., Paredes-Vallés, F., De Croon, G.: Self-supervised learning of event-based optical flow with spiking neural networks. Adv. Neural Inf. Process. Syst. (NeurIPS) 34, 7167-7179 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.223, + 0.921, + 0.293 + ], + "angle": 0, + "content": "29. Hamann, F., Gallego, G.: Stereo co-capture system for recording and tracking fish with frame- and event cameras. In: 26th Int. Conf. on Pattern Recognition (ICPR), Visual observation and analysis of Vertebrate And Insect Behavior (VAIB) Workshop (2022). https://doi.org/10.48550/ARXIV.2207.07332" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.295, + 0.92, + 0.348 + ], + "angle": 0, + "content": "30. Hamann, F., Ghosh, S., Juarez-Martínez, I., Hart, T., Kacelnik, A., Gallego, G.: Low-power, continuous remote behavioral localization with event cameras. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR) (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.35, + 0.92, + 0.403 + ], + "angle": 0, + "content": "31. Harley, A.W., Fang, Z., Fragkiadaki, K.: Particle video revisited: Tracking through occlusions using point trajectories. In: Eur. Conf. Comput. Vis. (ECCV). pp. 59-75 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.405, + 0.92, + 0.457 + ], + "angle": 0, + "content": "32. Huang, Z., Shi, X., Zhang, C., Wang, Q., Cheung, K.C., Qin, H., Dai, J., Li, H.: Flowformer: A transformer architecture for optical flow. In: Eur. Conf. Comput. Vis. (ECCV). pp. 668-685 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.459, + 0.92, + 0.528 + ], + "angle": 0, + "content": "33. Ilg, E., Mayer, N., Saikia, T., Keuper, M., Dosovitskiy, A., Brox, T.: FlowNet 2.0: Evolution of optical flow estimation with deep networks. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR). pp. 1647-1655 (2017). https://doi.org/10.1109/cvpr.2017.179" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.531, + 0.921, + 0.584 + ], + "angle": 0, + "content": "34. Janai, J., Guney, F., Ranjan, A., Black, M., Geiger, A.: Unsupervised learning of multi-frame optical flow with occlusions. In: Eur. Conf. Comput. Vis. (ECCV). pp. 690-706 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.586, + 0.92, + 0.639 + ], + "angle": 0, + "content": "35. Jonschkowski, R., Stone, A., Barron, J.T., Gordon, A., Konolige, K., Angelova, A.: What matters in unsupervised optical flow. In: Eur. Conf. Comput. Vis. (ECCV). pp. 557-572 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.641, + 0.92, + 0.676 + ], + "angle": 0, + "content": "36. Karaev, N., Rocco, I., Graham, B., Neverova, N., Vedaldi, A., Rupprecht, C.: Cotracker: It is better to track together. arXiv preprint arXiv:2307.07635 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.678, + 0.92, + 0.731 + ], + "angle": 0, + "content": "37. Kim, H., Kim, H.J.: Real-time rotational motion estimation with contrast maximization over globally aligned events. IEEE Robot. Autom. Lett. 6(3), 6016-6023 (2021). https://doi.org/10.1109/LRA.2021.3088793" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.732, + 0.92, + 0.785 + ], + "angle": 0, + "content": "38. Lee, C., Kosta, A., Zhu, A.Z., Chaney, K., Daniilidis, K., Roy, K.: Spike-flownet: Event-based optical flow estimation with energy-efficient hybrid neural networks. In: Eur. Conf. Comput. Vis. (ECCV). pp. 366-382 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.787, + 0.92, + 0.84 + ], + "angle": 0, + "content": "39. Lichtsteiner, P., Posch, C., Delbruck, T.: A \\(128 \\times 128\\) 120 dB \\(15~\\mu \\mathrm{s}\\) latency asynchronous temporal contrast vision sensor. IEEE J. Solid-State Circuits 43(2), 566-576 (2008). https://doi.org/10.1109/JSSC.2007.914337" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.842, + 0.92, + 0.911 + ], + "angle": 0, + "content": "40. Liu, H., Chen, G., Qu, S., Zhang, Y., Li, Z., Knoll, A., Jiang, C.: TMA: Temporal motion aggregation for event-based optical flow. In: Int. Conf. Comput. Vis. (ICCV). pp. 9651-9660 (Oct 2023). https://doi.org/10.1109/ICCV51070.2023.00888" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.913, + 0.92, + 0.95 + ], + "angle": 0, + "content": "41. Liu, M., Delbruck, T.: Adaptive time-slice block-matching optical flow algorithm for dynamic vision sensors. In: British Mach. Vis. Conf. (BMVC). pp. 1-12 (2018)" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.057, + 0.921, + 0.95 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.08, + 0.017, + 0.105, + 0.03 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.41, + 0.016, + 0.582, + 0.032 + ], + "angle": 0, + "content": "F. Hamann et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.057, + 0.92, + 0.129 + ], + "angle": 0, + "content": "42. Mayer, N., Ilg, E., Hausser, P., Fischer, P., Cremers, D., Dosovitskiy, A., Brox, T.: A large dataset to train convolutional networks for disparity, optical flow, and scene flow estimation. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR). pp. 4040-4048 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.129, + 0.92, + 0.197 + ], + "angle": 0, + "content": "43. Orchard, G., Benosman, R., Etienne-Cummings, R., Thakor, N.V.: A spiking neural network architecture for visual motion estimation. In: IEEE Biomed. Circuits Syst. Conf. (BioCAS). pp. 298-301 (2013). https://doi.org/10.1109/biocas.2013.6679698" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.199, + 0.92, + 0.268 + ], + "angle": 0, + "content": "44. Paredes-Valles, F., Scheper, K.Y., De Wagter, C., de Croon, G.C.: Taming contrast maximization for learning sequential, low-latency, event-based optical flow. In: Int. Conf. Comput. Vis. (ICCV). pp. 9661-9671 (Oct 2023). https://doi.org/10.1109/ICCV51070.2023.00889" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.269, + 0.92, + 0.338 + ], + "angle": 0, + "content": "45. Paredes-Valles, F., Hagenaars, J.J., Dupeyroux, J., Stroobants, S., Xu, Y., de Croon, G.C.H.E.: Fully neuromorphic vision and control for autonomous drone flight. Science Robotics 9(90), eadi0591 (2024). https://doi.org/10.1126/scirobotics.adio591" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.339, + 0.918, + 0.407 + ], + "angle": 0, + "content": "46. Posch, C., Serrano-Gotarredona, T., Linares-Barranco, B., Delbruck, T.: Retinomorphic event-based vision sensors: Bioinspired cameras with spiking output. Proc. IEEE 102(10), 1470-1484 (Oct 2014). https://doi.org/10.1109/jproc.2014.2346153" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.409, + 0.918, + 0.462 + ], + "angle": 0, + "content": "47. Rebecq, H., Gehrig, D., Scaramuzza, D.: ESIM: an open event camera simulator. In: Conf. on Robotics Learning (CoRL). Proc. Machine Learning Research, vol. 87, pp. 969-982. PMLR (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.462, + 0.918, + 0.496 + ], + "angle": 0, + "content": "48. Ren, Z., Yan, J., Ni, B., Liu, B., Yang, X., Zha, H.: Unsupervised deep learning for optical flow estimation. In: AAAI Conf. Artificial Intell. vol. 31 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.496, + 0.918, + 0.548 + ], + "angle": 0, + "content": "49. Ronneberger, O., Fischer, P., Brox, T.: U-Net: Convolutional networks for biomedical image segmentation. In: Int. Conf. Medical Image Computing and Computer-Assisted Intervention (MICCAI). pp. 234-241 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.549, + 0.918, + 0.583 + ], + "angle": 0, + "content": "50. Sand, P., Teller, S.: Particle video: Long-range motion estimation using point trajectories. Int. J. Comput. Vis. 80, 72-91 (2008)" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.583, + 0.918, + 0.635 + ], + "angle": 0, + "content": "51. Seok, H., Lim, J.: Robust feature tracking in dvs event stream using Bezier mapping. In: IEEE Winter Conf. Appl. Comput. Vis. (WACV). pp. 1647-1656 (2020). https://doi.org/10.1109/WACV45572.2020.9093607" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.635, + 0.918, + 0.669 + ], + "angle": 0, + "content": "52. Shiba, S., Aoki, Y., Gallego, G.: Event collapse in contrast maximization frameworks. Sensors 22(14), 1-20 (2022). https://doi.org/10.3390/s22145190" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.67, + 0.918, + 0.722 + ], + "angle": 0, + "content": "53. Shiba, S., Aoki, Y., Gallego, G.: A fast geometric regularizer to mitigate event collapse in the contrast maximization framework. Adv. Intell. Syst. p. 2200251 (2022). https://doi.org/10.1002/aisy.202200251" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.722, + 0.918, + 0.773 + ], + "angle": 0, + "content": "54. Shiba, S., Aoki, Y., Gallego, G.: Secrets of event-based optical flow. In: Eur. Conf. Comput. Vis. (ECCV). pp. 628-645 (2022). https://doi.org/10.1007/978-3-031-19797-0 36" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.774, + 0.918, + 0.826 + ], + "angle": 0, + "content": "55. Shiba, S., Hamann, F., Aoki, Y., Gallego, G.: Event-based background oriented schlieren. IEEE Trans. Pattern Anal. Mach. Intell. 46(4), 2011-2026 (2024). https://doi.org/10.1109/TPAMI.2023.3328188" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.826, + 0.918, + 0.879 + ], + "angle": 0, + "content": "56. Shiba, S., Klose, Y., Aoki, Y., Gallego, G.: Secrets of event-based optical flow, depth, and ego-motion by contrast maximization. IEEE Trans. Pattern Anal. Mach. Intell. pp. 1-18 (2024). https://doi.org/10.1109/TPAMI.2024.3396116" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.879, + 0.918, + 0.95 + ], + "angle": 0, + "content": "57. Stoffregen, T., Scheerlinck, C., Scaramuzza, D., Drummond, T., Barnes, N., Kleeman, L., Mahony, R.: Reducing the sim-to-real gap for event cameras. In: Eur. Conf. Comput. Vis. (ECCV). pp. 534-549 (2020). https://doi.org/https://doi.org/10.1007/978-3-030-58583-9_32" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.057, + 0.92, + 0.95 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.476, + 0.016, + 0.84, + 0.032 + ], + "angle": 0, + "content": "Motion-prior Contrast Maximization" + }, + { + "type": "page_number", + "bbox": [ + 0.894, + 0.017, + 0.92, + 0.031 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.057, + 0.918, + 0.111 + ], + "angle": 0, + "content": "58. Stone, A., Maurer, D., Ayvaci, A., Angelova, A., Jonschkowski, R.: Smurf: Self-teaching multi-frame unsupervised raft with full-image warping. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR). pp. 3887-3896 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.112, + 0.918, + 0.164 + ], + "angle": 0, + "content": "59. Sun, D., Vlasic, D., Herrmann, C., Jampani, V., Krainin, M., Chang, H., Zabih, R., Freeman, W.T., Liu, C.: Autoflow: Learning a better training set for optical flow. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR). pp. 10093-10102 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.165, + 0.918, + 0.219 + ], + "angle": 0, + "content": "60. Sun, D., Yang, X., Liu, M.Y., Kautz, J.: PWC-Net: CNNs for optical flow using pyramid, warping, and cost volume. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR) (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.219, + 0.918, + 0.254 + ], + "angle": 0, + "content": "61. Sun, X., Harley, A.W., Guibas, L.J.: Refining pre-trained motion models. arXiv preprint arXiv:2401.00850 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.254, + 0.918, + 0.307 + ], + "angle": 0, + "content": "62. Teed, Z., Deng, J.: RAFT: Recurrent all pairs field transforms for optical flow. In: Eur. Conf. Comput. Vis. (ECCV). pp. 402-419 (2020). https://doi.org/10.1007/978-3-030-58536-5_24" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.308, + 0.918, + 0.378 + ], + "angle": 0, + "content": "63. Tulyakov, S., Bochicchio, A., Gehrig, D., Georgoulis, S., Li, Y., Scaramuzza, D.: Time lens++: Event-based frame interpolation with parametric non-linear flow and multi-scale fusion. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR). pp. 17755-17764 (Jun 2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.379, + 0.918, + 0.414 + ], + "angle": 0, + "content": "64. Valmadre, J., Lucey, S.: General trajectory prior for non-rigid reconstruction. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR). pp. 1394-1401 (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.415, + 0.918, + 0.45 + ], + "angle": 0, + "content": "65. Wan, Z., Dai, Y., Mao, Y.: Learning dense and continuous optical flow from an event camera. IEEE Trans. Image Process. 31, 7237-7251 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.451, + 0.918, + 0.485 + ], + "angle": 0, + "content": "66. Wang, C., Eckart, B., Lucey, S., Gallo, O.: Neural trajectory fields for dynamic novel view synthesis. arXiv preprint arXiv:2105.05994 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.486, + 0.918, + 0.539 + ], + "angle": 0, + "content": "67. Wang, Z., Hamann, F., Chaney, K., Jiang, W., Gallego, G., Daniilidis, K.: Event-based continuous color video decompression from single frames. arXiv preprint arXiv:2312.00113 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.539, + 0.918, + 0.592 + ], + "angle": 0, + "content": "68. Weikersdorfer, D., Hoffmann, R., Conradt, J.: Simultaneous localization and mapping for event-based vision systems. In: Int. Conf. Comput. Vis. Syst. (ICVS). pp. 133-142 (2013). https://doi.org/10.1007/978-3-642-39402-7_14" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.593, + 0.918, + 0.628 + ], + "angle": 0, + "content": "69. Wu, Y., Paredes-Valles, F., de Croon, G.C.: Lightweight event-based optical flow estimation via iterative deblurring. arXiv preprint arXiv:2211.13726 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.629, + 0.918, + 0.699 + ], + "angle": 0, + "content": "70. Ye, C., Mitrokhin, A., Parameshwara, C., Fermüller, C., Yorke, J.A., Aloimonos, Y.: Unsupervised learning of dense optical flow, depth and egomotion with event-based sensors. In: IEEE/RSJ Int. Conf. Intell. Robot. Syst. (IROS). pp. 5831-5838 (2020). https://doi.org/10.1109/IROS45743.2020.9341224" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.7, + 0.918, + 0.753 + ], + "angle": 0, + "content": "71. Young, P., Lai, A., Hodosh, M., Hockenmaier, J.: From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. Trans. Assoc. Computational Linguistics 2, 67-78 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.754, + 0.918, + 0.807 + ], + "angle": 0, + "content": "72. Yu, J.J., Harley, A.W., Derpanis, K.G.: Back to basics: Unsupervised learning of optical flow via brightness constancy and motion smoothness. In: Eur. Conf. Comput. Vis. Workshops (ECCVW). pp. 3-10 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.808, + 0.918, + 0.859 + ], + "angle": 0, + "content": "73. Zen, G., Ricci, E., Sebe, N.: Exploiting sparse representations for robust analysis of noisy complex video scenes. In: Eur. Conf. Comput. Vis. (ECCV). pp. 199-213 (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.918, + 0.914 + ], + "angle": 0, + "content": "74. Zheng, Y., Harley, A.W., Shen, B., Wetzstein, G., Guibas, L.J.: PointOdyssey: A large-scale synthetic dataset for long-term point tracking. In: Int. Conf. Comput. Vis. (ICCV). pp. 19855-19865 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.914, + 0.918, + 0.947 + ], + "angle": 0, + "content": "75. Zhu, A.Z., Thakur, D., Ozaslan, T., Pfrommer, B., Kumar, V., Daniilidis, K.: The multivehicle stereo event camera dataset: An event camera dataset for 3D" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.057, + 0.918, + 0.947 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.079, + 0.017, + 0.105, + 0.031 + ], + "angle": 0, + "content": "20" + }, + { + "type": "header", + "bbox": [ + 0.41, + 0.016, + 0.584, + 0.032 + ], + "angle": 0, + "content": "F. Hamann et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.118, + 0.057, + 0.916, + 0.093 + ], + "angle": 0, + "content": "perception. IEEE Robot. Autom. Lett. 3(3), 2032-2039 (Jul 2018). https://doi.org/10.1109/lra.2018.2800793" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.094, + 0.92, + 0.147 + ], + "angle": 0, + "content": "76. Zhu, A.Z., Yuan, L., Chaney, K., Daniilidis, K.: EV-FlowNet: Self-supervised optical flow estimation for event-based cameras. In: Robotics: Science and Systems (RSS). pp. 1-9 (2018). https://doi.org/10.15607/RSS.2018.XIV.062" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.148, + 0.921, + 0.2 + ], + "angle": 0, + "content": "77. Zhu, A.Z., Yuan, L., Chaney, K., Daniilidis, K.: Unsupervised event-based learning of optical flow, depth, and egomotion. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR). pp. 989-997 (2019). https://doi.org/10.1109/CVPR.2019.00108" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.201, + 0.92, + 0.236 + ], + "angle": 0, + "content": "78. Zhu, Y., Lucey, S.: Convolutional sparse coding for trajectory reconstruction. IEEE Trans. Pattern Anal. Mach. Intell. 37(3), 529-540 (2013)" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.057, + 0.921, + 0.236 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Motion-prior Contrast Maximization for Dense Continuous-Time Motion Estimation/34d0d1c3-d92d-4348-bbb4-52308320fede_origin.pdf b/2024/Motion-prior Contrast Maximization for Dense Continuous-Time Motion Estimation/34d0d1c3-d92d-4348-bbb4-52308320fede_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..b3e93cd5e3adb75ef42cbb602b2c609ad6678d90 --- /dev/null +++ b/2024/Motion-prior Contrast Maximization for Dense Continuous-Time Motion Estimation/34d0d1c3-d92d-4348-bbb4-52308320fede_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5447f840b0acb7a1bfeaba4c4d35665f180db7e92c475668567810bd3684cd05 +size 9276255 diff --git a/2024/Motion-prior Contrast Maximization for Dense Continuous-Time Motion Estimation/full.md b/2024/Motion-prior Contrast Maximization for Dense Continuous-Time Motion Estimation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..2fa6825a9f45c18fa79a565de070ae4787c1c88d --- /dev/null +++ b/2024/Motion-prior Contrast Maximization for Dense Continuous-Time Motion Estimation/full.md @@ -0,0 +1,326 @@ +# Motion-prior Contrast Maximization for Dense Continuous-Time Motion Estimation + +Friedhelm Hamann1, Ziyun Wang2, Ioannis Asmanis2, Kenneth Chaney2, Guillermo Gallego1,3, and Kostas Daniilidis2,4 + +$^{1}$ TU Berlin and SCIolI Excellence Cluster, Berlin, Germany + $^{2}$ University of Pennsylvania, Philadelphia, US + +3 Einstein Center Digital Future and Robotics Institute Germany, Berlin, Germany +4 Archimedes, Athena RC, Greece + +Abstract. Current optical flow and point-tracking methods rely heavily on synthetic datasets. Event cameras are novel vision sensors with advantages in challenging visual conditions, but state-of-the-art frame-based methods cannot be easily adapted to event data due to the limitations of current event simulators. We introduce a novel self-supervised loss combining the Contrast Maximization framework with a non-linear motion prior in the form of pixel-level trajectories and propose an efficient solution to solve the high-dimensional assignment problem between non-linear trajectories and events. Their effectiveness is demonstrated in two scenarios: In dense continuous-time motion estimation, our method improves the zero-shot performance of a synthetically trained model on the real-world dataset EVIMO2 by $29\%$ . In optical flow estimation, our method elevates a simple UNet to achieve state-of-the-art performance among self-supervised methods on the DSEC optical flow benchmark. Our code is available at https://github.com/tub-rip/MotionPriorCMax. + +# 1 Introduction + +Determining the motion of arbitrary projected world points on the image plane over long time intervals is a difficult low-level computer vision problem. Researchers have studied it as optical flow and lately as point-tracking, with many practical applications in robotics, computational photography, video compression, and object-level tracking [26]. The best-performing frame-based methods for optical flow estimation and point tracking use large-scale synthetic datasets. Synthetic data is diverse and has highly accurate ground truth (GT), but is unrealistic and methods trained on synthetic data show a sim-to-real gap. + +Novel vision sensors called event cameras [16,17,39] have emerged as promising alternatives to take on the problem. Inspired by the transient visual pathway, which is responsible for motion perception, they are particularly fit for capturing scene motion in the form of asynchronous pixel-wise intensity changes. This working principle endows them with advantages, such as high speed, high dynamic range (HDR) and low power consumption. + +![](images/ad075927e22d4427da496a0fe9385f4e82a447e7725d122fa1ec6d2c8fe85a49.jpg) +a) Contrast Maximization with non-linear Trajectories + +![](images/d638bfaa03eae5a1b184f8a61befaab34ee71241f48f2df2c3d0192aca61e1ea.jpg) +b) Overview Method + +![](images/cfeaf3a68064fe3946cc8d75b9e32aa7998fc0aa1ee96b69fef645e7c0d3a039.jpg) + +![](images/bf7fe060a130b599dab3cda7f4387aaecc61c3e23db6b9d5519899c052f77b54.jpg) + +![](images/0501c97cc26b68f2996af6e510a69a55123e58ba6d3ced7a145af22b0bc68f36.jpg) + +![](images/9f2bed5b336b044fd57819770283e1c5d2b689e4a3744ec0883bd4a41d948e52.jpg) +c) Application to Trajectory Estimation + +![](images/a20f3b923ecfeb1822e64d8465bd12fc9a11527a90ac13c3a7adddd36ef984d4.jpg) +Fig.1: Summary. a) We present an approach to combine Contrast Maximization with dense non-linear trajectories. b) We show how it can be used for self-supervised learning in a pipeline to predict dense point trajectories, and c) evaluate it on the EVIMO2 dataset, for which we generate dense point tracks. d) Additionally, our approach provides state-of-the-art performance on self-supervised optical flow prediction. + +![](images/d53aca7b0ce0b63c52044b1ab88d10fd1de4c853c12c1383a1978fe5ef6ed691.jpg) + +![](images/b3684fcdca0d6303e409a626861c564e772277f4431be555e5fb81235c4b8979.jpg) +d) Application to Optical Flow + +Event-based motion estimation methods can be categorized according to the complexity of the motion considered (i.e., number of degrees-of-freedom (DOF)) and to the solution strategy: model-based or learning-based. Low-DOF motions arise in sparse feature tracking and ego-motion estimation, while high-DOF motions describe more complex scenes, e.g., via per-pixel displacement (i.e., densely, over the whole image plane). Focusing on the latter, the event-based optical flow problem has been extensively studied on mobile robotic datasets [23, 75], where GT is calculated as the motion field (from known depth and poses). Most approaches use this GT for direct supervision of dense flow prediction [22, 24, 40, 57, 65]. Alternatively, several works use a contrast loss [18, 19, 77], which allows training in a self-supervised manner. However, the contrast loss is prone to undesired local optima, called event collapse [52, 53], where many events are warped into a few pixels or lines. Evaluation of event-based optical flow has been mostly done on the MVSEC [75] and DSEC [23] datasets, which comprise largely uniform motions and test intervals from 0.022s to 0.1s. Recently, [25] has proposed a supervised method to predict pixel displacement over a larger duration (0.5s), by leveraging synthetic data. It relies on the generation of training event data from images, where current tools [21] are not as mature as frame-based simulation, therefore it suffers from a sim-to-real performance gap. + +In short, progress has been made and the research field is moving towards predicting motion over longer time intervals on the whole image plane, i.e., taking on more complex (i.e., nonlinear) motion problems. This is precisely the + +problem tackled in this paper (Fig. 1): long-time and dense event-based motion estimation, reducing the domain adaptation gap of previous approaches. It comes with several associated challenges, mainly overcoming the lack of large labeled datasets, and dealing with event noise and data association (events depend on motion, and for large motions, the appearance of "corresponding" events can be wildly different due to changes in motion direction, occlusions, etc.) while leveraging the space-time characteristics of event data. + +To this end, we propose tackling the problem in two stages, by leveraging both supervised and self-supervised strategies: first, using supervised learning on synthetic data to provide initial model weights for motion estimation, and secondly, fine-tuning the network on real data via a self-supervised loss to reduce the domain adaptation gap. Our technical contributions involve extending the contrast loss framework to regress continuous-time trajectories over long time intervals via motion priors (parametric functions that provide a good balance between motion generality and regularization [64, 66]). This includes a solution to accurately and efficiently associate events to the trajectories (Fig. 2). + +More specifically, finding the association between events and motion trajectories (to warp corresponding events and achieve event alignment) is a high-dimensional problem (e.g., for a time window of 0.3 s one can consider about ten million events, and as many trajectories as pixels), which needs to be implemented in a differentiable and parallelizable manner. We propose two actions to cope with these technical challenges. First, we relax the problem by interpolating over a coarse spatio-temporal displacement field, which serves as a lookup table. Secondly, we use a symbolic matrix framework [15] to calculate the $K = N_{\mathrm{traj}}$ nearest neighbor (KNN) trajectories for each grid point, thereby solving KNN in a memory-efficient and differentiable way on GPU. The warp displacement at the grid point is then set to the average of the neighboring trajectories. + +Our approach is versatile, allowing for different types of networks and trajectories. Hence, we evaluate its performance on two applications: dense continuous-time motion estimation and optical flow. The results on EVIMO2 [7] show that fine-tuning with our self-supervised loss improves the zero-shot performance of a model pre-trained on synthetic data by $29\%$ . On DSEC, our model shows state-of-the-art performance among self-supervised methods, on average improving the angular error by $19\%$ and the percentage of inliers by $14\%$ , while having a $5\times$ faster inference time. In summary, our contributions are (Fig. 1): + +1. We introduce motion priors (parametric functions with a good balance between generality and regularization) in the event-based contrast maximization framework for continuous-time and dense motion estimation. +2. We combine the self-supervised loss with Bflow [25], the current top-performing supervised model for dense continuous-time event-based motion estimation trained on synthetic data, and show that it can be used to improve over zero-shot performance on unseen real data (EVIMO2). +3. We show that the combination of our loss with a simple U-Net architecture (à la EV-FlowNet [76]) achieves state-of-the-art performance among the self-supervised methods on the DSEC Optical Flow benchmark. + +# 2 Related Work + +Frame-based Motion Estimation. Progress in deep learning triggered a large series of learned methods for optical flow estimation in classical, frame-based computer vision [2, 14, 32, 33, 60, 62]. The solutions rely on large-scale synthetic datasets [14, 42, 59, 73]. Similarly related is the task of point tracking, which has shown impressive progress on frame-based data [13, 31, 36, 50, 74], equally relying on simulated data [8, 12, 74]. This approach is unsatisfying, as it is prone to out-of-distribution (OOD) problems that cannot be easily overcome. As an alternative, self-supervised methods have been explored [34, 35, 48, 72] and shown to improve models pre-trained on simulated data for motion estimation tasks [58, 61]. + +Event-based Optical Flow. Event cameras [39, 46] are a relatively new technology, and have found applications in various computer vision domains, like mobile robotics [27, 45, 68], scene understanding [29, 30], and computational imaging [55, 63]. Their exploration for low-level vision tasks has taken a similar trajectory in a compressed timeline as previously frame-based methods [17]. The first event-based optical flow methods were model-based [3, 4, 6, 41, 43, 54], followed by learning-based approaches [11, 24, 28, 38, 44, 76, 77]. + +The scarce availability of event data and less mature simulation technology compared to frame-based cameras are major obstacles [21]. Therefore, event-based optical flow has been mostly evaluated on data acquired through egomotion, such as the robotics dataset MVSEC [75], the driving dataset DSEC [23] or the M3ED dataset [9]. However, GT is calculated using the motion field equation (with data from a synchronized depth sensor) and therefore does not provide an accurate flow at the event rate, nor spatially at occlusions and independently moving objects (IMOs). This GT is used to train supervised learning approaches [22, 24, 40, 57, 65], which inherit limitations from the ground truth. + +Self-supervised methods for event-based optical flow lessen the dependency on GT labels by leveraging an event alignment error [28,54,70,77]. Notably, variations of the contrast loss have been proposed [18], however, so far the application has been limited to either low-DOF problems (e.g., feature tracking [10, 19, 51], ego-motion [19, 20, 37]) or high-DOF but short-time optical flow problem [44, 54] (max. 0.1 s). We expand the frontier to the challenging problem of long-time and high-DOF (complex) motion estimation, by exploiting trajectory priors. + +Trajectory Prior. Using trajectories as motion priors has been a widely explored scheme in computer vision. Specifically, [1, 64, 78] propose a linear combination of basis functions for structure from motion. More recently it has been explored for dynamic novel view synthesis [66]. In the context of events, [63] uses cubic motion splines for video frame interpolation, [25] uses B-splines for supervised learning of non-linear optical flow, [67] use learned basis functions for event-based video decompression, and [10, 51] show event-based feature tracking (low-DOF) using Bezier or B-spline curves in combination with a contrast loss. + +![](images/67d6593e4476ffb8773fdbff0d1dee02a851723bc05e26dadce9f3780c710844.jpg) +Fig. 2: Pipeline overview. (a) Input events in a time interval are (b) voxelized and (c) passed to an artificial neural network that predicts per-pixel coefficients for continuous-time trajectories (d). The raw events and predicted trajectories are fed to the loss module (e). Here, a dense spatio-temporal displacement map is interpolated, and events are warped according to their looked-up displacement. Lastly, an image of warped events (IWE) is built at a random reference time and its gradient magnitude acts as training loss. Note that the prediction method displayed here is specific to the used Bflow backbone [25], with additional events before the prediction start time $t_s$ as input. + +# 3 Methodology + +# 3.1 Contrast Loss + +Contrast Maximization (CM) [18, 19] is used for self-supervised training. The idea is to find the point trajectories on the image plane best aligning with the events by maximizing the sharpness of warped events. It is an iterative approach, with three steps per iteration: (i) Each event $e_k = (\mathbf{x}_k, t_k, p_k)$ contains the pixel coordinates $\mathbf{x}_k = (x_k, y_k)^\top$ , timestamp $t_k$ and polarity $p_k$ of a brightness change of predefined size $C$ (contrast sensitivity). Events $\mathcal{E} = \{e_k\}_{k=1}^{N_e}$ are displaced according to a candidate motion hypothesis $\mathbf{x}_k' = \mathbf{W}(\mathbf{x}_k, t_k; \boldsymbol{\theta})$ , with parameters $\boldsymbol{\theta}$ , to a reference time $t_\mathrm{ref}$ , producing a set of warped events $\mathcal{E}_{t_\mathrm{ref}}' = \{e_k'\}_{k=1}^{N_e}$ : + +$$ +e _ {k} \doteq \left(\mathbf {x} _ {k}, t _ {k}, p _ {k}\right) \stackrel {\mathrm {W}} {\mapsto} e _ {k} ^ {\prime} \doteq \left(\mathbf {x} _ {k} ^ {\prime}, t _ {\text {r e f}}, p _ {k}\right). \tag {1} +$$ + +Afterwards, $(ii)$ the events are summed into an image of warped events (IWE), + +$$ +I (\mathbf {x}; \mathcal {E} _ {t _ {\text {r e f}}} ^ {\prime}, \boldsymbol {\theta}) \doteq \sum_ {k = 1} ^ {N _ {e}} \mathcal {N} (\mathbf {x}; \mathbf {x} _ {k} ^ {\prime}, \sigma^ {2}), \tag {2} +$$ + +which essentially counts the number of warped events $e_k'$ per pixel. Lastly (iii) the contrast or sharpness of the IWE is computed (e.g., the gradient magnitude of the IWE, $G \doteq \int \| \nabla I(\mathbf{x}) \| d\mathbf{x}$ ), which serves as a proxy for how well the model (motion hypothesis and $\pmb{\theta}$ ) fit with the events produced by the true motion. + +Table 1: Comparison of the contrast loss formulation in the most recent methods. + +
StepsShiba et. al [54]Paredes et. al [44]Ours
Warp displacement (3):Δxk=(tref−tk)v(ek)Δxk=∑i(Δti v_i)(ek)Δxk=1/Ntraj ∑n Δxn
linear trajectoryflow concatenationnon-linear trajectory
IWEI(x)=∑kN(x; x′k, σ2)T±(x)=∑ktkN(x; x′k, σ2)/∑kN(x; x′k, σ2)I(x)=∑kN(x; x′k, σ2)
Reference time tref:{t0, t0.5, t1}multi-partition p, multi-time-scale s~U(0,1)
Contrast objectiveG(t0)+2G(t0.5)+G(t1)/4G(v=0)1/S ∑s=0S-11/2s ∑p=02s-1LRCM,pG(tref)
+ +# 3.2 Estimating Continuous-Time and Dense Motion Trajectories + +The CM framework has been extended in several works. Table 1 compares recent extensions for optical flow estimation, where events are warped as + +$$ +\mathbf {x} _ {k} ^ {\prime} = \mathbf {x} _ {k} + \Delta \mathbf {x} _ {k}. \tag {3} +$$ + +The main differences between these approaches lie in the event displacement (i.e., warp) model, the reference times used, and the loss function (i.e., event-alignment metric), with the design choices having two goals: fitting the event data and regularizing the solution (e.g., avoiding event collapse [52]). Previous methods used mostly a linear model [54,56] or partitioned the inference interval into smaller flow intervals so that its concatenation need not be linear [44]. By contrast, we introduce an explicit continuous-time non-linear trajectory model, a random reference time per iteration, and an easier contrast loss. + +Trajectory Representation. Figure 2 shows an overview of the training pipeline. Events in the time interval $[t_s, t_e]$ are fed into a neural network that makes per-pixel predictions of the parameters (coefficients) $\pmb{\theta} \equiv \pmb{\alpha}_n = (\alpha_{n,1}, \dots, \alpha_{n,N_c})^\top \in \mathbb{R}^{N_c}$ of a continuous-time trajectory $\mathbf{q}_n(t; \pmb{\alpha}_n) \equiv \mathbf{q}_n(t) = (x_n(t), y_n(t))^\top$ . There is one trajectory per pixel (i.e., "dense" character), $n = 1, \dots, N_p$ , where $N_p = h w$ is the number of pixels (image height $h$ and width $w$ ). Thus, $n$ is the spatial index of the trajectory, identifying it among all trajectories on the image plane. + +We model trajectories as weighted combinations of basis functions, $\mathbf{q}_n(t) = \sum_{j=1}^{N_c} g_j(t) \mathbf{p}_{n,j}$ , where $g_j(t)$ are temporal basis shares by all trajectories, and $\mathbf{p}_{n,j} = (\alpha_{n,j}^x, \alpha_{n,j}^y)^\top$ are "control points" (we write $x$ and $y$ components explicitly with separate coefficients in $\alpha_n$ ). We investigate polynomial basis $g_j(t) = t^j$ , Bézier curves with basis $g_j(t) = \binom{N_c}{j}(1 - t)^{N_c - j}t^j$ , as well as a learned basis. + +Spatio-Temporal Event Warping. The trajectories can be used to warp events. This would require finding the trajectory that passes through the spacetime coordinates of the event and then finding the value of the trajectory at the reference time (i.e., the warped event location). However, the association between events and trajectories is unknown and needs to be estimated simultaneously with the parameters (i.e., shape) of the trajectories. We circumvent the problem by using a soft association between events and trajectories. Each event $e_k$ is associated with its $N_{\mathrm{traj}}$ nearest neighboring trajectories $\{\mathbf{q}_n\}_{n=1}^{N_{\mathrm{traj}}}$ , + +and the event displacement $\varDelta\mathbf{x}_k$ in (3) is computed as the average of the respective trajectory displacements $\{\varDelta\mathbf{x}_k^n\}_{n=1}^{N_{\mathrm{traj}}}$ . The trajectory displacement $\varDelta\mathbf{x}_k^n$ is defined as the difference of trajectory locations at the time of the event and the reference time. The event displacement is defined as the mean: + +$$ +\Delta \mathbf {x} _ {k} \doteq \frac {1}{N _ {\mathrm {t r a j}}} \sum_ {n = 1} ^ {N _ {\mathrm {t r a j}}} \Delta \mathbf {x} _ {k} ^ {n}, \quad \text {w i t h} \quad \Delta \mathbf {x} _ {k} ^ {n} \doteq \mathbf {q} _ {n} \left(t _ {\mathrm {r e f}}\right) - \mathbf {q} _ {n} \left(t _ {k}\right). \tag {4} +$$ + +Loss Calculation. Once the events are warped using (3), it is straightforward to compute the IWE. We adopt the magnitude of the IWE gradient as loss function [18,56]. Moreover, we use the contrast loss in a self-supervised learning setting, choosing a different reference time $t_{\mathrm{ref}}$ , uniformly sampled in the observation interval, for every batch during training. This simplifies the objective to a single warping operation and loss calculation (e.g., compared to the three warping operations and loss calculations in the optimization-based approach by [54,56]), while it has added regularization benefits, as mentioned below. + +Memory Effective Computation of the Displacement Field. The number of events and trajectories can be very large, and even more their combination. Calculating KNN for every event can quickly become computationally unfeasible, and traditional algorithms cannot be efficiently implemented in deep-learning frameworks. We propose calculating the per-event displacement $\Delta \mathbf{x}_k$ in (3) by first interpolating a dense but coarser spatio-temporal displacement field, and then looking up the per-event displacement in such space-time volume. Moreover, we relax the problem by solving the KNN search in 2D instead of in the volume. + +Figure 2 shows an overview of the interpolation process. The displacement map is a tensor of shape $[N_{\mathrm{bins}}, h/4, w/4]$ , where $N_{\mathrm{bins}}$ is the number of temporal bins, and $h$ and $w$ are the sensor's height and width, respectively. For each temporal bin-center $t_{\mathrm{bins}}$ the trajectories $\mathbf{q}_n(t = t_{\mathrm{bins}})$ are calculated and the KNN interpolation is performed between each pixel of a channel $\mathbf{x}[t_{\mathrm{bins}}]$ and $\mathbf{q}_n(t = t_{\mathrm{bins}})$ . We implement the KNN approach using KeOps [15], a framework for symbolic matrix computation, which provides a memory-efficient and differentiable solution. We perform a KNN search for every voxel in the table, and events are warped after a lookup operation. Please note that this step is entirely independent of the voxel grid passed as input to the backbone, and all raw events are used to calculate the contrast loss. From the calculated displacement map, we can directly look up per-event displacements $\Delta \mathbf{x}_k$ . + +Regularization. Objectives based on event alignment are prone to undesired local minima called "event collapse" [52,53]. Our formulation inherently provides regularization, temporally by the smoothness of the motion prior, and spatially by interpolating the event flow from several trajectories via a soft assignment. Our framework uses two additional regularization sources. Firstly, we penalize the magnitude of the spatial gradient of the interpolated displacement field between consecutive timesteps, $R \doteq \| \nabla (\Delta \mathbf{q}_{*}(t))\|_{L^{1}(\Omega)}$ . To interpolate $\mathbf{q}_{*}$ from the sparse trajectories $\mathbf{q}_n$ , we use the same KNN indices as for the main volume. This loss encourages spatial smoothness of the trajectories. + +Table 2: Results on EVIMO2 [7]. + +
MethodTEPE ↓TAE ↓%Out ↓
Paredes et. al [44]21.6951.910.63
E-RAFT (linear) [24]19.3874.520.66
BFlow (zero-shot) [25]8.6319.940.36
BFlow (in-domain)3.3811.680.17
Ours (self-supervised)6.1416.980.25
+ +Secondly, we use a multi-reference formulation by randomly sampling a reference time at every training step, while previous work used multiple and fixed timestamps (Tab. 1). This has the advantage that $(i)$ IWEs are required to be sharp at truly any time and $(ii)$ the memory requirement is lowered as it scales linearly with the number of reference times used during a training step. + +In summary, the training loss is + +$$ +\mathcal {L} = 1 / G + \lambda R, \tag {5} +$$ + +with $\lambda > 0$ as regularization weight. Further details are in the supplementary. Pre-processing and Prediction-Module Architecture. Events record brightness changes asynchronously, in the form of a sparse spatio-temporal signal. As input to the Prediction Module, events are customarily converted to voxel grids [77] for compatibility with conventional artificial neural networks. An event volume is discretized in the time dimension, and each voxel counts the number of events within it (bi-linearly voted for undistorted, rectified events). + +In principle, our self-supervised loss module can be paired with any segmentation or optical flow network architecture. We use a U-Net architecture for the experiments on DSEC (Sec. 4.2), and the architecture in the recent Bflow method [25], inspired by RAFT [62], for the experiments on EVIMO2 (Sec. 4.1). + +# 4 Experiments + +We test our method on two applications. First, we test the capabilities for nonlinear trajectory estimation on the real-world dataset EVIMO2 [7], with additional results on the synthetic MultiFlow dataset [25], which we use for pretraining (Sec. 4.1). Secondly, we evaluate our method on the DSEC dataset [23, 24] because this has been the previous frontier for event-based optical flow estimation and it allows for direct comparison with prior work (Sec. 4.2). Additional results, including on the MVSEC dataset [75], are given in the supplement. + +# 4.1 Results on Non-Linear Trajectory Estimation + +EVIMO2 Continuous Flow Dataset (CF-EVIMO2). MultiFlow [25] was first proposed to evaluate long-term "optical flow" methods. The dataset consists + +![](images/b0e3aa4cd2694299a90e822a42492a054bad75bcf13382c394eb69df9c30d443.jpg) +Fig. 3: Visualization of predicted trajectories on EVIMO2 data. $GT$ : Ground truth. In-domain: fine-tuned on EVIMO2 using GT (supervised). Zero-shot: network trained only on synthetic data (out-of-domain prediction). Ours: Pre-trained on synthetic data, fine-tuned with self-supervised loss. Note that supervision in-domain is often impossible in practice because dense trajectory labels for real data are difficult to obtain. + +of synthetic videos generated with Internet images as foreground and Flickr30K [71] images as background. Events are generated with a simulator [47] by rendering the synthetic scenes at 1000 frames/s. However, due to the sim-to-real gap between real and synthetic events, it is challenging to infer the actual performance of the proposed and baseline methods on real datasets. Furthermore, Multiflow generates trajectories by pasting 2D foreground objects onto background images, which lacks the challenging cases of self-occlusion due to 3D rotation. + +To address these data limitations, we present the EVIMO2 Continuous Flow Dataset based on the full 3D data and event data provided by the meticulously designed EVIMO2 [7] dataset. The EVIMO2 GT data provides high-quality 3D scans of the objects, camera poses, object poses, and camera intrinsics, which we use to compute optical flow GT as follows. Given a point $P_{t}^{i}$ on the object represented in the camera coordinate system at time $t$ , we project it onto a 2D point + +![](images/f8b3fb6b3d8b470457f6bb9fd0a8c874f50901f8ef3b49c2b3056ef6155a147c.jpg) +Fig. 4: End-point-error vs. prediction time span for three methods: in-distribution, out-of-distribution and self-supervised (Ours). Using the Bézier curve results from Tab. 3. + +Table 3: Sensitivity with respect to motion prior on EVIMO2 [7]. $SL$ , $ID$ : supervised, in-distribution, $SL$ , $OOD$ : supervised, out of distribution, $SSL$ : self-supervised. + +
MethodTEPE ↓TAE ↓%Out ↓
SL, IDBFlow (polyn.)3.5113.260.18
BFlow (learned)3.7813.870.19
BFlow (Bézier)3.3811.680.17
SL, OODBFlow (polyn.)9.3620.880.36
BFlow (learned)8.6619.640.35
BFlow (Bézier) [25]8.6319.940.36
SSLOurs, polyn.6.7819.760.27
Ours, learned7.4619.780.28
Ours, Bézier6.1416.980.25
+ +$p_{t}^{i} = \pi (P_{t}^{i})$ , where the $\pi (\cdot)$ function represents the perspective projection, including the camera intrinsics. The object poses at time $t$ and $t + 1$ are provided as $T_{o}^{c_{t}}$ , $T_{o}^{c_{t + 1}}$ in the GT data. The rigid-body transformation $T_{c_t}^{t + 1} = (T_o^{c_t + 1})^{-1}T_o^{c_t}$ maps 3D points in the camera frame at time $t$ to the camera frame at time $t + 1$ by aligning the shared object coordinate frame. The "flow" (or dense motion, since it may not be linear) is defined as $\Delta p_{t\rightarrow t + 1}^i\doteq \pi (T_{c_t}^{c_{t + 1}}P_t^i) - \pi (P_t^i)$ . For each sequence, we generate GT dense motion every 10ms for 300ms. We mask out areas where GT object masks are unavailable. Experiments are carried out on the IMO subset of EVIMO2 using the official train and test splits. + +Metrics. Ground truth is provided as dense motion from $t = 0$ to several timesteps in increasing order. We consider $N_{s} = 6$ timestamps (i.e., subintervals) in a 300ms window and evaluate the quality of predictions with direct extension of the common optical flow metrics end-point-error (EPE), angular error (AE), and percentage of predicted vectors above a specific EPE threshold (\%Out). We term the corresponding trajectory metrics TEPE and TAE, where + +$$ +\mathrm {T E P E} = \frac {1}{N _ {s}} \sum_ {k = 1} ^ {N _ {s}} \mathrm {E P E} (\varDelta \mathbf {x} _ {\mathrm {p r e d}} (t _ {k}), \varDelta \mathbf {x} _ {\mathrm {g t}} (t _ {k})). \tag {6} +$$ + +The number of outliers is calculated on TEPE with a threshold of 3px. + +Implementation Details. We use the BFlow network architecture and test different motion priors. Specifically, we report results with polynomial, Bezier (with $N_{c} = 10$ ) and learned basis $g_{j}(t)$ (a small dense neural network with three layers and hidden dimension 64 is trained alongside the main network). + +The contrast loss module uses one trajectory per $4 \times 4$ px, and similarly $4 \times 4$ px in the displacement volume. The weight for the spatial smoothness term is $\lambda = 0.003$ , and the number of neighbors in the KNN approach is $N_{\mathrm{traj}} = 32$ . The loss function $G$ is the $L^1$ norm of the IWE's gradient magnitude [18,56]. + +Training Schedule. The model is first pre-trained on MultiFlow ( $\approx 12000$ samples) for 50 epochs, with a batch size of 10 with an $L^1$ loss on the GT trajectory flow. We include data augmentation by flipping (horizontal and vertical) and cropping and use the training/test split provided in [25]. These weights are what we refer to as zero-shot or out-of-distribution (OOD) for the main comparison on EVIMO2. Afterwards, the pre-trained model is fine-tuned with two different losses on EVIMO2. Training using our self-supervised loss is carried out for 15 epochs with a batch size of 6. We refer to this method as Ours. Lastly, a second version is trained directly on the GT flow of EVIMO for 50 epochs, which we refer to as in-distribution (ID). All experiments are performed on Nvidia RTX A6000 GPUs with an AdamW optimizer and a learning rate of $10^{-4}$ . + +Baselines. This is the first usage of EVIMO2 for dense non-linear flow/motion. For comparison, we provide additional baselines. We use ERAFT [24] with the provided DSEC weights to infer linear flow over the whole prediction time. The timestamps of the intermediate flow are interpolated from the linear prediction. Additionally, we provide results for the prediction of Paredes et. al. [44]. The network is trained self-supervised on DSEC and performs recurrent prediction steps in time in intervals of $10\mathrm{ms}$ , inferring the total flow at every step by accumulation of the shorter flows. We unroll this prediction over the whole interval and take the accumulated flow after each $50\mathrm{ms}$ step as the predicted flow. + +Results. Table 2 presents the comparison of the three different training modes. The results of our self-supervised loss can improve the zero-shot performance by nearly $30\%$ . While direct supervision on EVIMO2 GT data delivers the best results, note that this is only possible because we are comparing on a dataset that was recorded with a motion capture setup and exact object models. Under real conditions, pixel-level annotations of motion tracks are very difficult to obtain. Therefore, our loss delivers a promising tool in combination with synthetic pretraining, substantially helping in overcoming the sim-to-real gap. + +Figure 3 shows the qualitative comparison of GT tracks and the three training schedules. It delivers insights into the failure cases that our method can overcome. Specifically, the zero-shot model shows errors like overseen object motion (Examples 2 and 3), wrong scale (Ex. 5), and prediction of non-existing motion (Ex. 6), which are improved with self-supervised domain training. + +Additionally, Fig. 4 shows the error for different prediction time spans. Intuitively, the error increases for longer intervals. The visualization shows that our loss improves zero-shot performance, especially for the long prediction times. + +Table 4: Results on the DSEC optical flow benchmark [24]. + +
Method\( {t}_{\text{inf }}\left\lbrack \mathrm{{ms}}\right\rbrack \)Allinterlaken_00_binterlaken_01_athun_01_a
EPE \( \downarrow \)AE \( \downarrow \)%Out ↓FWL ↑EPE \( \downarrow \)AE \( \downarrow \)%Out ↓FWL ↑EPE \( \downarrow \)AE \( \downarrow \)%Out ↓FWL ↑EPE \( \downarrow \)AE \( \downarrow \)%Out ↓FWL ↑
E-RAFT (SL) [24]46.330.7910.562.681.291.396.226.121.320.906.883.911.420.659.751.871.20
IDNet (SL) [69]0.722.722.04-1.252.114.35-0.772.252.60-0.572.661.47-
Paredes et al. (SSL) [44]40.102.3310.5617.77-3.346.2225.72-2.496.8819.15-1.739.7510.39-
RTEF (MB) [5]4.88-41.952.518.59-59.842.895.94-47.332.923.01-29.702.39
Ev-FlowNet (SSL) [27]3.86-31.451.306.32-47.951.464.91-36.071.422.33-20.921.32
MultiCM (MB) [54]\( {9.9} \cdot {10}^{3} \)3.4713.9830.861.375.749.1938.931.503.749.7731.371.512.1211.0617.681.24
Ours (poly, \( k = 1 \) )7.273.208.5315.211.463.214.8920.451.582.385.4617.401.701.396.997.361.30
+ +
Methodthun_01_bzurich_city_12_azurich_city_14_czurich_city_15_a
EPE ↓AE ↓%Out ↓FWL ↑EPE ↓AE ↓%Out ↓FWL ↑EPE ↓AE ↓%Out ↓FWL ↑EPE ↓AE ↓%Out ↓FWL ↑
E-RAFT (SL) [24]0.588.411.521.180.6123.161.061.120.7110.231.911.470.598.881.301.34
IDNet (SL) [69]0.552.071.35-0.604.561.16-0.763.742.74-0.552.551.02-
Paredes et al. (SSL) [44]1.668.419.34-2.7223.1626.65-2.6410.2323.01-1.698.889.98-
RTEF (MB) [5]3.91-34.692.483.14-34.081.424.00-45.672.673.78-37.992.82
EV-FlowNet (SSL) [77]3.04-25.411.332.62-25.801.033.36-36.341.242.97-25.531.33
MultiCM (MB) [54]2.4812.0523.561.243.8628.6143.961.142.7212.6230.531.502.3511.8220.991.41
Ours (poly, k = 1)1.546.559.691.338.3320.1622.391.131.788.7912.991.561.456.278.341.51
+ +Sensitivity to motion prior. Table 3 compares our method for different motion priors. While Bézier curves show slightly improved performance over the two basis function methods, self-supervised domain training can improve performance in all cases, which proves the robustness and generality of our method. + +# 4.2 Results on Optical Flow Estimation + +Dataset, Metrics and Details. We compare our method on the DSEC optical flow benchmark [24]. The dataset consists of sequences from a Prophesee Gen3 event camera, with a resolution of $640 \times 480$ px, mounted on a driving car. Ground truth is calculated as the motion field from a co-deployed depth sensor. + +We compare the common flow metrics EPE, AE, and percentage of outliers (%Out), where an outlier is a flow vector with EPE $>3\mathrm{px}$ . Additionally, we evaluate the Flow Warp Loss (FWL) [57]. FWL $>1$ means that the IWE is sharper than pixel-wise event accumulation (i.e., IWE with zero optical flow). + +We use a simple U-Net [49] as the backbone in the linear flow experiments (similar architecture like EV-FlowNet [76]) and choose polynomial basis $g_{j}(t) = t^{j}$ of degree 1, which effectively leads to a linear motion prior. We use $N_{\mathrm{traj}} = 32$ nearest neighbors and $N_{\mathrm{bins}} = 15$ time bins in the displacement map. The weight for the regularizer is $\lambda = 0.003$ . The network trains for 50 epochs with and Adam optimizer, a learning rate of $10^{-4}$ , and a total batch size of 28 on two RTX A6000. Results. Table 4 shows the DSEC benchmark results, confirming that our model can generate high-quality optical flow results without having access to any GT. It furthermore achieves the best results among all contrast-maximization-based methods on six out of seven sequences, and only methods directly supervised on GT perform better. Within the self-supervised methods, it improves the AE by $19\%$ and the number of inliers by $14\%$ . The only exception is the night sequence zurich_city_14. Contrast maximization is based on the brightness constancy assumption, which is violated here by flickering street lights. Hence, in these areas, the contrast loss does not provide reasonable flow predictions. While our methods still provide a better AE and %Out at night, higher outliers in the + +![](images/eaa3d82bf0778e152eaf6607b1c9b2320ebd76e2302e39fd4a869b50b4749b38.jpg) +Fig. 5: Results on DSEC. Image of warped events and predicted flow by three methods. + +flickering regions prevent the difference in the average EPE from reflecting the improved performance of our method on the other six sequences. + +Figure 5 shows qualitative results of our method. In comparison to another self-supervised optical flow method (d), the visualizations are noticeably sharper, allowing for a more precise representation of the motion in the scene. Moreover, the model performs well in delineating the contours of foreground objects, without the over-smoothing effect often observable in flow methods. The IWE (Fig. 5b) shows sharp results, aligning with the overall high FWL values reported in Tab. 4. While the FWL metric also increases in event collapse (e.g., for RTEF [5]), the IWE visualization reveals that here the model is well-regularized, confirming that the predicted motion aligns the events correctly. The third example in Fig. 5 highlights how our model has improved predictions with a single forward pass, where models relying on temporal recurrence (see (d)) fail at the beginning and need a warm-up stage resulting in higher latency. + +Inference time. Additionally, Tab. 4 provides inference times $t_{\mathrm{inf}}$ for several methods. Note that our method is about $5 \times$ faster than the competitive baselines. The reason is that we do not rely on any recurrence in this optical flow application, such as recurrence in time or RAFT-inspired refinement steps. + +Ablation and Sensitivity. Table 5 lists the performance under different loss settings. It confirms that the number of neighboring trajectories $N_{\mathrm{traj}}$ in the KNN approach has a regularizing effect and is important for good performance. The decreased performance on the benchmark for a lower $N_{\mathrm{traj}}$ aligns with the visual impression in Fig. 6: models with lower $N_{\mathrm{traj}}$ show artifacts and a stronger susceptibility to the aperture problem, as is observable on the rock pattern. + +Table 5 confirms that the multi-reference approach is crucial for regularizing the contrast loss. While using three instead of one $t_{\mathrm{ref}}$ improves performance, our approach using a randomized $t_{\mathrm{ref}}$ leads to an additional performance boost. Lastly, the non-linear prior showed no performance improvement on DSEC. + +Table 5: Sensitivity and Ablation Study for DSEC. Ours corresponds to "Ours" in Tab. 4. Configurations marked with “-” are unchanged from our main result. $\Delta$ specifies the change with respect to the original configuration. + +
NtrajNrefNcMotion priorEPE ↓ΔEPEAE ↓ΔAE%Out ↓Δ%Out
Ours32~U(0,1)1polynomial3.208.5315.21
Number of neighbor trajectories1---4.581.0713.735.2027.9012.69
8---3.510.3113.204.6723.398.18
Number of reference times-3--4.461.2614.435.9028.6013.39
-1--7.264.0618.7210.1944.7529.54
Type and degree of motion prior--5learned3.220.028.590.0615.340.13
--5polynomial3.280.068.610.0815.680.47
+ +![](images/4dd03fe367e6c8c39a80585d3f16b81aedc791444beabe2da998261ca097fdc5.jpg) +Fig. 6: Visual effect of the number of neighbors $N_{\mathrm{traj}}$ on the predicted flow. + +![](images/36fcbb9a333d2945ec606ba165b3dd2a9bdd499f72d6e37366ac886719d386cd.jpg) + +# 5 Limitations + +Like all CM-based methods, ours is based on the brightness constancy assumption. Therefore, it shows limitations in estimating flow from events that are not caused by motion, e.g., from flickering lights. + +Our KNN interpolation shows a good trade-off between granularity and regularization; nevertheless, our method is limited by the aperture problem inherent to optical flow. + +Lastly, during training, all raw events are passed to the loss module, increasing training time. Furthermore, gradient calculation is performed through the additional step of event warping. These steps increase memory requirements and training time of our loss compared to supervised methods. + +# 6 Conclusion + +We introduced a new loss formulation based on the contrast maximization framework, by combining it with a non-linear trajectory prior. It is a versatile tool that works with a variety of model architectures and trajectory representations. We presented an efficient method to make the high-dimensional assignment between millions of events and thousands of trajectories feasible. Our experiments show clear advantages against supervised methods on unseen data, where under real-world circumstances no GT is available. Additionally, a U-Net trained with our loss shows state-of-the-art performance on the DSEC optical flow benchmark, while being substantially faster than the previously best methods. + +# Acknowledgments + +We thank Dr. Cornelia Fermüller and the NeuroPAC network for fostering collaborations within the event-based community (NSF OISE 2020624). Funded by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) under Germany's Excellence Strategy - EXC 2002/1 "Science of Intelligence" - project number 390523135. We furthermore gratefully acknowledge the support by the following grants: NSF FRR 2220868, NSF IIS-RI 2212433, NSF TRIPODS 1934960, ONR N00014-22-1-2677. + +# References + +1. Akhter, I., Sheikh, Y., Khan, S., Kanade, T.: Trajectory space: A dual representation for nonrigid structure from motion. IEEE Trans. Pattern Anal. Mach. Intell. 33(7), 1442-1456 (jul 2011). https://doi.org/10.1109/TPAMI.2010.201 +2. Bailer, C., Varanasi, K., Stricker, D.: CNN-based patch matching for optical flow with thresholded hinge embedding loss. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR). pp. 2710-2719 (2017). https://doi.org/10.1109/CVPR.2017.290 +3. Benosman, R., Clercq, C., Lagorce, X., Ieng, S.H., Bartolozzi, C.: Event-based visual flow. IEEE Trans. Neural Netw. Learn. Syst. 25(2), 407-417 (2014). https://doi.org/10.1109/TNNLS.2013.2273537 +4. Benosman, R., Ieng, S.H., Clercq, C., Bartolozzi, C., Srinivasan, M.: Asynchronous frameless event-based optical flow. Neural Netw. 27, 32-37 (2012). https://doi.org/10.1016/j.neunet.2011.11.001 +5. Brebion, V., Moreau, J., Davoine, F.: Real-time optical flow for vehicular perception with low- and high-resolution event cameras. IEEE Trans. Intell. Transport. Syst. pp. 1-13 (2021). https://doi.org/10.1109/TITS.2021.3136358 +6. Brosch, T., Tschechne, S., Neumann, H.: On event-based optical flow detection. Front. Neurosci. 9(137) (Apr 2015). https://doi.org/10.3389/fnins.2015.00137 +7. Burner, L., Mitrokhin, A., Fermüller, C., Aloimonos, Y.: EVIMO2: An event camera dataset for motion segmentation, optical flow, structure from motion, and visual inertial odometry in indoor scenes with monocular or stereo algorithms. arXiv e-prints (May 2022). https://doi.org/10.48550/arXiv.2205.03467 +8. Butler, D.J., Wulff, J., Stanley, G.B., Black, M.J.: A naturalistic open source movie for optical flow evaluation. In: Eur. Conf. Comput. Vis. (ECCV). pp. 611-625 (2012). https://doi.org/10.1007/978-3-642-33783-3_44 +9. Chaney, K., Cladera Ojeda, F., Wang, Z., Bisulco, A., Hsieh, M.A., Korpela, C., Kumar, V., Taylor, C.J., Daniilidis, K.: M3ED: Multi-robot, multi-sensor, multi-environment event dataset. In: IEEE Conf. Comput. Vis. Pattern Recog. Workshops (CVPRW). pp. 4016-4023 (2023). https://doi.org/10.1109/CVPRW59228.2023.00419 +10. Chui, J., Klenk, S., Cremers, D.: Event-based feature tracking in continuous time with sliding window optimization. In: arXiv e-prints (2021). https://doi.org/10.48550/arXiv.2107.04536 +11. Ding, Z., Zhao, R., Zhang, J., Gao, T., Xiong, R., Yu, Z., Huang, T.: Spatiotemporal recurrent networks for event-based optical flow estimation. In: AAAI Conf. Artificial Intell. vol. 36, pp. 525-533 (2022). https://doi.org/10.1609/aaai.v36i1.19931 + +12. Doersch, C., Gupta, A., Markeeva, L., Recasens, A., Smaira, L., Aytar, Y., Carreira, J., Zisserman, A., Yang, Y.: Tap-vid: A benchmark for tracking any point in a video. Adv. Neural Inf. Process. Syst. (NeurIPS) 35, 13610-13626 (2022) +13. Doersch, C., Yang, Y., Vecerik, M., Gokay, D., Gupta, A., Aytar, Y., Carreira, J., Zisserman, A.: Tapir: Tracking any point with per-frame initialization and temporal refinement. arXiv preprint arXiv:2306.08637 (2023) +14. Dosovitskiy, A., Fischer, P., Ilg, E., Häusser, P., Hazirbaş, C., Golkov, V., van der Smagt, P., Cremers, D., Brox, T.: FlowNet: Learning optical flow with convolutional networks. In: Int. Conf. Comput. Vis. (ICCV). pp. 2758-2766 (2015). https://doi.org/10.1109/ICCV.2015.316 +15. Feydy, J., Glaunès, A., Charlier, B., Bronstein, M.: Fast geometric learning with symbolic matrices. Adv. Neural Inf. Process. Syst. (NeurIPS) 33, 14448-14462 (2020) +16. Finateu, T., Niwa, A., Matolin, D., Tsuchimoto, K., Mascheroni, A., Reynaud, E., Mostafalu, P., Brady, F., Chotard, L., LeGoff, F., Takahashi, H., Wakabayashi, H., Oike, Y., Posch, C.: A 1280x720 back-illuminated stacked temporal contrast event-based vision sensor with $4.86\mu \mathrm{m}$ pixels, 1.066Geps readout, programmable event-rate controller and compressive data-formating pipeline. In: IEEE Int. Solid-State Circuits Conf. (ISSCC). pp. 112-114 (2020). https://doi.org/10.1109/ISSCC19947.2020.9063149 +17. Gallego, G., Delbruck, T., Orchard, G., Bartolozzi, C., Taba, B., Censi, A., Leutenegger, S., Davison, A., Conradt, J., Daniilidis, K., Scaramuzza, D.: Event-based vision: A survey. IEEE Trans. Pattern Anal. Mach. Intell. 44(1), 154-180 (2022). https://doi.org/10.1109/TPAMI.2020.3008413 +18. Gallego, G., Gehrig, M., Scaramuzza, D.: Focus is all you need: Loss functions for event-based vision. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR). pp. 12272-12281 (2019). https://doi.org/10.1109/CVPR.2019.01256 +19. Gallego, G., Rebecq, H., Scaramuzza, D.: A unifying contrast maximization framework for event cameras, with applications to motion, depth, and optical flow estimation. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR). pp. 3867-3876 (2018). https://doi.org/10.1109/CVPR.2018.00407 +20. Gallego, G., Scaramuzza, D.: Accurate angular velocity estimation with an event camera. IEEE Robot. Autom. Lett. 2(2), 632-639 (2017). https://doi.org/10.1109/LRA.2016.2647639 +21. Gehrig, D., Gehrig, M., Hidalgo-Carrió, J., Scaramuzza, D.: Video to Events: Recycl- cling video datasets for event cameras. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR). pp. 3583-3592 (2020). https://doi.org/10.1109/CVPR42600.2020.00364 +22. Gehrig, D., Loquercio, A., Derpanis, K.G., Scaramuzza, D.: End-to-end learning of representations for asynchronous event-based data. In: Int. Conf. Comput. Vis. (ICCV). pp. 5632-5642 (2019). https://doi.org/10.1109/ICCV.2019.00573 +23. Gehrig, M., Aarents, W., Gehrig, D., Scaramuzza, D.: DSEC: A stereo event camera dataset for driving scenarios. IEEE Robot. Autom. Lett. 6(3), 4947-4954 (2021). https://doi.org/10.1109/LRA.2021.3068942 +24. Gehrig, M., Millhäsler, M., Gehrig, D., Scaramuzza, D.: E-RAFT: Dense optical flow from event cameras. In: Int. Conf. 3D Vision (3DV). pp. 197-206 (2021). https://doi.org/10.1109/3DV53792.2021.00030 +25. Gehrig, M., Muglikar, M., Scaramuzza, D.: Dense continuous-time optical flow from event cameras. IEEE Trans. Pattern Anal. Mach. Intell. pp. 1-12 (2024). https://doi.org/10.1109/TPAMI.2024.3361671 + +26. Guney, F., Sevilla-Lara, L., Sun, D., Wulff, J.: "what is optical flow for?": Workshop results and summary. In: Eur. Conf. Comput. Vis. Workshops (ECCVW). pp. 0-0 (2018) +27. Guo, S., Gallego, G.: CMax-SLAM: Event-based rotational-motion bundle adjustment and SLAM system using contrast maximization. IEEE Trans. Robot. 40, 2442-2461 (2024). https://doi.org/10.1109/TRO.2024.3378443 +28. Hagenaars, J., Paredes-Vallés, F., De Croon, G.: Self-supervised learning of event-based optical flow with spiking neural networks. Adv. Neural Inf. Process. Syst. (NeurIPS) 34, 7167-7179 (2021) +29. Hamann, F., Gallego, G.: Stereo co-capture system for recording and tracking fish with frame- and event cameras. In: 26th Int. Conf. on Pattern Recognition (ICPR), Visual observation and analysis of Vertebrate And Insect Behavior (VAIB) Workshop (2022). https://doi.org/10.48550/ARXIV.2207.07332 +30. Hamann, F., Ghosh, S., Juarez-Martínez, I., Hart, T., Kacelnik, A., Gallego, G.: Low-power, continuous remote behavioral localization with event cameras. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR) (2024) +31. Harley, A.W., Fang, Z., Fragkiadaki, K.: Particle video revisited: Tracking through occlusions using point trajectories. In: Eur. Conf. Comput. Vis. (ECCV). pp. 59-75 (2022) +32. Huang, Z., Shi, X., Zhang, C., Wang, Q., Cheung, K.C., Qin, H., Dai, J., Li, H.: Flowformer: A transformer architecture for optical flow. In: Eur. Conf. Comput. Vis. (ECCV). pp. 668-685 (2022) +33. Ilg, E., Mayer, N., Saikia, T., Keuper, M., Dosovitskiy, A., Brox, T.: FlowNet 2.0: Evolution of optical flow estimation with deep networks. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR). pp. 1647-1655 (2017). https://doi.org/10.1109/cvpr.2017.179 +34. Janai, J., Guney, F., Ranjan, A., Black, M., Geiger, A.: Unsupervised learning of multi-frame optical flow with occlusions. In: Eur. Conf. Comput. Vis. (ECCV). pp. 690-706 (2018) +35. Jonschkowski, R., Stone, A., Barron, J.T., Gordon, A., Konolige, K., Angelova, A.: What matters in unsupervised optical flow. In: Eur. Conf. Comput. Vis. (ECCV). pp. 557-572 (2020) +36. Karaev, N., Rocco, I., Graham, B., Neverova, N., Vedaldi, A., Rupprecht, C.: Cotracker: It is better to track together. arXiv preprint arXiv:2307.07635 (2023) +37. Kim, H., Kim, H.J.: Real-time rotational motion estimation with contrast maximization over globally aligned events. IEEE Robot. Autom. Lett. 6(3), 6016-6023 (2021). https://doi.org/10.1109/LRA.2021.3088793 +38. Lee, C., Kosta, A., Zhu, A.Z., Chaney, K., Daniilidis, K., Roy, K.: Spike-flownet: Event-based optical flow estimation with energy-efficient hybrid neural networks. In: Eur. Conf. Comput. Vis. (ECCV). pp. 366-382 (2020) +39. Lichtsteiner, P., Posch, C., Delbruck, T.: A $128 \times 128$ 120 dB $15~\mu \mathrm{s}$ latency asynchronous temporal contrast vision sensor. IEEE J. Solid-State Circuits 43(2), 566-576 (2008). https://doi.org/10.1109/JSSC.2007.914337 +40. Liu, H., Chen, G., Qu, S., Zhang, Y., Li, Z., Knoll, A., Jiang, C.: TMA: Temporal motion aggregation for event-based optical flow. In: Int. Conf. Comput. Vis. (ICCV). pp. 9651-9660 (Oct 2023). https://doi.org/10.1109/ICCV51070.2023.00888 +41. Liu, M., Delbruck, T.: Adaptive time-slice block-matching optical flow algorithm for dynamic vision sensors. In: British Mach. Vis. Conf. (BMVC). pp. 1-12 (2018) + +42. Mayer, N., Ilg, E., Hausser, P., Fischer, P., Cremers, D., Dosovitskiy, A., Brox, T.: A large dataset to train convolutional networks for disparity, optical flow, and scene flow estimation. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR). pp. 4040-4048 (2016) +43. Orchard, G., Benosman, R., Etienne-Cummings, R., Thakor, N.V.: A spiking neural network architecture for visual motion estimation. In: IEEE Biomed. Circuits Syst. Conf. (BioCAS). pp. 298-301 (2013). https://doi.org/10.1109/biocas.2013.6679698 +44. Paredes-Valles, F., Scheper, K.Y., De Wagter, C., de Croon, G.C.: Taming contrast maximization for learning sequential, low-latency, event-based optical flow. In: Int. Conf. Comput. Vis. (ICCV). pp. 9661-9671 (Oct 2023). https://doi.org/10.1109/ICCV51070.2023.00889 +45. Paredes-Valles, F., Hagenaars, J.J., Dupeyroux, J., Stroobants, S., Xu, Y., de Croon, G.C.H.E.: Fully neuromorphic vision and control for autonomous drone flight. Science Robotics 9(90), eadi0591 (2024). https://doi.org/10.1126/scirobotics.adio591 +46. Posch, C., Serrano-Gotarredona, T., Linares-Barranco, B., Delbruck, T.: Retinomorphic event-based vision sensors: Bioinspired cameras with spiking output. Proc. IEEE 102(10), 1470-1484 (Oct 2014). https://doi.org/10.1109/jproc.2014.2346153 +47. Rebecq, H., Gehrig, D., Scaramuzza, D.: ESIM: an open event camera simulator. In: Conf. on Robotics Learning (CoRL). Proc. Machine Learning Research, vol. 87, pp. 969-982. PMLR (2018) +48. Ren, Z., Yan, J., Ni, B., Liu, B., Yang, X., Zha, H.: Unsupervised deep learning for optical flow estimation. In: AAAI Conf. Artificial Intell. vol. 31 (2017) +49. Ronneberger, O., Fischer, P., Brox, T.: U-Net: Convolutional networks for biomedical image segmentation. In: Int. Conf. Medical Image Computing and Computer-Assisted Intervention (MICCAI). pp. 234-241 (2015) +50. Sand, P., Teller, S.: Particle video: Long-range motion estimation using point trajectories. Int. J. Comput. Vis. 80, 72-91 (2008) +51. Seok, H., Lim, J.: Robust feature tracking in dvs event stream using Bezier mapping. In: IEEE Winter Conf. Appl. Comput. Vis. (WACV). pp. 1647-1656 (2020). https://doi.org/10.1109/WACV45572.2020.9093607 +52. Shiba, S., Aoki, Y., Gallego, G.: Event collapse in contrast maximization frameworks. Sensors 22(14), 1-20 (2022). https://doi.org/10.3390/s22145190 +53. Shiba, S., Aoki, Y., Gallego, G.: A fast geometric regularizer to mitigate event collapse in the contrast maximization framework. Adv. Intell. Syst. p. 2200251 (2022). https://doi.org/10.1002/aisy.202200251 +54. Shiba, S., Aoki, Y., Gallego, G.: Secrets of event-based optical flow. In: Eur. Conf. Comput. Vis. (ECCV). pp. 628-645 (2022). https://doi.org/10.1007/978-3-031-19797-0 36 +55. Shiba, S., Hamann, F., Aoki, Y., Gallego, G.: Event-based background oriented schlieren. IEEE Trans. Pattern Anal. Mach. Intell. 46(4), 2011-2026 (2024). https://doi.org/10.1109/TPAMI.2023.3328188 +56. Shiba, S., Klose, Y., Aoki, Y., Gallego, G.: Secrets of event-based optical flow, depth, and ego-motion by contrast maximization. IEEE Trans. Pattern Anal. Mach. Intell. pp. 1-18 (2024). https://doi.org/10.1109/TPAMI.2024.3396116 +57. Stoffregen, T., Scheerlinck, C., Scaramuzza, D., Drummond, T., Barnes, N., Kleeman, L., Mahony, R.: Reducing the sim-to-real gap for event cameras. In: Eur. Conf. Comput. Vis. (ECCV). pp. 534-549 (2020). https://doi.org/https://doi.org/10.1007/978-3-030-58583-9_32 + +58. Stone, A., Maurer, D., Ayvaci, A., Angelova, A., Jonschkowski, R.: Smurf: Self-teaching multi-frame unsupervised raft with full-image warping. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR). pp. 3887-3896 (2021) +59. Sun, D., Vlasic, D., Herrmann, C., Jampani, V., Krainin, M., Chang, H., Zabih, R., Freeman, W.T., Liu, C.: Autoflow: Learning a better training set for optical flow. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR). pp. 10093-10102 (2021) +60. Sun, D., Yang, X., Liu, M.Y., Kautz, J.: PWC-Net: CNNs for optical flow using pyramid, warping, and cost volume. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR) (2018) +61. Sun, X., Harley, A.W., Guibas, L.J.: Refining pre-trained motion models. arXiv preprint arXiv:2401.00850 (2024) +62. Teed, Z., Deng, J.: RAFT: Recurrent all pairs field transforms for optical flow. In: Eur. Conf. Comput. Vis. (ECCV). pp. 402-419 (2020). https://doi.org/10.1007/978-3-030-58536-5_24 +63. Tulyakov, S., Bochicchio, A., Gehrig, D., Georgoulis, S., Li, Y., Scaramuzza, D.: Time lens++: Event-based frame interpolation with parametric non-linear flow and multi-scale fusion. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR). pp. 17755-17764 (Jun 2022) +64. Valmadre, J., Lucey, S.: General trajectory prior for non-rigid reconstruction. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR). pp. 1394-1401 (2012) +65. Wan, Z., Dai, Y., Mao, Y.: Learning dense and continuous optical flow from an event camera. IEEE Trans. Image Process. 31, 7237-7251 (2022) +66. Wang, C., Eckart, B., Lucey, S., Gallo, O.: Neural trajectory fields for dynamic novel view synthesis. arXiv preprint arXiv:2105.05994 (2021) +67. Wang, Z., Hamann, F., Chaney, K., Jiang, W., Gallego, G., Daniilidis, K.: Event-based continuous color video decompression from single frames. arXiv preprint arXiv:2312.00113 (2023) +68. Weikersdorfer, D., Hoffmann, R., Conradt, J.: Simultaneous localization and mapping for event-based vision systems. In: Int. Conf. Comput. Vis. Syst. (ICVS). pp. 133-142 (2013). https://doi.org/10.1007/978-3-642-39402-7_14 +69. Wu, Y., Paredes-Valles, F., de Croon, G.C.: Lightweight event-based optical flow estimation via iterative deblurring. arXiv preprint arXiv:2211.13726 (2022) +70. Ye, C., Mitrokhin, A., Parameshwara, C., Fermüller, C., Yorke, J.A., Aloimonos, Y.: Unsupervised learning of dense optical flow, depth and egomotion with event-based sensors. In: IEEE/RSJ Int. Conf. Intell. Robot. Syst. (IROS). pp. 5831-5838 (2020). https://doi.org/10.1109/IROS45743.2020.9341224 +71. Young, P., Lai, A., Hodosh, M., Hockenmaier, J.: From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. Trans. Assoc. Computational Linguistics 2, 67-78 (2014) +72. Yu, J.J., Harley, A.W., Derpanis, K.G.: Back to basics: Unsupervised learning of optical flow via brightness constancy and motion smoothness. In: Eur. Conf. Comput. Vis. Workshops (ECCVW). pp. 3-10 (2016) +73. Zen, G., Ricci, E., Sebe, N.: Exploiting sparse representations for robust analysis of noisy complex video scenes. In: Eur. Conf. Comput. Vis. (ECCV). pp. 199-213 (2012) +74. Zheng, Y., Harley, A.W., Shen, B., Wetzstein, G., Guibas, L.J.: PointOdyssey: A large-scale synthetic dataset for long-term point tracking. In: Int. Conf. Comput. Vis. (ICCV). pp. 19855-19865 (2023) +75. Zhu, A.Z., Thakur, D., Ozaslan, T., Pfrommer, B., Kumar, V., Daniilidis, K.: The multivehicle stereo event camera dataset: An event camera dataset for 3D + +perception. IEEE Robot. Autom. Lett. 3(3), 2032-2039 (Jul 2018). https://doi.org/10.1109/lra.2018.2800793 +76. Zhu, A.Z., Yuan, L., Chaney, K., Daniilidis, K.: EV-FlowNet: Self-supervised optical flow estimation for event-based cameras. In: Robotics: Science and Systems (RSS). pp. 1-9 (2018). https://doi.org/10.15607/RSS.2018.XIV.062 +77. Zhu, A.Z., Yuan, L., Chaney, K., Daniilidis, K.: Unsupervised event-based learning of optical flow, depth, and egomotion. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR). pp. 989-997 (2019). https://doi.org/10.1109/CVPR.2019.00108 +78. Zhu, Y., Lucey, S.: Convolutional sparse coding for trajectory reconstruction. IEEE Trans. Pattern Anal. Mach. Intell. 37(3), 529-540 (2013) \ No newline at end of file diff --git a/2024/Motion-prior Contrast Maximization for Dense Continuous-Time Motion Estimation/images.zip b/2024/Motion-prior Contrast Maximization for Dense Continuous-Time Motion Estimation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..c21cd8126f01f89491122f357899765a770a6f48 --- /dev/null +++ b/2024/Motion-prior Contrast Maximization for Dense Continuous-Time Motion Estimation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9091b654d39024e314c7dd01f052b66f29638be13821ec08e8240cab2844e47c +size 689206 diff --git a/2024/Motion-prior Contrast Maximization for Dense Continuous-Time Motion Estimation/layout.json b/2024/Motion-prior Contrast Maximization for Dense Continuous-Time Motion Estimation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..431e1d12b401c72a96619abc2fff3bdc2647342e --- /dev/null +++ b/2024/Motion-prior Contrast Maximization for Dense Continuous-Time Motion Estimation/layout.json @@ -0,0 +1,10748 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 54, + 30, + 358, + 64 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 30, + 358, + 64 + ], + "spans": [ + { + "bbox": [ + 54, + 30, + 358, + 64 + ], + "type": "text", + "content": "Motion-prior Contrast Maximization for Dense Continuous-Time Motion Estimation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 86, + 357, + 112 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 86, + 357, + 112 + ], + "spans": [ + { + "bbox": [ + 55, + 86, + 357, + 112 + ], + "type": "text", + "content": "Friedhelm Hamann1, Ziyun Wang2, Ioannis Asmanis2, Kenneth Chaney2, Guillermo Gallego1,3, and Kostas Daniilidis2,4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 82, + 120, + 328, + 143 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 120, + 328, + 143 + ], + "spans": [ + { + "bbox": [ + 82, + 120, + 328, + 143 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 82, + 120, + 328, + 143 + ], + "type": "text", + "content": " TU Berlin and SCIolI Excellence Cluster, Berlin, Germany \n" + }, + { + "bbox": [ + 82, + 120, + 328, + 143 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 82, + 120, + 328, + 143 + ], + "type": "text", + "content": " University of Pennsylvania, Philadelphia, US" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 35, + 143, + 376, + 164 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 35, + 143, + 376, + 164 + ], + "spans": [ + { + "bbox": [ + 35, + 143, + 376, + 164 + ], + "type": "text", + "content": "3 Einstein Center Digital Future and Robotics Institute Germany, Berlin, Germany \n4 Archimedes, Athena RC, Greece" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 59, + 199, + 353, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 199, + 353, + 365 + ], + "spans": [ + { + "bbox": [ + 59, + 199, + 353, + 365 + ], + "type": "text", + "content": "Abstract. Current optical flow and point-tracking methods rely heavily on synthetic datasets. Event cameras are novel vision sensors with advantages in challenging visual conditions, but state-of-the-art frame-based methods cannot be easily adapted to event data due to the limitations of current event simulators. We introduce a novel self-supervised loss combining the Contrast Maximization framework with a non-linear motion prior in the form of pixel-level trajectories and propose an efficient solution to solve the high-dimensional assignment problem between non-linear trajectories and events. Their effectiveness is demonstrated in two scenarios: In dense continuous-time motion estimation, our method improves the zero-shot performance of a synthetically trained model on the real-world dataset EVIMO2 by " + }, + { + "bbox": [ + 59, + 199, + 353, + 365 + ], + "type": "inline_equation", + "content": "29\\%" + }, + { + "bbox": [ + 59, + 199, + 353, + 365 + ], + "type": "text", + "content": ". In optical flow estimation, our method elevates a simple UNet to achieve state-of-the-art performance among self-supervised methods on the DSEC optical flow benchmark. Our code is available at https://github.com/tub-rip/MotionPriorCMax." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 32, + 388, + 128, + 399 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 388, + 128, + 399 + ], + "spans": [ + { + "bbox": [ + 32, + 388, + 128, + 399 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 30, + 415, + 381, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 415, + 381, + 511 + ], + "spans": [ + { + "bbox": [ + 30, + 415, + 381, + 511 + ], + "type": "text", + "content": "Determining the motion of arbitrary projected world points on the image plane over long time intervals is a difficult low-level computer vision problem. Researchers have studied it as optical flow and lately as point-tracking, with many practical applications in robotics, computational photography, video compression, and object-level tracking [26]. The best-performing frame-based methods for optical flow estimation and point tracking use large-scale synthetic datasets. Synthetic data is diverse and has highly accurate ground truth (GT), but is unrealistic and methods trained on synthetic data show a sim-to-real gap." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 30, + 512, + 381, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 512, + 381, + 584 + ], + "spans": [ + { + "bbox": [ + 30, + 512, + 381, + 584 + ], + "type": "text", + "content": "Novel vision sensors called event cameras [16,17,39] have emerged as promising alternatives to take on the problem. Inspired by the transient visual pathway, which is responsible for motion perception, they are particularly fit for capturing scene motion in the form of asynchronous pixel-wise intensity changes. This working principle endows them with advantages, such as high speed, high dynamic range (HDR) and low power consumption." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [], + "page_size": [ + 413, + 615 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 68, + 46, + 169, + 110 + ], + "blocks": [ + { + "bbox": [ + 68, + 46, + 169, + 110 + ], + "lines": [ + { + "bbox": [ + 68, + 46, + 169, + 110 + ], + "spans": [ + { + "bbox": [ + 68, + 46, + 169, + 110 + ], + "type": "image", + "image_path": "ad075927e22d4427da496a0fe9385f4e82a447e7725d122fa1ec6d2c8fe85a49.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 49, + 114, + 188, + 122 + ], + "lines": [ + { + "bbox": [ + 49, + 114, + 188, + 122 + ], + "spans": [ + { + "bbox": [ + 49, + 114, + 188, + 122 + ], + "type": "text", + "content": "a) Contrast Maximization with non-linear Trajectories" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 213, + 43, + 373, + 113 + ], + "blocks": [ + { + "bbox": [ + 213, + 43, + 373, + 113 + ], + "lines": [ + { + "bbox": [ + 213, + 43, + 373, + 113 + ], + "spans": [ + { + "bbox": [ + 213, + 43, + 373, + 113 + ], + "type": "image", + "image_path": "d638bfaa03eae5a1b184f8a61befaab34ee71241f48f2df2c3d0192aca61e1ea.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 220, + 114, + 275, + 121 + ], + "lines": [ + { + "bbox": [ + 220, + 114, + 275, + 121 + ], + "spans": [ + { + "bbox": [ + 220, + 114, + 275, + 121 + ], + "type": "text", + "content": "b) Overview Method" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 43, + 134, + 94, + 172 + ], + "blocks": [ + { + "bbox": [ + 43, + 134, + 94, + 172 + ], + "lines": [ + { + "bbox": [ + 43, + 134, + 94, + 172 + ], + "spans": [ + { + "bbox": [ + 43, + 134, + 94, + 172 + ], + "type": "image", + "image_path": "cfeaf3a68064fe3946cc8d75b9e32aa7998fc0aa1ee96b69fef645e7c0d3a039.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 95, + 134, + 147, + 172 + ], + "blocks": [ + { + "bbox": [ + 95, + 134, + 147, + 172 + ], + "lines": [ + { + "bbox": [ + 95, + 134, + 147, + 172 + ], + "spans": [ + { + "bbox": [ + 95, + 134, + 147, + 172 + ], + "type": "image", + "image_path": "bf7fe060a130b599dab3cda7f4387aaecc61c3e23db6b9d5519899c052f77b54.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 147, + 134, + 199, + 172 + ], + "blocks": [ + { + "bbox": [ + 147, + 134, + 199, + 172 + ], + "lines": [ + { + "bbox": [ + 147, + 134, + 199, + 172 + ], + "spans": [ + { + "bbox": [ + 147, + 134, + 199, + 172 + ], + "type": "image", + "image_path": "0501c97cc26b68f2996af6e510a69a55123e58ba6d3ced7a145af22b0bc68f36.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 43, + 172, + 94, + 210 + ], + "blocks": [ + { + "bbox": [ + 43, + 172, + 94, + 210 + ], + "lines": [ + { + "bbox": [ + 43, + 172, + 94, + 210 + ], + "spans": [ + { + "bbox": [ + 43, + 172, + 94, + 210 + ], + "type": "image", + "image_path": "9f2bed5b336b044fd57819770283e1c5d2b689e4a3744ec0883bd4a41d948e52.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 49, + 214, + 151, + 223 + ], + "lines": [ + { + "bbox": [ + 49, + 214, + 151, + 223 + ], + "spans": [ + { + "bbox": [ + 49, + 214, + 151, + 223 + ], + "type": "text", + "content": "c) Application to Trajectory Estimation" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 95, + 172, + 147, + 210 + ], + "blocks": [ + { + "bbox": [ + 95, + 172, + 147, + 210 + ], + "lines": [ + { + "bbox": [ + 95, + 172, + 147, + 210 + ], + "spans": [ + { + "bbox": [ + 95, + 172, + 147, + 210 + ], + "type": "image", + "image_path": "a20f3b923ecfeb1822e64d8465bd12fc9a11527a90ac13c3a7adddd36ef984d4.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 31, + 236, + 380, + 291 + ], + "lines": [ + { + "bbox": [ + 31, + 236, + 380, + 291 + ], + "spans": [ + { + "bbox": [ + 31, + 236, + 380, + 291 + ], + "type": "text", + "content": "Fig.1: Summary. a) We present an approach to combine Contrast Maximization with dense non-linear trajectories. b) We show how it can be used for self-supervised learning in a pipeline to predict dense point trajectories, and c) evaluate it on the EVIMO2 dataset, for which we generate dense point tracks. d) Additionally, our approach provides state-of-the-art performance on self-supervised optical flow prediction." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 147, + 172, + 199, + 210 + ], + "blocks": [ + { + "bbox": [ + 147, + 172, + 199, + 210 + ], + "lines": [ + { + "bbox": [ + 147, + 172, + 199, + 210 + ], + "spans": [ + { + "bbox": [ + 147, + 172, + 199, + 210 + ], + "type": "image", + "image_path": "d53aca7b0ce0b63c52044b1ab88d10fd1de4c853c12c1383a1978fe5ef6ed691.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 221, + 127, + 364, + 214 + ], + "blocks": [ + { + "bbox": [ + 221, + 127, + 364, + 214 + ], + "lines": [ + { + "bbox": [ + 221, + 127, + 364, + 214 + ], + "spans": [ + { + "bbox": [ + 221, + 127, + 364, + 214 + ], + "type": "image", + "image_path": "b3684fcdca0d6303e409a626861c564e772277f4431be555e5fb81235c4b8979.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 220, + 215, + 299, + 222 + ], + "lines": [ + { + "bbox": [ + 220, + 215, + 299, + 222 + ], + "spans": [ + { + "bbox": [ + 220, + 215, + 299, + 222 + ], + "type": "text", + "content": "d) Application to Optical Flow" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 31, + 319, + 379, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 319, + 379, + 546 + ], + "spans": [ + { + "bbox": [ + 31, + 319, + 379, + 546 + ], + "type": "text", + "content": "Event-based motion estimation methods can be categorized according to the complexity of the motion considered (i.e., number of degrees-of-freedom (DOF)) and to the solution strategy: model-based or learning-based. Low-DOF motions arise in sparse feature tracking and ego-motion estimation, while high-DOF motions describe more complex scenes, e.g., via per-pixel displacement (i.e., densely, over the whole image plane). Focusing on the latter, the event-based optical flow problem has been extensively studied on mobile robotic datasets [23, 75], where GT is calculated as the motion field (from known depth and poses). Most approaches use this GT for direct supervision of dense flow prediction [22, 24, 40, 57, 65]. Alternatively, several works use a contrast loss [18, 19, 77], which allows training in a self-supervised manner. However, the contrast loss is prone to undesired local optima, called event collapse [52, 53], where many events are warped into a few pixels or lines. Evaluation of event-based optical flow has been mostly done on the MVSEC [75] and DSEC [23] datasets, which comprise largely uniform motions and test intervals from 0.022s to 0.1s. Recently, [25] has proposed a supervised method to predict pixel displacement over a larger duration (0.5s), by leveraging synthetic data. It relies on the generation of training event data from images, where current tools [21] are not as mature as frame-based simulation, therefore it suffers from a sim-to-real performance gap." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 31, + 547, + 380, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 547, + 380, + 584 + ], + "spans": [ + { + "bbox": [ + 31, + 547, + 380, + 584 + ], + "type": "text", + "content": "In short, progress has been made and the research field is moving towards predicting motion over longer time intervals on the whole image plane, i.e., taking on more complex (i.e., nonlinear) motion problems. This is precisely the" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 33, + 10, + 39, + 18 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 10, + 39, + 18 + ], + "spans": [ + { + "bbox": [ + 33, + 10, + 39, + 18 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 169, + 9, + 239, + 19 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 169, + 9, + 239, + 19 + ], + "spans": [ + { + "bbox": [ + 169, + 9, + 239, + 19 + ], + "type": "text", + "content": "F. Hamann et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 31, + 34, + 380, + 118 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 34, + 380, + 118 + ], + "spans": [ + { + "bbox": [ + 31, + 34, + 380, + 118 + ], + "type": "text", + "content": "problem tackled in this paper (Fig. 1): long-time and dense event-based motion estimation, reducing the domain adaptation gap of previous approaches. It comes with several associated challenges, mainly overcoming the lack of large labeled datasets, and dealing with event noise and data association (events depend on motion, and for large motions, the appearance of \"corresponding\" events can be wildly different due to changes in motion direction, occlusions, etc.) while leveraging the space-time characteristics of event data." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 31, + 119, + 380, + 226 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 119, + 380, + 226 + ], + "spans": [ + { + "bbox": [ + 31, + 119, + 380, + 226 + ], + "type": "text", + "content": "To this end, we propose tackling the problem in two stages, by leveraging both supervised and self-supervised strategies: first, using supervised learning on synthetic data to provide initial model weights for motion estimation, and secondly, fine-tuning the network on real data via a self-supervised loss to reduce the domain adaptation gap. Our technical contributions involve extending the contrast loss framework to regress continuous-time trajectories over long time intervals via motion priors (parametric functions that provide a good balance between motion generality and regularization [64, 66]). This includes a solution to accurately and efficiently associate events to the trajectories (Fig. 2)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 31, + 227, + 380, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 227, + 380, + 358 + ], + "spans": [ + { + "bbox": [ + 31, + 227, + 380, + 358 + ], + "type": "text", + "content": "More specifically, finding the association between events and motion trajectories (to warp corresponding events and achieve event alignment) is a high-dimensional problem (e.g., for a time window of 0.3 s one can consider about ten million events, and as many trajectories as pixels), which needs to be implemented in a differentiable and parallelizable manner. We propose two actions to cope with these technical challenges. First, we relax the problem by interpolating over a coarse spatio-temporal displacement field, which serves as a lookup table. Secondly, we use a symbolic matrix framework [15] to calculate the " + }, + { + "bbox": [ + 31, + 227, + 380, + 358 + ], + "type": "inline_equation", + "content": "K = N_{\\mathrm{traj}}" + }, + { + "bbox": [ + 31, + 227, + 380, + 358 + ], + "type": "text", + "content": " nearest neighbor (KNN) trajectories for each grid point, thereby solving KNN in a memory-efficient and differentiable way on GPU. The warp displacement at the grid point is then set to the average of the neighboring trajectories." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 31, + 359, + 380, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 359, + 380, + 455 + ], + "spans": [ + { + "bbox": [ + 31, + 359, + 380, + 455 + ], + "type": "text", + "content": "Our approach is versatile, allowing for different types of networks and trajectories. Hence, we evaluate its performance on two applications: dense continuous-time motion estimation and optical flow. The results on EVIMO2 [7] show that fine-tuning with our self-supervised loss improves the zero-shot performance of a model pre-trained on synthetic data by " + }, + { + "bbox": [ + 31, + 359, + 380, + 455 + ], + "type": "inline_equation", + "content": "29\\%" + }, + { + "bbox": [ + 31, + 359, + 380, + 455 + ], + "type": "text", + "content": ". On DSEC, our model shows state-of-the-art performance among self-supervised methods, on average improving the angular error by " + }, + { + "bbox": [ + 31, + 359, + 380, + 455 + ], + "type": "inline_equation", + "content": "19\\%" + }, + { + "bbox": [ + 31, + 359, + 380, + 455 + ], + "type": "text", + "content": " and the percentage of inliers by " + }, + { + "bbox": [ + 31, + 359, + 380, + 455 + ], + "type": "inline_equation", + "content": "14\\%" + }, + { + "bbox": [ + 31, + 359, + 380, + 455 + ], + "type": "text", + "content": ", while having a " + }, + { + "bbox": [ + 31, + 359, + 380, + 455 + ], + "type": "inline_equation", + "content": "5\\times" + }, + { + "bbox": [ + 31, + 359, + 380, + 455 + ], + "type": "text", + "content": " faster inference time. In summary, our contributions are (Fig. 1):" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 35, + 463, + 387, + 582 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 35, + 463, + 379, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 35, + 463, + 379, + 498 + ], + "spans": [ + { + "bbox": [ + 35, + 463, + 379, + 498 + ], + "type": "text", + "content": "1. We introduce motion priors (parametric functions with a good balance between generality and regularization) in the event-based contrast maximization framework for continuous-time and dense motion estimation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 35, + 499, + 387, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 35, + 499, + 387, + 546 + ], + "spans": [ + { + "bbox": [ + 35, + 499, + 387, + 546 + ], + "type": "text", + "content": "2. We combine the self-supervised loss with Bflow [25], the current top-performing supervised model for dense continuous-time event-based motion estimation trained on synthetic data, and show that it can be used to improve over zero-shot performance on unseen real data (EVIMO2)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 35, + 547, + 380, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 35, + 547, + 380, + 582 + ], + "spans": [ + { + "bbox": [ + 35, + 547, + 380, + 582 + ], + "type": "text", + "content": "3. We show that the combination of our loss with a simple U-Net architecture (à la EV-FlowNet [76]) achieves state-of-the-art performance among the self-supervised methods on the DSEC Optical Flow benchmark." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 196, + 9, + 346, + 20 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 196, + 9, + 346, + 20 + ], + "spans": [ + { + "bbox": [ + 196, + 9, + 346, + 20 + ], + "type": "text", + "content": "Motion-prior Contrast Maximization" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 373, + 10, + 380, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 10, + 380, + 19 + ], + "spans": [ + { + "bbox": [ + 373, + 10, + 380, + 19 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 33, + 33, + 136, + 46 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 33, + 136, + 46 + ], + "spans": [ + { + "bbox": [ + 33, + 33, + 136, + 46 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 33, + 84, + 379, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 84, + 379, + 192 + ], + "spans": [ + { + "bbox": [ + 33, + 84, + 379, + 192 + ], + "type": "text", + "content": "Frame-based Motion Estimation. Progress in deep learning triggered a large series of learned methods for optical flow estimation in classical, frame-based computer vision [2, 14, 32, 33, 60, 62]. The solutions rely on large-scale synthetic datasets [14, 42, 59, 73]. Similarly related is the task of point tracking, which has shown impressive progress on frame-based data [13, 31, 36, 50, 74], equally relying on simulated data [8, 12, 74]. This approach is unsatisfying, as it is prone to out-of-distribution (OOD) problems that cannot be easily overcome. As an alternative, self-supervised methods have been explored [34, 35, 48, 72] and shown to improve models pre-trained on simulated data for motion estimation tasks [58, 61]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 33, + 197, + 379, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 197, + 379, + 281 + ], + "spans": [ + { + "bbox": [ + 33, + 197, + 379, + 281 + ], + "type": "text", + "content": "Event-based Optical Flow. Event cameras [39, 46] are a relatively new technology, and have found applications in various computer vision domains, like mobile robotics [27, 45, 68], scene understanding [29, 30], and computational imaging [55, 63]. Their exploration for low-level vision tasks has taken a similar trajectory in a compressed timeline as previously frame-based methods [17]. The first event-based optical flow methods were model-based [3, 4, 6, 41, 43, 54], followed by learning-based approaches [11, 24, 28, 38, 44, 76, 77]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 33, + 285, + 379, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 285, + 379, + 393 + ], + "spans": [ + { + "bbox": [ + 33, + 285, + 379, + 393 + ], + "type": "text", + "content": "The scarce availability of event data and less mature simulation technology compared to frame-based cameras are major obstacles [21]. Therefore, event-based optical flow has been mostly evaluated on data acquired through egomotion, such as the robotics dataset MVSEC [75], the driving dataset DSEC [23] or the M3ED dataset [9]. However, GT is calculated using the motion field equation (with data from a synchronized depth sensor) and therefore does not provide an accurate flow at the event rate, nor spatially at occlusions and independently moving objects (IMOs). This GT is used to train supervised learning approaches [22, 24, 40, 57, 65], which inherit limitations from the ground truth." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 33, + 398, + 379, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 398, + 379, + 482 + ], + "spans": [ + { + "bbox": [ + 33, + 398, + 379, + 482 + ], + "type": "text", + "content": "Self-supervised methods for event-based optical flow lessen the dependency on GT labels by leveraging an event alignment error [28,54,70,77]. Notably, variations of the contrast loss have been proposed [18], however, so far the application has been limited to either low-DOF problems (e.g., feature tracking [10, 19, 51], ego-motion [19, 20, 37]) or high-DOF but short-time optical flow problem [44, 54] (max. 0.1 s). We expand the frontier to the challenging problem of long-time and high-DOF (complex) motion estimation, by exploiting trajectory priors." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 33, + 487, + 379, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 487, + 379, + 583 + ], + "spans": [ + { + "bbox": [ + 33, + 487, + 379, + 583 + ], + "type": "text", + "content": "Trajectory Prior. Using trajectories as motion priors has been a widely explored scheme in computer vision. Specifically, [1, 64, 78] propose a linear combination of basis functions for structure from motion. More recently it has been explored for dynamic novel view synthesis [66]. In the context of events, [63] uses cubic motion splines for video frame interpolation, [25] uses B-splines for supervised learning of non-linear optical flow, [67] use learned basis functions for event-based video decompression, and [10, 51] show event-based feature tracking (low-DOF) using Bezier or B-spline curves in combination with a contrast loss." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 33, + 11, + 39, + 18 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 11, + 39, + 18 + ], + "spans": [ + { + "bbox": [ + 33, + 11, + 39, + 18 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 170, + 10, + 239, + 19 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 170, + 10, + 239, + 19 + ], + "spans": [ + { + "bbox": [ + 170, + 10, + 239, + 19 + ], + "type": "text", + "content": "F. Hamann et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 33, + 31, + 381, + 218 + ], + "blocks": [ + { + "bbox": [ + 33, + 31, + 381, + 218 + ], + "lines": [ + { + "bbox": [ + 33, + 31, + 381, + 218 + ], + "spans": [ + { + "bbox": [ + 33, + 31, + 381, + 218 + ], + "type": "image", + "image_path": "67d6593e4476ffb8773fdbff0d1dee02a851723bc05e26dadce9f3780c710844.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 31, + 225, + 381, + 315 + ], + "lines": [ + { + "bbox": [ + 31, + 225, + 381, + 315 + ], + "spans": [ + { + "bbox": [ + 31, + 225, + 381, + 315 + ], + "type": "text", + "content": "Fig. 2: Pipeline overview. (a) Input events in a time interval are (b) voxelized and (c) passed to an artificial neural network that predicts per-pixel coefficients for continuous-time trajectories (d). The raw events and predicted trajectories are fed to the loss module (e). Here, a dense spatio-temporal displacement map is interpolated, and events are warped according to their looked-up displacement. Lastly, an image of warped events (IWE) is built at a random reference time and its gradient magnitude acts as training loss. Note that the prediction method displayed here is specific to the used Bflow backbone [25], with additional events before the prediction start time " + }, + { + "bbox": [ + 31, + 225, + 381, + 315 + ], + "type": "inline_equation", + "content": "t_s" + }, + { + "bbox": [ + 31, + 225, + 381, + 315 + ], + "type": "text", + "content": " as input." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 31, + 333, + 132, + 348 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 333, + 132, + 348 + ], + "spans": [ + { + "bbox": [ + 31, + 333, + 132, + 348 + ], + "type": "text", + "content": "3 Methodology" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 31, + 356, + 130, + 367 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 356, + 130, + 367 + ], + "spans": [ + { + "bbox": [ + 31, + 356, + 130, + 367 + ], + "type": "text", + "content": "3.1 Contrast Loss" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 30, + 372, + 381, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 372, + 381, + 472 + ], + "spans": [ + { + "bbox": [ + 30, + 372, + 381, + 472 + ], + "type": "text", + "content": "Contrast Maximization (CM) [18, 19] is used for self-supervised training. The idea is to find the point trajectories on the image plane best aligning with the events by maximizing the sharpness of warped events. It is an iterative approach, with three steps per iteration: (i) Each event " + }, + { + "bbox": [ + 30, + 372, + 381, + 472 + ], + "type": "inline_equation", + "content": "e_k = (\\mathbf{x}_k, t_k, p_k)" + }, + { + "bbox": [ + 30, + 372, + 381, + 472 + ], + "type": "text", + "content": " contains the pixel coordinates " + }, + { + "bbox": [ + 30, + 372, + 381, + 472 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_k = (x_k, y_k)^\\top" + }, + { + "bbox": [ + 30, + 372, + 381, + 472 + ], + "type": "text", + "content": ", timestamp " + }, + { + "bbox": [ + 30, + 372, + 381, + 472 + ], + "type": "inline_equation", + "content": "t_k" + }, + { + "bbox": [ + 30, + 372, + 381, + 472 + ], + "type": "text", + "content": " and polarity " + }, + { + "bbox": [ + 30, + 372, + 381, + 472 + ], + "type": "inline_equation", + "content": "p_k" + }, + { + "bbox": [ + 30, + 372, + 381, + 472 + ], + "type": "text", + "content": " of a brightness change of predefined size " + }, + { + "bbox": [ + 30, + 372, + 381, + 472 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 30, + 372, + 381, + 472 + ], + "type": "text", + "content": " (contrast sensitivity). Events " + }, + { + "bbox": [ + 30, + 372, + 381, + 472 + ], + "type": "inline_equation", + "content": "\\mathcal{E} = \\{e_k\\}_{k=1}^{N_e}" + }, + { + "bbox": [ + 30, + 372, + 381, + 472 + ], + "type": "text", + "content": " are displaced according to a candidate motion hypothesis " + }, + { + "bbox": [ + 30, + 372, + 381, + 472 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_k' = \\mathbf{W}(\\mathbf{x}_k, t_k; \\boldsymbol{\\theta})" + }, + { + "bbox": [ + 30, + 372, + 381, + 472 + ], + "type": "text", + "content": ", with parameters " + }, + { + "bbox": [ + 30, + 372, + 381, + 472 + ], + "type": "inline_equation", + "content": "\\boldsymbol{\\theta}" + }, + { + "bbox": [ + 30, + 372, + 381, + 472 + ], + "type": "text", + "content": ", to a reference time " + }, + { + "bbox": [ + 30, + 372, + 381, + 472 + ], + "type": "inline_equation", + "content": "t_\\mathrm{ref}" + }, + { + "bbox": [ + 30, + 372, + 381, + 472 + ], + "type": "text", + "content": ", producing a set of warped events " + }, + { + "bbox": [ + 30, + 372, + 381, + 472 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{t_\\mathrm{ref}}' = \\{e_k'\\}_{k=1}^{N_e}" + }, + { + "bbox": [ + 30, + 372, + 381, + 472 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 118, + 476, + 379, + 493 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 476, + 379, + 493 + ], + "spans": [ + { + "bbox": [ + 118, + 476, + 379, + 493 + ], + "type": "interline_equation", + "content": "e _ {k} \\doteq \\left(\\mathbf {x} _ {k}, t _ {k}, p _ {k}\\right) \\stackrel {\\mathrm {W}} {\\mapsto} e _ {k} ^ {\\prime} \\doteq \\left(\\mathbf {x} _ {k} ^ {\\prime}, t _ {\\text {r e f}}, p _ {k}\\right). \\tag {1}", + "image_path": "ad071a0f7504628dc0b25fa532539200343ec0087fda723168a074e157795473.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 31, + 498, + 376, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 498, + 376, + 511 + ], + "spans": [ + { + "bbox": [ + 31, + 498, + 376, + 511 + ], + "type": "text", + "content": "Afterwards, " + }, + { + "bbox": [ + 31, + 498, + 376, + 511 + ], + "type": "inline_equation", + "content": "(ii)" + }, + { + "bbox": [ + 31, + 498, + 376, + 511 + ], + "type": "text", + "content": " the events are summed into an image of warped events (IWE)," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 515, + 379, + 531 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 515, + 379, + 531 + ], + "spans": [ + { + "bbox": [ + 130, + 515, + 379, + 531 + ], + "type": "interline_equation", + "content": "I (\\mathbf {x}; \\mathcal {E} _ {t _ {\\text {r e f}}} ^ {\\prime}, \\boldsymbol {\\theta}) \\doteq \\sum_ {k = 1} ^ {N _ {e}} \\mathcal {N} (\\mathbf {x}; \\mathbf {x} _ {k} ^ {\\prime}, \\sigma^ {2}), \\tag {2}", + "image_path": "5ea4b7bda4216a7fc8e51b60a39eaf2c3d6e06c6a78a8f89a5467652b1a845c7.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 31, + 535, + 381, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 535, + 381, + 584 + ], + "spans": [ + { + "bbox": [ + 31, + 535, + 381, + 584 + ], + "type": "text", + "content": "which essentially counts the number of warped events " + }, + { + "bbox": [ + 31, + 535, + 381, + 584 + ], + "type": "inline_equation", + "content": "e_k'" + }, + { + "bbox": [ + 31, + 535, + 381, + 584 + ], + "type": "text", + "content": " per pixel. Lastly (iii) the contrast or sharpness of the IWE is computed (e.g., the gradient magnitude of the IWE, " + }, + { + "bbox": [ + 31, + 535, + 381, + 584 + ], + "type": "inline_equation", + "content": "G \\doteq \\int \\| \\nabla I(\\mathbf{x}) \\| d\\mathbf{x}" + }, + { + "bbox": [ + 31, + 535, + 381, + 584 + ], + "type": "text", + "content": "), which serves as a proxy for how well the model (motion hypothesis and " + }, + { + "bbox": [ + 31, + 535, + 381, + 584 + ], + "type": "inline_equation", + "content": "\\pmb{\\theta}" + }, + { + "bbox": [ + 31, + 535, + 381, + 584 + ], + "type": "text", + "content": ") fit with the events produced by the true motion." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 197, + 9, + 346, + 20 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 9, + 346, + 20 + ], + "spans": [ + { + "bbox": [ + 197, + 9, + 346, + 20 + ], + "type": "text", + "content": "Motion-prior Contrast Maximization" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 373, + 10, + 379, + 18 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 10, + 379, + 18 + ], + "spans": [ + { + "bbox": [ + 373, + 10, + 379, + 18 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 35, + 54, + 378, + 144 + ], + "blocks": [ + { + "bbox": [ + 37, + 33, + 373, + 44 + ], + "lines": [ + { + "bbox": [ + 37, + 33, + 373, + 44 + ], + "spans": [ + { + "bbox": [ + 37, + 33, + 373, + 44 + ], + "type": "text", + "content": "Table 1: Comparison of the contrast loss formulation in the most recent methods." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 35, + 54, + 378, + 144 + ], + "lines": [ + { + "bbox": [ + 35, + 54, + 378, + 144 + ], + "spans": [ + { + "bbox": [ + 35, + 54, + 378, + 144 + ], + "type": "table", + "html": "
StepsShiba et. al [54]Paredes et. al [44]Ours
Warp displacement (3):Δxk=(tref−tk)v(ek)Δxk=∑i(Δti v_i)(ek)Δxk=1/Ntraj ∑n Δxn
linear trajectoryflow concatenationnon-linear trajectory
IWEI(x)=∑kN(x; x′k, σ2)T±(x)=∑ktkN(x; x′k, σ2)/∑kN(x; x′k, σ2)I(x)=∑kN(x; x′k, σ2)
Reference time tref:{t0, t0.5, t1}multi-partition p, multi-time-scale s~U(0,1)
Contrast objectiveG(t0)+2G(t0.5)+G(t1)/4G(v=0)1/S ∑s=0S-11/2s ∑p=02s-1LRCM,pG(tref)
", + "image_path": "6d8962e708d63c5de902dfc62b06940f37f5db416b89a1ae20f70e597fea79bd.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 31, + 167, + 364, + 180 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 167, + 364, + 180 + ], + "spans": [ + { + "bbox": [ + 31, + 167, + 364, + 180 + ], + "type": "text", + "content": "3.2 Estimating Continuous-Time and Dense Motion Trajectories" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 31, + 186, + 379, + 210 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 186, + 379, + 210 + ], + "spans": [ + { + "bbox": [ + 31, + 186, + 379, + 210 + ], + "type": "text", + "content": "The CM framework has been extended in several works. Table 1 compares recent extensions for optical flow estimation, where events are warped as" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 168, + 220, + 379, + 233 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 168, + 220, + 379, + 233 + ], + "spans": [ + { + "bbox": [ + 168, + 220, + 379, + 233 + ], + "type": "interline_equation", + "content": "\\mathbf {x} _ {k} ^ {\\prime} = \\mathbf {x} _ {k} + \\Delta \\mathbf {x} _ {k}. \\tag {3}", + "image_path": "7bb9503ab9cb28dfad5d3d2702e4c0f087530925f31658f5a671544cc7e45cec.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 30, + 242, + 379, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 242, + 379, + 337 + ], + "spans": [ + { + "bbox": [ + 30, + 242, + 379, + 337 + ], + "type": "text", + "content": "The main differences between these approaches lie in the event displacement (i.e., warp) model, the reference times used, and the loss function (i.e., event-alignment metric), with the design choices having two goals: fitting the event data and regularizing the solution (e.g., avoiding event collapse [52]). Previous methods used mostly a linear model [54,56] or partitioned the inference interval into smaller flow intervals so that its concatenation need not be linear [44]. By contrast, we introduce an explicit continuous-time non-linear trajectory model, a random reference time per iteration, and an easier contrast loss." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 30, + 338, + 385, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 338, + 385, + 421 + ], + "spans": [ + { + "bbox": [ + 30, + 338, + 385, + 421 + ], + "type": "text", + "content": "Trajectory Representation. Figure 2 shows an overview of the training pipeline. Events in the time interval " + }, + { + "bbox": [ + 30, + 338, + 385, + 421 + ], + "type": "inline_equation", + "content": "[t_s, t_e]" + }, + { + "bbox": [ + 30, + 338, + 385, + 421 + ], + "type": "text", + "content": " are fed into a neural network that makes per-pixel predictions of the parameters (coefficients) " + }, + { + "bbox": [ + 30, + 338, + 385, + 421 + ], + "type": "inline_equation", + "content": "\\pmb{\\theta} \\equiv \\pmb{\\alpha}_n = (\\alpha_{n,1}, \\dots, \\alpha_{n,N_c})^\\top \\in \\mathbb{R}^{N_c}" + }, + { + "bbox": [ + 30, + 338, + 385, + 421 + ], + "type": "text", + "content": " of a continuous-time trajectory " + }, + { + "bbox": [ + 30, + 338, + 385, + 421 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_n(t; \\pmb{\\alpha}_n) \\equiv \\mathbf{q}_n(t) = (x_n(t), y_n(t))^\\top" + }, + { + "bbox": [ + 30, + 338, + 385, + 421 + ], + "type": "text", + "content": ". There is one trajectory per pixel (i.e., \"dense\" character), " + }, + { + "bbox": [ + 30, + 338, + 385, + 421 + ], + "type": "inline_equation", + "content": "n = 1, \\dots, N_p" + }, + { + "bbox": [ + 30, + 338, + 385, + 421 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 30, + 338, + 385, + 421 + ], + "type": "inline_equation", + "content": "N_p = h w" + }, + { + "bbox": [ + 30, + 338, + 385, + 421 + ], + "type": "text", + "content": " is the number of pixels (image height " + }, + { + "bbox": [ + 30, + 338, + 385, + 421 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 30, + 338, + 385, + 421 + ], + "type": "text", + "content": " and width " + }, + { + "bbox": [ + 30, + 338, + 385, + 421 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 30, + 338, + 385, + 421 + ], + "type": "text", + "content": "). Thus, " + }, + { + "bbox": [ + 30, + 338, + 385, + 421 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 30, + 338, + 385, + 421 + ], + "type": "text", + "content": " is the spatial index of the trajectory, identifying it among all trajectories on the image plane." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 30, + 421, + 380, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 421, + 380, + 487 + ], + "spans": [ + { + "bbox": [ + 30, + 421, + 380, + 487 + ], + "type": "text", + "content": "We model trajectories as weighted combinations of basis functions, " + }, + { + "bbox": [ + 30, + 421, + 380, + 487 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_n(t) = \\sum_{j=1}^{N_c} g_j(t) \\mathbf{p}_{n,j}" + }, + { + "bbox": [ + 30, + 421, + 380, + 487 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 30, + 421, + 380, + 487 + ], + "type": "inline_equation", + "content": "g_j(t)" + }, + { + "bbox": [ + 30, + 421, + 380, + 487 + ], + "type": "text", + "content": " are temporal basis shares by all trajectories, and " + }, + { + "bbox": [ + 30, + 421, + 380, + 487 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_{n,j} = (\\alpha_{n,j}^x, \\alpha_{n,j}^y)^\\top" + }, + { + "bbox": [ + 30, + 421, + 380, + 487 + ], + "type": "text", + "content": " are \"control points\" (we write " + }, + { + "bbox": [ + 30, + 421, + 380, + 487 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 30, + 421, + 380, + 487 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 30, + 421, + 380, + 487 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 30, + 421, + 380, + 487 + ], + "type": "text", + "content": " components explicitly with separate coefficients in " + }, + { + "bbox": [ + 30, + 421, + 380, + 487 + ], + "type": "inline_equation", + "content": "\\alpha_n" + }, + { + "bbox": [ + 30, + 421, + 380, + 487 + ], + "type": "text", + "content": "). We investigate polynomial basis " + }, + { + "bbox": [ + 30, + 421, + 380, + 487 + ], + "type": "inline_equation", + "content": "g_j(t) = t^j" + }, + { + "bbox": [ + 30, + 421, + 380, + 487 + ], + "type": "text", + "content": ", Bézier curves with basis " + }, + { + "bbox": [ + 30, + 421, + 380, + 487 + ], + "type": "inline_equation", + "content": "g_j(t) = \\binom{N_c}{j}(1 - t)^{N_c - j}t^j" + }, + { + "bbox": [ + 30, + 421, + 380, + 487 + ], + "type": "text", + "content": ", as well as a learned basis." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 30, + 487, + 380, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 487, + 380, + 585 + ], + "spans": [ + { + "bbox": [ + 30, + 487, + 380, + 585 + ], + "type": "text", + "content": "Spatio-Temporal Event Warping. The trajectories can be used to warp events. This would require finding the trajectory that passes through the spacetime coordinates of the event and then finding the value of the trajectory at the reference time (i.e., the warped event location). However, the association between events and trajectories is unknown and needs to be estimated simultaneously with the parameters (i.e., shape) of the trajectories. We circumvent the problem by using a soft association between events and trajectories. Each event " + }, + { + "bbox": [ + 30, + 487, + 380, + 585 + ], + "type": "inline_equation", + "content": "e_k" + }, + { + "bbox": [ + 30, + 487, + 380, + 585 + ], + "type": "text", + "content": " is associated with its " + }, + { + "bbox": [ + 30, + 487, + 380, + 585 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{traj}}" + }, + { + "bbox": [ + 30, + 487, + 380, + 585 + ], + "type": "text", + "content": " nearest neighboring trajectories " + }, + { + "bbox": [ + 30, + 487, + 380, + 585 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{q}_n\\}_{n=1}^{N_{\\mathrm{traj}}}" + }, + { + "bbox": [ + 30, + 487, + 380, + 585 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 33, + 10, + 39, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 10, + 39, + 19 + ], + "spans": [ + { + "bbox": [ + 33, + 10, + 39, + 19 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 169, + 9, + 240, + 19 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 169, + 9, + 240, + 19 + ], + "spans": [ + { + "bbox": [ + 169, + 9, + 240, + 19 + ], + "type": "text", + "content": "F. Hamann et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 31, + 35, + 380, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 35, + 380, + 84 + ], + "spans": [ + { + "bbox": [ + 31, + 35, + 380, + 84 + ], + "type": "text", + "content": "and the event displacement " + }, + { + "bbox": [ + 31, + 35, + 380, + 84 + ], + "type": "inline_equation", + "content": "\\varDelta\\mathbf{x}_k" + }, + { + "bbox": [ + 31, + 35, + 380, + 84 + ], + "type": "text", + "content": " in (3) is computed as the average of the respective trajectory displacements " + }, + { + "bbox": [ + 31, + 35, + 380, + 84 + ], + "type": "inline_equation", + "content": "\\{\\varDelta\\mathbf{x}_k^n\\}_{n=1}^{N_{\\mathrm{traj}}}" + }, + { + "bbox": [ + 31, + 35, + 380, + 84 + ], + "type": "text", + "content": ". The trajectory displacement " + }, + { + "bbox": [ + 31, + 35, + 380, + 84 + ], + "type": "inline_equation", + "content": "\\varDelta\\mathbf{x}_k^n" + }, + { + "bbox": [ + 31, + 35, + 380, + 84 + ], + "type": "text", + "content": " is defined as the difference of trajectory locations at the time of the event and the reference time. The event displacement is defined as the mean:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 81, + 95, + 379, + 128 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 95, + 379, + 128 + ], + "spans": [ + { + "bbox": [ + 81, + 95, + 379, + 128 + ], + "type": "interline_equation", + "content": "\\Delta \\mathbf {x} _ {k} \\doteq \\frac {1}{N _ {\\mathrm {t r a j}}} \\sum_ {n = 1} ^ {N _ {\\mathrm {t r a j}}} \\Delta \\mathbf {x} _ {k} ^ {n}, \\quad \\text {w i t h} \\quad \\Delta \\mathbf {x} _ {k} ^ {n} \\doteq \\mathbf {q} _ {n} \\left(t _ {\\mathrm {r e f}}\\right) - \\mathbf {q} _ {n} \\left(t _ {k}\\right). \\tag {4}", + "image_path": "779800de1bb6c6c27779777bff6c7258251fe744ea6757bc3a1fdb330cef7233.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 30, + 139, + 380, + 234 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 139, + 380, + 234 + ], + "spans": [ + { + "bbox": [ + 30, + 139, + 380, + 234 + ], + "type": "text", + "content": "Loss Calculation. Once the events are warped using (3), it is straightforward to compute the IWE. We adopt the magnitude of the IWE gradient as loss function [18,56]. Moreover, we use the contrast loss in a self-supervised learning setting, choosing a different reference time " + }, + { + "bbox": [ + 30, + 139, + 380, + 234 + ], + "type": "inline_equation", + "content": "t_{\\mathrm{ref}}" + }, + { + "bbox": [ + 30, + 139, + 380, + 234 + ], + "type": "text", + "content": ", uniformly sampled in the observation interval, for every batch during training. This simplifies the objective to a single warping operation and loss calculation (e.g., compared to the three warping operations and loss calculations in the optimization-based approach by [54,56]), while it has added regularization benefits, as mentioned below." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 30, + 235, + 379, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 235, + 379, + 331 + ], + "spans": [ + { + "bbox": [ + 30, + 235, + 379, + 331 + ], + "type": "text", + "content": "Memory Effective Computation of the Displacement Field. The number of events and trajectories can be very large, and even more their combination. Calculating KNN for every event can quickly become computationally unfeasible, and traditional algorithms cannot be efficiently implemented in deep-learning frameworks. We propose calculating the per-event displacement " + }, + { + "bbox": [ + 30, + 235, + 379, + 331 + ], + "type": "inline_equation", + "content": "\\Delta \\mathbf{x}_k" + }, + { + "bbox": [ + 30, + 235, + 379, + 331 + ], + "type": "text", + "content": " in (3) by first interpolating a dense but coarser spatio-temporal displacement field, and then looking up the per-event displacement in such space-time volume. Moreover, we relax the problem by solving the KNN search in 2D instead of in the volume." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 30, + 332, + 379, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 332, + 379, + 475 + ], + "spans": [ + { + "bbox": [ + 30, + 332, + 379, + 475 + ], + "type": "text", + "content": "Figure 2 shows an overview of the interpolation process. The displacement map is a tensor of shape " + }, + { + "bbox": [ + 30, + 332, + 379, + 475 + ], + "type": "inline_equation", + "content": "[N_{\\mathrm{bins}}, h/4, w/4]" + }, + { + "bbox": [ + 30, + 332, + 379, + 475 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 30, + 332, + 379, + 475 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{bins}}" + }, + { + "bbox": [ + 30, + 332, + 379, + 475 + ], + "type": "text", + "content": " is the number of temporal bins, and " + }, + { + "bbox": [ + 30, + 332, + 379, + 475 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 30, + 332, + 379, + 475 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 30, + 332, + 379, + 475 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 30, + 332, + 379, + 475 + ], + "type": "text", + "content": " are the sensor's height and width, respectively. For each temporal bin-center " + }, + { + "bbox": [ + 30, + 332, + 379, + 475 + ], + "type": "inline_equation", + "content": "t_{\\mathrm{bins}}" + }, + { + "bbox": [ + 30, + 332, + 379, + 475 + ], + "type": "text", + "content": " the trajectories " + }, + { + "bbox": [ + 30, + 332, + 379, + 475 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_n(t = t_{\\mathrm{bins}})" + }, + { + "bbox": [ + 30, + 332, + 379, + 475 + ], + "type": "text", + "content": " are calculated and the KNN interpolation is performed between each pixel of a channel " + }, + { + "bbox": [ + 30, + 332, + 379, + 475 + ], + "type": "inline_equation", + "content": "\\mathbf{x}[t_{\\mathrm{bins}}]" + }, + { + "bbox": [ + 30, + 332, + 379, + 475 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 30, + 332, + 379, + 475 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_n(t = t_{\\mathrm{bins}})" + }, + { + "bbox": [ + 30, + 332, + 379, + 475 + ], + "type": "text", + "content": ". We implement the KNN approach using KeOps [15], a framework for symbolic matrix computation, which provides a memory-efficient and differentiable solution. We perform a KNN search for every voxel in the table, and events are warped after a lookup operation. Please note that this step is entirely independent of the voxel grid passed as input to the backbone, and all raw events are used to calculate the contrast loss. From the calculated displacement map, we can directly look up per-event displacements " + }, + { + "bbox": [ + 30, + 332, + 379, + 475 + ], + "type": "inline_equation", + "content": "\\Delta \\mathbf{x}_k" + }, + { + "bbox": [ + 30, + 332, + 379, + 475 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 30, + 476, + 379, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 476, + 379, + 583 + ], + "spans": [ + { + "bbox": [ + 30, + 476, + 379, + 583 + ], + "type": "text", + "content": "Regularization. Objectives based on event alignment are prone to undesired local minima called \"event collapse\" [52,53]. Our formulation inherently provides regularization, temporally by the smoothness of the motion prior, and spatially by interpolating the event flow from several trajectories via a soft assignment. Our framework uses two additional regularization sources. Firstly, we penalize the magnitude of the spatial gradient of the interpolated displacement field between consecutive timesteps, " + }, + { + "bbox": [ + 30, + 476, + 379, + 583 + ], + "type": "inline_equation", + "content": "R \\doteq \\| \\nabla (\\Delta \\mathbf{q}_{*}(t))\\|_{L^{1}(\\Omega)}" + }, + { + "bbox": [ + 30, + 476, + 379, + 583 + ], + "type": "text", + "content": ". To interpolate " + }, + { + "bbox": [ + 30, + 476, + 379, + 583 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_{*}" + }, + { + "bbox": [ + 30, + 476, + 379, + 583 + ], + "type": "text", + "content": " from the sparse trajectories " + }, + { + "bbox": [ + 30, + 476, + 379, + 583 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_n" + }, + { + "bbox": [ + 30, + 476, + 379, + 583 + ], + "type": "text", + "content": ", we use the same KNN indices as for the main volume. This loss encourages spatial smoothness of the trajectories." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 196, + 9, + 346, + 20 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 196, + 9, + 346, + 20 + ], + "spans": [ + { + "bbox": [ + 196, + 9, + 346, + 20 + ], + "type": "text", + "content": "Motion-prior Contrast Maximization" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 373, + 10, + 379, + 18 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 10, + 379, + 18 + ], + "spans": [ + { + "bbox": [ + 373, + 10, + 379, + 18 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 82, + 54, + 334, + 133 + ], + "blocks": [ + { + "bbox": [ + 135, + 32, + 275, + 44 + ], + "lines": [ + { + "bbox": [ + 135, + 32, + 275, + 44 + ], + "spans": [ + { + "bbox": [ + 135, + 32, + 275, + 44 + ], + "type": "text", + "content": "Table 2: Results on EVIMO2 [7]." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 82, + 54, + 334, + 133 + ], + "lines": [ + { + "bbox": [ + 82, + 54, + 334, + 133 + ], + "spans": [ + { + "bbox": [ + 82, + 54, + 334, + 133 + ], + "type": "table", + "html": "
MethodTEPE ↓TAE ↓%Out ↓
Paredes et. al [44]21.6951.910.63
E-RAFT (linear) [24]19.3874.520.66
BFlow (zero-shot) [25]8.6319.940.36
BFlow (in-domain)3.3811.680.17
Ours (self-supervised)6.1416.980.25
", + "image_path": "6ee93ed40e4de9b5753f70848db2571291c9b4b9be11be7ae3d5c123df5fb00f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 30, + 156, + 380, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 156, + 380, + 216 + ], + "spans": [ + { + "bbox": [ + 30, + 156, + 380, + 216 + ], + "type": "text", + "content": "Secondly, we use a multi-reference formulation by randomly sampling a reference time at every training step, while previous work used multiple and fixed timestamps (Tab. 1). This has the advantage that " + }, + { + "bbox": [ + 30, + 156, + 380, + 216 + ], + "type": "inline_equation", + "content": "(i)" + }, + { + "bbox": [ + 30, + 156, + 380, + 216 + ], + "type": "text", + "content": " IWEs are required to be sharp at truly any time and " + }, + { + "bbox": [ + 30, + 156, + 380, + 216 + ], + "type": "inline_equation", + "content": "(ii)" + }, + { + "bbox": [ + 30, + 156, + 380, + 216 + ], + "type": "text", + "content": " the memory requirement is lowered as it scales linearly with the number of reference times used during a training step." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 45, + 217, + 188, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 217, + 188, + 228 + ], + "spans": [ + { + "bbox": [ + 45, + 217, + 188, + 228 + ], + "type": "text", + "content": "In summary, the training loss is" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 170, + 238, + 379, + 251 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 170, + 238, + 379, + 251 + ], + "spans": [ + { + "bbox": [ + 170, + 238, + 379, + 251 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = 1 / G + \\lambda R, \\tag {5}", + "image_path": "7038a2b1019543f85413cedbb6e0db413d9511da1fc9d4e32fb4ebdcd592aa64.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 30, + 260, + 385, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 260, + 385, + 344 + ], + "spans": [ + { + "bbox": [ + 30, + 260, + 385, + 344 + ], + "type": "text", + "content": "with " + }, + { + "bbox": [ + 30, + 260, + 385, + 344 + ], + "type": "inline_equation", + "content": "\\lambda > 0" + }, + { + "bbox": [ + 30, + 260, + 385, + 344 + ], + "type": "text", + "content": " as regularization weight. Further details are in the supplementary. Pre-processing and Prediction-Module Architecture. Events record brightness changes asynchronously, in the form of a sparse spatio-temporal signal. As input to the Prediction Module, events are customarily converted to voxel grids [77] for compatibility with conventional artificial neural networks. An event volume is discretized in the time dimension, and each voxel counts the number of events within it (bi-linearly voted for undistorted, rectified events)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 30, + 345, + 380, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 345, + 380, + 393 + ], + "spans": [ + { + "bbox": [ + 30, + 345, + 380, + 393 + ], + "type": "text", + "content": "In principle, our self-supervised loss module can be paired with any segmentation or optical flow network architecture. We use a U-Net architecture for the experiments on DSEC (Sec. 4.2), and the architecture in the recent Bflow method [25], inspired by RAFT [62], for the experiments on EVIMO2 (Sec. 4.1)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 31, + 412, + 128, + 424 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 412, + 128, + 424 + ], + "spans": [ + { + "bbox": [ + 31, + 412, + 128, + 424 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 30, + 436, + 379, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 436, + 379, + 522 + ], + "spans": [ + { + "bbox": [ + 30, + 436, + 379, + 522 + ], + "type": "text", + "content": "We test our method on two applications. First, we test the capabilities for nonlinear trajectory estimation on the real-world dataset EVIMO2 [7], with additional results on the synthetic MultiFlow dataset [25], which we use for pretraining (Sec. 4.1). Secondly, we evaluate our method on the DSEC dataset [23, 24] because this has been the previous frontier for event-based optical flow estimation and it allows for direct comparison with prior work (Sec. 4.2). Additional results, including on the MVSEC dataset [75], are given in the supplement." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 31, + 539, + 288, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 539, + 288, + 551 + ], + "spans": [ + { + "bbox": [ + 31, + 539, + 288, + 551 + ], + "type": "text", + "content": "4.1 Results on Non-Linear Trajectory Estimation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 30, + 559, + 380, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 559, + 380, + 584 + ], + "spans": [ + { + "bbox": [ + 30, + 559, + 380, + 584 + ], + "type": "text", + "content": "EVIMO2 Continuous Flow Dataset (CF-EVIMO2). MultiFlow [25] was first proposed to evaluate long-term \"optical flow\" methods. The dataset consists" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 33, + 10, + 39, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 10, + 39, + 19 + ], + "spans": [ + { + "bbox": [ + 33, + 10, + 39, + 19 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 169, + 9, + 240, + 19 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 169, + 9, + 240, + 19 + ], + "spans": [ + { + "bbox": [ + 169, + 9, + 240, + 19 + ], + "type": "text", + "content": "F. Hamann et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 63, + 31, + 348, + 339 + ], + "blocks": [ + { + "bbox": [ + 63, + 31, + 348, + 339 + ], + "lines": [ + { + "bbox": [ + 63, + 31, + 348, + 339 + ], + "spans": [ + { + "bbox": [ + 63, + 31, + 348, + 339 + ], + "type": "image", + "image_path": "b0e3aa4cd2694299a90e822a42492a054bad75bcf13382c394eb69df9c30d443.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 31, + 346, + 380, + 402 + ], + "lines": [ + { + "bbox": [ + 31, + 346, + 380, + 402 + ], + "spans": [ + { + "bbox": [ + 31, + 346, + 380, + 402 + ], + "type": "text", + "content": "Fig. 3: Visualization of predicted trajectories on EVIMO2 data. " + }, + { + "bbox": [ + 31, + 346, + 380, + 402 + ], + "type": "inline_equation", + "content": "GT" + }, + { + "bbox": [ + 31, + 346, + 380, + 402 + ], + "type": "text", + "content": ": Ground truth. In-domain: fine-tuned on EVIMO2 using GT (supervised). Zero-shot: network trained only on synthetic data (out-of-domain prediction). Ours: Pre-trained on synthetic data, fine-tuned with self-supervised loss. Note that supervision in-domain is often impossible in practice because dense trajectory labels for real data are difficult to obtain." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 30, + 427, + 387, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 427, + 387, + 511 + ], + "spans": [ + { + "bbox": [ + 30, + 427, + 387, + 511 + ], + "type": "text", + "content": "of synthetic videos generated with Internet images as foreground and Flickr30K [71] images as background. Events are generated with a simulator [47] by rendering the synthetic scenes at 1000 frames/s. However, due to the sim-to-real gap between real and synthetic events, it is challenging to infer the actual performance of the proposed and baseline methods on real datasets. Furthermore, Multiflow generates trajectories by pasting 2D foreground objects onto background images, which lacks the challenging cases of self-occlusion due to 3D rotation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 31, + 511, + 382, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 511, + 382, + 584 + ], + "spans": [ + { + "bbox": [ + 31, + 511, + 382, + 584 + ], + "type": "text", + "content": "To address these data limitations, we present the EVIMO2 Continuous Flow Dataset based on the full 3D data and event data provided by the meticulously designed EVIMO2 [7] dataset. The EVIMO2 GT data provides high-quality 3D scans of the objects, camera poses, object poses, and camera intrinsics, which we use to compute optical flow GT as follows. Given a point " + }, + { + "bbox": [ + 31, + 511, + 382, + 584 + ], + "type": "inline_equation", + "content": "P_{t}^{i}" + }, + { + "bbox": [ + 31, + 511, + 382, + 584 + ], + "type": "text", + "content": " on the object represented in the camera coordinate system at time " + }, + { + "bbox": [ + 31, + 511, + 382, + 584 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 31, + 511, + 382, + 584 + ], + "type": "text", + "content": ", we project it onto a 2D point" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 197, + 9, + 346, + 20 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 9, + 346, + 20 + ], + "spans": [ + { + "bbox": [ + 197, + 9, + 346, + 20 + ], + "type": "text", + "content": "Motion-prior Contrast Maximization" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 373, + 10, + 379, + 18 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 10, + 379, + 18 + ], + "spans": [ + { + "bbox": [ + 373, + 10, + 379, + 18 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 33, + 281, + 153 + ], + "blocks": [ + { + "bbox": [ + 133, + 33, + 281, + 153 + ], + "lines": [ + { + "bbox": [ + 133, + 33, + 281, + 153 + ], + "spans": [ + { + "bbox": [ + 133, + 33, + 281, + 153 + ], + "type": "image", + "image_path": "f8b3fb6b3d8b470457f6bb9fd0a8c874f50901f8ef3b49c2b3056ef6155a147c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 31, + 162, + 380, + 186 + ], + "lines": [ + { + "bbox": [ + 31, + 162, + 380, + 186 + ], + "spans": [ + { + "bbox": [ + 31, + 162, + 380, + 186 + ], + "type": "text", + "content": "Fig. 4: End-point-error vs. prediction time span for three methods: in-distribution, out-of-distribution and self-supervised (Ours). Using the Bézier curve results from Tab. 3." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 84, + 228, + 331, + 362 + ], + "blocks": [ + { + "bbox": [ + 31, + 196, + 380, + 218 + ], + "lines": [ + { + "bbox": [ + 31, + 196, + 380, + 218 + ], + "spans": [ + { + "bbox": [ + 31, + 196, + 380, + 218 + ], + "type": "text", + "content": "Table 3: Sensitivity with respect to motion prior on EVIMO2 [7]. " + }, + { + "bbox": [ + 31, + 196, + 380, + 218 + ], + "type": "inline_equation", + "content": "SL" + }, + { + "bbox": [ + 31, + 196, + 380, + 218 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 31, + 196, + 380, + 218 + ], + "type": "inline_equation", + "content": "ID" + }, + { + "bbox": [ + 31, + 196, + 380, + 218 + ], + "type": "text", + "content": ": supervised, in-distribution, " + }, + { + "bbox": [ + 31, + 196, + 380, + 218 + ], + "type": "inline_equation", + "content": "SL" + }, + { + "bbox": [ + 31, + 196, + 380, + 218 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 31, + 196, + 380, + 218 + ], + "type": "inline_equation", + "content": "OOD" + }, + { + "bbox": [ + 31, + 196, + 380, + 218 + ], + "type": "text", + "content": ": supervised, out of distribution, " + }, + { + "bbox": [ + 31, + 196, + 380, + 218 + ], + "type": "inline_equation", + "content": "SSL" + }, + { + "bbox": [ + 31, + 196, + 380, + 218 + ], + "type": "text", + "content": ": self-supervised." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 84, + 228, + 331, + 362 + ], + "lines": [ + { + "bbox": [ + 84, + 228, + 331, + 362 + ], + "spans": [ + { + "bbox": [ + 84, + 228, + 331, + 362 + ], + "type": "table", + "html": "
MethodTEPE ↓TAE ↓%Out ↓
SL, IDBFlow (polyn.)3.5113.260.18
BFlow (learned)3.7813.870.19
BFlow (Bézier)3.3811.680.17
SL, OODBFlow (polyn.)9.3620.880.36
BFlow (learned)8.6619.640.35
BFlow (Bézier) [25]8.6319.940.36
SSLOurs, polyn.6.7819.760.27
Ours, learned7.4619.780.28
Ours, Bézier6.1416.980.25
", + "image_path": "1ebf7d29a0880d3a5b8df3c60225558b9508dcaa2a447b1206aadc81ff35e6bc.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 30, + 382, + 379, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 382, + 379, + 491 + ], + "spans": [ + { + "bbox": [ + 30, + 382, + 379, + 491 + ], + "type": "inline_equation", + "content": "p_{t}^{i} = \\pi (P_{t}^{i})" + }, + { + "bbox": [ + 30, + 382, + 379, + 491 + ], + "type": "text", + "content": ", where the " + }, + { + "bbox": [ + 30, + 382, + 379, + 491 + ], + "type": "inline_equation", + "content": "\\pi (\\cdot)" + }, + { + "bbox": [ + 30, + 382, + 379, + 491 + ], + "type": "text", + "content": " function represents the perspective projection, including the camera intrinsics. The object poses at time " + }, + { + "bbox": [ + 30, + 382, + 379, + 491 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 30, + 382, + 379, + 491 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 30, + 382, + 379, + 491 + ], + "type": "inline_equation", + "content": "t + 1" + }, + { + "bbox": [ + 30, + 382, + 379, + 491 + ], + "type": "text", + "content": " are provided as " + }, + { + "bbox": [ + 30, + 382, + 379, + 491 + ], + "type": "inline_equation", + "content": "T_{o}^{c_{t}}" + }, + { + "bbox": [ + 30, + 382, + 379, + 491 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 30, + 382, + 379, + 491 + ], + "type": "inline_equation", + "content": "T_{o}^{c_{t + 1}}" + }, + { + "bbox": [ + 30, + 382, + 379, + 491 + ], + "type": "text", + "content": " in the GT data. The rigid-body transformation " + }, + { + "bbox": [ + 30, + 382, + 379, + 491 + ], + "type": "inline_equation", + "content": "T_{c_t}^{t + 1} = (T_o^{c_t + 1})^{-1}T_o^{c_t}" + }, + { + "bbox": [ + 30, + 382, + 379, + 491 + ], + "type": "text", + "content": " maps 3D points in the camera frame at time " + }, + { + "bbox": [ + 30, + 382, + 379, + 491 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 30, + 382, + 379, + 491 + ], + "type": "text", + "content": " to the camera frame at time " + }, + { + "bbox": [ + 30, + 382, + 379, + 491 + ], + "type": "inline_equation", + "content": "t + 1" + }, + { + "bbox": [ + 30, + 382, + 379, + 491 + ], + "type": "text", + "content": " by aligning the shared object coordinate frame. The \"flow\" (or dense motion, since it may not be linear) is defined as " + }, + { + "bbox": [ + 30, + 382, + 379, + 491 + ], + "type": "inline_equation", + "content": "\\Delta p_{t\\rightarrow t + 1}^i\\doteq \\pi (T_{c_t}^{c_{t + 1}}P_t^i) - \\pi (P_t^i)" + }, + { + "bbox": [ + 30, + 382, + 379, + 491 + ], + "type": "text", + "content": ". For each sequence, we generate GT dense motion every 10ms for 300ms. We mask out areas where GT object masks are unavailable. Experiments are carried out on the IMO subset of EVIMO2 using the official train and test splits." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 30, + 492, + 380, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 492, + 380, + 564 + ], + "spans": [ + { + "bbox": [ + 30, + 492, + 380, + 564 + ], + "type": "text", + "content": "Metrics. Ground truth is provided as dense motion from " + }, + { + "bbox": [ + 30, + 492, + 380, + 564 + ], + "type": "inline_equation", + "content": "t = 0" + }, + { + "bbox": [ + 30, + 492, + 380, + 564 + ], + "type": "text", + "content": " to several timesteps in increasing order. We consider " + }, + { + "bbox": [ + 30, + 492, + 380, + 564 + ], + "type": "inline_equation", + "content": "N_{s} = 6" + }, + { + "bbox": [ + 30, + 492, + 380, + 564 + ], + "type": "text", + "content": " timestamps (i.e., subintervals) in a 300ms window and evaluate the quality of predictions with direct extension of the common optical flow metrics end-point-error (EPE), angular error (AE), and percentage of predicted vectors above a specific EPE threshold (\\%Out). We term the corresponding trajectory metrics TEPE and TAE, where" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 102, + 569, + 379, + 587 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 569, + 379, + 587 + ], + "spans": [ + { + "bbox": [ + 102, + 569, + 379, + 587 + ], + "type": "interline_equation", + "content": "\\mathrm {T E P E} = \\frac {1}{N _ {s}} \\sum_ {k = 1} ^ {N _ {s}} \\mathrm {E P E} (\\varDelta \\mathbf {x} _ {\\mathrm {p r e d}} (t _ {k}), \\varDelta \\mathbf {x} _ {\\mathrm {g t}} (t _ {k})). \\tag {6}", + "image_path": "51efad43ca878d174e0a5e758f1e6f262f64b429fa3162164d8b2d3a79216081.jpg" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 33, + 9, + 43, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 9, + 43, + 19 + ], + "spans": [ + { + "bbox": [ + 33, + 9, + 43, + 19 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 169, + 9, + 241, + 19 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 169, + 9, + 241, + 19 + ], + "spans": [ + { + "bbox": [ + 169, + 9, + 241, + 19 + ], + "type": "text", + "content": "F. Hamann et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 32, + 34, + 343, + 47 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 34, + 343, + 47 + ], + "spans": [ + { + "bbox": [ + 32, + 34, + 343, + 47 + ], + "type": "text", + "content": "The number of outliers is calculated on TEPE with a threshold of 3px." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 32, + 48, + 379, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 48, + 379, + 96 + ], + "spans": [ + { + "bbox": [ + 32, + 48, + 379, + 96 + ], + "type": "text", + "content": "Implementation Details. We use the BFlow network architecture and test different motion priors. Specifically, we report results with polynomial, Bezier (with " + }, + { + "bbox": [ + 32, + 48, + 379, + 96 + ], + "type": "inline_equation", + "content": "N_{c} = 10" + }, + { + "bbox": [ + 32, + 48, + 379, + 96 + ], + "type": "text", + "content": ") and learned basis " + }, + { + "bbox": [ + 32, + 48, + 379, + 96 + ], + "type": "inline_equation", + "content": "g_{j}(t)" + }, + { + "bbox": [ + 32, + 48, + 379, + 96 + ], + "type": "text", + "content": " (a small dense neural network with three layers and hidden dimension 64 is trained alongside the main network)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 32, + 97, + 379, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 97, + 379, + 145 + ], + "spans": [ + { + "bbox": [ + 32, + 97, + 379, + 145 + ], + "type": "text", + "content": "The contrast loss module uses one trajectory per " + }, + { + "bbox": [ + 32, + 97, + 379, + 145 + ], + "type": "inline_equation", + "content": "4 \\times 4" + }, + { + "bbox": [ + 32, + 97, + 379, + 145 + ], + "type": "text", + "content": " px, and similarly " + }, + { + "bbox": [ + 32, + 97, + 379, + 145 + ], + "type": "inline_equation", + "content": "4 \\times 4" + }, + { + "bbox": [ + 32, + 97, + 379, + 145 + ], + "type": "text", + "content": " px in the displacement volume. The weight for the spatial smoothness term is " + }, + { + "bbox": [ + 32, + 97, + 379, + 145 + ], + "type": "inline_equation", + "content": "\\lambda = 0.003" + }, + { + "bbox": [ + 32, + 97, + 379, + 145 + ], + "type": "text", + "content": ", and the number of neighbors in the KNN approach is " + }, + { + "bbox": [ + 32, + 97, + 379, + 145 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{traj}} = 32" + }, + { + "bbox": [ + 32, + 97, + 379, + 145 + ], + "type": "text", + "content": ". The loss function " + }, + { + "bbox": [ + 32, + 97, + 379, + 145 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 32, + 97, + 379, + 145 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 32, + 97, + 379, + 145 + ], + "type": "inline_equation", + "content": "L^1" + }, + { + "bbox": [ + 32, + 97, + 379, + 145 + ], + "type": "text", + "content": " norm of the IWE's gradient magnitude [18,56]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 32, + 146, + 379, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 146, + 379, + 277 + ], + "spans": [ + { + "bbox": [ + 32, + 146, + 379, + 277 + ], + "type": "text", + "content": "Training Schedule. The model is first pre-trained on MultiFlow (" + }, + { + "bbox": [ + 32, + 146, + 379, + 277 + ], + "type": "inline_equation", + "content": "\\approx 12000" + }, + { + "bbox": [ + 32, + 146, + 379, + 277 + ], + "type": "text", + "content": " samples) for 50 epochs, with a batch size of 10 with an " + }, + { + "bbox": [ + 32, + 146, + 379, + 277 + ], + "type": "inline_equation", + "content": "L^1" + }, + { + "bbox": [ + 32, + 146, + 379, + 277 + ], + "type": "text", + "content": " loss on the GT trajectory flow. We include data augmentation by flipping (horizontal and vertical) and cropping and use the training/test split provided in [25]. These weights are what we refer to as zero-shot or out-of-distribution (OOD) for the main comparison on EVIMO2. Afterwards, the pre-trained model is fine-tuned with two different losses on EVIMO2. Training using our self-supervised loss is carried out for 15 epochs with a batch size of 6. We refer to this method as Ours. Lastly, a second version is trained directly on the GT flow of EVIMO for 50 epochs, which we refer to as in-distribution (ID). All experiments are performed on Nvidia RTX A6000 GPUs with an AdamW optimizer and a learning rate of " + }, + { + "bbox": [ + 32, + 146, + 379, + 277 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 32, + 146, + 379, + 277 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 32, + 279, + 379, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 279, + 379, + 387 + ], + "spans": [ + { + "bbox": [ + 32, + 279, + 379, + 387 + ], + "type": "text", + "content": "Baselines. This is the first usage of EVIMO2 for dense non-linear flow/motion. For comparison, we provide additional baselines. We use ERAFT [24] with the provided DSEC weights to infer linear flow over the whole prediction time. The timestamps of the intermediate flow are interpolated from the linear prediction. Additionally, we provide results for the prediction of Paredes et. al. [44]. The network is trained self-supervised on DSEC and performs recurrent prediction steps in time in intervals of " + }, + { + "bbox": [ + 32, + 279, + 379, + 387 + ], + "type": "inline_equation", + "content": "10\\mathrm{ms}" + }, + { + "bbox": [ + 32, + 279, + 379, + 387 + ], + "type": "text", + "content": ", inferring the total flow at every step by accumulation of the shorter flows. We unroll this prediction over the whole interval and take the accumulated flow after each " + }, + { + "bbox": [ + 32, + 279, + 379, + 387 + ], + "type": "inline_equation", + "content": "50\\mathrm{ms}" + }, + { + "bbox": [ + 32, + 279, + 379, + 387 + ], + "type": "text", + "content": " step as the predicted flow." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 32, + 389, + 379, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 389, + 379, + 485 + ], + "spans": [ + { + "bbox": [ + 32, + 389, + 379, + 485 + ], + "type": "text", + "content": "Results. Table 2 presents the comparison of the three different training modes. The results of our self-supervised loss can improve the zero-shot performance by nearly " + }, + { + "bbox": [ + 32, + 389, + 379, + 485 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 32, + 389, + 379, + 485 + ], + "type": "text", + "content": ". While direct supervision on EVIMO2 GT data delivers the best results, note that this is only possible because we are comparing on a dataset that was recorded with a motion capture setup and exact object models. Under real conditions, pixel-level annotations of motion tracks are very difficult to obtain. Therefore, our loss delivers a promising tool in combination with synthetic pretraining, substantially helping in overcoming the sim-to-real gap." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 32, + 486, + 379, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 486, + 379, + 546 + ], + "spans": [ + { + "bbox": [ + 32, + 486, + 379, + 546 + ], + "type": "text", + "content": "Figure 3 shows the qualitative comparison of GT tracks and the three training schedules. It delivers insights into the failure cases that our method can overcome. Specifically, the zero-shot model shows errors like overseen object motion (Examples 2 and 3), wrong scale (Ex. 5), and prediction of non-existing motion (Ex. 6), which are improved with self-supervised domain training." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 32, + 547, + 379, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 547, + 379, + 583 + ], + "spans": [ + { + "bbox": [ + 32, + 547, + 379, + 583 + ], + "type": "text", + "content": "Additionally, Fig. 4 shows the error for different prediction time spans. Intuitively, the error increases for longer intervals. The visualization shows that our loss improves zero-shot performance, especially for the long prediction times." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 196, + 9, + 346, + 20 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 196, + 9, + 346, + 20 + ], + "spans": [ + { + "bbox": [ + 196, + 9, + 346, + 20 + ], + "type": "text", + "content": "Motion-prior Contrast Maximization" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 369, + 10, + 379, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 10, + 379, + 19 + ], + "spans": [ + { + "bbox": [ + 369, + 10, + 379, + 19 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 33, + 54, + 379, + 119 + ], + "blocks": [ + { + "bbox": [ + 83, + 33, + 328, + 44 + ], + "lines": [ + { + "bbox": [ + 83, + 33, + 328, + 44 + ], + "spans": [ + { + "bbox": [ + 83, + 33, + 328, + 44 + ], + "type": "text", + "content": "Table 4: Results on the DSEC optical flow benchmark [24]." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 33, + 54, + 379, + 119 + ], + "lines": [ + { + "bbox": [ + 33, + 54, + 379, + 119 + ], + "spans": [ + { + "bbox": [ + 33, + 54, + 379, + 119 + ], + "type": "table", + "html": "
Method\\( {t}_{\\text{inf }}\\left\\lbrack \\mathrm{{ms}}\\right\\rbrack \\)Allinterlaken_00_binterlaken_01_athun_01_a
EPE \\( \\downarrow \\)AE \\( \\downarrow \\)%Out ↓FWL ↑EPE \\( \\downarrow \\)AE \\( \\downarrow \\)%Out ↓FWL ↑EPE \\( \\downarrow \\)AE \\( \\downarrow \\)%Out ↓FWL ↑EPE \\( \\downarrow \\)AE \\( \\downarrow \\)%Out ↓FWL ↑
E-RAFT (SL) [24]46.330.7910.562.681.291.396.226.121.320.906.883.911.420.659.751.871.20
IDNet (SL) [69]0.722.722.04-1.252.114.35-0.772.252.60-0.572.661.47-
Paredes et al. (SSL) [44]40.102.3310.5617.77-3.346.2225.72-2.496.8819.15-1.739.7510.39-
RTEF (MB) [5]4.88-41.952.518.59-59.842.895.94-47.332.923.01-29.702.39
Ev-FlowNet (SSL) [27]3.86-31.451.306.32-47.951.464.91-36.071.422.33-20.921.32
MultiCM (MB) [54]\\( {9.9} \\cdot {10}^{3} \\)3.4713.9830.861.375.749.1938.931.503.749.7731.371.512.1211.0617.681.24
Ours (poly, \\( k = 1 \\) )7.273.208.5315.211.463.214.8920.451.582.385.4617.401.701.396.997.361.30
", + "image_path": "a95bbfb2baa25093f29689cf475f4a5ad1e572edb33ea343d24ade742bde5cac.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 33, + 119, + 379, + 181 + ], + "blocks": [ + { + "bbox": [ + 33, + 119, + 379, + 181 + ], + "lines": [ + { + "bbox": [ + 33, + 119, + 379, + 181 + ], + "spans": [ + { + "bbox": [ + 33, + 119, + 379, + 181 + ], + "type": "table", + "html": "
Methodthun_01_bzurich_city_12_azurich_city_14_czurich_city_15_a
EPE ↓AE ↓%Out ↓FWL ↑EPE ↓AE ↓%Out ↓FWL ↑EPE ↓AE ↓%Out ↓FWL ↑EPE ↓AE ↓%Out ↓FWL ↑
E-RAFT (SL) [24]0.588.411.521.180.6123.161.061.120.7110.231.911.470.598.881.301.34
IDNet (SL) [69]0.552.071.35-0.604.561.16-0.763.742.74-0.552.551.02-
Paredes et al. (SSL) [44]1.668.419.34-2.7223.1626.65-2.6410.2323.01-1.698.889.98-
RTEF (MB) [5]3.91-34.692.483.14-34.081.424.00-45.672.673.78-37.992.82
EV-FlowNet (SSL) [77]3.04-25.411.332.62-25.801.033.36-36.341.242.97-25.531.33
MultiCM (MB) [54]2.4812.0523.561.243.8628.6143.961.142.7212.6230.531.502.3511.8220.991.41
Ours (poly, k = 1)1.546.559.691.338.3320.1622.391.131.788.7912.991.561.456.278.341.51
", + "image_path": "759680ad81dd9ec551822c0b65370591bd65ae8c4f0b9060cdd4c17cfc57c95e.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 31, + 204, + 380, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 204, + 380, + 253 + ], + "spans": [ + { + "bbox": [ + 31, + 204, + 380, + 253 + ], + "type": "text", + "content": "Sensitivity to motion prior. Table 3 compares our method for different motion priors. While Bézier curves show slightly improved performance over the two basis function methods, self-supervised domain training can improve performance in all cases, which proves the robustness and generality of our method." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 31, + 273, + 239, + 285 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 273, + 239, + 285 + ], + "spans": [ + { + "bbox": [ + 31, + 273, + 239, + 285 + ], + "type": "text", + "content": "4.2 Results on Optical Flow Estimation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 30, + 295, + 379, + 343 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 295, + 379, + 343 + ], + "spans": [ + { + "bbox": [ + 30, + 295, + 379, + 343 + ], + "type": "text", + "content": "Dataset, Metrics and Details. We compare our method on the DSEC optical flow benchmark [24]. The dataset consists of sequences from a Prophesee Gen3 event camera, with a resolution of " + }, + { + "bbox": [ + 30, + 295, + 379, + 343 + ], + "type": "inline_equation", + "content": "640 \\times 480" + }, + { + "bbox": [ + 30, + 295, + 379, + 343 + ], + "type": "text", + "content": " px, mounted on a driving car. Ground truth is calculated as the motion field from a co-deployed depth sensor." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 31, + 343, + 379, + 391 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 343, + 379, + 391 + ], + "spans": [ + { + "bbox": [ + 31, + 343, + 379, + 391 + ], + "type": "text", + "content": "We compare the common flow metrics EPE, AE, and percentage of outliers (%Out), where an outlier is a flow vector with EPE " + }, + { + "bbox": [ + 31, + 343, + 379, + 391 + ], + "type": "inline_equation", + "content": ">3\\mathrm{px}" + }, + { + "bbox": [ + 31, + 343, + 379, + 391 + ], + "type": "text", + "content": ". Additionally, we evaluate the Flow Warp Loss (FWL) [57]. FWL " + }, + { + "bbox": [ + 31, + 343, + 379, + 391 + ], + "type": "inline_equation", + "content": ">1" + }, + { + "bbox": [ + 31, + 343, + 379, + 391 + ], + "type": "text", + "content": " means that the IWE is sharper than pixel-wise event accumulation (i.e., IWE with zero optical flow)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 30, + 391, + 380, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 391, + 380, + 584 + ], + "spans": [ + { + "bbox": [ + 30, + 391, + 380, + 584 + ], + "type": "text", + "content": "We use a simple U-Net [49] as the backbone in the linear flow experiments (similar architecture like EV-FlowNet [76]) and choose polynomial basis " + }, + { + "bbox": [ + 30, + 391, + 380, + 584 + ], + "type": "inline_equation", + "content": "g_{j}(t) = t^{j}" + }, + { + "bbox": [ + 30, + 391, + 380, + 584 + ], + "type": "text", + "content": " of degree 1, which effectively leads to a linear motion prior. We use " + }, + { + "bbox": [ + 30, + 391, + 380, + 584 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{traj}} = 32" + }, + { + "bbox": [ + 30, + 391, + 380, + 584 + ], + "type": "text", + "content": " nearest neighbors and " + }, + { + "bbox": [ + 30, + 391, + 380, + 584 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{bins}} = 15" + }, + { + "bbox": [ + 30, + 391, + 380, + 584 + ], + "type": "text", + "content": " time bins in the displacement map. The weight for the regularizer is " + }, + { + "bbox": [ + 30, + 391, + 380, + 584 + ], + "type": "inline_equation", + "content": "\\lambda = 0.003" + }, + { + "bbox": [ + 30, + 391, + 380, + 584 + ], + "type": "text", + "content": ". The network trains for 50 epochs with and Adam optimizer, a learning rate of " + }, + { + "bbox": [ + 30, + 391, + 380, + 584 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 30, + 391, + 380, + 584 + ], + "type": "text", + "content": ", and a total batch size of 28 on two RTX A6000. Results. Table 4 shows the DSEC benchmark results, confirming that our model can generate high-quality optical flow results without having access to any GT. It furthermore achieves the best results among all contrast-maximization-based methods on six out of seven sequences, and only methods directly supervised on GT perform better. Within the self-supervised methods, it improves the AE by " + }, + { + "bbox": [ + 30, + 391, + 380, + 584 + ], + "type": "inline_equation", + "content": "19\\%" + }, + { + "bbox": [ + 30, + 391, + 380, + 584 + ], + "type": "text", + "content": " and the number of inliers by " + }, + { + "bbox": [ + 30, + 391, + 380, + 584 + ], + "type": "inline_equation", + "content": "14\\%" + }, + { + "bbox": [ + 30, + 391, + 380, + 584 + ], + "type": "text", + "content": ". The only exception is the night sequence zurich_city_14. Contrast maximization is based on the brightness constancy assumption, which is violated here by flickering street lights. Hence, in these areas, the contrast loss does not provide reasonable flow predictions. While our methods still provide a better AE and %Out at night, higher outliers in the" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 33, + 9, + 43, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 9, + 43, + 19 + ], + "spans": [ + { + "bbox": [ + 33, + 9, + 43, + 19 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 169, + 9, + 241, + 19 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 169, + 9, + 241, + 19 + ], + "spans": [ + { + "bbox": [ + 169, + 9, + 241, + 19 + ], + "type": "text", + "content": "F. Hamann et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 35, + 32, + 377, + 202 + ], + "blocks": [ + { + "bbox": [ + 35, + 32, + 377, + 202 + ], + "lines": [ + { + "bbox": [ + 35, + 32, + 377, + 202 + ], + "spans": [ + { + "bbox": [ + 35, + 32, + 377, + 202 + ], + "type": "image", + "image_path": "eaa3d82bf0778e152eaf6607b1c9b2320ebd76e2302e39fd4a869b50b4749b38.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 31, + 210, + 379, + 222 + ], + "lines": [ + { + "bbox": [ + 31, + 210, + 379, + 222 + ], + "spans": [ + { + "bbox": [ + 31, + 210, + 379, + 222 + ], + "type": "text", + "content": "Fig. 5: Results on DSEC. Image of warped events and predicted flow by three methods." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 30, + 247, + 379, + 271 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 247, + 379, + 271 + ], + "spans": [ + { + "bbox": [ + 30, + 247, + 379, + 271 + ], + "type": "text", + "content": "flickering regions prevent the difference in the average EPE from reflecting the improved performance of our method on the other six sequences." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 30, + 271, + 380, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 271, + 380, + 415 + ], + "spans": [ + { + "bbox": [ + 30, + 271, + 380, + 415 + ], + "type": "text", + "content": "Figure 5 shows qualitative results of our method. In comparison to another self-supervised optical flow method (d), the visualizations are noticeably sharper, allowing for a more precise representation of the motion in the scene. Moreover, the model performs well in delineating the contours of foreground objects, without the over-smoothing effect often observable in flow methods. The IWE (Fig. 5b) shows sharp results, aligning with the overall high FWL values reported in Tab. 4. While the FWL metric also increases in event collapse (e.g., for RTEF [5]), the IWE visualization reveals that here the model is well-regularized, confirming that the predicted motion aligns the events correctly. The third example in Fig. 5 highlights how our model has improved predictions with a single forward pass, where models relying on temporal recurrence (see (d)) fail at the beginning and need a warm-up stage resulting in higher latency." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 30, + 415, + 379, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 415, + 379, + 463 + ], + "spans": [ + { + "bbox": [ + 30, + 415, + 379, + 463 + ], + "type": "text", + "content": "Inference time. Additionally, Tab. 4 provides inference times " + }, + { + "bbox": [ + 30, + 415, + 379, + 463 + ], + "type": "inline_equation", + "content": "t_{\\mathrm{inf}}" + }, + { + "bbox": [ + 30, + 415, + 379, + 463 + ], + "type": "text", + "content": " for several methods. Note that our method is about " + }, + { + "bbox": [ + 30, + 415, + 379, + 463 + ], + "type": "inline_equation", + "content": "5 \\times" + }, + { + "bbox": [ + 30, + 415, + 379, + 463 + ], + "type": "text", + "content": " faster than the competitive baselines. The reason is that we do not rely on any recurrence in this optical flow application, such as recurrence in time or RAFT-inspired refinement steps." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 30, + 463, + 380, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 463, + 380, + 535 + ], + "spans": [ + { + "bbox": [ + 30, + 463, + 380, + 535 + ], + "type": "text", + "content": "Ablation and Sensitivity. Table 5 lists the performance under different loss settings. It confirms that the number of neighboring trajectories " + }, + { + "bbox": [ + 30, + 463, + 380, + 535 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{traj}}" + }, + { + "bbox": [ + 30, + 463, + 380, + 535 + ], + "type": "text", + "content": " in the KNN approach has a regularizing effect and is important for good performance. The decreased performance on the benchmark for a lower " + }, + { + "bbox": [ + 30, + 463, + 380, + 535 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{traj}}" + }, + { + "bbox": [ + 30, + 463, + 380, + 535 + ], + "type": "text", + "content": " aligns with the visual impression in Fig. 6: models with lower " + }, + { + "bbox": [ + 30, + 463, + 380, + 535 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{traj}}" + }, + { + "bbox": [ + 30, + 463, + 380, + 535 + ], + "type": "text", + "content": " show artifacts and a stronger susceptibility to the aperture problem, as is observable on the rock pattern." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 30, + 536, + 380, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 536, + 380, + 584 + ], + "spans": [ + { + "bbox": [ + 30, + 536, + 380, + 584 + ], + "type": "text", + "content": "Table 5 confirms that the multi-reference approach is crucial for regularizing the contrast loss. While using three instead of one " + }, + { + "bbox": [ + 30, + 536, + 380, + 584 + ], + "type": "inline_equation", + "content": "t_{\\mathrm{ref}}" + }, + { + "bbox": [ + 30, + 536, + 380, + 584 + ], + "type": "text", + "content": " improves performance, our approach using a randomized " + }, + { + "bbox": [ + 30, + 536, + 380, + 584 + ], + "type": "inline_equation", + "content": "t_{\\mathrm{ref}}" + }, + { + "bbox": [ + 30, + 536, + 380, + 584 + ], + "type": "text", + "content": " leads to an additional performance boost. Lastly, the non-linear prior showed no performance improvement on DSEC." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 197, + 9, + 346, + 20 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 9, + 346, + 20 + ], + "spans": [ + { + "bbox": [ + 197, + 9, + 346, + 20 + ], + "type": "text", + "content": "Motion-prior Contrast Maximization" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 369, + 10, + 379, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 10, + 379, + 19 + ], + "spans": [ + { + "bbox": [ + 369, + 10, + 379, + 19 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 36, + 76, + 378, + 166 + ], + "blocks": [ + { + "bbox": [ + 31, + 33, + 380, + 67 + ], + "lines": [ + { + "bbox": [ + 31, + 33, + 380, + 67 + ], + "spans": [ + { + "bbox": [ + 31, + 33, + 380, + 67 + ], + "type": "text", + "content": "Table 5: Sensitivity and Ablation Study for DSEC. Ours corresponds to \"Ours\" in Tab. 4. Configurations marked with “-” are unchanged from our main result. " + }, + { + "bbox": [ + 31, + 33, + 380, + 67 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 31, + 33, + 380, + 67 + ], + "type": "text", + "content": " specifies the change with respect to the original configuration." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 36, + 76, + 378, + 166 + ], + "lines": [ + { + "bbox": [ + 36, + 76, + 378, + 166 + ], + "spans": [ + { + "bbox": [ + 36, + 76, + 378, + 166 + ], + "type": "table", + "html": "
NtrajNrefNcMotion priorEPE ↓ΔEPEAE ↓ΔAE%Out ↓Δ%Out
Ours32~U(0,1)1polynomial3.208.5315.21
Number of neighbor trajectories1---4.581.0713.735.2027.9012.69
8---3.510.3113.204.6723.398.18
Number of reference times-3--4.461.2614.435.9028.6013.39
-1--7.264.0618.7210.1944.7529.54
Type and degree of motion prior--5learned3.220.028.590.0615.340.13
--5polynomial3.280.068.610.0815.680.47
", + "image_path": "4624f4afb0b623fcb8cb27e259b4472b59a262b94b03f51e8881bf60b59f093f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 34, + 174, + 202, + 248 + ], + "blocks": [ + { + "bbox": [ + 34, + 174, + 202, + 248 + ], + "lines": [ + { + "bbox": [ + 34, + 174, + 202, + 248 + ], + "spans": [ + { + "bbox": [ + 34, + 174, + 202, + 248 + ], + "type": "image", + "image_path": "4dd03fe367e6c8c39a80585d3f16b81aedc791444beabe2da998261ca097fdc5.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 50, + 257, + 359, + 269 + ], + "lines": [ + { + "bbox": [ + 50, + 257, + 359, + 269 + ], + "spans": [ + { + "bbox": [ + 50, + 257, + 359, + 269 + ], + "type": "text", + "content": "Fig. 6: Visual effect of the number of neighbors " + }, + { + "bbox": [ + 50, + 257, + 359, + 269 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{traj}}" + }, + { + "bbox": [ + 50, + 257, + 359, + 269 + ], + "type": "text", + "content": " on the predicted flow." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 207, + 174, + 375, + 248 + ], + "blocks": [ + { + "bbox": [ + 207, + 174, + 375, + 248 + ], + "lines": [ + { + "bbox": [ + 207, + 174, + 375, + 248 + ], + "spans": [ + { + "bbox": [ + 207, + 174, + 375, + 248 + ], + "type": "image", + "image_path": "36fcbb9a333d2945ec606ba165b3dd2a9bdd499f72d6e37366ac886719d386cd.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 31, + 290, + 122, + 302 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 290, + 122, + 302 + ], + "spans": [ + { + "bbox": [ + 31, + 290, + 122, + 302 + ], + "type": "text", + "content": "5 Limitations" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 30, + 314, + 379, + 349 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 314, + 379, + 349 + ], + "spans": [ + { + "bbox": [ + 30, + 314, + 379, + 349 + ], + "type": "text", + "content": "Like all CM-based methods, ours is based on the brightness constancy assumption. Therefore, it shows limitations in estimating flow from events that are not caused by motion, e.g., from flickering lights." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 30, + 350, + 379, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 350, + 379, + 385 + ], + "spans": [ + { + "bbox": [ + 30, + 350, + 379, + 385 + ], + "type": "text", + "content": "Our KNN interpolation shows a good trade-off between granularity and regularization; nevertheless, our method is limited by the aperture problem inherent to optical flow." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 30, + 386, + 380, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 386, + 380, + 434 + ], + "spans": [ + { + "bbox": [ + 30, + 386, + 380, + 434 + ], + "type": "text", + "content": "Lastly, during training, all raw events are passed to the loss module, increasing training time. Furthermore, gradient calculation is performed through the additional step of event warping. These steps increase memory requirements and training time of our loss compared to supervised methods." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 31, + 452, + 118, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 452, + 118, + 464 + ], + "spans": [ + { + "bbox": [ + 31, + 452, + 118, + 464 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 30, + 476, + 380, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 476, + 380, + 584 + ], + "spans": [ + { + "bbox": [ + 30, + 476, + 380, + 584 + ], + "type": "text", + "content": "We introduced a new loss formulation based on the contrast maximization framework, by combining it with a non-linear trajectory prior. It is a versatile tool that works with a variety of model architectures and trajectory representations. We presented an efficient method to make the high-dimensional assignment between millions of events and thousands of trajectories feasible. Our experiments show clear advantages against supervised methods on unseen data, where under real-world circumstances no GT is available. Additionally, a U-Net trained with our loss shows state-of-the-art performance on the DSEC optical flow benchmark, while being substantially faster than the previously best methods." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 33, + 10, + 43, + 18 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 10, + 43, + 18 + ], + "spans": [ + { + "bbox": [ + 33, + 10, + 43, + 18 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 169, + 9, + 240, + 19 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 169, + 9, + 240, + 19 + ], + "spans": [ + { + "bbox": [ + 169, + 9, + 240, + 19 + ], + "type": "text", + "content": "F. Hamann et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 32, + 33, + 139, + 47 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 33, + 139, + 47 + ], + "spans": [ + { + "bbox": [ + 32, + 33, + 139, + 47 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 31, + 59, + 381, + 143 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 59, + 381, + 143 + ], + "spans": [ + { + "bbox": [ + 31, + 59, + 381, + 143 + ], + "type": "text", + "content": "We thank Dr. Cornelia Fermüller and the NeuroPAC network for fostering collaborations within the event-based community (NSF OISE 2020624). Funded by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) under Germany's Excellence Strategy - EXC 2002/1 \"Science of Intelligence\" - project number 390523135. We furthermore gratefully acknowledge the support by the following grants: NSF FRR 2220868, NSF IIS-RI 2212433, NSF TRIPODS 1934960, ONR N00014-22-1-2677." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 32, + 162, + 98, + 174 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 162, + 98, + 174 + ], + "spans": [ + { + "bbox": [ + 32, + 162, + 98, + 174 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 33, + 186, + 379, + 583 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 36, + 186, + 379, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 36, + 186, + 379, + 220 + ], + "spans": [ + { + "bbox": [ + 36, + 186, + 379, + 220 + ], + "type": "text", + "content": "1. Akhter, I., Sheikh, Y., Khan, S., Kanade, T.: Trajectory space: A dual representation for nonrigid structure from motion. IEEE Trans. Pattern Anal. Mach. Intell. 33(7), 1442-1456 (jul 2011). https://doi.org/10.1109/TPAMI.2010.201" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 36, + 221, + 379, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 36, + 221, + 379, + 253 + ], + "spans": [ + { + "bbox": [ + 36, + 221, + 379, + 253 + ], + "type": "text", + "content": "2. Bailer, C., Varanasi, K., Stricker, D.: CNN-based patch matching for optical flow with thresholded hinge embedding loss. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR). pp. 2710-2719 (2017). https://doi.org/10.1109/CVPR.2017.290" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 36, + 254, + 379, + 287 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 36, + 254, + 379, + 287 + ], + "spans": [ + { + "bbox": [ + 36, + 254, + 379, + 287 + ], + "type": "text", + "content": "3. Benosman, R., Clercq, C., Lagorce, X., Ieng, S.H., Bartolozzi, C.: Event-based visual flow. IEEE Trans. Neural Netw. Learn. Syst. 25(2), 407-417 (2014). https://doi.org/10.1109/TNNLS.2013.2273537" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 37, + 287, + 379, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 287, + 379, + 319 + ], + "spans": [ + { + "bbox": [ + 37, + 287, + 379, + 319 + ], + "type": "text", + "content": "4. Benosman, R., Ieng, S.H., Clercq, C., Bartolozzi, C., Srinivasan, M.: Asynchronous frameless event-based optical flow. Neural Netw. 27, 32-37 (2012). https://doi.org/10.1016/j.neunet.2011.11.001" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 37, + 320, + 379, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 320, + 379, + 352 + ], + "spans": [ + { + "bbox": [ + 37, + 320, + 379, + 352 + ], + "type": "text", + "content": "5. Brebion, V., Moreau, J., Davoine, F.: Real-time optical flow for vehicular perception with low- and high-resolution event cameras. IEEE Trans. Intell. Transport. Syst. pp. 1-13 (2021). https://doi.org/10.1109/TITS.2021.3136358" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 37, + 353, + 378, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 353, + 378, + 374 + ], + "spans": [ + { + "bbox": [ + 37, + 353, + 378, + 374 + ], + "type": "text", + "content": "6. Brosch, T., Tschechne, S., Neumann, H.: On event-based optical flow detection. Front. Neurosci. 9(137) (Apr 2015). https://doi.org/10.3389/fnins.2015.00137" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 37, + 375, + 379, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 375, + 379, + 418 + ], + "spans": [ + { + "bbox": [ + 37, + 375, + 379, + 418 + ], + "type": "text", + "content": "7. Burner, L., Mitrokhin, A., Fermüller, C., Aloimonos, Y.: EVIMO2: An event camera dataset for motion segmentation, optical flow, structure from motion, and visual inertial odometry in indoor scenes with monocular or stereo algorithms. arXiv e-prints (May 2022). https://doi.org/10.48550/arXiv.2205.03467" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 37, + 419, + 379, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 419, + 379, + 452 + ], + "spans": [ + { + "bbox": [ + 37, + 419, + 379, + 452 + ], + "type": "text", + "content": "8. Butler, D.J., Wulff, J., Stanley, G.B., Black, M.J.: A naturalistic open source movie for optical flow evaluation. In: Eur. Conf. Comput. Vis. (ECCV). pp. 611-625 (2012). https://doi.org/10.1007/978-3-642-33783-3_44" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 37, + 452, + 379, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 452, + 379, + 506 + ], + "spans": [ + { + "bbox": [ + 37, + 452, + 379, + 506 + ], + "type": "text", + "content": "9. Chaney, K., Cladera Ojeda, F., Wang, Z., Bisulco, A., Hsieh, M.A., Korpela, C., Kumar, V., Taylor, C.J., Daniilidis, K.: M3ED: Multi-robot, multi-sensor, multi-environment event dataset. In: IEEE Conf. Comput. Vis. Pattern Recog. Workshops (CVPRW). pp. 4016-4023 (2023). https://doi.org/10.1109/CVPRW59228.2023.00419" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 33, + 506, + 379, + 539 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 506, + 379, + 539 + ], + "spans": [ + { + "bbox": [ + 33, + 506, + 379, + 539 + ], + "type": "text", + "content": "10. Chui, J., Klenk, S., Cremers, D.: Event-based feature tracking in continuous time with sliding window optimization. In: arXiv e-prints (2021). https://doi.org/10.48550/arXiv.2107.04536" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 33, + 539, + 379, + 583 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 539, + 379, + 583 + ], + "spans": [ + { + "bbox": [ + 33, + 539, + 379, + 583 + ], + "type": "text", + "content": "11. Ding, Z., Zhao, R., Zhang, J., Gao, T., Xiong, R., Yu, Z., Huang, T.: Spatiotemporal recurrent networks for event-based optical flow estimation. In: AAAI Conf. Artificial Intell. vol. 36, pp. 525-533 (2022). https://doi.org/10.1609/aaai.v36i1.19931" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 196, + 9, + 346, + 20 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 196, + 9, + 346, + 20 + ], + "spans": [ + { + "bbox": [ + 196, + 9, + 346, + 20 + ], + "type": "text", + "content": "Motion-prior Contrast Maximization" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 369, + 10, + 379, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 10, + 379, + 19 + ], + "spans": [ + { + "bbox": [ + 369, + 10, + 379, + 19 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 33, + 35, + 379, + 584 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 33, + 35, + 379, + 68 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 35, + 379, + 68 + ], + "spans": [ + { + "bbox": [ + 33, + 35, + 379, + 68 + ], + "type": "text", + "content": "12. Doersch, C., Gupta, A., Markeeva, L., Recasens, A., Smaira, L., Aytar, Y., Carreira, J., Zisserman, A., Yang, Y.: Tap-vid: A benchmark for tracking any point in a video. Adv. Neural Inf. Process. Syst. (NeurIPS) 35, 13610-13626 (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 33, + 69, + 379, + 102 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 69, + 379, + 102 + ], + "spans": [ + { + "bbox": [ + 33, + 69, + 379, + 102 + ], + "type": "text", + "content": "13. Doersch, C., Yang, Y., Vecerik, M., Gokay, D., Gupta, A., Aytar, Y., Carreira, J., Zisserman, A.: Tapir: Tracking any point with per-frame initialization and temporal refinement. arXiv preprint arXiv:2306.08637 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 33, + 103, + 379, + 146 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 103, + 379, + 146 + ], + "spans": [ + { + "bbox": [ + 33, + 103, + 379, + 146 + ], + "type": "text", + "content": "14. Dosovitskiy, A., Fischer, P., Ilg, E., Häusser, P., Hazirbaş, C., Golkov, V., van der Smagt, P., Cremers, D., Brox, T.: FlowNet: Learning optical flow with convolutional networks. In: Int. Conf. Comput. Vis. (ICCV). pp. 2758-2766 (2015). https://doi.org/10.1109/ICCV.2015.316" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 33, + 148, + 379, + 180 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 148, + 379, + 180 + ], + "spans": [ + { + "bbox": [ + 33, + 148, + 379, + 180 + ], + "type": "text", + "content": "15. Feydy, J., Glaunès, A., Charlier, B., Bronstein, M.: Fast geometric learning with symbolic matrices. Adv. Neural Inf. Process. Syst. (NeurIPS) 33, 14448-14462 (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 33, + 181, + 379, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 181, + 379, + 257 + ], + "spans": [ + { + "bbox": [ + 33, + 181, + 379, + 257 + ], + "type": "text", + "content": "16. Finateu, T., Niwa, A., Matolin, D., Tsuchimoto, K., Mascheroni, A., Reynaud, E., Mostafalu, P., Brady, F., Chotard, L., LeGoff, F., Takahashi, H., Wakabayashi, H., Oike, Y., Posch, C.: A 1280x720 back-illuminated stacked temporal contrast event-based vision sensor with " + }, + { + "bbox": [ + 33, + 181, + 379, + 257 + ], + "type": "inline_equation", + "content": "4.86\\mu \\mathrm{m}" + }, + { + "bbox": [ + 33, + 181, + 379, + 257 + ], + "type": "text", + "content": " pixels, 1.066Geps readout, programmable event-rate controller and compressive data-formating pipeline. In: IEEE Int. Solid-State Circuits Conf. (ISSCC). pp. 112-114 (2020). https://doi.org/10.1109/ISSCC19947.2020.9063149" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 33, + 258, + 379, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 258, + 379, + 303 + ], + "spans": [ + { + "bbox": [ + 33, + 258, + 379, + 303 + ], + "type": "text", + "content": "17. Gallego, G., Delbruck, T., Orchard, G., Bartolozzi, C., Taba, B., Censi, A., Leutenegger, S., Davison, A., Conradt, J., Daniilidis, K., Scaramuzza, D.: Event-based vision: A survey. IEEE Trans. Pattern Anal. Mach. Intell. 44(1), 154-180 (2022). https://doi.org/10.1109/TPAMI.2020.3008413" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 33, + 303, + 379, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 303, + 379, + 336 + ], + "spans": [ + { + "bbox": [ + 33, + 303, + 379, + 336 + ], + "type": "text", + "content": "18. Gallego, G., Gehrig, M., Scaramuzza, D.: Focus is all you need: Loss functions for event-based vision. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR). pp. 12272-12281 (2019). https://doi.org/10.1109/CVPR.2019.01256" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 33, + 337, + 379, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 337, + 379, + 381 + ], + "spans": [ + { + "bbox": [ + 33, + 337, + 379, + 381 + ], + "type": "text", + "content": "19. Gallego, G., Rebecq, H., Scaramuzza, D.: A unifying contrast maximization framework for event cameras, with applications to motion, depth, and optical flow estimation. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR). pp. 3867-3876 (2018). https://doi.org/10.1109/CVPR.2018.00407" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 33, + 382, + 379, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 382, + 379, + 413 + ], + "spans": [ + { + "bbox": [ + 33, + 382, + 379, + 413 + ], + "type": "text", + "content": "20. Gallego, G., Scaramuzza, D.: Accurate angular velocity estimation with an event camera. IEEE Robot. Autom. Lett. 2(2), 632-639 (2017). https://doi.org/10.1109/LRA.2016.2647639" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 33, + 415, + 379, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 415, + 379, + 449 + ], + "spans": [ + { + "bbox": [ + 33, + 415, + 379, + 449 + ], + "type": "text", + "content": "21. Gehrig, D., Gehrig, M., Hidalgo-Carrió, J., Scaramuzza, D.: Video to Events: Recycl- cling video datasets for event cameras. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR). pp. 3583-3592 (2020). https://doi.org/10.1109/CVPR42600.2020.00364" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 33, + 449, + 379, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 449, + 379, + 482 + ], + "spans": [ + { + "bbox": [ + 33, + 449, + 379, + 482 + ], + "type": "text", + "content": "22. Gehrig, D., Loquercio, A., Derpanis, K.G., Scaramuzza, D.: End-to-end learning of representations for asynchronous event-based data. In: Int. Conf. Comput. Vis. (ICCV). pp. 5632-5642 (2019). https://doi.org/10.1109/ICCV.2019.00573" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 33, + 483, + 379, + 516 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 483, + 379, + 516 + ], + "spans": [ + { + "bbox": [ + 33, + 483, + 379, + 516 + ], + "type": "text", + "content": "23. Gehrig, M., Aarents, W., Gehrig, D., Scaramuzza, D.: DSEC: A stereo event camera dataset for driving scenarios. IEEE Robot. Autom. Lett. 6(3), 4947-4954 (2021). https://doi.org/10.1109/LRA.2021.3068942" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 33, + 517, + 379, + 550 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 517, + 379, + 550 + ], + "spans": [ + { + "bbox": [ + 33, + 517, + 379, + 550 + ], + "type": "text", + "content": "24. Gehrig, M., Millhäsler, M., Gehrig, D., Scaramuzza, D.: E-RAFT: Dense optical flow from event cameras. In: Int. Conf. 3D Vision (3DV). pp. 197-206 (2021). https://doi.org/10.1109/3DV53792.2021.00030" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 33, + 551, + 379, + 584 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 551, + 379, + 584 + ], + "spans": [ + { + "bbox": [ + 33, + 551, + 379, + 584 + ], + "type": "text", + "content": "25. Gehrig, M., Muglikar, M., Scaramuzza, D.: Dense continuous-time optical flow from event cameras. IEEE Trans. Pattern Anal. Mach. Intell. pp. 1-12 (2024). https://doi.org/10.1109/TPAMI.2024.3361671" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 33, + 9, + 43, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 9, + 43, + 19 + ], + "spans": [ + { + "bbox": [ + 33, + 9, + 43, + 19 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 169, + 9, + 241, + 19 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 169, + 9, + 241, + 19 + ], + "spans": [ + { + "bbox": [ + 169, + 9, + 241, + 19 + ], + "type": "text", + "content": "F. Hamann et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 32, + 35, + 380, + 584 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 32, + 35, + 379, + 68 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 35, + 379, + 68 + ], + "spans": [ + { + "bbox": [ + 32, + 35, + 379, + 68 + ], + "type": "text", + "content": "26. Guney, F., Sevilla-Lara, L., Sun, D., Wulff, J.: \"what is optical flow for?\": Workshop results and summary. In: Eur. Conf. Comput. Vis. Workshops (ECCVW). pp. 0-0 (2018)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 33, + 69, + 379, + 102 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 69, + 379, + 102 + ], + "spans": [ + { + "bbox": [ + 33, + 69, + 379, + 102 + ], + "type": "text", + "content": "27. Guo, S., Gallego, G.: CMax-SLAM: Event-based rotational-motion bundle adjustment and SLAM system using contrast maximization. IEEE Trans. Robot. 40, 2442-2461 (2024). https://doi.org/10.1109/TRO.2024.3378443" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 33, + 103, + 379, + 136 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 103, + 379, + 136 + ], + "spans": [ + { + "bbox": [ + 33, + 103, + 379, + 136 + ], + "type": "text", + "content": "28. Hagenaars, J., Paredes-Vallés, F., De Croon, G.: Self-supervised learning of event-based optical flow with spiking neural networks. Adv. Neural Inf. Process. Syst. (NeurIPS) 34, 7167-7179 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 33, + 137, + 380, + 180 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 137, + 380, + 180 + ], + "spans": [ + { + "bbox": [ + 33, + 137, + 380, + 180 + ], + "type": "text", + "content": "29. Hamann, F., Gallego, G.: Stereo co-capture system for recording and tracking fish with frame- and event cameras. In: 26th Int. Conf. on Pattern Recognition (ICPR), Visual observation and analysis of Vertebrate And Insect Behavior (VAIB) Workshop (2022). https://doi.org/10.48550/ARXIV.2207.07332" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 33, + 181, + 379, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 181, + 379, + 214 + ], + "spans": [ + { + "bbox": [ + 33, + 181, + 379, + 214 + ], + "type": "text", + "content": "30. Hamann, F., Ghosh, S., Juarez-Martínez, I., Hart, T., Kacelnik, A., Gallego, G.: Low-power, continuous remote behavioral localization with event cameras. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR) (2024)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 33, + 215, + 379, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 215, + 379, + 247 + ], + "spans": [ + { + "bbox": [ + 33, + 215, + 379, + 247 + ], + "type": "text", + "content": "31. Harley, A.W., Fang, Z., Fragkiadaki, K.: Particle video revisited: Tracking through occlusions using point trajectories. In: Eur. Conf. Comput. Vis. (ECCV). pp. 59-75 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 33, + 249, + 379, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 249, + 379, + 281 + ], + "spans": [ + { + "bbox": [ + 33, + 249, + 379, + 281 + ], + "type": "text", + "content": "32. Huang, Z., Shi, X., Zhang, C., Wang, Q., Cheung, K.C., Qin, H., Dai, J., Li, H.: Flowformer: A transformer architecture for optical flow. In: Eur. Conf. Comput. Vis. (ECCV). pp. 668-685 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 33, + 282, + 379, + 324 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 282, + 379, + 324 + ], + "spans": [ + { + "bbox": [ + 33, + 282, + 379, + 324 + ], + "type": "text", + "content": "33. Ilg, E., Mayer, N., Saikia, T., Keuper, M., Dosovitskiy, A., Brox, T.: FlowNet 2.0: Evolution of optical flow estimation with deep networks. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR). pp. 1647-1655 (2017). https://doi.org/10.1109/cvpr.2017.179" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 33, + 326, + 380, + 359 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 326, + 380, + 359 + ], + "spans": [ + { + "bbox": [ + 33, + 326, + 380, + 359 + ], + "type": "text", + "content": "34. Janai, J., Guney, F., Ranjan, A., Black, M., Geiger, A.: Unsupervised learning of multi-frame optical flow with occlusions. In: Eur. Conf. Comput. Vis. (ECCV). pp. 690-706 (2018)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 33, + 360, + 379, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 360, + 379, + 392 + ], + "spans": [ + { + "bbox": [ + 33, + 360, + 379, + 392 + ], + "type": "text", + "content": "35. Jonschkowski, R., Stone, A., Barron, J.T., Gordon, A., Konolige, K., Angelova, A.: What matters in unsupervised optical flow. In: Eur. Conf. Comput. Vis. (ECCV). pp. 557-572 (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 33, + 394, + 379, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 394, + 379, + 415 + ], + "spans": [ + { + "bbox": [ + 33, + 394, + 379, + 415 + ], + "type": "text", + "content": "36. Karaev, N., Rocco, I., Graham, B., Neverova, N., Vedaldi, A., Rupprecht, C.: Cotracker: It is better to track together. arXiv preprint arXiv:2307.07635 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 33, + 416, + 379, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 416, + 379, + 449 + ], + "spans": [ + { + "bbox": [ + 33, + 416, + 379, + 449 + ], + "type": "text", + "content": "37. Kim, H., Kim, H.J.: Real-time rotational motion estimation with contrast maximization over globally aligned events. IEEE Robot. Autom. Lett. 6(3), 6016-6023 (2021). https://doi.org/10.1109/LRA.2021.3088793" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 33, + 450, + 379, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 450, + 379, + 482 + ], + "spans": [ + { + "bbox": [ + 33, + 450, + 379, + 482 + ], + "type": "text", + "content": "38. Lee, C., Kosta, A., Zhu, A.Z., Chaney, K., Daniilidis, K., Roy, K.: Spike-flownet: Event-based optical flow estimation with energy-efficient hybrid neural networks. In: Eur. Conf. Comput. Vis. (ECCV). pp. 366-382 (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 33, + 484, + 379, + 516 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 484, + 379, + 516 + ], + "spans": [ + { + "bbox": [ + 33, + 484, + 379, + 516 + ], + "type": "text", + "content": "39. Lichtsteiner, P., Posch, C., Delbruck, T.: A " + }, + { + "bbox": [ + 33, + 484, + 379, + 516 + ], + "type": "inline_equation", + "content": "128 \\times 128" + }, + { + "bbox": [ + 33, + 484, + 379, + 516 + ], + "type": "text", + "content": " 120 dB " + }, + { + "bbox": [ + 33, + 484, + 379, + 516 + ], + "type": "inline_equation", + "content": "15~\\mu \\mathrm{s}" + }, + { + "bbox": [ + 33, + 484, + 379, + 516 + ], + "type": "text", + "content": " latency asynchronous temporal contrast vision sensor. IEEE J. Solid-State Circuits 43(2), 566-576 (2008). https://doi.org/10.1109/JSSC.2007.914337" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 33, + 517, + 379, + 560 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 517, + 379, + 560 + ], + "spans": [ + { + "bbox": [ + 33, + 517, + 379, + 560 + ], + "type": "text", + "content": "40. Liu, H., Chen, G., Qu, S., Zhang, Y., Li, Z., Knoll, A., Jiang, C.: TMA: Temporal motion aggregation for event-based optical flow. In: Int. Conf. Comput. Vis. (ICCV). pp. 9651-9660 (Oct 2023). https://doi.org/10.1109/ICCV51070.2023.00888" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 33, + 561, + 379, + 584 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 561, + 379, + 584 + ], + "spans": [ + { + "bbox": [ + 33, + 561, + 379, + 584 + ], + "type": "text", + "content": "41. Liu, M., Delbruck, T.: Adaptive time-slice block-matching optical flow algorithm for dynamic vision sensors. In: British Mach. Vis. Conf. (BMVC). pp. 1-12 (2018)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 196, + 9, + 346, + 19 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 196, + 9, + 346, + 19 + ], + "spans": [ + { + "bbox": [ + 196, + 9, + 346, + 19 + ], + "type": "text", + "content": "Motion-prior Contrast Maximization" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 369, + 10, + 379, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 10, + 379, + 19 + ], + "spans": [ + { + "bbox": [ + 369, + 10, + 379, + 19 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 33, + 35, + 379, + 584 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 33, + 35, + 379, + 79 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 35, + 379, + 79 + ], + "spans": [ + { + "bbox": [ + 33, + 35, + 379, + 79 + ], + "type": "text", + "content": "42. Mayer, N., Ilg, E., Hausser, P., Fischer, P., Cremers, D., Dosovitskiy, A., Brox, T.: A large dataset to train convolutional networks for disparity, optical flow, and scene flow estimation. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR). pp. 4040-4048 (2016)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 33, + 79, + 379, + 121 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 79, + 379, + 121 + ], + "spans": [ + { + "bbox": [ + 33, + 79, + 379, + 121 + ], + "type": "text", + "content": "43. Orchard, G., Benosman, R., Etienne-Cummings, R., Thakor, N.V.: A spiking neural network architecture for visual motion estimation. In: IEEE Biomed. Circuits Syst. Conf. (BioCAS). pp. 298-301 (2013). https://doi.org/10.1109/biocas.2013.6679698" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 33, + 122, + 379, + 164 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 122, + 379, + 164 + ], + "spans": [ + { + "bbox": [ + 33, + 122, + 379, + 164 + ], + "type": "text", + "content": "44. Paredes-Valles, F., Scheper, K.Y., De Wagter, C., de Croon, G.C.: Taming contrast maximization for learning sequential, low-latency, event-based optical flow. In: Int. Conf. Comput. Vis. (ICCV). pp. 9661-9671 (Oct 2023). https://doi.org/10.1109/ICCV51070.2023.00889" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 33, + 165, + 379, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 165, + 379, + 207 + ], + "spans": [ + { + "bbox": [ + 33, + 165, + 379, + 207 + ], + "type": "text", + "content": "45. Paredes-Valles, F., Hagenaars, J.J., Dupeyroux, J., Stroobants, S., Xu, Y., de Croon, G.C.H.E.: Fully neuromorphic vision and control for autonomous drone flight. Science Robotics 9(90), eadi0591 (2024). https://doi.org/10.1126/scirobotics.adio591" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 33, + 208, + 379, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 208, + 379, + 250 + ], + "spans": [ + { + "bbox": [ + 33, + 208, + 379, + 250 + ], + "type": "text", + "content": "46. Posch, C., Serrano-Gotarredona, T., Linares-Barranco, B., Delbruck, T.: Retinomorphic event-based vision sensors: Bioinspired cameras with spiking output. Proc. IEEE 102(10), 1470-1484 (Oct 2014). https://doi.org/10.1109/jproc.2014.2346153" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 33, + 251, + 379, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 251, + 379, + 284 + ], + "spans": [ + { + "bbox": [ + 33, + 251, + 379, + 284 + ], + "type": "text", + "content": "47. Rebecq, H., Gehrig, D., Scaramuzza, D.: ESIM: an open event camera simulator. In: Conf. on Robotics Learning (CoRL). Proc. Machine Learning Research, vol. 87, pp. 969-982. PMLR (2018)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 33, + 284, + 379, + 305 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 284, + 379, + 305 + ], + "spans": [ + { + "bbox": [ + 33, + 284, + 379, + 305 + ], + "type": "text", + "content": "48. Ren, Z., Yan, J., Ni, B., Liu, B., Yang, X., Zha, H.: Unsupervised deep learning for optical flow estimation. In: AAAI Conf. Artificial Intell. vol. 31 (2017)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 33, + 305, + 379, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 305, + 379, + 337 + ], + "spans": [ + { + "bbox": [ + 33, + 305, + 379, + 337 + ], + "type": "text", + "content": "49. Ronneberger, O., Fischer, P., Brox, T.: U-Net: Convolutional networks for biomedical image segmentation. In: Int. Conf. Medical Image Computing and Computer-Assisted Intervention (MICCAI). pp. 234-241 (2015)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 33, + 337, + 379, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 337, + 379, + 358 + ], + "spans": [ + { + "bbox": [ + 33, + 337, + 379, + 358 + ], + "type": "text", + "content": "50. Sand, P., Teller, S.: Particle video: Long-range motion estimation using point trajectories. Int. J. Comput. Vis. 80, 72-91 (2008)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 33, + 358, + 379, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 358, + 379, + 390 + ], + "spans": [ + { + "bbox": [ + 33, + 358, + 379, + 390 + ], + "type": "text", + "content": "51. Seok, H., Lim, J.: Robust feature tracking in dvs event stream using Bezier mapping. In: IEEE Winter Conf. Appl. Comput. Vis. (WACV). pp. 1647-1656 (2020). https://doi.org/10.1109/WACV45572.2020.9093607" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 33, + 390, + 379, + 411 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 390, + 379, + 411 + ], + "spans": [ + { + "bbox": [ + 33, + 390, + 379, + 411 + ], + "type": "text", + "content": "52. Shiba, S., Aoki, Y., Gallego, G.: Event collapse in contrast maximization frameworks. Sensors 22(14), 1-20 (2022). https://doi.org/10.3390/s22145190" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 33, + 412, + 379, + 444 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 412, + 379, + 444 + ], + "spans": [ + { + "bbox": [ + 33, + 412, + 379, + 444 + ], + "type": "text", + "content": "53. Shiba, S., Aoki, Y., Gallego, G.: A fast geometric regularizer to mitigate event collapse in the contrast maximization framework. Adv. Intell. Syst. p. 2200251 (2022). https://doi.org/10.1002/aisy.202200251" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 33, + 444, + 379, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 444, + 379, + 475 + ], + "spans": [ + { + "bbox": [ + 33, + 444, + 379, + 475 + ], + "type": "text", + "content": "54. Shiba, S., Aoki, Y., Gallego, G.: Secrets of event-based optical flow. In: Eur. Conf. Comput. Vis. (ECCV). pp. 628-645 (2022). https://doi.org/10.1007/978-3-031-19797-0 36" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 33, + 476, + 379, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 476, + 379, + 507 + ], + "spans": [ + { + "bbox": [ + 33, + 476, + 379, + 507 + ], + "type": "text", + "content": "55. Shiba, S., Hamann, F., Aoki, Y., Gallego, G.: Event-based background oriented schlieren. IEEE Trans. Pattern Anal. Mach. Intell. 46(4), 2011-2026 (2024). https://doi.org/10.1109/TPAMI.2023.3328188" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 33, + 507, + 379, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 507, + 379, + 540 + ], + "spans": [ + { + "bbox": [ + 33, + 507, + 379, + 540 + ], + "type": "text", + "content": "56. Shiba, S., Klose, Y., Aoki, Y., Gallego, G.: Secrets of event-based optical flow, depth, and ego-motion by contrast maximization. IEEE Trans. Pattern Anal. Mach. Intell. pp. 1-18 (2024). https://doi.org/10.1109/TPAMI.2024.3396116" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 33, + 540, + 379, + 584 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 540, + 379, + 584 + ], + "spans": [ + { + "bbox": [ + 33, + 540, + 379, + 584 + ], + "type": "text", + "content": "57. Stoffregen, T., Scheerlinck, C., Scaramuzza, D., Drummond, T., Barnes, N., Kleeman, L., Mahony, R.: Reducing the sim-to-real gap for event cameras. In: Eur. Conf. Comput. Vis. (ECCV). pp. 534-549 (2020). https://doi.org/https://doi.org/10.1007/978-3-030-58583-9_32" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 33, + 10, + 43, + 18 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 10, + 43, + 18 + ], + "spans": [ + { + "bbox": [ + 33, + 10, + 43, + 18 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 169, + 9, + 240, + 19 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 169, + 9, + 240, + 19 + ], + "spans": [ + { + "bbox": [ + 169, + 9, + 240, + 19 + ], + "type": "text", + "content": "F. Hamann et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 33, + 35, + 379, + 582 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 33, + 35, + 379, + 68 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 35, + 379, + 68 + ], + "spans": [ + { + "bbox": [ + 33, + 35, + 379, + 68 + ], + "type": "text", + "content": "58. Stone, A., Maurer, D., Ayvaci, A., Angelova, A., Jonschkowski, R.: Smurf: Self-teaching multi-frame unsupervised raft with full-image warping. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR). pp. 3887-3896 (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 33, + 68, + 379, + 100 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 68, + 379, + 100 + ], + "spans": [ + { + "bbox": [ + 33, + 68, + 379, + 100 + ], + "type": "text", + "content": "59. Sun, D., Vlasic, D., Herrmann, C., Jampani, V., Krainin, M., Chang, H., Zabih, R., Freeman, W.T., Liu, C.: Autoflow: Learning a better training set for optical flow. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR). pp. 10093-10102 (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 33, + 101, + 379, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 101, + 379, + 134 + ], + "spans": [ + { + "bbox": [ + 33, + 101, + 379, + 134 + ], + "type": "text", + "content": "60. Sun, D., Yang, X., Liu, M.Y., Kautz, J.: PWC-Net: CNNs for optical flow using pyramid, warping, and cost volume. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR) (2018)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 33, + 134, + 379, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 134, + 379, + 156 + ], + "spans": [ + { + "bbox": [ + 33, + 134, + 379, + 156 + ], + "type": "text", + "content": "61. Sun, X., Harley, A.W., Guibas, L.J.: Refining pre-trained motion models. arXiv preprint arXiv:2401.00850 (2024)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 33, + 156, + 379, + 188 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 156, + 379, + 188 + ], + "spans": [ + { + "bbox": [ + 33, + 156, + 379, + 188 + ], + "type": "text", + "content": "62. Teed, Z., Deng, J.: RAFT: Recurrent all pairs field transforms for optical flow. In: Eur. Conf. Comput. Vis. (ECCV). pp. 402-419 (2020). https://doi.org/10.1007/978-3-030-58536-5_24" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 33, + 189, + 379, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 189, + 379, + 232 + ], + "spans": [ + { + "bbox": [ + 33, + 189, + 379, + 232 + ], + "type": "text", + "content": "63. Tulyakov, S., Bochicchio, A., Gehrig, D., Georgoulis, S., Li, Y., Scaramuzza, D.: Time lens++: Event-based frame interpolation with parametric non-linear flow and multi-scale fusion. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR). pp. 17755-17764 (Jun 2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 33, + 233, + 379, + 254 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 233, + 379, + 254 + ], + "spans": [ + { + "bbox": [ + 33, + 233, + 379, + 254 + ], + "type": "text", + "content": "64. Valmadre, J., Lucey, S.: General trajectory prior for non-rigid reconstruction. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR). pp. 1394-1401 (2012)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 33, + 255, + 379, + 276 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 255, + 379, + 276 + ], + "spans": [ + { + "bbox": [ + 33, + 255, + 379, + 276 + ], + "type": "text", + "content": "65. Wan, Z., Dai, Y., Mao, Y.: Learning dense and continuous optical flow from an event camera. IEEE Trans. Image Process. 31, 7237-7251 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 33, + 277, + 379, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 277, + 379, + 298 + ], + "spans": [ + { + "bbox": [ + 33, + 277, + 379, + 298 + ], + "type": "text", + "content": "66. Wang, C., Eckart, B., Lucey, S., Gallo, O.: Neural trajectory fields for dynamic novel view synthesis. arXiv preprint arXiv:2105.05994 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 33, + 298, + 379, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 298, + 379, + 331 + ], + "spans": [ + { + "bbox": [ + 33, + 298, + 379, + 331 + ], + "type": "text", + "content": "67. Wang, Z., Hamann, F., Chaney, K., Jiang, W., Gallego, G., Daniilidis, K.: Event-based continuous color video decompression from single frames. arXiv preprint arXiv:2312.00113 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 33, + 331, + 379, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 331, + 379, + 364 + ], + "spans": [ + { + "bbox": [ + 33, + 331, + 379, + 364 + ], + "type": "text", + "content": "68. Weikersdorfer, D., Hoffmann, R., Conradt, J.: Simultaneous localization and mapping for event-based vision systems. In: Int. Conf. Comput. Vis. Syst. (ICVS). pp. 133-142 (2013). https://doi.org/10.1007/978-3-642-39402-7_14" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 33, + 364, + 379, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 364, + 379, + 386 + ], + "spans": [ + { + "bbox": [ + 33, + 364, + 379, + 386 + ], + "type": "text", + "content": "69. Wu, Y., Paredes-Valles, F., de Croon, G.C.: Lightweight event-based optical flow estimation via iterative deblurring. arXiv preprint arXiv:2211.13726 (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 33, + 386, + 379, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 386, + 379, + 429 + ], + "spans": [ + { + "bbox": [ + 33, + 386, + 379, + 429 + ], + "type": "text", + "content": "70. Ye, C., Mitrokhin, A., Parameshwara, C., Fermüller, C., Yorke, J.A., Aloimonos, Y.: Unsupervised learning of dense optical flow, depth and egomotion with event-based sensors. In: IEEE/RSJ Int. Conf. Intell. Robot. Syst. (IROS). pp. 5831-5838 (2020). https://doi.org/10.1109/IROS45743.2020.9341224" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 33, + 430, + 379, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 430, + 379, + 463 + ], + "spans": [ + { + "bbox": [ + 33, + 430, + 379, + 463 + ], + "type": "text", + "content": "71. Young, P., Lai, A., Hodosh, M., Hockenmaier, J.: From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. Trans. Assoc. Computational Linguistics 2, 67-78 (2014)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 33, + 463, + 379, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 463, + 379, + 496 + ], + "spans": [ + { + "bbox": [ + 33, + 463, + 379, + 496 + ], + "type": "text", + "content": "72. Yu, J.J., Harley, A.W., Derpanis, K.G.: Back to basics: Unsupervised learning of optical flow via brightness constancy and motion smoothness. In: Eur. Conf. Comput. Vis. Workshops (ECCVW). pp. 3-10 (2016)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 33, + 496, + 379, + 528 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 496, + 379, + 528 + ], + "spans": [ + { + "bbox": [ + 33, + 496, + 379, + 528 + ], + "type": "text", + "content": "73. Zen, G., Ricci, E., Sebe, N.: Exploiting sparse representations for robust analysis of noisy complex video scenes. In: Eur. Conf. Comput. Vis. (ECCV). pp. 199-213 (2012)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 33, + 528, + 379, + 562 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 528, + 379, + 562 + ], + "spans": [ + { + "bbox": [ + 33, + 528, + 379, + 562 + ], + "type": "text", + "content": "74. Zheng, Y., Harley, A.W., Shen, B., Wetzstein, G., Guibas, L.J.: PointOdyssey: A large-scale synthetic dataset for long-term point tracking. In: Int. Conf. Comput. Vis. (ICCV). pp. 19855-19865 (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 33, + 562, + 379, + 582 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 562, + 379, + 582 + ], + "spans": [ + { + "bbox": [ + 33, + 562, + 379, + 582 + ], + "type": "text", + "content": "75. Zhu, A.Z., Thakur, D., Ozaslan, T., Pfrommer, B., Kumar, V., Daniilidis, K.: The multivehicle stereo event camera dataset: An event camera dataset for 3D" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 196, + 9, + 346, + 19 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 196, + 9, + 346, + 19 + ], + "spans": [ + { + "bbox": [ + 196, + 9, + 346, + 19 + ], + "type": "text", + "content": "Motion-prior Contrast Maximization" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 369, + 10, + 379, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 10, + 379, + 19 + ], + "spans": [ + { + "bbox": [ + 369, + 10, + 379, + 19 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 18 + }, + { + "para_blocks": [ + { + "bbox": [ + 32, + 35, + 380, + 145 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 48, + 35, + 378, + 57 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 35, + 378, + 57 + ], + "spans": [ + { + "bbox": [ + 48, + 35, + 378, + 57 + ], + "type": "text", + "content": "perception. IEEE Robot. Autom. Lett. 3(3), 2032-2039 (Jul 2018). https://doi.org/10.1109/lra.2018.2800793" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 32, + 57, + 379, + 90 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 57, + 379, + 90 + ], + "spans": [ + { + "bbox": [ + 32, + 57, + 379, + 90 + ], + "type": "text", + "content": "76. Zhu, A.Z., Yuan, L., Chaney, K., Daniilidis, K.: EV-FlowNet: Self-supervised optical flow estimation for event-based cameras. In: Robotics: Science and Systems (RSS). pp. 1-9 (2018). https://doi.org/10.15607/RSS.2018.XIV.062" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 33, + 91, + 380, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 91, + 380, + 123 + ], + "spans": [ + { + "bbox": [ + 33, + 91, + 380, + 123 + ], + "type": "text", + "content": "77. Zhu, A.Z., Yuan, L., Chaney, K., Daniilidis, K.: Unsupervised event-based learning of optical flow, depth, and egomotion. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR). pp. 989-997 (2019). https://doi.org/10.1109/CVPR.2019.00108" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 33, + 123, + 379, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 123, + 379, + 145 + ], + "spans": [ + { + "bbox": [ + 33, + 123, + 379, + 145 + ], + "type": "text", + "content": "78. Zhu, Y., Lucey, S.: Convolutional sparse coding for trajectory reconstruction. IEEE Trans. Pattern Anal. Mach. Intell. 37(3), 529-540 (2013)" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 32, + 10, + 43, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 10, + 43, + 19 + ], + "spans": [ + { + "bbox": [ + 32, + 10, + 43, + 19 + ], + "type": "text", + "content": "20" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 169, + 9, + 241, + 19 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 169, + 9, + 241, + 19 + ], + "spans": [ + { + "bbox": [ + 169, + 9, + 241, + 19 + ], + "type": "text", + "content": "F. Hamann et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 19 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/MotionChain_ Conversational Motion Controllers via Multimodal Prompts/24d4fe30-81b0-462c-a698-38fed56b9fc4_content_list.json b/2024/MotionChain_ Conversational Motion Controllers via Multimodal Prompts/24d4fe30-81b0-462c-a698-38fed56b9fc4_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d4ddf19e83961e16b16fe1003f3430592032b040 --- /dev/null +++ b/2024/MotionChain_ Conversational Motion Controllers via Multimodal Prompts/24d4fe30-81b0-462c-a698-38fed56b9fc4_content_list.json @@ -0,0 +1,1771 @@ +[ + { + "type": "text", + "text": "MotionChain: Conversational Motion Controllers via Multimodal Prompts", + "text_level": 1, + "bbox": [ + 220, + 140, + 782, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Biao Jiang $^{1,2\\star}$ , Xin Chen $^{2\\star\\star}$ , Chi Zhang $^{2}$ , Fukun Yin $^{1}$ , Zhuoyuan Li $^{1}$ , Gang Yu $^{2}$ , and Jiayuan Fan $^{1\\star\\star\\star}$", + "bbox": [ + 217, + 210, + 784, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Fudan University", + "bbox": [ + 434, + 253, + 566, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 Tencent", + "bbox": [ + 467, + 268, + 535, + 281 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Recent advancements in language models have demonstrated their adeptness in conducting multi-turn dialogues and retaining conversational context. However, this proficiency remains largely unexplored in other multimodal generative models, particularly in human motion models. By integrating multi-turn conversations in controlling continuous virtual human movements, generative human motion models can achieve an intuitive and step-by-step process of human task execution for humanoid robotics, game agents, or other embodied systems. In this work, we present MotionChain, a conversational human motion controller that generates continuous and long-term human motion through multimodal prompts. Specifically, MotionChain consists of multi-modal tokenizers that transform various data types such as text, image, and motion, into discrete tokens, coupled with a Vision-Motion-aware Language model. By leveraging large-scale language, vision-language, and vision-motion data to assist motion-related generation tasks, MotionChain thus comprehends each instruction in multi-turn conversation and generates human motions followed by these prompts. Extensive experiments validate the efficacy of MotionChain, demonstrating state-of-the-art performance in conversational motion generation, as well as more intuitive manners of controlling and interacting with virtual humans.", + "bbox": [ + 261, + 320, + 743, + 598 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: 3D Motion $\\cdot$ Motion Generation $\\cdot$ Text-to-Motion", + "bbox": [ + 261, + 611, + 679, + 625 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 651, + 375, + 667 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The success of large language models (LLMs) [60, 62, 93, 94, 122] has sparked significant interest in the development of multi-modal language models. These models aim to transfer instruction-following and zero-shot abilities to other modalities tasks, such as image-language models [1, 50, 107, 124], video-language models [1, 43, 44, 113], and 3D-language models [12, 26, 28, 108]. However, a comprehensive model that can perceive visual input and generate continuous motion through multi-turn conversations has not yet been developed. Such a", + "bbox": [ + 212, + 681, + 787, + 789 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Work done while Biao Jiang was a Research Intern with Tencent.", + "bbox": [ + 215, + 797, + 666, + 811 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "\\*\\* Project lead.", + "bbox": [ + 215, + 811, + 320, + 825 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "\\*\\*\\* Corresponding author.", + "bbox": [ + 205, + 825, + 383, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/86a343de4f67d818efac918a57e0a0271f3ff81233309de547b2138df43313c3.jpg", + "image_caption": [ + "Fig. 1: MotionChain can interpret instructions from multi-turn conversations and generate human motions or textual answers based on text, motion, or image inputs. We provide the conversation results in image-conditioned motion generation (1st column), motion reasoning (second column), motion editing (third column), and motion translation (third column), with each subsequent turn informed by all previous conversations. Left-to-right represents the temporal order." + ], + "image_footnote": [], + "bbox": [ + 222, + 150, + 781, + 252 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "multi-modal model would have wide-ranging applications in fields like humanoid robotics, virtual assistants, game agents and so on.", + "bbox": [ + 212, + 383, + 784, + 412 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Previous research on human motion has explored various tasks, including motion generation [23, 32, 66, 92, 102, 116], motion captioning [20, 24, 32], motion prediction [32, 56, 111, 121], and motion composition [3]. Recent works in text-to-motion [67, 92, 102, 117] have involved pre-trained language models [16, 73] for motion generation. For instance, TEMOS [67] employs BERT [16] text embeddings in an end-to-end transformer architecture, while MDM [92] and MLD [102] both utilize text embeddings from CLIP [73] during the conditional diffusion process. On the other hand, MotionCLIP [91] and TMR [68] focus on modeling the coupled relationship between motion and text description, and MotionGPT [32] introduces a motion-language model that represents human motion and language in one unified vocabulary. However, these above methods treat all tasks as a one-turn conditioned generation, lacking contextual understanding and multi-turn continuous generation abilities. Therefore, we construct a Vision-Motion language model, integrating multi-turn conversations and continuous human motions.", + "bbox": [ + 212, + 415, + 787, + 641 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Two crucial challenges need to be addressed in this conversational motion generation. The first challenge is to contextually generate human motion in a continuous manner, resembling the way real humans move. The second challenge is the scarcity of text-motion paired datasets compared to datasets with pairs of image-language [11, 81], image-pose [37, 58, 61, 64] and video-motion [4, 5, 8, 30, 96]. Fortunately, both human motion and language are sequential and can be continuously \"written\". Building upon this observation, we employ the general vision-language instruction-tuning approach [50, 123] to enable conversational motion generation and question-answering through multi-modal instructions. By integrating image, motion, and language data and encoding them into tokens, the relationship between these three modalities becomes more evident. Therefore, with the advent of large vision-motion and vision-language data, Vision-Motion-language pre-training can enhance the performance of motion-related tasks.", + "bbox": [ + 212, + 643, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Biao Jiang, Xin Chen et al.", + "bbox": [ + 271, + 114, + 455, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this study, we introduce MotionChain, a comprehensive framework that integrates vision, motion, and language. MotionChain leverages large-scale vision-language data, vision-motion data, and pre-trained language models' strong language generation abilities to assist in motion-related generation tasks. To enable MotionChain to comprehend and generate human-like motions, we first train a motion-specific vector quantized variational autoencoder (VQ-VAE) model. This model constructs a \"motion vocabulary\" similar to the English word vocabulary and converts raw motion data into a sequence of motion tokens. To incorporate vision inputs into MotionChain, we then introduce a specialized vision tokenizer that connects a pre-trained vision encoder to the language model. This tokenizer converts image data into visual tokens within the language-motion \"words\" embedding space. These tokens are then processed by a pre-trained language model [14,74,93,94], which learns the relationship between image, motion and language. To enable conversational generation, we construct a multi-modal motion conversation dataset based on the existing text-motion dataset [23] and vision-motion dataset [5]. We then train the language model using our multimodal conversation dataset to learn the correlation between the three modalities. Extensive experiments demonstrate that MotionChain achieves state-of-the-art performance in multiple motion-related tasks.", + "bbox": [ + 212, + 145, + 787, + 431 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We summarize our contributions as follows: (1) We propose MotionChain, a unified vision-motion-language generative pre-trained model, which performs conversational generation tasks via multi-modal inputs with language models. (2) We introduce a motion composition technique, to generate 3D human motions following the temporal order of instructions. (3) We propose a multi-modal motion conversation benchmark, wherein MotionChain achieves competitive performance across diverse motion tasks.", + "bbox": [ + 212, + 433, + 787, + 540 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 564, + 388, + 580 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Human Motion Modeling. There have been numerous attempts to model the relationship between 3D human motion and multiple modalities including incomplete motion [56, 111, 121], action [3, 25, 40, 66, 99, 102], text [22-24, 32, 55, 67, 82, 92, 117-119], image [19, 21, 33, 114, 115] and video [5, 13, 19, 36, 75]. Text-to-motion is one of the most important motion generation tasks, due to the user-friendly and convenient language input. MDM [92], MotionDiffuse [117] and MLD [102] proposes a diffusion-based generative model [27, 77, 85] to generate motions conditioned on different inputs. TM2T [24] and T2M-GPT [116] investigate a generative framework based on VQ-VAE [76, 95] and generative transformer for motion generation. Motion completion task generates motion conditioning on partial motions, such as classical motion prediction [56, 111, 121] or motion in-between [92], which generates the intermediate motion while the first and last parts are fixed. TEACH [3] proposes a past-conditioned transformer model that generates motion from a sequence of actions autoregressively. Apart from motion generation, there is also work investigating other modalities of generation from motion. Two statistical models [90] and recurrent networks [70, 105]", + "bbox": [ + 212, + 597, + 787, + 843 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "MotionChain", + "bbox": [ + 640, + 114, + 730, + 127 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/2ece0ff6c53ff2ed370c8c580373637a05cb859bee80aa4474ce9fa6307beb0f.jpg", + "image_caption": [ + "Fig. 2: Method overview: MotionChain consists of a motion tokenizer $\\mathcal{V}_{\\mathcal{M}}$ (Sec. 3.2), a vision tokenizer $\\mathcal{V}_{\\mathcal{I}}$ (r Sec. 3.2) and a vision-motion-aware language model (Sec. 3.3). By leveraging motion tokens generated by $\\mathcal{V}_{\\mathcal{M}}$ , alongside visual language token embeddings projected by vision tokenizer $\\mathcal{V}_{\\mathcal{I}}$ , and text tokens by text tokenizer, MotionChian achieves a unified learning paradigm for both motion and linguistic data." + ], + "image_footnote": [], + "bbox": [ + 228, + 145, + 784, + 318 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "are learned in mapping motions to language. TM2T [24] proposed a new motion representation that compresses motions into a short sequence of discrete variables and then uses a neural translation network to build mappings between two modalities. In contrast to the above methods limited to only several tasks, MotionGPT [32] treats human motion as a foreign language and leverages language understanding and zero-shot transfer abilities of pre-trained language models.", + "bbox": [ + 212, + 431, + 784, + 521 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Character Control and Animation. Character control involves generating interactive motion sequences based on user instruction signals. One kind of approach [38, 59, 79] is to construct a graph representing transitions between motion clips and plan motion using graph search. Considering the limitations of these graph-based approaches in coarse discreteness, alternative methods like frame blending and concatenation [41], low-dimensional latent space learning [42], motion matching [15] proposed for embedding the task in the feature and [88] do the similar thing through hierarchical setup. Although the control signals for motion control and character animation are different from the instructions in text-to-motion tasks, we still recognize textual commands of conventional human motion generation as a boost for intuitive character control.", + "bbox": [ + 212, + 523, + 787, + 688 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Multi-Modal Language Models. In the field of computer vision, there has been a recent surge of interest in multi-modal models that can process text along with other modalities, including images [18, 29, 44, 104], audio [6, 80, 101], and 3D [12, 45, 54, 83, 108]. CLIP [73] is an example of such a model, which learns a semantic latent representation that connects images with corresponding language descriptions. While language models have achieved success in various tasks, the development of multi-modal language models capable of handling human motion is still limited. Existing works in computer vision can be broadly categorized into two classes. The first consists of end-to-end trained models explored separately for specific research topics. For example, tasks like vision-language navi", + "bbox": [ + 212, + 689, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Biao Jiang, Xin Chen et al.", + "bbox": [ + 271, + 114, + 454, + 128 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/5b5fc2e2f5bbf4400f66ff844044bc3d92ff303935cbe7943668403b496b8640.jpg", + "image_caption": [ + "Fig. 3: Data collection overview: Our initial step in collecting the motion reasoning data involves the utilization of human motion captions derived from an existing text-motion dataset. Subsequent to this, the text-motion retrieval model TMR [68] aids in the segmentation of motion pairs into categories based on the similarity between them. With the assistance of ChatGPT, we proceed to craft motion editing task data that correspond to these categorized similarity levels. Incorporating both motion reasoning and editing single-turn tasks, as well as the extensive 14 tasks delineated in [32], we construct a rich multi-modal multi-turn conversation dataset." + ], + "image_footnote": [], + "bbox": [ + 326, + 156, + 676, + 329 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "gation [2, 7] and Habitat [89] require embodied AI agents to follow natural language instructions and take actions to accomplish goals in visual environments. InstructPix2Pix [7] in image translation enables agents to edit images based on human instructions. The second involves systems that coordinate various models using approaches like LangChain or LLMs [51, 62, 103]. Examples of such systems include Visual ChatGPT [100], X-GPT [125], and MMREACT [110]. While these methods focus on building instruction-following agents, we aim to develop an end-to-end trained multimodal model that can perform conversational motion generation tasks via multi-modal inputs with language models.", + "bbox": [ + 212, + 497, + 787, + 633 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Methods", + "text_level": 1, + "bbox": [ + 215, + 656, + 338, + 672 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To leverage large language data, vision-language data, and vision-motion data for assisting motion-related tasks, we propose a motion-language-vision framework called MotionChain. The framework, as depicted in Fig. 2, consists of a multi-modal tokenizer that converts various types of data (text, image, and motion) into discrete tokens (Sec. 3.2), and a vision-motion-aware Language model that comprehends information from different modalities and generates corresponding answers based on input instructions (Sec. 3.3). Additionally, to simultaneously understand data from multiple modalities, we employ a multi-stage training strategy (Sec. 3.4) for the training of the multi-modal tokenizer and the motion-language-vision framework.", + "bbox": [ + 212, + 689, + 787, + 839 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "MotionChain", + "bbox": [ + 640, + 114, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We first introduced the multi-modal tokenizer, which comprises three branches for processing textual, image, and motion inputs. For textual inputs $w^{1:N} = \\{w^i\\}$ of length $N$ that describes a motion-related question or demand, we employ the SentencePiece model [39] used in previous works [14, 74, 93, 94], which has a vocabulary size of $K_{t}$ and is trained on a large number of language datasets. The motion branch consists of a motion encoder $\\mathcal{E}_{\\mathcal{M}}$ that encodes a motion sequence $m^{1:M} = \\{x^i\\}$ of $M$ frames into $L$ motion tokens $z^{1:L} = \\{z^i\\}$ , where $L = M / l$ and $l$ represents the temporal downsampling rate on motion frames. It also includes a motion decoder $\\mathcal{D}_{\\mathcal{M}}$ that can decode motion tokens back to human motion $\\hat{m}^{1:M}$ . The vision branch processes the input image $X$ with a pre-trained CLIP visual encoder and a learnable linear projection that follows it, converted into language token embeddings $H_{q}$ . Given a textual sentence $w^{1:N}$ , a sequence of motion $m^{1:M}$ , and an image condition $X$ , all encoded as language tokens, our vision-motion-aware language model is designed to produce an answer comprising $L$ tokens, denoted as $\\hat{x}^{1:L} = \\{\\hat{x}^i\\}$ . These output tokens can represent either motion sequences $\\hat{x}_{m}^{1:L}$ or textual descriptions $\\hat{x}_{t}^{1:L}$ , which integrate both human motion $\\hat{m}^{1:M}$ , and text $\\hat{w}^{1:L}$ within the given context.", + "bbox": [ + 212, + 146, + 787, + 402 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.1 Data Collection", + "text_level": 1, + "bbox": [ + 215, + 430, + 392, + 444 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "With the emergence of text-conditioned motion generation tasks, datasets like KIT [69], BABEL [71], HumanML3D [23] and the more recent Motion-X [49] have been developed. However, these datasets predominantly offer text labels as simple action phrases or captions. Building upon these foundations, MotionGPT [32] introduces an instruction-based motion-language dataset that encapsulates 14 core tasks, including motion prediction, translation, and editing, through thousands of instruction templates in a unified format. Despite this advancement, MotionGPT's data lack a deep engagement with the nuances of human motion analysis and are limited to single-turn generation tasks without incorporating contextual memory. Inspired by the recent success of GPT models across text-annotation tasks [17], image-annotation tasks [50], 3D-annotation tasks [28], we propose a data collection methodology integrates the capabilities of existing LLMs like ChatGPT [62], with the text-motion retrieval model TMR [68] to facilitate motion conversation data collection. In addition to the 14 motion-related tasks in MotionGPT [32], we introduce tasks centered around motion reasoning and motion editing, leveraging contextual insights for a deeper motion analysis.", + "bbox": [ + 212, + 460, + 787, + 717 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Utilizing ChatGPT [62], we initiate the collection of motion reasoning data using human motion captions from the text-motion dataset [23], starting with manually designed example queries that explore the contextual scenarios surrounding motions, possible preceding or succeeding actions, the subjects' roles, and the tools or equipment involved, etc. Following this, we employ TMR [68] for categorizing motions from the dataset into varying similarity levels. For medium-similarity motion pairs, we utilize ChatGPT [62] to generate motion editing directives that enable the transformation of one motion to another. For motions of", + "bbox": [ + 212, + 719, + 787, + 839 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Biao Jiang, Xin Chen et al.", + "bbox": [ + 271, + 114, + 454, + 128 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "high similarity, we manually devise tasks aimed at editing their lengths, further enriching the dataset's versatility and analytical scope.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "After the collection of single-turn generation tasks, we progress to develop multi-turn conversation data. This involves the deliberate association of initial motion generation tasks with a variety of follow-up tasks randomly chosen among motion translation, reasoning, editing, etc. Following [122] we construct our conversation data in a structured format, as depicted below:", + "bbox": [ + 212, + 176, + 784, + 252 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "$X_{\\mathrm{system - message}}$", + "bbox": [ + 243, + 262, + 343, + 273 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "USER: $X_{v}$ $\\dot{X}_s^1$ ASSISTANT: $X_{a}^{1} < / s>$", + "bbox": [ + 243, + 273, + 503, + 287 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "USER: $X_{v} X_{s}^{2}$ ASSISTANT: $X_{a}^{2} < / \\mathrm{s}>$", + "bbox": [ + 245, + 287, + 503, + 301 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "USER: $X_{v} X_{s}^{3}$ ASSISTANT: $X_{a}^{3} < / \\mathrm{s}>$ ...", + "bbox": [ + 245, + 301, + 521, + 315 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Where $X_{v}$ is defined as the vision language token embeddings, processed via the visual tokenizer. $X_{s}^{i}$ and $X_{a}^{i}$ are used to denote the source inputs and target answers for each round $i$ , respectively. Both sets of tokens originate from the integrated motion-language vocabulary $V$ , which includes motions, texts, or a blend thereof. The dataset exhibits variability in the number of generation turns up to 10; for the sake of clarity, we present only three examples herein. MotionChain is trained to predict answers, incorporating a learning mechanism that determines whether to stop generation by outputting end of sentence flag $< / \\mathrm{s} >$ based on the current instruction and all preceding questions and answers. In the computation of the loss, as defined in Eq. (5), only the green tokens are utilized.", + "bbox": [ + 212, + 325, + 782, + 477 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.2 Multi-modal Tokenizer", + "text_level": 1, + "bbox": [ + 214, + 497, + 450, + 512 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Motion tokenizer, denoted as $\\mathcal{V}_{\\mathcal{M}}$ , is based on the architecture of Vector Quantized Variational Autoencoders (VQ-VAE) utilized in previous studies [22,24,32, 84,95,98,106,109,116]. Once pre-trained, it can represent motion using discrete tokens, facilitating the integration of motion and language. The Motion tokenizer consists of a motion encoder $\\mathcal{E}_{\\mathcal{M}}$ and a motion decoder $\\mathcal{D}_{\\mathcal{M}}$ . Initially, the motion encoder $\\mathcal{E}$ applies 1D convolutions to the motion features $m^{1:M}$ along the temporal dimension to obtain latent vectors $\\hat{z}^{1:L} = \\mathcal{E}_{\\mathcal{M}}(m^{1:M})$ . Subsequently, the latent vectors $\\hat{z}$ are quantized and transformed into a collection of codebook entries $z$ . The learnable codebook $Z = \\{z_{i}\\}_{i=1}^{K} \\subset \\mathbb{R}^{d}$ comprises $K$ latent embedding vectors, each with a dimension of $d$ . The quantization process $Q(\\cdot)$ replaces each row vector $\\hat{z}$ with its nearest codebook entry $z_{k}$ in $Z$ , which can be expressed as:", + "bbox": [ + 212, + 521, + 784, + 702 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nz _ {i} = Q \\left(\\hat {z} ^ {i}\\right) := \\arg \\min _ {z _ {k} \\in Z} \\| \\hat {z} _ {i} - z _ {k} \\| _ {2}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 367, + 718, + 784, + 734 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We assign $s^i$ as the index number of motion tokens $z^{1:L}$ , so motion tokens $z^{1:L}$ can be represented as a sequence of indices $s^{1:L} = \\{s^i\\}_{i=1}^L$ . The motion decoder $\\mathcal{D}_{\\mathcal{M}}$ can project $z^{1:L} = \\{z^i\\}_{i=1}^L$ back to the raw motion space, resulting in the motion $\\hat{m}^{1:M}$ with $M$ frames. Following [22,24,32,98,116], we adopt three distinct loss functions when training the motion tokenizer:", + "bbox": [ + 212, + 738, + 782, + 814 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathcal {V}} = \\mathcal {L} _ {r} + \\mathcal {L} _ {e} + \\mathcal {L} _ {c} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 431, + 825, + 784, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "MotionChain", + "bbox": [ + 640, + 114, + 730, + 127 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7a41c3c7f2ec84c74344d9d49bbf140e657db0b17849fef6493b0d086748291c.jpg", + "image_caption": [ + "(a) Independent" + ], + "image_footnote": [], + "bbox": [ + 228, + 141, + 400, + 234 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/65da011d9b7b6a97fa4383c8d2c5d774a27ccef12991c5a264f7d1c7b3fbefe4.jpg", + "image_caption": [ + "(b) Past-condition", + "Fig. 4: Motion Composition Variants: We illustrate the baselines for motion composition during multi-turn motion generation (a). independent decoding each turn (b). separate decoding conditioned on the last few tokens from the prior turn (c). decoding with joint motion tokens. Green tokens stand for image condition, blue tokens stand for textual instruction, and orange tokens stand for human motions." + ], + "image_footnote": [], + "bbox": [ + 416, + 141, + 586, + 234 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7e2d5e43eaa0c0fd055ee3e0f8fd642875832808805ab1728c09fdf3a0dd7f5f.jpg", + "image_caption": [ + "(c) Tokens-joint" + ], + "image_footnote": [], + "bbox": [ + 602, + 141, + 776, + 234 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_r$ denotes reconstruction loss, $\\mathcal{L}_e$ denotes the embedding loss, and $\\mathcal{L}_c$ denotes commitment loss.", + "bbox": [ + 212, + 363, + 784, + 391 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "During multi-turn motion generation, the motion continuity between turns is achieved through our motion decoder, which links the motion of the current turn with that of the preceding ones. Taking the composition of two motions as an example: we concatenate the past motion tokens, denoted as $z_{p}^{1:L_{p}}$ , with the tokens representing the current motion, $z_{c}^{1:L_{c}}$ . This concatenated sequence of tokens is subsequently decoded into a comprehensive set of continuous motion features, represented as $m_{\\mathrm{whole}}^{1:M_{\\mathrm{whole}}}$ , as depicted below:", + "bbox": [ + 212, + 393, + 787, + 503 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nz _ {\\text {w h o l e}} ^ {1: \\left(L _ {p} + L _ {c}\\right)} = \\left[ z _ {p} ^ {1: L _ {p}}, z _ {c} ^ {1: L _ {c}} \\right]. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 406, + 513, + 785, + 536 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Similarly, this framework is adept at executing composition tasks involving an array of motions. The comparison results in Tab. 3 demonstrate that our motion tokenizer could effectively perform motion composition tasks.", + "bbox": [ + 212, + 546, + 784, + 590 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Visual Tokenizer accepts images $X_{I}$ as inputs. We employ the CLIP visual encoder that is pre-trained on image-text pairs to derive visual feature $Z_{I}$ . These features are then projected into language token embeddings $X_{v}$ via a linear layer like previous work [50], maintaining consistency in the dimensionality with the language model's word embedding space.", + "bbox": [ + 212, + 592, + 785, + 667 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.3 Motion-aware Language Model", + "text_level": 1, + "bbox": [ + 214, + 691, + 517, + 707 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Language models such as Llama [93, 94] and T5 [14, 74] employ the Sentence-Piece [39] model to encode textual inputs into WordPiece tokens, utilizing a $K_{t}$ word piece vocabulary. Unlike prior text-to-motion [24, 102, 116, 118] and motion-to-text [24] methods that process text and motion separately, we merge the text vocabulary $V_{t} = \\{v_{t}^{i}\\}_{i=1}^{K_{t}}$ with the motion vocabulary $V_{m} = \\{v_{m}^{i}\\}_{i=1}^{K_{m}}$ , maintaining the motion tokenizer's codebook $Z$ order and including special tokens for boundary demarcation. This creates a unified vocabulary $V = \\{V_{t}, V_{m}\\}$ , enabling the formulation of motion-centric tasks in a universal template, where inputs and", + "bbox": [ + 212, + 718, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Biao Jiang, Xin Chen et al.", + "bbox": [ + 271, + 114, + 454, + 128 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "outputs share the same vocabulary. For visual input, our visual tokenizer converts images into visual token embeddings $X_{v}$ , aligning with the language model [39, 62, 74, 93] token space for integrated representation.", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "For single conditioned generation tasks, our input comprises a sequence of $N$ length tokens $X_{s} = \\{x_{s}^{i}\\}_{i = 1}^{N}$ , where $x_{s}\\in \\{V_{t},V_{m}\\}$ representing either text, motion, or a combination thereof, drawn from the unified vocabularies. In cases involving image inputs, visual tokens $X_{v}$ are interspersed at the beginning of the source tokens sequence, forming $[X_v,X_s]$ . Subsequent interaction rounds generate target answer tokens $X_{a}$ . To facilitate iterative result generation and content retention, our framework generates multi-turn conversation data $(X_{v},X_{s}^{1},X_{a}^{1},X_{s}^{2},X_{a}^{2},\\dots ,X_{s}^{T},X_{a}^{T})$ , with $T$ indicating the total turn count. Notably, visual tokens are consistently placed at the forefront of the initial turn's source tokens. The processing sequence is organized such that to predict target answer tokens autoregressively, as shown in Fig. 2. Source tokens are processed by the transformer to predict the next token's probability distribution, formulated as:", + "bbox": [ + 212, + 191, + 784, + 385 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\theta} \\left(X _ {a} \\mid X _ {v}, X _ {s}\\right) = \\prod_ {i} p _ {\\theta} \\left(x _ {a} ^ {i} \\mid X _ {v}, X _ {s, < i}, X _ {a, < i}\\right) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 331, + 386, + 784, + 415 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "with $\\theta$ indicating trainable parameters, and $X_{s, < i}, X_{a, < i}$ representing the sequences of source and preceding target tokens. The training objective is maximizing the log-likelihood of distribution:", + "bbox": [ + 212, + 421, + 782, + 467 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {L M} = - \\sum_ {i = 0} ^ {L _ {t} - 1} \\log p _ {\\theta} \\left(x _ {a} ^ {i} \\mid X _ {v}, X _ {s, < i}, X _ {a, < i}\\right). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 343, + 479, + 784, + 518 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "By optimizing this objective, MotionChain captures the complex interrelations among images, motion, and text, facilitating accurate target \"word\" generation.", + "bbox": [ + 212, + 529, + 782, + 559 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "During the inference phase, target tokens are recursively sampled from the model's predicted distribution $p_{\\theta}\\left(\\hat{x}_a^i \\mid X_v, X_s, \\hat{X}_{a, < i}\\right)$ , ceasing with the appearance of a special end token. This strategy facilitates a step-by-step target sequence generation, where each token's probability is conditioned on all previous turns' sources and targets and current source input.", + "bbox": [ + 212, + 560, + 784, + 641 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "3.4 Training Strategy", + "text_level": 1, + "bbox": [ + 214, + 662, + 408, + 679 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "To facilitate the integration of image and motion comprehension within the language modeling context, we adopt a 3-stage training strategy. (1) The initial stage involves pre-training the motion tokenizer on a corpus of human motion data, in line with [32]. This process establishes the motion vocabulary $V_{m}$ , which serves as a foundation for encoding human motions as a series of discrete tokens. (2) Subsequently, the motion tokenizer remains frozen while we connect the visual tokenizer to the language model framework. This integration is supported by a suite of supervised objectives, including text-to-motion, motion-to-text, and image-based motion generation, aiming to learn the intricate relationships between images, motion, and language. (3) The final stage involves instruction", + "bbox": [ + 212, + 689, + 784, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "MotionChain", + "bbox": [ + 640, + 114, + 730, + 127 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "tuning, and refines the model's capabilities through the application of prompt-based instructions. These instructions are framed within multi-turn conversation sequences, as detailed in Sec. 3.3, to expanded range of motion-related tasks.", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Training of Motion Tokenizer. The initial step involves training the motion tokenizer, guided by the loss objective in Equation 2. This stage enables the tokenizer to represent human motion sequences $\\hat{x}^{1:L}$ as discrete motion tokens, a key step for merging motion data with textual information seamlessly. Once optimized, the motion tokenizer remains frozen.", + "bbox": [ + 212, + 195, + 787, + 271 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Motion-language Pre-training Stage. Leveraging recent developments in language modeling [14, 74, 93, 94, 122] pre-trained on natural language datasets and then fine-tuned with instruction-based phrasing [14, 62]. To augment the model's ability to discern relationships between images and human motions, we first pre-train our MotionChain using a mix of language, image, and motion datasets. Following the stage 1 training of the motion tokenizer, we have established a unified motion-language vocabulary $V = V_{t}, V_{m}$ , capable of representing motions in discrete token form. Moreover, we maintain the visual encoder's weights in the visual tokenizer as fixed, while the linear projection weight $W$ is jointly optimized with the language model. During this stage, the model undertakes three fundamental single-turn modality translation tasks: text-to-motion, motion-to-text, and image-conditioned motion generation, as outlined in Sec. 3.1. The primary objective is to maximize the likelihood of the model according to the loss function specified in Eq. (5), thereby letting the model understand the relationship between language, vision conditions, and motions.", + "bbox": [ + 212, + 275, + 787, + 501 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Instruction Tuning Stage. As described in Sec. 3.1, we construct a multimodal, multi-task, and multi-turn motion conversation dataset by augmenting existing text-to-motion [23] and human mesh reconstruction datasets [5] with targeted instructional prompts and leverage the capabilities of LLMs [62] and the text-motion retrieval model [68] for motion reasoning and editing tasks. The efficacy of instruction tuning, as evidenced across language models [14, 50, 62, 122], is well-established, yielding enhancements in model performance across a wide range of tasks. After instruction tuning, MotionChain can handle more motion-related tasks including the proficient handling of previously unseen tasks", + "bbox": [ + 212, + 505, + 787, + 642 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 683, + 375, + 700 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We evaluate the proposed MotionChain encompasses comprehensive comparisons across both one-turn motion-related tasks and multi-turn motion generation tasks. Firstly, we provide details of the dataset settings, evaluation criteria, and implementation details as specified in Sec. 4.1. Subsequently, comparative analyses are presented, focusing on the motion reasoning task (Sec. 4.2) and the temporal motion composition task (Sec. 4.3). In Sec. 4.4, we evaluate the choice of motion composition technique and different architectures of vision tokenizer.", + "bbox": [ + 212, + 734, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Biao Jiang, Xin Chen et al.", + "bbox": [ + 271, + 114, + 454, + 128 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.1 Experimental Setup", + "text_level": 1, + "bbox": [ + 215, + 146, + 426, + 162 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Datasets. For one-turn motion reasoning tasks, the study employs our proposed multi-modal multi-turn conversation dataset upon HumanML3D [23] with 44,970 sequence-level textual descriptions for 14,616 motion sequences obtained from AMASS [57] and HumanAct12 [25]. The datasets are divided into training, testing, and validation sets with a ratio of $0.8:0.15:0.05$ . To evaluate the multi-turn motion generation task, we focus on BABEL [71] that provides textual descriptions for the motions in the AMASS [57] with annotated segments that overlap in each sequence, which allows evaluating generation of a sequence of motion or actions. We adopt the processed text labels by [3] and motion representation of HumanML3D [23] which combines joint velocities, positions, and rotations. Following [3] we consider pairs of actions for simplicity but Motion-Chain applies to a sequence of actions or motion of arbitrary length. For the image-conditioned motion generation task, we mainly focus on BEDLAM [5], a large synthetic dataset of realistic moving 3D humans containing more than 200 subjects and 380K frames video and motion pair.", + "bbox": [ + 212, + 174, + 787, + 400 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Evaluation Metrics are summarized as four parts. (1) Motion quality: We adopt Frechet Inception Distance (FID) as the primary metric, FID quantifies the divergence in feature distributions between generated and actual motion sequences. Utilizing feature extractors from prior studies [23,49,68], FID measures the distance of feature distributions between the generated and real motions. Following [5,32,33,65,102], we also adopt MPJPE, PA-MPJPE to measure global and local errors in millimeters and ACCL for acceleration errors, to evaluate the quality of the reconstructed motions. (2) Motion Diversity: Utilizing the Diversity (DIV) metric, we calculate variance across motion features to evaluate generation diversity. (3) Text matching: The precision of text-to-motion matches is quantified by the R Precision metric, based on the feature evaluator [23,49,68], and includes an analysis of Top 1/2/3 retrieval accuracy. The Multi-modal Distance (MM Dist) quantifies the semantic gap between motions and texts. (4) Linguistic quality: We follow [24] utilizing linguistic metrics from natural language studies, including BLUE [63], Rouge [48], Cider [97], and BertScore [120] to evaluate the quality of generated motion captions. More detailed benchmark information is provided in the supplementary materials.", + "bbox": [ + 212, + 401, + 787, + 657 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Implementation Details. We set the codebook of the motion tokenizer as $K \\in \\mathbb{R}^{512 \\times 1024}$ for most experiments. The motion encoder, denoted as $\\mathcal{E}_{\\mathcal{M}}$ , integrates a temporal downsampling rate, $l = 4$ . Our vision tokenizer incorporates a frozen Vision Transformer (ViT-L/14) [73] as visual encoder for most experiments. Additionally, for comprehensive ablation studies, we explored the use of both a frozen vision encoder and a Q-former from BLIP-2 [43] as a vision tokenizer. We mainly utilize Flan-T5-base [14] as the underlying architecture for our language model. Moreover, all our models employ the AdamW [53] optimizer with $[\\beta_1, \\beta_2] = [0.9, 0.99]$ for training. The motion tokenizers are trained to utilize a $10^{-4}$ learning rate employing cosine annealing scheduler and a 256 mini-batch size. Our language models based on Flan-T5-base [14] have a $10^{-4}$ learning rate with cosine annealing scheduler and 16 mini-batch sizes in both the", + "bbox": [ + 212, + 659, + 787, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "MotionChain", + "bbox": [ + 640, + 114, + 730, + 127 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/a3ebed76ae76380f8af1425574d735e86c2456539d427e0eba7e7a08c28ca116.jpg", + "table_caption": [ + "Table 1: Comparison of motion reasoning on the test set of our conversation dataset. Our proposed MotionChain is fine-tuned on motion reasoning tasks while other methods' results are generated by their pre-trained weight. $Length_{avg}$ represents the average words in generated answers to all questions. We adopt metrics commonly used in natural language processing tasks for evaluation." + ], + "table_footnote": [], + "table_body": "
MethodsParamsLengthavgBleu@1↑Bleu@4↑Rouge↑Cider↑BertScore↑
Flan-t5-base [14]250M8.344.641.7815.3215.933.45
Flan-t5-large [14]780M11.9512.184.8322.8115.0214.19
Flan-t5-xl [14]3B9.098.544.0124.8915.0318.34
Llama-2-7b [94]7B130.8411.123.6719.141.046.81
Vicuna-1.5-7b [122]7B71.4919.277.3925.755.4419.05
Vicuna-1.5-13b [122]13B84.7417.206.5324.187.7718.00
MotionChain (Ours)573M22.1737.9219.1938.0524.5332.24
", + "bbox": [ + 220, + 226, + 785, + 361 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "pre-train stage and the instruction tuning stage. The motion tokenizer undergoes 10000 epochs of training, while the language model undergoes 500 epochs during the pre-train stage and another 50 epochs during the instruction tuning stage. Most models are trained on 8 Tesla V100 GPUs.", + "bbox": [ + 215, + 387, + 785, + 448 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.2 Comparisons on Motion Reasoning.", + "text_level": 1, + "bbox": [ + 217, + 468, + 553, + 484 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "In Sec. 3.1, we introduce a multi-modal motion conversation dataset, enriched with motion reasoning data facilitated by ChatGPT [62]. This task evaluates the model's reasoning capabilities with motion reasoning tasks, where a motion sequence or its corresponding textual descriptions serve as inputs. Our evaluation compares our MotionChain, which integrates motion perception, against contemporary Large Language Models (LLMs) that possess solely textual processing capabilities. The compared LLMs are assessed using their original pretrained weight. Results in Tab. 1, illustrate that MotionChain exhibits superior motion reasoning proficiency, benefiting from its integrated motion perception.", + "bbox": [ + 215, + 492, + 785, + 628 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.3 Comparisons on Temporal Composition.", + "text_level": 1, + "bbox": [ + 217, + 648, + 591, + 665 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "The temporal motion composition task involves generating a continuous motion sequence from two actions in a time series. We conducted our experiments following the settings in TEACH [3] and used the Amass [57] subset BABEL [71] validation set. Additionally, we processed the motion in AMASS [57] into the format proposed by HumanML3D [23] and trained our MotionChain on the action-to-motion task. To compare with TEACH, we initially used an officially provided pre-trained model to sample motion on the validation set 20 times. Subsequently, we post-processed their motion into the HumanML3D format, represented in SMPL [52]. The performance of our MotionChain is summarized in Table 2. As evaluating generative models quantitatively is challenging, we also provide qualitative comparisons in the supplementary materials.", + "bbox": [ + 215, + 672, + 785, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Biao Jiang, Xin Chen et al.", + "bbox": [ + 271, + 114, + 454, + 128 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/bd4c4bd18243fa1ca18d6361086ab82baf9ef9bea4279cd54c24af155416cb04.jpg", + "table_caption": [ + "Table 2: Comparison of temporal motion composition on Babel [71]. We evaluate the state-of-the-art motion temporal composition method Teach [3] under the $95\\%$ confidence interval from 20 times running. (cf. Sec. 4.1 for notations.)" + ], + "table_footnote": [], + "table_body": "
MethodsDiversityMPJPE↓PA-MPJPE↓ACCL↓
Real15.74±.149---
Teach [3]27.11±.159979.21±.215933.32±.25423.02±.018
MotionChain (Ours)43.25±.159276.05±6.7253.72±.5807.11±0.100
", + "bbox": [ + 269, + 198, + 733, + 277 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/f6c80ae33c1d635aad20d36ec9842fb327502938d709b845b4ab7e49179f44d2.jpg", + "table_caption": [ + "Table 3: Evaluation of motion composition methods on HumanML3D [23]. Here Independent, Past-condition, and Tokens-joint stand for different motion composition variants during multi-turn motion conversation, as illustrated in Fig. 4." + ], + "table_footnote": [], + "table_body": "
MethodMPJPE↓PA-MPJPE↓ACCL↓Diversity
Independent350.79102.9711.406.47
Past-condition232.4646.156.186.01
Tokens-joint108.7718.852.265.56
", + "bbox": [ + 313, + 345, + 689, + 419 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.4 Ablation Studies", + "text_level": 1, + "bbox": [ + 215, + 446, + 401, + 462 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "MotionChain enables multi-modal motion conversation using two main techniques. The first technique involves generating a smooth sequence of motions by concatenating motion tokens which are then decoded back to motion by motion decoder $\\mathcal{D}_{\\mathcal{M}}$ . The second technique involves processing multi-modal visual input through a vision tokenizer, which consists of a frozen vision encoder and a trainable linear projection. To evaluate the effectiveness of these two designs, we compare them with other variants. For a more comprehensive analysis, detailed ablation studies can be found in the supplementary materials.", + "bbox": [ + 212, + 476, + 787, + 595 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Motion Composition Mechanism Apart from the jointly token concatenating mechanism, we also evaluate the performance of temporal motion composition through the other motion temporal composition variants Motion-cat: concatenating the motion in final motion level rather than token level. Experimental results in Tab. 3 show that jointly concatenating motion tokens achieved remarkable performance compared to the other variants. For further information regarding the implementation of the aforementioned vision tokenizer, please refer to the supplementary materials.", + "bbox": [ + 212, + 598, + 787, + 717 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Image Tokenizer Architecture. MotionChain connects the frozen vision encoder to the language model through a linear layer. However, previous vision-language [1,43] works also demonstrate the effectiveness of other kinds of visual-aligning modules. Here we consider the other two vision tokenizer variants: (a) inspired by [1,10,31], we introduce a perceiver module that incorporates a transformer receiving a predefined number of latent input queries. These queries cross-attend to the visual features, enabling effective information exchange. (b) We directly adopt the pre-trained Q-former from BLIP-2 [43] to align visual", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "MotionChain", + "bbox": [ + 640, + 114, + 730, + 127 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/20efcd51b2d5d3e8c826b95b148cfe69c873d871d2168248ffc94b3fc89d8732.jpg", + "table_caption": [ + "Table 4: Evaluation of vision tokenizer architecture on Bedlam [49]. We implement three different architectures, including Q-former, Perceiver, and Linear. We evaluate these results with the metrics in motion reconstruction. Additional information regarding the implementation is in the supplementary materials. (cf. Tab. 2 for notations.)" + ], + "table_footnote": [], + "table_body": "
ArchitectureFirst-frameLast-frame
MPJPE ↓PA-MPJPE ↓MPJPE ↓PA-MPJPE ↓
Q-former195.4986.56134.7357.17
Perceiver185.6199.21134.8957.58
Linear144.3776.48133.7356.73
", + "bbox": [ + 292, + 212, + 709, + 306 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "inputs with the language model. We evaluate the different architectures under the single human image as the first frame condition and the last frame condition separately. Experimental results in Tab. 4 show that a lightweight linear projection is sufficient for comprehending the human pose from visual input. Additional details about the implementation of the above vision tokenizer can be found in the supplements.", + "bbox": [ + 212, + 338, + 787, + 430 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion and Limitation", + "text_level": 1, + "bbox": [ + 215, + 462, + 509, + 479 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Limitation. As the trial to explore conversational human motion generation with visual language models, the proposed MotionChain still has limitations as follows. MotionChain utilizes indeterministic generative models, similar to other language models, but other traditional or neural motion controllers [86, 87] are mostly deterministic and sensitive to control signals. Besides, our method can only generate motion on articulated human bodies, excluding many other human parts such as faces [9, 34, 72] and hands [46, 46, 47, 47, 78]. Although we utilize vision, language, and motion as multimodal conditional inputs akin to human perception, MotionChain is still restricted to the collision signals for human-object and human-scene interactions [35, 82, 112].", + "bbox": [ + 212, + 505, + 787, + 656 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Conclusion. We summarize the proposed MotionChain as a conversational human motion controller to generate continuous and long-term human motion through multimodal prompts. Compared to these one-turn motion generation methods [32, 92, 117], our MotionChain produces more contextually rich generation and can achieve the step-by-step process of human task execution for humanoid robotics and game agents. By leveraging large-scale language, vision-language, and vision-motion data to assist motion-related generation tasks, MotionChain thus comprehends each instruction in multi-turn conversation and generates human motions followed by these prompts. Extensive experiments validate the efficacy of MotionChain, demonstrating state-of-the-art performance in conversational motion generation, as well as more intuitive manners of controlling and interacting with virtual humans.", + "bbox": [ + 212, + 657, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Biao Jiang, Xin Chen et al.", + "bbox": [ + 271, + 114, + 454, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work is supported by National Natural Science Foundation of China (No. 62101137), National Key Research and Development Program of China (No. 2022ZD0160101), Shanghai Natural Science Foundation (No. 23ZR1402900), Shanghai Municipal Science and Technology Major Project (No.2021SHZDZX0103). The computations in this research were performed using the CFFF platform of Fudan University.", + "bbox": [ + 212, + 174, + 807, + 265 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 287, + 321, + 303 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., et al.: Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems 35, 23716-23736 (2022)", + "2. Anderson, P., Wu, Q., Teney, D., Bruce, J., Johnson, M., Sünderhauf, N., Reid, I., Gould, S., Van Den Hengel, A.: Vision-and-language navigation: Interpreting visually-grounded navigation instructions in real environments. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3674-3683 (2018)", + "3. Athanasiou, N., Petrovich, M., Black, M.J., Varol, G.: Teach: Temporal action compositions for 3d humans. In: International Conference on 3D Vision (3DV) (September 2022)", + "4. Bazavan, E.G., Zanfir, A., Zanfir, M., Freeman, W.T., Sukthankar, R., Sminchisescu, C.: Hspace: Synthetic parametric humans animated in complex environments. arXiv preprint arXiv:2112.12867 (2021)", + "5. Black, M.J., Patel, P., Tesch, J., Yang, J.: Bedlam: A synthetic dataset of bodies exhibiting detailed lifelike animated motion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8726-8737 (2023)", + "6. Borsos, Z., Marinier, R., Vincent, D., Kharitonov, E., Pietquin, O., Sharifi, M., Roblek, D., Teboul, O., Grangier, D., Tagliasacchi, M., et al.: Audiolm: a language modeling approach to audio generation. IEEE/ACM transactions on audio, speech, and language processing 31, 2523-2533 (2023)", + "7. Brooks, T., Holynski, A., Efros, A.A.: Instructpix2pix: Learning to follow image editing instructions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18392-18402 (2023)", + "8. Cai, Z., Zhang, M., Ren, J., Wei, C., Ren, D., Lin, Z., Zhao, H., Yang, L., Loy, C.C., Liu, Z.: Playing for 3d human recovery. arXiv preprint arXiv:2110.07588 (2021)", + "9. Cao, X., Chen, Z., Chen, A., Chen, X., Li, S., Yu, J.: Sparse photometric 3d face reconstruction guided by morphable models. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 4635-4644 (2018)", + "10. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: European conference on computer vision. pp. 213-229. Springer (2020)", + "11. Changpinyo, S., Sharma, P., Ding, N., Soricut, R.: Conceptual 12m: Pushing web-scale image-text pre-training to recognize long-tail visual concepts. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3558-3568 (2021)" + ], + "bbox": [ + 225, + 316, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "MotionChain", + "bbox": [ + 640, + 114, + 730, + 127 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "12. Chen, S., Chen, X., Zhang, C., Li, M., Yu, G., Fei, H., Zhu, H., Fan, J., Chen, T.: Ll3da: Visual interactive instruction tuning for omni-3d understanding reasoning and planning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 26428-26438 (2024)", + "13. Choudhury, R., Kitani, K.M., Jeni, L.A.: Tempo: Efficient multi-view pose estimation, tracking, and forecasting. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 14750-14760 (2023)", + "14. Chung, H.W., Hou, L., Longpre, S., Zoph, B., Tay, Y., Fedus, W., Li, E., Wang, X., Dehghani, M., Brahma, S., et al.: Scaling instruction-finetuned language models. arXiv preprint arXiv:2210.11416 (2022)", + "15. Clavet, S.: Motion matching and the road to next-gen animation. In: Proc. of GDC. vol. 2, p. 9 (2016)", + "16. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018)", + "17. Gilardi, F., Alizadeh, M., Kubli, M.: Chatgpt outperforms crowd-workers for text-annotation tasks. arXiv preprint arXiv:2303.15056 (2023)", + "18. Girdhar, R., El-Nouby, A., Liu, Z., Singh, M., Alwala, K.V., Joulin, A., Misra, I.: Imagebind: One embedding space to bind them all. arXiv preprint arXiv:2305.05665 (2023)", + "19. Goel, S., Pavlakos, G., Rajasegaran, J., Kanazawa, A., Malik, J.: Humans in 4d: Reconstructing and tracking humans with transformers. arXiv preprint arXiv:2305.20091 (2023)", + "20. Goutsu, Y., Inamura, T.: Linguistic descriptions of human motion with generative adversarial seq2seq learning. In: 2021 IEEE International Conference on Robotics and Automation (ICRA). pp. 4281-4287. IEEE (2021)", + "21. Guler, R.A., Kokkinos, I.: Holopose: Holistic 3d human reconstruction in-the-wild. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10884-10894 (2019)", + "22. Guo, C., Mu, Y., Javed, M.G., Wang, S., Cheng, L.: Momask: Generative masked modeling of 3d human motions. arXiv preprint arXiv:2312.00063 (2023)", + "23. Guo, C., Zou, S., Zuo, X., Wang, S., Ji, W., Li, X., Cheng, L.: Generating diverse and natural 3d human motions from text. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5152-5161 (June 2022)", + "24. Guo, C., Zuo, X., Wang, S., Cheng, L.: Tm2t: Stochastic and tokenized modeling for the reciprocal generation of 3d human motions and texts. In: ECCV (2022)", + "25. Guo, C., Zuo, X., Wang, S., Zou, S., Sun, Q., Deng, A., Gong, M., Cheng, L.: Action2motion: Conditioned generation of 3d human motions. In: Proceedings of the 28th ACM International Conference on Multimedia. pp. 2021-2029 (2020)", + "26. Guo, Z., Zhang, R., Zhu, X., Tang, Y., Ma, X., Han, J., Chen, K., Gao, P., Li, X., Li, H., et al.: Point-bind & point-llm: Aligning point cloud with multimodality for 3d understanding, generation, and instruction following. arXiv preprint arXiv:2309.00615 (2023)", + "27. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems 33, 6840-6851 (2020)", + "28. Hong, Y., Zhen, H., Chen, P., Zheng, S., Du, Y., Chen, Z., Gan, C.: 3d-lm: Injecting the 3d world into large language models. arXiv preprint arXiv:2307.12981 (2023)" + ], + "bbox": [ + 225, + 146, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Biao Jiang, Xin Chen et al.", + "bbox": [ + 271, + 114, + 454, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "29. Huang, S., Dong, L., Wang, W., Hao, Y., Singhal, S., Ma, S., Lv, T., Cui, L., Mohammed, O.K., Liu, Q., et al.: Language is not all you need: Aligning perception with language models. arXiv preprint arXiv:2302.14045 (2023)", + "30. Ionescu, C., Papava, D., Olaru, V., Sminchisescu, C.: Human3. 6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. IEEE transactions on pattern analysis and machine intelligence 36(7), 1325-1339 (2013)", + "31. Jaegle, A., Gimeno, F., Brock, A., Vinyals, O., Zisserman, A., Carreira, J.: Perceiver: General perception with iterative attention. In: International conference on machine learning. pp. 4651-4664. PMLR (2021)", + "32. Jiang, B., Chen, X., Liu, W., Yu, J., Yu, G., Chen, T.: Motiongpt: Human motion as a foreign language. arXiv preprint arXiv:2306.14795 (2023)", + "33. Kanazawa, A., Black, M.J., Jacobs, D.W., Malik, J.: End-to-end recovery of human shape and pose. In: Computer Vision and Pattern Regognition (CVPR) (2018)", + "34. Karras, T., Aila, T., Laine, S., Herva, A., Lehtinen, J.: Audio-driven facial animation by joint end-to-end learning of pose and emotion. ACM Transactions on Graphics (TOG) 36(4), 1-12 (2017)", + "35. Karunratanakul, K., Preechakul, K., Suwajanakorn, S., Tang, S.: Guided motion diffusion for controllable human motion synthesis. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2151-2162 (2023)", + "36. Kocabas, M., Athanasiou, N., Black, M.J.: Vibe: Video inference for human body pose and shape estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2020)", + "37. Kocabas, M., Huang, C.H.P., Tesch, J., Müller, L., Hilliges, O., Black, M.J.: Spec: Seeing people in the wild with an estimated camera. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 11035-11045 (2021)", + "38. Kovar, L., Gleicher, M., Pighin, F.: Motion graphs. In: Seminal Graphics Papers: Pushing the Boundaries, Volume 2, pp. 723-732 (2023)", + "39. Kudo, T., Richardson, J.: Sentencepiece: A simple and language independent subword tokenizer and detokenizer for neural text processing. arXiv preprint arXiv:1808.06226 (2018)", + "40. Lee, T., Moon, G., Lee, K.M.: Multiact: Long-term 3d human motion generation from multiple action labels. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 1231-1239 (2023)", + "41. Lee, Y., Wampler, K., Bernstein, G., Popovic, J., Popovic, Z.: Motion fields for interactive character locomotion. In: ACM SIGGRAPH Asia 2010 papers, pp. 1-8 (2010)", + "42. Levine, S., Wang, J.M., Haraux, A., Popovic, Z., Koltun, V.: Continuous character control with low-dimensional embeddings. ACM Transactions on Graphics (TOG) 31(4), 1-10 (2012)", + "43. Li, J., Li, D., Savarese, S., Hoi, S.: Blip-2: Bootstrapping language-image pretraining with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597 (2023)", + "44. Li, J., Li, D., Xiong, C., Hoi, S.: Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In: International Conference on Machine Learning. pp. 12888-12900. PMLR (2022)", + "45. Li, M., Chen, X., Zhang, C., Chen, S., Zhu, H., Yin, F., Yu, G., Chen, T.: M3dbench: Let's instruct large models with multi-modal 3d prompts. arXiv preprint arXiv:2312.10763 (2023)" + ], + "bbox": [ + 225, + 146, + 785, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "MotionChain", + "bbox": [ + 640, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "46. Li, Y., Wu, M., Zhang, Y., Xu, L., Yu, J.: Piano: A parametric hand bone model from magnetic resonance imaging. arXiv preprint arXiv:2106.10893 (2021)", + "47. Li, Y., Zhang, L., Qiu, Z., Jiang, Y., Li, N., Ma, Y., Zhang, Y., Xu, L., Yu, J.: Nimble: a non-rigid hand model with bones and muscles. ACM Transactions on Graphics (TOG) 41(4), 1-16 (2022)", + "48. Lin, C.Y.: Rouge: A package for automatic evaluation of summaries. In: Text summarization branches out. pp. 74-81 (2004)", + "49. Lin, J., Zeng, A., Lu, S., Cai, Y., Zhang, R., Wang, H., Zhang, L.: Motion-x: A large-scale 3d expressive whole-body human motion dataset. arXiv preprint arXiv:2307.00818 (2023)", + "50. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. arXiv preprint arXiv:2304.08485 (2023)", + "51. Liu, X., Yan, H., Zhang, S., An, C., Qiu, X., Lin, D.: Scaling laws of rope-based extrapolation. arXiv preprint arXiv:2310.05209 (2023)", + "52. Loper, M., Mahmood, N., Romero, J., Pons-Moll, G., Black, M.J.: Smpl: A skinned multi-person linear model. ACM Trans. Graph. 34(6), 248:1-248:16 (Oct 2015)", + "53. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101 (2017)", + "54. Lu, C., Yin, F., Chen, X., Liu, W., Chen, T., Yu, G., Fan, J.: A large-scale outdoor multi-modal dataset and benchmark for novel view synthesis and implicit scene reconstruction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7557-7567 (2023)", + "55. Lu, S., Chen, L.H., Zeng, A., Lin, J., Zhang, R., Zhang, L., Shum, H.Y.: Humantomato: Text-aligned whole-body motion generation. arXiv preprint arXiv:2310.12978 (2023)", + "56. Ma, H., Li, J., Hosseini, R., Tomizuka, M., Choi, C.: Multi-objective diverse human motion prediction with knowledge distillation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8161-8171 (2022)", + "57. Mahmood, N., Ghorbani, N., Troje, N.F., Pons-Moll, G., Black, M.J.: Amass: Archive of motion capture as surface shapes. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (October 2019)", + "58. Mehta, D., Rhodin, H., Casas, D., Fua, P., Sotnychenko, O., Xu, W., Theobalt, C.: Monocular 3d human pose estimation in the wild using improved cnn supervision. In: International Conference on 3D Vision (3DV) (2017). https://doi.org/10.1109/3dv.2017.00064, http://gvv.mpi-inf.mpg.de/3dhp_dataset", + "59. Min, J., Chai, J.: Motion graphs++ a compact generative model for semantic motion analysis and synthesis. ACM Transactions on Graphics (TOG) 31(6), 1-12 (2012)", + "60. OpenAI: Gpt-4 technical report (2023)", + "61. Ordonez, V., Kulkarni, G., Berg, T.: Im2text: Describing images using 1 million captioned photographs. Advances in neural information processing systems 24 (2011)", + "62. Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., et al.: Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems 35, 27730-27744 (2022)", + "63. Papineni, K., Roukos, S., Ward, T., Zhu, W.J.: Bleu: a method for automatic evaluation of machine translation. In: Proceedings of the 40th annual meeting of the Association for Computational Linguistics. pp. 311-318 (2002)" + ], + "bbox": [ + 225, + 146, + 784, + 839 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Biao Jiang, Xin Chen et al.", + "bbox": [ + 271, + 114, + 454, + 127 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "64. Patel, P., Huang, C.H.P., Tesch, J., Hoffmann, D.T., Tripathi, S., Black, M.J.: Agora: Avatars in geography optimized for regression analysis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13468-13478 (2021)", + "65. Pavlakos, G., Choutas, V., Ghorbani, N., Bolkart, T., Osman, A.A.A., Tzionas, D., Black, M.J.: Expressive body capture: 3d hands, face, and body from a single image. In: Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR) (2019)", + "66. Petrovich, M., Black, M.J., Varol, G.: Action-conditioned 3D human motion synthesis with transformer VAE. In: International Conference on Computer Vision (ICCV) (2021)", + "67. Petrovich, M., Black, M.J., Varol, G.: TEMOS: Generating diverse human motions from textual descriptions. In: European Conference on Computer Vision (ECCV) (2022)", + "68. Petrovich, M., Black, M.J., Varol, G.: Tmr: Text-to-motion retrieval using contrastive 3d human motion synthesis. arXiv preprint arXiv:2305.00976 (2023)", + "69. Plappert, M., Mandery, C., Asfour, T.: The kit motion-language dataset. Big Data 4(4), 236-252 (dec 2016). https://doi.org/10.1089/big.2016.0028, http://dx.doi.org/10.1089/big.2016.0028", + "70. Plappert, M., Mandery, C., Asfour, T.: Learning a bidirectional mapping between human whole-body motion and natural language using deep recurrent neural networks. Robotics and Autonomous Systems 109, 13-26 (2018)", + "71. Punnakkal, A.R., Chandrasekaran, A., Athanasiou, N., Quiros-Ramirez, A., Black, M.J.: BABEL: Bodies, action and behavior with english labels. In: Proceedings IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR). pp. 722-731 (Jun 2021)", + "72. Qiu, Z., Li, Y., He, D., Zhang, Q., Zhang, L., Zhang, Y., Wang, J., Xu, L., Wang, X., Zhang, Y., et al.: Sculptor: Skeleton-consistent face creation using a learned parametric generator. ACM Transactions on Graphics (TOG) 41(6), 1-17 (2022)", + "73. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International Conference on Machine Learning. pp. 8748-8763. PMLR (2021)", + "74. Raffel, C., Shazeer, N., Roberts, A., Lee, K., Narang, S., Matena, M., Zhou, Y., Li, W., Liu, P.J.: Exploring the limits of transfer learning with a unified text-to-text transformer. The Journal of Machine Learning Research 21(1), 5485-5551 (2020)", + "75. Rajasegaran, J., Pavlakos, G., Kanazawa, A., Malik, J.: Tracking people by predicting 3d appearance, location and pose. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2740-2749 (2022)", + "76. Razavi, A., Van den Oord, A., Vinyals, O.: Generating diverse high-fidelity images with vq-vae-2. Advances in neural information processing systems 32 (2019)", + "77. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2022), https://github.com/CompVis/latent-diffusionhttps://arxiv.org/abs/2112.10752", + "78. Romero, J., Tzionas, D., Black, M.J.: Embodied hands: Modeling and capturing hands and bodies together. arXiv preprint arXiv:2201.02610 (2022)", + "79. Rose, C., Cohen, M.F., Bodenheimer, B.: Verbs and adverbs: Multidimensional motion interpolation. IEEE Computer Graphics and Applications 18(5), 32-40 (1998)" + ], + "bbox": [ + 225, + 146, + 784, + 839 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "MotionChain", + "bbox": [ + 640, + 114, + 730, + 127 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 18 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "80. Rubenstein, P.K., Asawaroengchai, C., Nguyen, D.D., Bapna, A., Borsos, Z., Quitry, F.d.C., Chen, P., Badawy, D.E., Han, W., Kharitonov, E., et al.: Audiopalm: A large language model that can speak and listen. arXiv preprint arXiv:2306.12925 (2023)", + "81. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems 35, 25278-25294 (2022)", + "82. Shafir, Y., Tevet, G., Kapon, R., Bermano, A.H.: Human motion diffusion as a generative prior. arXiv preprint arXiv:2303.01418 (2023)", + "83. Siddiqui, Y., Alliegro, A., Artemov, A., Tommasi, T., Sirigatti, D., Rosov, V., Dai, A., Nießner, M.: Meshgpt: Generating triangle meshes with decoder-only transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19615-19625 (2024)", + "84. Siyao, L., Yu, W., Gu, T., Lin, C., Wang, Q., Qian, C., Loy, C.C., Liu, Z.: Bailando: 3d dance generation by actor-critic gpt with choreographic memory. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11050-11059 (2022)", + "85. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502 (2020)", + "86. Starke, S., Mason, I., Komura, T.: Deepphase: periodic autoencoders for learning motion phase manifolds. ACM Transactions on Graphics (TOG) 41(4), 1-13 (2022)", + "87. Starke, S., Zhang, H., Komura, T., Saito, J.: Neural state machine for character-scene interactions. ACM Trans. Graph. 38(6), 209-1 (2019)", + "88. Starke, S., Zhao, Y., Zinno, F., Komura, T.: Neural animation layering for synthesizing martial arts movements. ACM Transactions on Graphics (TOG) 40(4), 1-16 (2021)", + "89. Szot, A., Clegg, A., Undersander, E., Wijmans, E., Zhao, Y., Turner, J., Maestre, N., Mukadam, M., Chaplot, D.S., Maksymets, O., et al.: Habitat 2.0: Training home assistants to rearrange their habitat. Advances in Neural Information Processing Systems 34, 251-266 (2021)", + "90. Takano, W., Nakamura, Y.: Statistical mutual conversion between whole body motion primitives and linguistic sentences for human motions. The International Journal of Robotics Research 34(10), 1314-1328 (2015)", + "91. Tevet, G., Gordon, B., Hertz, A., Bermano, A.H., Cohen-Or, D.: Motionclip: Exposing human motion generation to clip space. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXII. pp. 358-374. Springer (2022)", + "92. Tevet, G., Raab, S., Gordon, B., Shafir, Y., Bermano, A.H., Cohen-Or, D.: Human motion diffusion model. arXiv preprint arXiv:2209.14916 (2022)", + "93. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., et al.: Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 (2023)", + "94. Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi, A., Babaei, Y., Bashlykov, N., Batra, S., Bhargava, P., Bhosale, S., et al.: Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288 (2023)", + "95. Van Den Oord, A., Vinyals, O., et al.: Neural discrete representation learning. Advances in neural information processing systems 30 (2017)" + ], + "bbox": [ + 225, + 146, + 784, + 839 + ], + "page_idx": 19 + }, + { + "type": "page_number", + "text": "20", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 19 + }, + { + "type": "header", + "text": "Biao Jiang, Xin Chen et al.", + "bbox": [ + 271, + 114, + 454, + 127 + ], + "page_idx": 19 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "96. Varol, G., Romero, J., Martin, X., Mahmood, N., Black, M., Laptev, I., Schmid, C.: Learning from synthetic humans. In: Computer Vision and Pattern Recognition (CVPR) (2017)", + "97. Vedantam, R., Lawrence Zitnick, C., Parikh, D.: Cider: Consensus-based image description evaluation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4566-4575 (2015)", + "98. Wang, C.: T2m-hifigpt: Generating high quality human motion from textual descriptions with residual discrete representations. arXiv preprint arXiv:2312.10628 (2023)", + "99. Wang, W., Zhe, X., Chen, H., Kang, D., Li, T., Chen, R., Bao, L.: Neural marionette: A transformer-based multi-action human motion synthesis system. arXiv preprint arXiv:2209.13204 (2022)", + "100. Wu, C., Yin, S., Qi, W., Wang, X., Tang, Z., Duan, N.: Visual chatgpt: Talking, drawing and editing with visual foundation models. arXiv preprint arXiv:2303.04671 (2023)", + "101. Wu, Y.C., Gebru, I.D., Markovic, D., Richard, A.: Audiodec: An open-source streaming high-fidelity neural audio codec. In: ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 1-5. IEEE (2023)", + "102. Xin, C., Jiang, B., Liu, W., Huang, Z., Fu, B., Chen, T., Yu, J., Yu, G.: Executing your commands via motion diffusion in latent space. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2023)", + "103. Xiong, W., Liu, J., Molybog, I., Zhang, H., Bhargava, P., Hou, R., Martin, L., Rungta, R., Sankararaman, K.A., Oguz, B., et al.: Effective long-context scaling of foundation models. arXiv preprint arXiv:2309.16039 (2023)", + "104. Xu, H., Ghosh, G., Huang, P.Y., Okhonko, D., Aghajanyan, A., Metze, F., Zettlemoyer, L., Feichtenhofer, C.: Videoclip: Contrastive pre-training for zero-shot video-text understanding. arXiv preprint arXiv:2109.14084 (2021)", + "105. Yamada, T., Matsunaga, H., Ogata, T.: Paired recurrent autoencoders for bidirectional translation between robot actions and linguistic descriptions. IEEE Robotics and Automation Letters 3(4), 3441-3448 (2018)", + "106. Yao, H., Song, Z., Zhou, Y., Ao, T., Chen, B., Liu, L.: Moconvq: Unified physics-based motion control via scalable discrete representations (2023)", + "107. Ye, Q., Xu, H., Xu, G., Ye, J., Yan, M., Zhou, Y., Wang, J., Hu, A., Shi, P., Shi, Y., et al.: mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178 (2023)", + "108. Yin, F., Chen, X., Zhang, C., Jiang, B., Zhao, Z., Fan, J., Yu, G., Li, T., Chen, T.: Shapeqpt: 3d shape generation with a unified multi-modal language model. arXiv preprint arXiv:2311.17618 (2023)", + "109. Yin, F., Liu, W., Huang, Z., Cheng, P., Chen, T., Yu, G.: Coordinates are not lonely-codebook prior helps implicit neural 3d representations. Advances in Neural Information Processing Systems 35, 12705-12717 (2022)", + "110. Yu, J., Xu, Y., Koh, J.Y., Luong, T., Baid, G., Wang, Z., Vasudevan, V., Ku, A., Yang, Y., Ayan, B.K., et al.: Scaling autoregressive models for content-rich text-to-image generation. arXiv preprint arXiv:2206.10789 (2022)", + "111. Yuan, Y., Kitani, K.: Dlow: Diversifying latent flows for diverse human motion prediction. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part IX 16. pp. 346-364. Springer (2020)" + ], + "bbox": [ + 218, + 146, + 784, + 839 + ], + "page_idx": 20 + }, + { + "type": "header", + "text": "MotionChain", + "bbox": [ + 640, + 114, + 730, + 127 + ], + "page_idx": 20 + }, + { + "type": "page_number", + "text": "21", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 20 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "112. Yuan, Y., Song, J., Iqbal, U., Vahdat, A., Kautz, J.: PhysDiff: Physics-guided human motion diffusion model. In: IEEE International Conference on Computer Vision (ICCV) (October 2023)", + "113. Zhang, H., Li, X., Bing, L.: Video-llama: An instruction-tuned audio-visual language model for video understanding. arXiv preprint arXiv:2306.02858 (2023)", + "114. Zhang, H., Cao, J., Lu, G., Ouyang, W., Sun, Z.: Danet: Decompose-and-aggregate network for 3d human shape and pose estimation. In: Proceedings of the 27th ACM International Conference on Multimedia. pp. 935-944 (2019)", + "115. Zhang, H., Tian, Y., Zhang, Y., Li, M., An, L., Sun, Z., Liu, Y.: Pymaf-x: Towards well-aligned full-body model regression from monocular images. IEEE Transactions on Pattern Analysis and Machine Intelligence (2023)", + "116. Zhang, J., Zhang, Y., Cun, X., Huang, S., Zhang, Y., Zhao, H., Lu, H., Shen, X.: T2m-gpt: Generating human motion from textual descriptions with discrete representations. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023)", + "117. Zhang, M., Cai, Z., Pan, L., Hong, F., Guo, X., Yang, L., Liu, Z.: Motiondiffuse: Text-driven human motion generation with diffusion model. arXiv preprint arXiv:2208.15001 (2022)", + "118. Zhang, M., Guo, X., Pan, L., Cai, Z., Hong, F., Li, H., Yang, L., Liu, Z.: Remodiffuse: Retrieval-augmented motion diffusion model. arXiv preprint arXiv:2304.01116 (2023)", + "119. Zhang, Q., Song, J., Huang, X., Chen, Y., Liu, M.Y.: Diffcollage: Parallel generation of large content with diffusion models. In: 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 10188-10198. IEEE (2023)", + "120. Zhang, T., Kishore, V., Wu, F., Weinberger, K.Q., Artzi, Y.: Bertscore: Evaluating text generation with bert. arXiv preprint arXiv:1904.09675 (2019)", + "121. Zhang, Y., Black, M.J., Tang, S.: We are more than our joints: Predicting how 3d bodies move. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3372-3382 (2021)", + "122. Zheng, L., Chiang, W.L., Sheng, Y., Zhuang, S., Wu, Z., Zhuang, Y., Lin, Z., Li, Z., Li, D., Xing, E.P., Zhang, H., Gonzalez, J.E., Stoica, I.: Judging lvm-as-a-judge with mt-bench and chatbot arena (2023)", + "123. Zhou, C., Liu, P., Xu, P., Iyer, S., Sun, J., Mao, Y., Ma, X., Efrat, A., Yu, P., Yu, L., et al.: Lima: Less is more for alignment. Advances in Neural Information Processing Systems 36 (2024)", + "124. Zhu, D., Chen, J., Shen, X., Li, X., Elhoseiny, M.: Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592 (2023)", + "125. Zou, X., Dou, Z.Y., Yang, J., Gan, Z., Li, L., Li, C., Dai, X., Behl, H., Wang, J., Yuan, L., et al.: Generalized decoding for pixel, image, and language. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15116-15127 (2023)" + ], + "bbox": [ + 217, + 146, + 784, + 744 + ], + "page_idx": 21 + }, + { + "type": "page_number", + "text": "22", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 21 + }, + { + "type": "header", + "text": "Biao Jiang, Xin Chen et al.", + "bbox": [ + 271, + 114, + 454, + 127 + ], + "page_idx": 21 + } +] \ No newline at end of file diff --git a/2024/MotionChain_ Conversational Motion Controllers via Multimodal Prompts/24d4fe30-81b0-462c-a698-38fed56b9fc4_model.json b/2024/MotionChain_ Conversational Motion Controllers via Multimodal Prompts/24d4fe30-81b0-462c-a698-38fed56b9fc4_model.json new file mode 100644 index 0000000000000000000000000000000000000000..695acf3853f9dbac54274406141aabdad3a6b69b --- /dev/null +++ b/2024/MotionChain_ Conversational Motion Controllers via Multimodal Prompts/24d4fe30-81b0-462c-a698-38fed56b9fc4_model.json @@ -0,0 +1,3115 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.221, + 0.141, + 0.783, + 0.187 + ], + "angle": 0, + "content": "MotionChain: Conversational Motion Controllers via Multimodal Prompts" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.212, + 0.785, + 0.245 + ], + "angle": 0, + "content": "Biao Jiang\\(^{1,2\\star}\\), Xin Chen\\(^{2\\star\\star}\\), Chi Zhang\\(^{2}\\), Fukun Yin\\(^{1}\\), Zhuoyuan Li\\(^{1}\\), Gang Yu\\(^{2}\\), and Jiayuan Fan\\(^{1\\star\\star\\star}\\)" + }, + { + "type": "text", + "bbox": [ + 0.435, + 0.255, + 0.568, + 0.269 + ], + "angle": 0, + "content": "1 Fudan University" + }, + { + "type": "text", + "bbox": [ + 0.468, + 0.27, + 0.537, + 0.282 + ], + "angle": 0, + "content": "2 Tencent" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.321, + 0.744, + 0.599 + ], + "angle": 0, + "content": "Abstract. Recent advancements in language models have demonstrated their adeptness in conducting multi-turn dialogues and retaining conversational context. However, this proficiency remains largely unexplored in other multimodal generative models, particularly in human motion models. By integrating multi-turn conversations in controlling continuous virtual human movements, generative human motion models can achieve an intuitive and step-by-step process of human task execution for humanoid robotics, game agents, or other embodied systems. In this work, we present MotionChain, a conversational human motion controller that generates continuous and long-term human motion through multimodal prompts. Specifically, MotionChain consists of multi-modal tokenizers that transform various data types such as text, image, and motion, into discrete tokens, coupled with a Vision-Motion-aware Language model. By leveraging large-scale language, vision-language, and vision-motion data to assist motion-related generation tasks, MotionChain thus comprehends each instruction in multi-turn conversation and generates human motions followed by these prompts. Extensive experiments validate the efficacy of MotionChain, demonstrating state-of-the-art performance in conversational motion generation, as well as more intuitive manners of controlling and interacting with virtual humans." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.612, + 0.681, + 0.626 + ], + "angle": 0, + "content": "Keywords: 3D Motion \\(\\cdot\\) Motion Generation \\(\\cdot\\) Text-to-Motion" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.652, + 0.377, + 0.668 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.683, + 0.788, + 0.79 + ], + "angle": 0, + "content": "The success of large language models (LLMs) [60, 62, 93, 94, 122] has sparked significant interest in the development of multi-modal language models. These models aim to transfer instruction-following and zero-shot abilities to other modalities tasks, such as image-language models [1, 50, 107, 124], video-language models [1, 43, 44, 113], and 3D-language models [12, 26, 28, 108]. However, a comprehensive model that can perceive visual input and generate continuous motion through multi-turn conversations has not yet been developed. Such a" + }, + { + "type": "page_footnote", + "bbox": [ + 0.216, + 0.798, + 0.668, + 0.812 + ], + "angle": 0, + "content": "* Work done while Biao Jiang was a Research Intern with Tencent." + }, + { + "type": "page_footnote", + "bbox": [ + 0.217, + 0.813, + 0.321, + 0.826 + ], + "angle": 0, + "content": "\\*\\* Project lead." + }, + { + "type": "page_footnote", + "bbox": [ + 0.206, + 0.827, + 0.384, + 0.84 + ], + "angle": 0, + "content": "\\*\\*\\* Corresponding author." + }, + { + "type": "list", + "bbox": [ + 0.206, + 0.798, + 0.668, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.456, + 0.129 + ], + "angle": 0, + "content": "Biao Jiang, Xin Chen et al." + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.151, + 0.782, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.269, + 0.788, + 0.352 + ], + "angle": 0, + "content": "Fig. 1: MotionChain can interpret instructions from multi-turn conversations and generate human motions or textual answers based on text, motion, or image inputs. We provide the conversation results in image-conditioned motion generation (1st column), motion reasoning (second column), motion editing (third column), and motion translation (third column), with each subsequent turn informed by all previous conversations. Left-to-right represents the temporal order." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.385, + 0.785, + 0.414 + ], + "angle": 0, + "content": "multi-modal model would have wide-ranging applications in fields like humanoid robotics, virtual assistants, game agents and so on." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.416, + 0.788, + 0.642 + ], + "angle": 0, + "content": "Previous research on human motion has explored various tasks, including motion generation [23, 32, 66, 92, 102, 116], motion captioning [20, 24, 32], motion prediction [32, 56, 111, 121], and motion composition [3]. Recent works in text-to-motion [67, 92, 102, 117] have involved pre-trained language models [16, 73] for motion generation. For instance, TEMOS [67] employs BERT [16] text embeddings in an end-to-end transformer architecture, while MDM [92] and MLD [102] both utilize text embeddings from CLIP [73] during the conditional diffusion process. On the other hand, MotionCLIP [91] and TMR [68] focus on modeling the coupled relationship between motion and text description, and MotionGPT [32] introduces a motion-language model that represents human motion and language in one unified vocabulary. However, these above methods treat all tasks as a one-turn conditioned generation, lacking contextual understanding and multi-turn continuous generation abilities. Therefore, we construct a Vision-Motion language model, integrating multi-turn conversations and continuous human motions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Two crucial challenges need to be addressed in this conversational motion generation. The first challenge is to contextually generate human motion in a continuous manner, resembling the way real humans move. The second challenge is the scarcity of text-motion paired datasets compared to datasets with pairs of image-language [11, 81], image-pose [37, 58, 61, 64] and video-motion [4, 5, 8, 30, 96]. Fortunately, both human motion and language are sequential and can be continuously \"written\". Building upon this observation, we employ the general vision-language instruction-tuning approach [50, 123] to enable conversational motion generation and question-answering through multi-modal instructions. By integrating image, motion, and language data and encoding them into tokens, the relationship between these three modalities becomes more evident. Therefore, with the advent of large vision-motion and vision-language data, Vision-Motion-language pre-training can enhance the performance of motion-related tasks." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.641, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "MotionChain" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.788, + 0.433 + ], + "angle": 0, + "content": "In this study, we introduce MotionChain, a comprehensive framework that integrates vision, motion, and language. MotionChain leverages large-scale vision-language data, vision-motion data, and pre-trained language models' strong language generation abilities to assist in motion-related generation tasks. To enable MotionChain to comprehend and generate human-like motions, we first train a motion-specific vector quantized variational autoencoder (VQ-VAE) model. This model constructs a \"motion vocabulary\" similar to the English word vocabulary and converts raw motion data into a sequence of motion tokens. To incorporate vision inputs into MotionChain, we then introduce a specialized vision tokenizer that connects a pre-trained vision encoder to the language model. This tokenizer converts image data into visual tokens within the language-motion \"words\" embedding space. These tokens are then processed by a pre-trained language model [14,74,93,94], which learns the relationship between image, motion and language. To enable conversational generation, we construct a multi-modal motion conversation dataset based on the existing text-motion dataset [23] and vision-motion dataset [5]. We then train the language model using our multimodal conversation dataset to learn the correlation between the three modalities. Extensive experiments demonstrate that MotionChain achieves state-of-the-art performance in multiple motion-related tasks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.434, + 0.788, + 0.541 + ], + "angle": 0, + "content": "We summarize our contributions as follows: (1) We propose MotionChain, a unified vision-motion-language generative pre-trained model, which performs conversational generation tasks via multi-modal inputs with language models. (2) We introduce a motion composition technique, to generate 3D human motions following the temporal order of instructions. (3) We propose a multi-modal motion conversation benchmark, wherein MotionChain achieves competitive performance across diverse motion tasks." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.565, + 0.389, + 0.581 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.598, + 0.788, + 0.844 + ], + "angle": 0, + "content": "Human Motion Modeling. There have been numerous attempts to model the relationship between 3D human motion and multiple modalities including incomplete motion [56, 111, 121], action [3, 25, 40, 66, 99, 102], text [22-24, 32, 55, 67, 82, 92, 117-119], image [19, 21, 33, 114, 115] and video [5, 13, 19, 36, 75]. Text-to-motion is one of the most important motion generation tasks, due to the user-friendly and convenient language input. MDM [92], MotionDiffuse [117] and MLD [102] proposes a diffusion-based generative model [27, 77, 85] to generate motions conditioned on different inputs. TM2T [24] and T2M-GPT [116] investigate a generative framework based on VQ-VAE [76, 95] and generative transformer for motion generation. Motion completion task generates motion conditioning on partial motions, such as classical motion prediction [56, 111, 121] or motion in-between [92], which generates the intermediate motion while the first and last parts are fixed. TEACH [3] proposes a past-conditioned transformer model that generates motion from a sequence of actions autoregressively. Apart from motion generation, there is also work investigating other modalities of generation from motion. Two statistical models [90] and recurrent networks [70, 105]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.455, + 0.129 + ], + "angle": 0, + "content": "Biao Jiang, Xin Chen et al." + }, + { + "type": "image", + "bbox": [ + 0.229, + 0.146, + 0.785, + 0.319 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.332, + 0.788, + 0.404 + ], + "angle": 0, + "content": "Fig. 2: Method overview: MotionChain consists of a motion tokenizer \\(\\mathcal{V}_{\\mathcal{M}}\\) (Sec. 3.2), a vision tokenizer \\(\\mathcal{V}_{\\mathcal{I}}\\) (r Sec. 3.2) and a vision-motion-aware language model (Sec. 3.3). By leveraging motion tokens generated by \\(\\mathcal{V}_{\\mathcal{M}}\\), alongside visual language token embeddings projected by vision tokenizer \\(\\mathcal{V}_{\\mathcal{I}}\\), and text tokens by text tokenizer, MotionChian achieves a unified learning paradigm for both motion and linguistic data." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.432, + 0.785, + 0.522 + ], + "angle": 0, + "content": "are learned in mapping motions to language. TM2T [24] proposed a new motion representation that compresses motions into a short sequence of discrete variables and then uses a neural translation network to build mappings between two modalities. In contrast to the above methods limited to only several tasks, MotionGPT [32] treats human motion as a foreign language and leverages language understanding and zero-shot transfer abilities of pre-trained language models." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.524, + 0.788, + 0.689 + ], + "angle": 0, + "content": "Character Control and Animation. Character control involves generating interactive motion sequences based on user instruction signals. One kind of approach [38, 59, 79] is to construct a graph representing transitions between motion clips and plan motion using graph search. Considering the limitations of these graph-based approaches in coarse discreteness, alternative methods like frame blending and concatenation [41], low-dimensional latent space learning [42], motion matching [15] proposed for embedding the task in the feature and [88] do the similar thing through hierarchical setup. Although the control signals for motion control and character animation are different from the instructions in text-to-motion tasks, we still recognize textual commands of conventional human motion generation as a boost for intuitive character control." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.69, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Multi-Modal Language Models. In the field of computer vision, there has been a recent surge of interest in multi-modal models that can process text along with other modalities, including images [18, 29, 44, 104], audio [6, 80, 101], and 3D [12, 45, 54, 83, 108]. CLIP [73] is an example of such a model, which learns a semantic latent representation that connects images with corresponding language descriptions. While language models have achieved success in various tasks, the development of multi-modal language models capable of handling human motion is still limited. Existing works in computer vision can be broadly categorized into two classes. The first consists of end-to-end trained models explored separately for specific research topics. For example, tasks like vision-language navi" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.642, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "MotionChain" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.328, + 0.157, + 0.677, + 0.33 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.357, + 0.788, + 0.468 + ], + "angle": 0, + "content": "Fig. 3: Data collection overview: Our initial step in collecting the motion reasoning data involves the utilization of human motion captions derived from an existing text-motion dataset. Subsequent to this, the text-motion retrieval model TMR [68] aids in the segmentation of motion pairs into categories based on the similarity between them. With the assistance of ChatGPT, we proceed to craft motion editing task data that correspond to these categorized similarity levels. Incorporating both motion reasoning and editing single-turn tasks, as well as the extensive 14 tasks delineated in [32], we construct a rich multi-modal multi-turn conversation dataset." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.498, + 0.788, + 0.634 + ], + "angle": 0, + "content": "gation [2, 7] and Habitat [89] require embodied AI agents to follow natural language instructions and take actions to accomplish goals in visual environments. InstructPix2Pix [7] in image translation enables agents to edit images based on human instructions. The second involves systems that coordinate various models using approaches like LangChain or LLMs [51, 62, 103]. Examples of such systems include Visual ChatGPT [100], X-GPT [125], and MMREACT [110]. While these methods focus on building instruction-following agents, we aim to develop an end-to-end trained multimodal model that can perform conversational motion generation tasks via multi-modal inputs with language models." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.657, + 0.339, + 0.673 + ], + "angle": 0, + "content": "3 Methods" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.69, + 0.788, + 0.84 + ], + "angle": 0, + "content": "To leverage large language data, vision-language data, and vision-motion data for assisting motion-related tasks, we propose a motion-language-vision framework called MotionChain. The framework, as depicted in Fig. 2, consists of a multi-modal tokenizer that converts various types of data (text, image, and motion) into discrete tokens (Sec. 3.2), and a vision-motion-aware Language model that comprehends information from different modalities and generates corresponding answers based on input instructions (Sec. 3.3). Additionally, to simultaneously understand data from multiple modalities, we employ a multi-stage training strategy (Sec. 3.4) for the training of the multi-modal tokenizer and the motion-language-vision framework." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.455, + 0.129 + ], + "angle": 0, + "content": "Biao Jiang, Xin Chen et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.789, + 0.404 + ], + "angle": 0, + "content": "We first introduced the multi-modal tokenizer, which comprises three branches for processing textual, image, and motion inputs. For textual inputs \\( w^{1:N} = \\{w^i\\} \\) of length \\( N \\) that describes a motion-related question or demand, we employ the SentencePiece model [39] used in previous works [14, 74, 93, 94], which has a vocabulary size of \\( K_{t} \\) and is trained on a large number of language datasets. The motion branch consists of a motion encoder \\( \\mathcal{E}_{\\mathcal{M}} \\) that encodes a motion sequence \\( m^{1:M} = \\{x^i\\} \\) of \\( M \\) frames into \\( L \\) motion tokens \\( z^{1:L} = \\{z^i\\} \\), where \\( L = M / l \\) and \\( l \\) represents the temporal downsampling rate on motion frames. It also includes a motion decoder \\( \\mathcal{D}_{\\mathcal{M}} \\) that can decode motion tokens back to human motion \\( \\hat{m}^{1:M} \\). The vision branch processes the input image \\( X \\) with a pre-trained CLIP visual encoder and a learnable linear projection that follows it, converted into language token embeddings \\( H_{q} \\). Given a textual sentence \\( w^{1:N} \\), a sequence of motion \\( m^{1:M} \\), and an image condition \\( X \\), all encoded as language tokens, our vision-motion-aware language model is designed to produce an answer comprising \\( L \\) tokens, denoted as \\( \\hat{x}^{1:L} = \\{\\hat{x}^i\\} \\). These output tokens can represent either motion sequences \\( \\hat{x}_{m}^{1:L} \\) or textual descriptions \\( \\hat{x}_{t}^{1:L} \\), which integrate both human motion \\( \\hat{m}^{1:M} \\), and text \\( \\hat{w}^{1:L} \\) within the given context." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.431, + 0.393, + 0.445 + ], + "angle": 0, + "content": "3.1 Data Collection" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.462, + 0.789, + 0.718 + ], + "angle": 0, + "content": "With the emergence of text-conditioned motion generation tasks, datasets like KIT [69], BABEL [71], HumanML3D [23] and the more recent Motion-X [49] have been developed. However, these datasets predominantly offer text labels as simple action phrases or captions. Building upon these foundations, MotionGPT [32] introduces an instruction-based motion-language dataset that encapsulates 14 core tasks, including motion prediction, translation, and editing, through thousands of instruction templates in a unified format. Despite this advancement, MotionGPT's data lack a deep engagement with the nuances of human motion analysis and are limited to single-turn generation tasks without incorporating contextual memory. Inspired by the recent success of GPT models across text-annotation tasks [17], image-annotation tasks [50], 3D-annotation tasks [28], we propose a data collection methodology integrates the capabilities of existing LLMs like ChatGPT [62], with the text-motion retrieval model TMR [68] to facilitate motion conversation data collection. In addition to the 14 motion-related tasks in MotionGPT [32], we introduce tasks centered around motion reasoning and motion editing, leveraging contextual insights for a deeper motion analysis." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.789, + 0.84 + ], + "angle": 0, + "content": "Utilizing ChatGPT [62], we initiate the collection of motion reasoning data using human motion captions from the text-motion dataset [23], starting with manually designed example queries that explore the contextual scenarios surrounding motions, possible preceding or succeeding actions, the subjects' roles, and the tools or equipment involved, etc. Following this, we employ TMR [68] for categorizing motions from the dataset into varying similarity levels. For medium-similarity motion pairs, we utilize ChatGPT [62] to generate motion editing directives that enable the transformation of one motion to another. For motions of" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.641, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "MotionChain" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "high similarity, we manually devise tasks aimed at editing their lengths, further enriching the dataset's versatility and analytical scope." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.785, + 0.253 + ], + "angle": 0, + "content": "After the collection of single-turn generation tasks, we progress to develop multi-turn conversation data. This involves the deliberate association of initial motion generation tasks with a variety of follow-up tasks randomly chosen among motion translation, reasoning, editing, etc. Following [122] we construct our conversation data in a structured format, as depicted below:" + }, + { + "type": "text", + "bbox": [ + 0.245, + 0.263, + 0.344, + 0.275 + ], + "angle": 0, + "content": "\\(X_{\\mathrm{system - message}}\\)" + }, + { + "type": "text", + "bbox": [ + 0.245, + 0.275, + 0.504, + 0.289 + ], + "angle": 0, + "content": "USER: \\(X_{v}\\) \\(\\dot{X}_s^1\\) ASSISTANT: \\(X_{a}^{1} < / s>\\)" + }, + { + "type": "text", + "bbox": [ + 0.246, + 0.289, + 0.504, + 0.303 + ], + "angle": 0, + "content": "USER: \\(X_{v} X_{s}^{2}\\) ASSISTANT: \\(X_{a}^{2} < / \\mathrm{s}>\\)" + }, + { + "type": "text", + "bbox": [ + 0.246, + 0.303, + 0.522, + 0.316 + ], + "angle": 0, + "content": "USER: \\(X_{v} X_{s}^{3}\\) ASSISTANT: \\(X_{a}^{3} < / \\mathrm{s}>\\) ..." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.327, + 0.784, + 0.478 + ], + "angle": 0, + "content": "Where \\( X_{v} \\) is defined as the vision language token embeddings, processed via the visual tokenizer. \\( X_{s}^{i} \\) and \\( X_{a}^{i} \\) are used to denote the source inputs and target answers for each round \\( i \\), respectively. Both sets of tokens originate from the integrated motion-language vocabulary \\( V \\), which includes motions, texts, or a blend thereof. The dataset exhibits variability in the number of generation turns up to 10; for the sake of clarity, we present only three examples herein. MotionChain is trained to predict answers, incorporating a learning mechanism that determines whether to stop generation by outputting end of sentence flag \\( < / \\mathrm{s} > \\) based on the current instruction and all preceding questions and answers. In the computation of the loss, as defined in Eq. (5), only the green tokens are utilized." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.498, + 0.451, + 0.513 + ], + "angle": 0, + "content": "3.2 Multi-modal Tokenizer" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.522, + 0.785, + 0.703 + ], + "angle": 0, + "content": "Motion tokenizer, denoted as \\(\\mathcal{V}_{\\mathcal{M}}\\), is based on the architecture of Vector Quantized Variational Autoencoders (VQ-VAE) utilized in previous studies [22,24,32, 84,95,98,106,109,116]. Once pre-trained, it can represent motion using discrete tokens, facilitating the integration of motion and language. The Motion tokenizer consists of a motion encoder \\(\\mathcal{E}_{\\mathcal{M}}\\) and a motion decoder \\(\\mathcal{D}_{\\mathcal{M}}\\). Initially, the motion encoder \\(\\mathcal{E}\\) applies 1D convolutions to the motion features \\(m^{1:M}\\) along the temporal dimension to obtain latent vectors \\(\\hat{z}^{1:L} = \\mathcal{E}_{\\mathcal{M}}(m^{1:M})\\). Subsequently, the latent vectors \\(\\hat{z}\\) are quantized and transformed into a collection of codebook entries \\(z\\). The learnable codebook \\(Z = \\{z_{i}\\}_{i=1}^{K} \\subset \\mathbb{R}^{d}\\) comprises \\(K\\) latent embedding vectors, each with a dimension of \\(d\\). The quantization process \\(Q(\\cdot)\\) replaces each row vector \\(\\hat{z}\\) with its nearest codebook entry \\(z_{k}\\) in \\(Z\\), which can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.368, + 0.719, + 0.785, + 0.736 + ], + "angle": 0, + "content": "\\[\nz _ {i} = Q \\left(\\hat {z} ^ {i}\\right) := \\arg \\min _ {z _ {k} \\in Z} \\| \\hat {z} _ {i} - z _ {k} \\| _ {2}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.739, + 0.784, + 0.815 + ], + "angle": 0, + "content": "We assign \\( s^i \\) as the index number of motion tokens \\( z^{1:L} \\), so motion tokens \\( z^{1:L} \\) can be represented as a sequence of indices \\( s^{1:L} = \\{s^i\\}_{i=1}^L \\). The motion decoder \\( \\mathcal{D}_{\\mathcal{M}} \\) can project \\( z^{1:L} = \\{z^i\\}_{i=1}^L \\) back to the raw motion space, resulting in the motion \\( \\hat{m}^{1:M} \\) with \\( M \\) frames. Following [22,24,32,98,116], we adopt three distinct loss functions when training the motion tokenizer:" + }, + { + "type": "equation", + "bbox": [ + 0.433, + 0.826, + 0.785, + 0.841 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathcal {V}} = \\mathcal {L} _ {r} + \\mathcal {L} _ {e} + \\mathcal {L} _ {c} \\tag {2}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.455, + 0.129 + ], + "angle": 0, + "content": "Biao Jiang, Xin Chen et al." + }, + { + "type": "image", + "bbox": [ + 0.229, + 0.142, + 0.401, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.265, + 0.24, + 0.362, + 0.252 + ], + "angle": 0, + "content": "(a) Independent" + }, + { + "type": "image", + "bbox": [ + 0.417, + 0.142, + 0.588, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.447, + 0.24, + 0.557, + 0.252 + ], + "angle": 0, + "content": "(b) Past-condition" + }, + { + "type": "image", + "bbox": [ + 0.604, + 0.142, + 0.777, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.643, + 0.24, + 0.739, + 0.252 + ], + "angle": 0, + "content": "(c) Tokens-joint" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.263, + 0.787, + 0.334 + ], + "angle": 0, + "content": "Fig. 4: Motion Composition Variants: We illustrate the baselines for motion composition during multi-turn motion generation (a). independent decoding each turn (b). separate decoding conditioned on the last few tokens from the prior turn (c). decoding with joint motion tokens. Green tokens stand for image condition, blue tokens stand for textual instruction, and orange tokens stand for human motions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.364, + 0.785, + 0.392 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_r\\) denotes reconstruction loss, \\(\\mathcal{L}_e\\) denotes the embedding loss, and \\(\\mathcal{L}_c\\) denotes commitment loss." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.395, + 0.788, + 0.504 + ], + "angle": 0, + "content": "During multi-turn motion generation, the motion continuity between turns is achieved through our motion decoder, which links the motion of the current turn with that of the preceding ones. Taking the composition of two motions as an example: we concatenate the past motion tokens, denoted as \\( z_{p}^{1:L_{p}} \\), with the tokens representing the current motion, \\( z_{c}^{1:L_{c}} \\). This concatenated sequence of tokens is subsequently decoded into a comprehensive set of continuous motion features, represented as \\( m_{\\mathrm{whole}}^{1:M_{\\mathrm{whole}}} \\), as depicted below:" + }, + { + "type": "equation", + "bbox": [ + 0.407, + 0.515, + 0.786, + 0.537 + ], + "angle": 0, + "content": "\\[\nz _ {\\text {w h o l e}} ^ {1: \\left(L _ {p} + L _ {c}\\right)} = \\left[ z _ {p} ^ {1: L _ {p}}, z _ {c} ^ {1: L _ {c}} \\right]. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.547, + 0.785, + 0.592 + ], + "angle": 0, + "content": "Similarly, this framework is adept at executing composition tasks involving an array of motions. The comparison results in Tab. 3 demonstrate that our motion tokenizer could effectively perform motion composition tasks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.593, + 0.787, + 0.669 + ], + "angle": 0, + "content": "Visual Tokenizer accepts images \\( X_{I} \\) as inputs. We employ the CLIP visual encoder that is pre-trained on image-text pairs to derive visual feature \\( Z_{I} \\). These features are then projected into language token embeddings \\( X_{v} \\) via a linear layer like previous work [50], maintaining consistency in the dimensionality with the language model's word embedding space." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.693, + 0.519, + 0.708 + ], + "angle": 0, + "content": "3.3 Motion-aware Language Model" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.719, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Language models such as Llama [93, 94] and T5 [14, 74] employ the Sentence-Piece [39] model to encode textual inputs into WordPiece tokens, utilizing a \\( K_{t} \\) word piece vocabulary. Unlike prior text-to-motion [24, 102, 116, 118] and motion-to-text [24] methods that process text and motion separately, we merge the text vocabulary \\( V_{t} = \\{v_{t}^{i}\\}_{i=1}^{K_{t}} \\) with the motion vocabulary \\( V_{m} = \\{v_{m}^{i}\\}_{i=1}^{K_{m}} \\), maintaining the motion tokenizer's codebook \\( Z \\) order and including special tokens for boundary demarcation. This creates a unified vocabulary \\( V = \\{V_{t}, V_{m}\\} \\), enabling the formulation of motion-centric tasks in a universal template, where inputs and" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.642, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "MotionChain" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.193 + ], + "angle": 0, + "content": "outputs share the same vocabulary. For visual input, our visual tokenizer converts images into visual token embeddings \\( X_{v} \\), aligning with the language model [39, 62, 74, 93] token space for integrated representation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.193, + 0.785, + 0.386 + ], + "angle": 0, + "content": "For single conditioned generation tasks, our input comprises a sequence of \\(N\\) length tokens \\(X_{s} = \\{x_{s}^{i}\\}_{i = 1}^{N}\\), where \\(x_{s}\\in \\{V_{t},V_{m}\\}\\) representing either text, motion, or a combination thereof, drawn from the unified vocabularies. In cases involving image inputs, visual tokens \\(X_{v}\\) are interspersed at the beginning of the source tokens sequence, forming \\([X_v,X_s]\\). Subsequent interaction rounds generate target answer tokens \\(X_{a}\\). To facilitate iterative result generation and content retention, our framework generates multi-turn conversation data \\((X_{v},X_{s}^{1},X_{a}^{1},X_{s}^{2},X_{a}^{2},\\dots ,X_{s}^{T},X_{a}^{T})\\), with \\(T\\) indicating the total turn count. Notably, visual tokens are consistently placed at the forefront of the initial turn's source tokens. The processing sequence is organized such that to predict target answer tokens autoregressively, as shown in Fig. 2. Source tokens are processed by the transformer to predict the next token's probability distribution, formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.333, + 0.387, + 0.785, + 0.416 + ], + "angle": 0, + "content": "\\[\np _ {\\theta} \\left(X _ {a} \\mid X _ {v}, X _ {s}\\right) = \\prod_ {i} p _ {\\theta} \\left(x _ {a} ^ {i} \\mid X _ {v}, X _ {s, < i}, X _ {a, < i}\\right) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.422, + 0.784, + 0.468 + ], + "angle": 0, + "content": "with \\(\\theta\\) indicating trainable parameters, and \\(X_{s, < i}, X_{a, < i}\\) representing the sequences of source and preceding target tokens. The training objective is maximizing the log-likelihood of distribution:" + }, + { + "type": "equation", + "bbox": [ + 0.344, + 0.48, + 0.785, + 0.52 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {L M} = - \\sum_ {i = 0} ^ {L _ {t} - 1} \\log p _ {\\theta} \\left(x _ {a} ^ {i} \\mid X _ {v}, X _ {s, < i}, X _ {a, < i}\\right). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.53, + 0.784, + 0.56 + ], + "angle": 0, + "content": "By optimizing this objective, MotionChain captures the complex interrelations among images, motion, and text, facilitating accurate target \"word\" generation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.561, + 0.785, + 0.642 + ], + "angle": 0, + "content": "During the inference phase, target tokens are recursively sampled from the model's predicted distribution \\( p_{\\theta}\\left(\\hat{x}_a^i \\mid X_v, X_s, \\hat{X}_{a, < i}\\right) \\), ceasing with the appearance of a special end token. This strategy facilitates a step-by-step target sequence generation, where each token's probability is conditioned on all previous turns' sources and targets and current source input." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.664, + 0.409, + 0.68 + ], + "angle": 0, + "content": "3.4 Training Strategy" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.69, + 0.785, + 0.841 + ], + "angle": 0, + "content": "To facilitate the integration of image and motion comprehension within the language modeling context, we adopt a 3-stage training strategy. (1) The initial stage involves pre-training the motion tokenizer on a corpus of human motion data, in line with [32]. This process establishes the motion vocabulary \\( V_{m} \\), which serves as a foundation for encoding human motions as a series of discrete tokens. (2) Subsequently, the motion tokenizer remains frozen while we connect the visual tokenizer to the language model framework. This integration is supported by a suite of supervised objectives, including text-to-motion, motion-to-text, and image-based motion generation, aiming to learn the intricate relationships between images, motion, and language. (3) The final stage involves instruction" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.455, + 0.129 + ], + "angle": 0, + "content": "Biao Jiang, Xin Chen et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.193 + ], + "angle": 0, + "content": "tuning, and refines the model's capabilities through the application of prompt-based instructions. These instructions are framed within multi-turn conversation sequences, as detailed in Sec. 3.3, to expanded range of motion-related tasks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.196, + 0.788, + 0.272 + ], + "angle": 0, + "content": "Training of Motion Tokenizer. The initial step involves training the motion tokenizer, guided by the loss objective in Equation 2. This stage enables the tokenizer to represent human motion sequences \\(\\hat{x}^{1:L}\\) as discrete motion tokens, a key step for merging motion data with textual information seamlessly. Once optimized, the motion tokenizer remains frozen." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.276, + 0.788, + 0.502 + ], + "angle": 0, + "content": "Motion-language Pre-training Stage. Leveraging recent developments in language modeling [14, 74, 93, 94, 122] pre-trained on natural language datasets and then fine-tuned with instruction-based phrasing [14, 62]. To augment the model's ability to discern relationships between images and human motions, we first pre-train our MotionChain using a mix of language, image, and motion datasets. Following the stage 1 training of the motion tokenizer, we have established a unified motion-language vocabulary \\( V = V_{t}, V_{m} \\), capable of representing motions in discrete token form. Moreover, we maintain the visual encoder's weights in the visual tokenizer as fixed, while the linear projection weight \\( W \\) is jointly optimized with the language model. During this stage, the model undertakes three fundamental single-turn modality translation tasks: text-to-motion, motion-to-text, and image-conditioned motion generation, as outlined in Sec. 3.1. The primary objective is to maximize the likelihood of the model according to the loss function specified in Eq. (5), thereby letting the model understand the relationship between language, vision conditions, and motions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.506, + 0.788, + 0.643 + ], + "angle": 0, + "content": "Instruction Tuning Stage. As described in Sec. 3.1, we construct a multimodal, multi-task, and multi-turn motion conversation dataset by augmenting existing text-to-motion [23] and human mesh reconstruction datasets [5] with targeted instructional prompts and leverage the capabilities of LLMs [62] and the text-motion retrieval model [68] for motion reasoning and editing tasks. The efficacy of instruction tuning, as evidenced across language models [14, 50, 62, 122], is well-established, yielding enhancements in model performance across a wide range of tasks. After instruction tuning, MotionChain can handle more motion-related tasks including the proficient handling of previously unseen tasks" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.684, + 0.377, + 0.701 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We evaluate the proposed MotionChain encompasses comprehensive comparisons across both one-turn motion-related tasks and multi-turn motion generation tasks. Firstly, we provide details of the dataset settings, evaluation criteria, and implementation details as specified in Sec. 4.1. Subsequently, comparative analyses are presented, focusing on the motion reasoning task (Sec. 4.2) and the temporal motion composition task (Sec. 4.3). In Sec. 4.4, we evaluate the choice of motion composition technique and different architectures of vision tokenizer." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.641, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "MotionChain" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.428, + 0.163 + ], + "angle": 0, + "content": "4.1 Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.175, + 0.788, + 0.401 + ], + "angle": 0, + "content": "Datasets. For one-turn motion reasoning tasks, the study employs our proposed multi-modal multi-turn conversation dataset upon HumanML3D [23] with 44,970 sequence-level textual descriptions for 14,616 motion sequences obtained from AMASS [57] and HumanAct12 [25]. The datasets are divided into training, testing, and validation sets with a ratio of \\(0.8:0.15:0.05\\). To evaluate the multi-turn motion generation task, we focus on BABEL [71] that provides textual descriptions for the motions in the AMASS [57] with annotated segments that overlap in each sequence, which allows evaluating generation of a sequence of motion or actions. We adopt the processed text labels by [3] and motion representation of HumanML3D [23] which combines joint velocities, positions, and rotations. Following [3] we consider pairs of actions for simplicity but Motion-Chain applies to a sequence of actions or motion of arbitrary length. For the image-conditioned motion generation task, we mainly focus on BEDLAM [5], a large synthetic dataset of realistic moving 3D humans containing more than 200 subjects and 380K frames video and motion pair." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.402, + 0.788, + 0.658 + ], + "angle": 0, + "content": "Evaluation Metrics are summarized as four parts. (1) Motion quality: We adopt Frechet Inception Distance (FID) as the primary metric, FID quantifies the divergence in feature distributions between generated and actual motion sequences. Utilizing feature extractors from prior studies [23,49,68], FID measures the distance of feature distributions between the generated and real motions. Following [5,32,33,65,102], we also adopt MPJPE, PA-MPJPE to measure global and local errors in millimeters and ACCL for acceleration errors, to evaluate the quality of the reconstructed motions. (2) Motion Diversity: Utilizing the Diversity (DIV) metric, we calculate variance across motion features to evaluate generation diversity. (3) Text matching: The precision of text-to-motion matches is quantified by the R Precision metric, based on the feature evaluator [23,49,68], and includes an analysis of Top 1/2/3 retrieval accuracy. The Multi-modal Distance (MM Dist) quantifies the semantic gap between motions and texts. (4) Linguistic quality: We follow [24] utilizing linguistic metrics from natural language studies, including BLUE [63], Rouge [48], Cider [97], and BertScore [120] to evaluate the quality of generated motion captions. More detailed benchmark information is provided in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.66, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Implementation Details. We set the codebook of the motion tokenizer as \\( K \\in \\mathbb{R}^{512 \\times 1024} \\) for most experiments. The motion encoder, denoted as \\( \\mathcal{E}_{\\mathcal{M}} \\), integrates a temporal downsampling rate, \\( l = 4 \\). Our vision tokenizer incorporates a frozen Vision Transformer (ViT-L/14) [73] as visual encoder for most experiments. Additionally, for comprehensive ablation studies, we explored the use of both a frozen vision encoder and a Q-former from BLIP-2 [43] as a vision tokenizer. We mainly utilize Flan-T5-base [14] as the underlying architecture for our language model. Moreover, all our models employ the AdamW [53] optimizer with \\( [\\beta_1, \\beta_2] = [0.9, 0.99] \\) for training. The motion tokenizers are trained to utilize a \\( 10^{-4} \\) learning rate employing cosine annealing scheduler and a 256 mini-batch size. Our language models based on Flan-T5-base [14] have a \\( 10^{-4} \\) learning rate with cosine annealing scheduler and 16 mini-batch sizes in both the" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.455, + 0.129 + ], + "angle": 0, + "content": "Biao Jiang, Xin Chen et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.145, + 0.787, + 0.215 + ], + "angle": 0, + "content": "Table 1: Comparison of motion reasoning on the test set of our conversation dataset. Our proposed MotionChain is fine-tuned on motion reasoning tasks while other methods' results are generated by their pre-trained weight. \\(Length_{avg}\\) represents the average words in generated answers to all questions. We adopt metrics commonly used in natural language processing tasks for evaluation." + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.227, + 0.787, + 0.362 + ], + "angle": 0, + "content": "
MethodsParamsLengthavgBleu@1↑Bleu@4↑Rouge↑Cider↑BertScore↑
Flan-t5-base [14]250M8.344.641.7815.3215.933.45
Flan-t5-large [14]780M11.9512.184.8322.8115.0214.19
Flan-t5-xl [14]3B9.098.544.0124.8915.0318.34
Llama-2-7b [94]7B130.8411.123.6719.141.046.81
Vicuna-1.5-7b [122]7B71.4919.277.3925.755.4419.05
Vicuna-1.5-13b [122]13B84.7417.206.5324.187.7718.00
MotionChain (Ours)573M22.1737.9219.1938.0524.5332.24
" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.388, + 0.787, + 0.449 + ], + "angle": 0, + "content": "pre-train stage and the instruction tuning stage. The motion tokenizer undergoes 10000 epochs of training, while the language model undergoes 500 epochs during the pre-train stage and another 50 epochs during the instruction tuning stage. Most models are trained on 8 Tesla V100 GPUs." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.469, + 0.555, + 0.485 + ], + "angle": 0, + "content": "4.2 Comparisons on Motion Reasoning." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.493, + 0.787, + 0.629 + ], + "angle": 0, + "content": "In Sec. 3.1, we introduce a multi-modal motion conversation dataset, enriched with motion reasoning data facilitated by ChatGPT [62]. This task evaluates the model's reasoning capabilities with motion reasoning tasks, where a motion sequence or its corresponding textual descriptions serve as inputs. Our evaluation compares our MotionChain, which integrates motion perception, against contemporary Large Language Models (LLMs) that possess solely textual processing capabilities. The compared LLMs are assessed using their original pretrained weight. Results in Tab. 1, illustrate that MotionChain exhibits superior motion reasoning proficiency, benefiting from its integrated motion perception." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.65, + 0.593, + 0.666 + ], + "angle": 0, + "content": "4.3 Comparisons on Temporal Composition." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.674, + 0.787, + 0.84 + ], + "angle": 0, + "content": "The temporal motion composition task involves generating a continuous motion sequence from two actions in a time series. We conducted our experiments following the settings in TEACH [3] and used the Amass [57] subset BABEL [71] validation set. Additionally, we processed the motion in AMASS [57] into the format proposed by HumanML3D [23] and trained our MotionChain on the action-to-motion task. To compare with TEACH, we initially used an officially provided pre-trained model to sample motion on the validation set 20 times. Subsequently, we post-processed their motion into the HumanML3D format, represented in SMPL [52]. The performance of our MotionChain is summarized in Table 2. As evaluating generative models quantitatively is challenging, we also provide qualitative comparisons in the supplementary materials." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.641, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "MotionChain" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.189 + ], + "angle": 0, + "content": "Table 2: Comparison of temporal motion composition on Babel [71]. We evaluate the state-of-the-art motion temporal composition method Teach [3] under the \\(95\\%\\) confidence interval from 20 times running. (cf. Sec. 4.1 for notations.)" + }, + { + "type": "table", + "bbox": [ + 0.271, + 0.199, + 0.734, + 0.279 + ], + "angle": 0, + "content": "
MethodsDiversityMPJPE↓PA-MPJPE↓ACCL↓
Real15.74±.149---
Teach [3]27.11±.159979.21±.215933.32±.25423.02±.018
MotionChain (Ours)43.25±.159276.05±6.7253.72±.5807.11±0.100
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.292, + 0.789, + 0.336 + ], + "angle": 0, + "content": "Table 3: Evaluation of motion composition methods on HumanML3D [23]. Here Independent, Past-condition, and Tokens-joint stand for different motion composition variants during multi-turn motion conversation, as illustrated in Fig. 4." + }, + { + "type": "table", + "bbox": [ + 0.314, + 0.346, + 0.691, + 0.42 + ], + "angle": 0, + "content": "
MethodMPJPE↓PA-MPJPE↓ACCL↓Diversity
Independent350.79102.9711.406.47
Past-condition232.4646.156.186.01
Tokens-joint108.7718.852.265.56
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.448, + 0.403, + 0.463 + ], + "angle": 0, + "content": "4.4 Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.477, + 0.788, + 0.597 + ], + "angle": 0, + "content": "MotionChain enables multi-modal motion conversation using two main techniques. The first technique involves generating a smooth sequence of motions by concatenating motion tokens which are then decoded back to motion by motion decoder \\(\\mathcal{D}_{\\mathcal{M}}\\). The second technique involves processing multi-modal visual input through a vision tokenizer, which consists of a frozen vision encoder and a trainable linear projection. To evaluate the effectiveness of these two designs, we compare them with other variants. For a more comprehensive analysis, detailed ablation studies can be found in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.788, + 0.718 + ], + "angle": 0, + "content": "Motion Composition Mechanism Apart from the jointly token concatenating mechanism, we also evaluate the performance of temporal motion composition through the other motion temporal composition variants Motion-cat: concatenating the motion in final motion level rather than token level. Experimental results in Tab. 3 show that jointly concatenating motion tokens achieved remarkable performance compared to the other variants. For further information regarding the implementation of the aforementioned vision tokenizer, please refer to the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Image Tokenizer Architecture. MotionChain connects the frozen vision encoder to the language model through a linear layer. However, previous vision-language [1,43] works also demonstrate the effectiveness of other kinds of visual-aligning modules. Here we consider the other two vision tokenizer variants: (a) inspired by [1,10,31], we introduce a perceiver module that incorporates a transformer receiving a predefined number of latent input queries. These queries cross-attend to the visual features, enabling effective information exchange. (b) We directly adopt the pre-trained Q-former from BLIP-2 [43] to align visual" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.455, + 0.129 + ], + "angle": 0, + "content": "Biao Jiang, Xin Chen et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.203 + ], + "angle": 0, + "content": "Table 4: Evaluation of vision tokenizer architecture on Bedlam [49]. We implement three different architectures, including Q-former, Perceiver, and Linear. We evaluate these results with the metrics in motion reconstruction. Additional information regarding the implementation is in the supplementary materials. (cf. Tab. 2 for notations.)" + }, + { + "type": "table", + "bbox": [ + 0.293, + 0.213, + 0.71, + 0.307 + ], + "angle": 0, + "content": "
ArchitectureFirst-frameLast-frame
MPJPE ↓PA-MPJPE ↓MPJPE ↓PA-MPJPE ↓
Q-former195.4986.56134.7357.17
Perceiver185.6199.21134.8957.58
Linear144.3776.48133.7356.73
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.339, + 0.788, + 0.431 + ], + "angle": 0, + "content": "inputs with the language model. We evaluate the different architectures under the single human image as the first frame condition and the last frame condition separately. Experimental results in Tab. 4 show that a lightweight linear projection is sufficient for comprehending the human pose from visual input. Additional details about the implementation of the above vision tokenizer can be found in the supplements." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.463, + 0.51, + 0.48 + ], + "angle": 0, + "content": "5 Conclusion and Limitation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.506, + 0.788, + 0.657 + ], + "angle": 0, + "content": "Limitation. As the trial to explore conversational human motion generation with visual language models, the proposed MotionChain still has limitations as follows. MotionChain utilizes indeterministic generative models, similar to other language models, but other traditional or neural motion controllers [86, 87] are mostly deterministic and sensitive to control signals. Besides, our method can only generate motion on articulated human bodies, excluding many other human parts such as faces [9, 34, 72] and hands [46, 46, 47, 47, 78]. Although we utilize vision, language, and motion as multimodal conditional inputs akin to human perception, MotionChain is still restricted to the collision signals for human-object and human-scene interactions [35, 82, 112]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Conclusion. We summarize the proposed MotionChain as a conversational human motion controller to generate continuous and long-term human motion through multimodal prompts. Compared to these one-turn motion generation methods [32, 92, 117], our MotionChain produces more contextually rich generation and can achieve the step-by-step process of human task execution for humanoid robotics and game agents. By leveraging large-scale language, vision-language, and vision-motion data to assist motion-related generation tasks, MotionChain thus comprehends each instruction in multi-turn conversation and generates human motions followed by these prompts. Extensive experiments validate the efficacy of MotionChain, demonstrating state-of-the-art performance in conversational motion generation, as well as more intuitive manners of controlling and interacting with virtual humans." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.641, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "MotionChain" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.175, + 0.808, + 0.266 + ], + "angle": 0, + "content": "This work is supported by National Natural Science Foundation of China (No. 62101137), National Key Research and Development Program of China (No. 2022ZD0160101), Shanghai Natural Science Foundation (No. 23ZR1402900), Shanghai Municipal Science and Technology Major Project (No.2021SHZDZX0103). The computations in this research were performed using the CFFF platform of Fudan University." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.288, + 0.323, + 0.304 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.232, + 0.317, + 0.787, + 0.373 + ], + "angle": 0, + "content": "1. Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., et al.: Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems 35, 23716-23736 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.232, + 0.373, + 0.787, + 0.442 + ], + "angle": 0, + "content": "2. Anderson, P., Wu, Q., Teney, D., Bruce, J., Johnson, M., Sünderhauf, N., Reid, I., Gould, S., Van Den Hengel, A.: Vision-and-language navigation: Interpreting visually-grounded navigation instructions in real environments. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3674-3683 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.232, + 0.443, + 0.787, + 0.483 + ], + "angle": 0, + "content": "3. Athanasiou, N., Petrovich, M., Black, M.J., Varol, G.: Teach: Temporal action compositions for 3d humans. In: International Conference on 3D Vision (3DV) (September 2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.232, + 0.484, + 0.787, + 0.525 + ], + "angle": 0, + "content": "4. Bazavan, E.G., Zanfir, A., Zanfir, M., Freeman, W.T., Sukthankar, R., Sminchisescu, C.: Hspace: Synthetic parametric humans animated in complex environments. arXiv preprint arXiv:2112.12867 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.232, + 0.525, + 0.787, + 0.567 + ], + "angle": 0, + "content": "5. Black, M.J., Patel, P., Tesch, J., Yang, J.: Bedlam: A synthetic dataset of bodies exhibiting detailed lifelike animated motion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8726-8737 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.232, + 0.567, + 0.787, + 0.621 + ], + "angle": 0, + "content": "6. Borsos, Z., Marinier, R., Vincent, D., Kharitonov, E., Pietquin, O., Sharifi, M., Roblek, D., Teboul, O., Grangier, D., Tagliasacchi, M., et al.: Audiolm: a language modeling approach to audio generation. IEEE/ACM transactions on audio, speech, and language processing 31, 2523-2533 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.232, + 0.621, + 0.787, + 0.662 + ], + "angle": 0, + "content": "7. Brooks, T., Holynski, A., Efros, A.A.: Instructpix2pix: Learning to follow image editing instructions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18392-18402 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.232, + 0.662, + 0.787, + 0.703 + ], + "angle": 0, + "content": "8. Cai, Z., Zhang, M., Ren, J., Wei, C., Ren, D., Lin, Z., Zhao, H., Yang, L., Loy, C.C., Liu, Z.: Playing for 3d human recovery. arXiv preprint arXiv:2110.07588 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.232, + 0.703, + 0.787, + 0.744 + ], + "angle": 0, + "content": "9. Cao, X., Chen, Z., Chen, A., Chen, X., Li, S., Yu, J.: Sparse photometric 3d face reconstruction guided by morphable models. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 4635-4644 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.744, + 0.787, + 0.786 + ], + "angle": 0, + "content": "10. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: European conference on computer vision. pp. 213-229. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.786, + 0.787, + 0.84 + ], + "angle": 0, + "content": "11. Changpinyo, S., Sharma, P., Ding, N., Soricut, R.: Conceptual 12m: Pushing web-scale image-text pre-training to recognize long-tail visual concepts. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3558-3568 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.317, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.455, + 0.128 + ], + "angle": 0, + "content": "Biao Jiang, Xin Chen et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.147, + 0.787, + 0.204 + ], + "angle": 0, + "content": "12. Chen, S., Chen, X., Zhang, C., Li, M., Yu, G., Fei, H., Zhu, H., Fan, J., Chen, T.: Ll3da: Visual interactive instruction tuning for omni-3d understanding reasoning and planning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 26428-26438 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.205, + 0.787, + 0.247 + ], + "angle": 0, + "content": "13. Choudhury, R., Kitani, K.M., Jeni, L.A.: Tempo: Efficient multi-view pose estimation, tracking, and forecasting. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 14750-14760 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.248, + 0.787, + 0.288 + ], + "angle": 0, + "content": "14. Chung, H.W., Hou, L., Longpre, S., Zoph, B., Tay, Y., Fedus, W., Li, E., Wang, X., Dehghani, M., Brahma, S., et al.: Scaling instruction-finetuned language models. arXiv preprint arXiv:2210.11416 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.289, + 0.787, + 0.317 + ], + "angle": 0, + "content": "15. Clavet, S.: Motion matching and the road to next-gen animation. In: Proc. of GDC. vol. 2, p. 9 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.318, + 0.787, + 0.359 + ], + "angle": 0, + "content": "16. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.36, + 0.787, + 0.388 + ], + "angle": 0, + "content": "17. Gilardi, F., Alizadeh, M., Kubli, M.: Chatgpt outperforms crowd-workers for text-annotation tasks. arXiv preprint arXiv:2303.15056 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.389, + 0.787, + 0.43 + ], + "angle": 0, + "content": "18. Girdhar, R., El-Nouby, A., Liu, Z., Singh, M., Alwala, K.V., Joulin, A., Misra, I.: Imagebind: One embedding space to bind them all. arXiv preprint arXiv:2305.05665 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.431, + 0.787, + 0.472 + ], + "angle": 0, + "content": "19. Goel, S., Pavlakos, G., Rajasegaran, J., Kanazawa, A., Malik, J.: Humans in 4d: Reconstructing and tracking humans with transformers. arXiv preprint arXiv:2305.20091 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.474, + 0.787, + 0.515 + ], + "angle": 0, + "content": "20. Goutsu, Y., Inamura, T.: Linguistic descriptions of human motion with generative adversarial seq2seq learning. In: 2021 IEEE International Conference on Robotics and Automation (ICRA). pp. 4281-4287. IEEE (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.516, + 0.787, + 0.558 + ], + "angle": 0, + "content": "21. Guler, R.A., Kokkinos, I.: Holopose: Holistic 3d human reconstruction in-the-wild. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10884-10894 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.559, + 0.787, + 0.586 + ], + "angle": 0, + "content": "22. Guo, C., Mu, Y., Javed, M.G., Wang, S., Cheng, L.: Momask: Generative masked modeling of 3d human motions. arXiv preprint arXiv:2312.00063 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.587, + 0.787, + 0.642 + ], + "angle": 0, + "content": "23. Guo, C., Zou, S., Zuo, X., Wang, S., Ji, W., Li, X., Cheng, L.: Generating diverse and natural 3d human motions from text. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5152-5161 (June 2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.643, + 0.787, + 0.671 + ], + "angle": 0, + "content": "24. Guo, C., Zuo, X., Wang, S., Cheng, L.: Tm2t: Stochastic and tokenized modeling for the reciprocal generation of 3d human motions and texts. In: ECCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.672, + 0.787, + 0.713 + ], + "angle": 0, + "content": "25. Guo, C., Zuo, X., Wang, S., Zou, S., Sun, Q., Deng, A., Gong, M., Cheng, L.: Action2motion: Conditioned generation of 3d human motions. In: Proceedings of the 28th ACM International Conference on Multimedia. pp. 2021-2029 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.714, + 0.787, + 0.769 + ], + "angle": 0, + "content": "26. Guo, Z., Zhang, R., Zhu, X., Tang, Y., Ma, X., Han, J., Chen, K., Gao, P., Li, X., Li, H., et al.: Point-bind & point-llm: Aligning point cloud with multimodality for 3d understanding, generation, and instruction following. arXiv preprint arXiv:2309.00615 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.77, + 0.787, + 0.798 + ], + "angle": 0, + "content": "27. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems 33, 6840-6851 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.799, + 0.787, + 0.84 + ], + "angle": 0, + "content": "28. Hong, Y., Zhen, H., Chen, P., Zheng, S., Du, Y., Chen, Z., Gan, C.: 3d-lm: Injecting the 3d world into large language models. arXiv preprint arXiv:2307.12981 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.642, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "MotionChain" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "29. Huang, S., Dong, L., Wang, W., Hao, Y., Singhal, S., Ma, S., Lv, T., Cui, L., Mohammed, O.K., Liu, Q., et al.: Language is not all you need: Aligning perception with language models. arXiv preprint arXiv:2302.14045 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.189, + 0.787, + 0.243 + ], + "angle": 0, + "content": "30. Ionescu, C., Papava, D., Olaru, V., Sminchisescu, C.: Human3. 6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. IEEE transactions on pattern analysis and machine intelligence 36(7), 1325-1339 (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.243, + 0.785, + 0.285 + ], + "angle": 0, + "content": "31. Jaegle, A., Gimeno, F., Brock, A., Vinyals, O., Zisserman, A., Carreira, J.: Perceiver: General perception with iterative attention. In: International conference on machine learning. pp. 4651-4664. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.285, + 0.785, + 0.311 + ], + "angle": 0, + "content": "32. Jiang, B., Chen, X., Liu, W., Yu, J., Yu, G., Chen, T.: Motiongpt: Human motion as a foreign language. arXiv preprint arXiv:2306.14795 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.311, + 0.785, + 0.352 + ], + "angle": 0, + "content": "33. Kanazawa, A., Black, M.J., Jacobs, D.W., Malik, J.: End-to-end recovery of human shape and pose. In: Computer Vision and Pattern Regognition (CVPR) (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.352, + 0.785, + 0.393 + ], + "angle": 0, + "content": "34. Karras, T., Aila, T., Laine, S., Herva, A., Lehtinen, J.: Audio-driven facial animation by joint end-to-end learning of pose and emotion. ACM Transactions on Graphics (TOG) 36(4), 1-12 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.393, + 0.785, + 0.434 + ], + "angle": 0, + "content": "35. Karunratanakul, K., Preechakul, K., Suwajanakorn, S., Tang, S.: Guided motion diffusion for controllable human motion synthesis. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2151-2162 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.434, + 0.785, + 0.473 + ], + "angle": 0, + "content": "36. Kocabas, M., Athanasiou, N., Black, M.J.: Vibe: Video inference for human body pose and shape estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.473, + 0.785, + 0.528 + ], + "angle": 0, + "content": "37. Kocabas, M., Huang, C.H.P., Tesch, J., Müller, L., Hilliges, O., Black, M.J.: Spec: Seeing people in the wild with an estimated camera. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 11035-11045 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.528, + 0.785, + 0.555 + ], + "angle": 0, + "content": "38. Kovar, L., Gleicher, M., Pighin, F.: Motion graphs. In: Seminal Graphics Papers: Pushing the Boundaries, Volume 2, pp. 723-732 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.555, + 0.785, + 0.596 + ], + "angle": 0, + "content": "39. Kudo, T., Richardson, J.: Sentencepiece: A simple and language independent subword tokenizer and detokenizer for neural text processing. arXiv preprint arXiv:1808.06226 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.596, + 0.785, + 0.637 + ], + "angle": 0, + "content": "40. Lee, T., Moon, G., Lee, K.M.: Multiact: Long-term 3d human motion generation from multiple action labels. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 1231-1239 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.637, + 0.785, + 0.677 + ], + "angle": 0, + "content": "41. Lee, Y., Wampler, K., Bernstein, G., Popovic, J., Popovic, Z.: Motion fields for interactive character locomotion. In: ACM SIGGRAPH Asia 2010 papers, pp. 1-8 (2010)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.677, + 0.785, + 0.717 + ], + "angle": 0, + "content": "42. Levine, S., Wang, J.M., Haraux, A., Popovic, Z., Koltun, V.: Continuous character control with low-dimensional embeddings. ACM Transactions on Graphics (TOG) 31(4), 1-10 (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.718, + 0.785, + 0.759 + ], + "angle": 0, + "content": "43. Li, J., Li, D., Savarese, S., Hoi, S.: Blip-2: Bootstrapping language-image pretraining with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.759, + 0.785, + 0.799 + ], + "angle": 0, + "content": "44. Li, J., Li, D., Xiong, C., Hoi, S.: Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In: International Conference on Machine Learning. pp. 12888-12900. PMLR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "45. Li, M., Chen, X., Zhang, C., Chen, S., Zhu, H., Yin, F., Yu, G., Chen, T.: M3dbench: Let's instruct large models with multi-modal 3d prompts. arXiv preprint arXiv:2312.10763 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.455, + 0.128 + ], + "angle": 0, + "content": "Biao Jiang, Xin Chen et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "46. Li, Y., Wu, M., Zhang, Y., Xu, L., Yu, J.: Piano: A parametric hand bone model from magnetic resonance imaging. arXiv preprint arXiv:2106.10893 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.175, + 0.785, + 0.216 + ], + "angle": 0, + "content": "47. Li, Y., Zhang, L., Qiu, Z., Jiang, Y., Li, N., Ma, Y., Zhang, Y., Xu, L., Yu, J.: Nimble: a non-rigid hand model with bones and muscles. ACM Transactions on Graphics (TOG) 41(4), 1-16 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.216, + 0.785, + 0.243 + ], + "angle": 0, + "content": "48. Lin, C.Y.: Rouge: A package for automatic evaluation of summaries. In: Text summarization branches out. pp. 74-81 (2004)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.243, + 0.785, + 0.283 + ], + "angle": 0, + "content": "49. Lin, J., Zeng, A., Lu, S., Cai, Y., Zhang, R., Wang, H., Zhang, L.: Motion-x: A large-scale 3d expressive whole-body human motion dataset. arXiv preprint arXiv:2307.00818 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.283, + 0.785, + 0.31 + ], + "angle": 0, + "content": "50. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. arXiv preprint arXiv:2304.08485 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.31, + 0.785, + 0.337 + ], + "angle": 0, + "content": "51. Liu, X., Yan, H., Zhang, S., An, C., Qiu, X., Lin, D.: Scaling laws of rope-based extrapolation. arXiv preprint arXiv:2310.05209 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.337, + 0.785, + 0.378 + ], + "angle": 0, + "content": "52. Loper, M., Mahmood, N., Romero, J., Pons-Moll, G., Black, M.J.: Smpl: A skinned multi-person linear model. ACM Trans. Graph. 34(6), 248:1-248:16 (Oct 2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.378, + 0.785, + 0.404 + ], + "angle": 0, + "content": "53. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.404, + 0.785, + 0.459 + ], + "angle": 0, + "content": "54. Lu, C., Yin, F., Chen, X., Liu, W., Chen, T., Yu, G., Fan, J.: A large-scale outdoor multi-modal dataset and benchmark for novel view synthesis and implicit scene reconstruction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7557-7567 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.459, + 0.785, + 0.5 + ], + "angle": 0, + "content": "55. Lu, S., Chen, L.H., Zeng, A., Lin, J., Zhang, R., Zhang, L., Shum, H.Y.: Humantomato: Text-aligned whole-body motion generation. arXiv preprint arXiv:2310.12978 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.5, + 0.785, + 0.554 + ], + "angle": 0, + "content": "56. Ma, H., Li, J., Hosseini, R., Tomizuka, M., Choi, C.: Multi-objective diverse human motion prediction with knowledge distillation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8161-8171 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.554, + 0.785, + 0.595 + ], + "angle": 0, + "content": "57. Mahmood, N., Ghorbani, N., Troje, N.F., Pons-Moll, G., Black, M.J.: Amass: Archive of motion capture as surface shapes. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (October 2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.595, + 0.785, + 0.65 + ], + "angle": 0, + "content": "58. Mehta, D., Rhodin, H., Casas, D., Fua, P., Sotnychenko, O., Xu, W., Theobalt, C.: Monocular 3d human pose estimation in the wild using improved cnn supervision. In: International Conference on 3D Vision (3DV) (2017). https://doi.org/10.1109/3dv.2017.00064, http://gvv.mpi-inf.mpg.de/3dhp_dataset" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.65, + 0.785, + 0.691 + ], + "angle": 0, + "content": "59. Min, J., Chai, J.: Motion graphs++ a compact generative model for semantic motion analysis and synthesis. ACM Transactions on Graphics (TOG) 31(6), 1-12 (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.691, + 0.51, + 0.704 + ], + "angle": 0, + "content": "60. OpenAI: Gpt-4 technical report (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.704, + 0.785, + 0.744 + ], + "angle": 0, + "content": "61. Ordonez, V., Kulkarni, G., Berg, T.: Im2text: Describing images using 1 million captioned photographs. Advances in neural information processing systems 24 (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.744, + 0.785, + 0.799 + ], + "angle": 0, + "content": "62. Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., et al.: Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems 35, 27730-27744 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "63. Papineni, K., Roukos, S., Ward, T., Zhu, W.J.: Bleu: a method for automatic evaluation of machine translation. In: Proceedings of the 40th annual meeting of the Association for Computational Linguistics. pp. 311-318 (2002)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.642, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "MotionChain" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.147, + 0.785, + 0.203 + ], + "angle": 0, + "content": "64. Patel, P., Huang, C.H.P., Tesch, J., Hoffmann, D.T., Tripathi, S., Black, M.J.: Agora: Avatars in geography optimized for regression analysis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13468-13478 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.203, + 0.785, + 0.258 + ], + "angle": 0, + "content": "65. Pavlakos, G., Choutas, V., Ghorbani, N., Bolkart, T., Osman, A.A.A., Tzionas, D., Black, M.J.: Expressive body capture: 3d hands, face, and body from a single image. In: Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR) (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.258, + 0.785, + 0.298 + ], + "angle": 0, + "content": "66. Petrovich, M., Black, M.J., Varol, G.: Action-conditioned 3D human motion synthesis with transformer VAE. In: International Conference on Computer Vision (ICCV) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.298, + 0.785, + 0.339 + ], + "angle": 0, + "content": "67. Petrovich, M., Black, M.J., Varol, G.: TEMOS: Generating diverse human motions from textual descriptions. In: European Conference on Computer Vision (ECCV) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.339, + 0.785, + 0.366 + ], + "angle": 0, + "content": "68. Petrovich, M., Black, M.J., Varol, G.: Tmr: Text-to-motion retrieval using contrastive 3d human motion synthesis. arXiv preprint arXiv:2305.00976 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.366, + 0.785, + 0.406 + ], + "angle": 0, + "content": "69. Plappert, M., Mandery, C., Asfour, T.: The kit motion-language dataset. Big Data 4(4), 236-252 (dec 2016). https://doi.org/10.1089/big.2016.0028, http://dx.doi.org/10.1089/big.2016.0028" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.406, + 0.785, + 0.447 + ], + "angle": 0, + "content": "70. Plappert, M., Mandery, C., Asfour, T.: Learning a bidirectional mapping between human whole-body motion and natural language using deep recurrent neural networks. Robotics and Autonomous Systems 109, 13-26 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.447, + 0.785, + 0.5 + ], + "angle": 0, + "content": "71. Punnakkal, A.R., Chandrasekaran, A., Athanasiou, N., Quiros-Ramirez, A., Black, M.J.: BABEL: Bodies, action and behavior with english labels. In: Proceedings IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR). pp. 722-731 (Jun 2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.501, + 0.785, + 0.542 + ], + "angle": 0, + "content": "72. Qiu, Z., Li, Y., He, D., Zhang, Q., Zhang, L., Zhang, Y., Wang, J., Xu, L., Wang, X., Zhang, Y., et al.: Sculptor: Skeleton-consistent face creation using a learned parametric generator. ACM Transactions on Graphics (TOG) 41(6), 1-17 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.542, + 0.785, + 0.597 + ], + "angle": 0, + "content": "73. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International Conference on Machine Learning. pp. 8748-8763. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.597, + 0.785, + 0.651 + ], + "angle": 0, + "content": "74. Raffel, C., Shazeer, N., Roberts, A., Lee, K., Narang, S., Matena, M., Zhou, Y., Li, W., Liu, P.J.: Exploring the limits of transfer learning with a unified text-to-text transformer. The Journal of Machine Learning Research 21(1), 5485-5551 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.651, + 0.785, + 0.692 + ], + "angle": 0, + "content": "75. Rajasegaran, J., Pavlakos, G., Kanazawa, A., Malik, J.: Tracking people by predicting 3d appearance, location and pose. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2740-2749 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.692, + 0.785, + 0.718 + ], + "angle": 0, + "content": "76. Razavi, A., Van den Oord, A., Vinyals, O.: Generating diverse high-fidelity images with vq-vae-2. Advances in neural information processing systems 32 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.718, + 0.785, + 0.772 + ], + "angle": 0, + "content": "77. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2022), https://github.com/CompVis/latent-diffusionhttps://arxiv.org/abs/2112.10752" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.772, + 0.785, + 0.8 + ], + "angle": 0, + "content": "78. Romero, J., Tzionas, D., Black, M.J.: Embodied hands: Modeling and capturing hands and bodies together. arXiv preprint arXiv:2201.02610 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "79. Rose, C., Cohen, M.F., Bodenheimer, B.: Verbs and adverbs: Multidimensional motion interpolation. IEEE Computer Graphics and Applications 18(5), 32-40 (1998)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "20" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.455, + 0.128 + ], + "angle": 0, + "content": "Biao Jiang, Xin Chen et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.147, + 0.785, + 0.203 + ], + "angle": 0, + "content": "80. Rubenstein, P.K., Asawaroengchai, C., Nguyen, D.D., Bapna, A., Borsos, Z., Quitry, F.d.C., Chen, P., Badawy, D.E., Han, W., Kharitonov, E., et al.: Audiopalm: A large language model that can speak and listen. arXiv preprint arXiv:2306.12925 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.204, + 0.785, + 0.26 + ], + "angle": 0, + "content": "81. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems 35, 25278-25294 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.261, + 0.785, + 0.288 + ], + "angle": 0, + "content": "82. Shafir, Y., Tevet, G., Kapon, R., Bermano, A.H.: Human motion diffusion as a generative prior. arXiv preprint arXiv:2303.01418 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.289, + 0.785, + 0.344 + ], + "angle": 0, + "content": "83. Siddiqui, Y., Alliegro, A., Artemov, A., Tommasi, T., Sirigatti, D., Rosov, V., Dai, A., Nießner, M.: Meshgpt: Generating triangle meshes with decoder-only transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19615-19625 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.346, + 0.785, + 0.401 + ], + "angle": 0, + "content": "84. Siyao, L., Yu, W., Gu, T., Lin, C., Wang, Q., Qian, C., Loy, C.C., Liu, Z.: Bailando: 3d dance generation by actor-critic gpt with choreographic memory. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11050-11059 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.402, + 0.785, + 0.43 + ], + "angle": 0, + "content": "85. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.431, + 0.785, + 0.471 + ], + "angle": 0, + "content": "86. Starke, S., Mason, I., Komura, T.: Deepphase: periodic autoencoders for learning motion phase manifolds. ACM Transactions on Graphics (TOG) 41(4), 1-13 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.473, + 0.785, + 0.501 + ], + "angle": 0, + "content": "87. Starke, S., Zhang, H., Komura, T., Saito, J.: Neural state machine for character-scene interactions. ACM Trans. Graph. 38(6), 209-1 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.502, + 0.785, + 0.543 + ], + "angle": 0, + "content": "88. Starke, S., Zhao, Y., Zinno, F., Komura, T.: Neural animation layering for synthesizing martial arts movements. ACM Transactions on Graphics (TOG) 40(4), 1-16 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.544, + 0.785, + 0.599 + ], + "angle": 0, + "content": "89. Szot, A., Clegg, A., Undersander, E., Wijmans, E., Zhao, Y., Turner, J., Maestre, N., Mukadam, M., Chaplot, D.S., Maksymets, O., et al.: Habitat 2.0: Training home assistants to rearrange their habitat. Advances in Neural Information Processing Systems 34, 251-266 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.6, + 0.785, + 0.641 + ], + "angle": 0, + "content": "90. Takano, W., Nakamura, Y.: Statistical mutual conversion between whole body motion primitives and linguistic sentences for human motions. The International Journal of Robotics Research 34(10), 1314-1328 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.642, + 0.785, + 0.698 + ], + "angle": 0, + "content": "91. Tevet, G., Gordon, B., Hertz, A., Bermano, A.H., Cohen-Or, D.: Motionclip: Exposing human motion generation to clip space. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXII. pp. 358-374. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.699, + 0.785, + 0.727 + ], + "angle": 0, + "content": "92. Tevet, G., Raab, S., Gordon, B., Shafir, Y., Bermano, A.H., Cohen-Or, D.: Human motion diffusion model. arXiv preprint arXiv:2209.14916 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.728, + 0.785, + 0.769 + ], + "angle": 0, + "content": "93. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., et al.: Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.77, + 0.785, + 0.812 + ], + "angle": 0, + "content": "94. Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi, A., Babaei, Y., Bashlykov, N., Batra, S., Bhargava, P., Bhosale, S., et al.: Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "95. Van Den Oord, A., Vinyals, O., et al.: Neural discrete representation learning. Advances in neural information processing systems 30 (2017)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.641, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "MotionChain" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "21" + }, + { + "type": "ref_text", + "bbox": [ + 0.223, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "96. Varol, G., Romero, J., Martin, X., Mahmood, N., Black, M., Laptev, I., Schmid, C.: Learning from synthetic humans. In: Computer Vision and Pattern Recognition (CVPR) (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.19, + 0.785, + 0.232 + ], + "angle": 0, + "content": "97. Vedantam, R., Lawrence Zitnick, C., Parikh, D.: Cider: Consensus-based image description evaluation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4566-4575 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.232, + 0.785, + 0.272 + ], + "angle": 0, + "content": "98. Wang, C.: T2m-hifigpt: Generating high quality human motion from textual descriptions with residual discrete representations. arXiv preprint arXiv:2312.10628 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.273, + 0.785, + 0.314 + ], + "angle": 0, + "content": "99. Wang, W., Zhe, X., Chen, H., Kang, D., Li, T., Chen, R., Bao, L.: Neural marionette: A transformer-based multi-action human motion synthesis system. arXiv preprint arXiv:2209.13204 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.315, + 0.785, + 0.355 + ], + "angle": 0, + "content": "100. Wu, C., Yin, S., Qi, W., Wang, X., Tang, Z., Duan, N.: Visual chatgpt: Talking, drawing and editing with visual foundation models. arXiv preprint arXiv:2303.04671 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.356, + 0.785, + 0.412 + ], + "angle": 0, + "content": "101. Wu, Y.C., Gebru, I.D., Markovic, D., Richard, A.: Audiodec: An open-source streaming high-fidelity neural audio codec. In: ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 1-5. IEEE (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.412, + 0.785, + 0.466 + ], + "angle": 0, + "content": "102. Xin, C., Jiang, B., Liu, W., Huang, Z., Fu, B., Chen, T., Yu, J., Yu, G.: Executing your commands via motion diffusion in latent space. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.467, + 0.785, + 0.508 + ], + "angle": 0, + "content": "103. Xiong, W., Liu, J., Molybog, I., Zhang, H., Bhargava, P., Hou, R., Martin, L., Rungta, R., Sankararaman, K.A., Oguz, B., et al.: Effective long-context scaling of foundation models. arXiv preprint arXiv:2309.16039 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.508, + 0.785, + 0.55 + ], + "angle": 0, + "content": "104. Xu, H., Ghosh, G., Huang, P.Y., Okhonko, D., Aghajanyan, A., Metze, F., Zettlemoyer, L., Feichtenhofer, C.: Videoclip: Contrastive pre-training for zero-shot video-text understanding. arXiv preprint arXiv:2109.14084 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.55, + 0.785, + 0.592 + ], + "angle": 0, + "content": "105. Yamada, T., Matsunaga, H., Ogata, T.: Paired recurrent autoencoders for bidirectional translation between robot actions and linguistic descriptions. IEEE Robotics and Automation Letters 3(4), 3441-3448 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.592, + 0.785, + 0.619 + ], + "angle": 0, + "content": "106. Yao, H., Song, Z., Zhou, Y., Ao, T., Chen, B., Liu, L.: Moconvq: Unified physics-based motion control via scalable discrete representations (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.62, + 0.785, + 0.66 + ], + "angle": 0, + "content": "107. Ye, Q., Xu, H., Xu, G., Ye, J., Yan, M., Zhou, Y., Wang, J., Hu, A., Shi, P., Shi, Y., et al.: mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.66, + 0.785, + 0.702 + ], + "angle": 0, + "content": "108. Yin, F., Chen, X., Zhang, C., Jiang, B., Zhao, Z., Fan, J., Yu, G., Li, T., Chen, T.: Shapeqpt: 3d shape generation with a unified multi-modal language model. arXiv preprint arXiv:2311.17618 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.702, + 0.785, + 0.743 + ], + "angle": 0, + "content": "109. Yin, F., Liu, W., Huang, Z., Cheng, P., Chen, T., Yu, G.: Coordinates are not lonely-codebook prior helps implicit neural 3d representations. Advances in Neural Information Processing Systems 35, 12705-12717 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.744, + 0.785, + 0.785 + ], + "angle": 0, + "content": "110. Yu, J., Xu, Y., Koh, J.Y., Luong, T., Baid, G., Wang, Z., Vasudevan, V., Ku, A., Yang, Y., Ayan, B.K., et al.: Scaling autoregressive models for content-rich text-to-image generation. arXiv preprint arXiv:2206.10789 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.785, + 0.785, + 0.84 + ], + "angle": 0, + "content": "111. Yuan, Y., Kitani, K.: Dlow: Diversifying latent flows for diverse human motion prediction. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part IX 16. pp. 346-364. Springer (2020)" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "22" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.455, + 0.128 + ], + "angle": 0, + "content": "Biao Jiang, Xin Chen et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "112. Yuan, Y., Song, J., Iqbal, U., Vahdat, A., Kautz, J.: PhysDiff: Physics-guided human motion diffusion model. In: IEEE International Conference on Computer Vision (ICCV) (October 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.785, + 0.218 + ], + "angle": 0, + "content": "113. Zhang, H., Li, X., Bing, L.: Video-llama: An instruction-tuned audio-visual language model for video understanding. arXiv preprint arXiv:2306.02858 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.218, + 0.785, + 0.259 + ], + "angle": 0, + "content": "114. Zhang, H., Cao, J., Lu, G., Ouyang, W., Sun, Z.: Danet: Decompose-and-aggregate network for 3d human shape and pose estimation. In: Proceedings of the 27th ACM International Conference on Multimedia. pp. 935-944 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.259, + 0.785, + 0.3 + ], + "angle": 0, + "content": "115. Zhang, H., Tian, Y., Zhang, Y., Li, M., An, L., Sun, Z., Liu, Y.: Pymaf-x: Towards well-aligned full-body model regression from monocular images. IEEE Transactions on Pattern Analysis and Machine Intelligence (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.3, + 0.785, + 0.357 + ], + "angle": 0, + "content": "116. Zhang, J., Zhang, Y., Cun, X., Huang, S., Zhang, Y., Zhao, H., Lu, H., Shen, X.: T2m-gpt: Generating human motion from textual descriptions with discrete representations. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.357, + 0.785, + 0.398 + ], + "angle": 0, + "content": "117. Zhang, M., Cai, Z., Pan, L., Hong, F., Guo, X., Yang, L., Liu, Z.: Motiondiffuse: Text-driven human motion generation with diffusion model. arXiv preprint arXiv:2208.15001 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.398, + 0.785, + 0.439 + ], + "angle": 0, + "content": "118. Zhang, M., Guo, X., Pan, L., Cai, Z., Hong, F., Li, H., Yang, L., Liu, Z.: Remodiffuse: Retrieval-augmented motion diffusion model. arXiv preprint arXiv:2304.01116 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.439, + 0.785, + 0.495 + ], + "angle": 0, + "content": "119. Zhang, Q., Song, J., Huang, X., Chen, Y., Liu, M.Y.: Diffcollage: Parallel generation of large content with diffusion models. In: 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 10188-10198. IEEE (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.494, + 0.785, + 0.522 + ], + "angle": 0, + "content": "120. Zhang, T., Kishore, V., Wu, F., Weinberger, K.Q., Artzi, Y.: Bertscore: Evaluating text generation with bert. arXiv preprint arXiv:1904.09675 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.522, + 0.785, + 0.563 + ], + "angle": 0, + "content": "121. Zhang, Y., Black, M.J., Tang, S.: We are more than our joints: Predicting how 3d bodies move. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3372-3382 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.564, + 0.785, + 0.605 + ], + "angle": 0, + "content": "122. Zheng, L., Chiang, W.L., Sheng, Y., Zhuang, S., Wu, Z., Zhuang, Y., Lin, Z., Li, Z., Li, D., Xing, E.P., Zhang, H., Gonzalez, J.E., Stoica, I.: Judging lvm-as-a-judge with mt-bench and chatbot arena (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.605, + 0.785, + 0.647 + ], + "angle": 0, + "content": "123. Zhou, C., Liu, P., Xu, P., Iyer, S., Sun, J., Mao, Y., Ma, X., Efrat, A., Yu, P., Yu, L., et al.: Lima: Less is more for alignment. Advances in Neural Information Processing Systems 36 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.647, + 0.785, + 0.688 + ], + "angle": 0, + "content": "124. Zhu, D., Chen, J., Shen, X., Li, X., Elhoseiny, M.: Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.688, + 0.785, + 0.745 + ], + "angle": 0, + "content": "125. Zou, X., Dou, Z.Y., Yang, J., Gan, Z., Li, L., Li, C., Dai, X., Behl, H., Wang, J., Yuan, L., et al.: Generalized decoding for pixel, image, and language. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15116-15127 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.745 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/MotionChain_ Conversational Motion Controllers via Multimodal Prompts/24d4fe30-81b0-462c-a698-38fed56b9fc4_origin.pdf b/2024/MotionChain_ Conversational Motion Controllers via Multimodal Prompts/24d4fe30-81b0-462c-a698-38fed56b9fc4_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..aa207c4257cf310d057eb4b603a9b1b41dbec99c --- /dev/null +++ b/2024/MotionChain_ Conversational Motion Controllers via Multimodal Prompts/24d4fe30-81b0-462c-a698-38fed56b9fc4_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5d90085ccf2dfef76b25358317d180cc25addbc05f39cc55402cfdb6b09f9ff6 +size 3111759 diff --git a/2024/MotionChain_ Conversational Motion Controllers via Multimodal Prompts/full.md b/2024/MotionChain_ Conversational Motion Controllers via Multimodal Prompts/full.md new file mode 100644 index 0000000000000000000000000000000000000000..52a05cc7b2772a626fea6afd1de21339feff4bdd --- /dev/null +++ b/2024/MotionChain_ Conversational Motion Controllers via Multimodal Prompts/full.md @@ -0,0 +1,335 @@ +# MotionChain: Conversational Motion Controllers via Multimodal Prompts + +Biao Jiang $^{1,2\star}$ , Xin Chen $^{2\star\star}$ , Chi Zhang $^{2}$ , Fukun Yin $^{1}$ , Zhuoyuan Li $^{1}$ , Gang Yu $^{2}$ , and Jiayuan Fan $^{1\star\star\star}$ + +1 Fudan University + +2 Tencent + +Abstract. Recent advancements in language models have demonstrated their adeptness in conducting multi-turn dialogues and retaining conversational context. However, this proficiency remains largely unexplored in other multimodal generative models, particularly in human motion models. By integrating multi-turn conversations in controlling continuous virtual human movements, generative human motion models can achieve an intuitive and step-by-step process of human task execution for humanoid robotics, game agents, or other embodied systems. In this work, we present MotionChain, a conversational human motion controller that generates continuous and long-term human motion through multimodal prompts. Specifically, MotionChain consists of multi-modal tokenizers that transform various data types such as text, image, and motion, into discrete tokens, coupled with a Vision-Motion-aware Language model. By leveraging large-scale language, vision-language, and vision-motion data to assist motion-related generation tasks, MotionChain thus comprehends each instruction in multi-turn conversation and generates human motions followed by these prompts. Extensive experiments validate the efficacy of MotionChain, demonstrating state-of-the-art performance in conversational motion generation, as well as more intuitive manners of controlling and interacting with virtual humans. + +Keywords: 3D Motion $\cdot$ Motion Generation $\cdot$ Text-to-Motion + +# 1 Introduction + +The success of large language models (LLMs) [60, 62, 93, 94, 122] has sparked significant interest in the development of multi-modal language models. These models aim to transfer instruction-following and zero-shot abilities to other modalities tasks, such as image-language models [1, 50, 107, 124], video-language models [1, 43, 44, 113], and 3D-language models [12, 26, 28, 108]. However, a comprehensive model that can perceive visual input and generate continuous motion through multi-turn conversations has not yet been developed. Such a + +![](images/86a343de4f67d818efac918a57e0a0271f3ff81233309de547b2138df43313c3.jpg) +Fig. 1: MotionChain can interpret instructions from multi-turn conversations and generate human motions or textual answers based on text, motion, or image inputs. We provide the conversation results in image-conditioned motion generation (1st column), motion reasoning (second column), motion editing (third column), and motion translation (third column), with each subsequent turn informed by all previous conversations. Left-to-right represents the temporal order. + +multi-modal model would have wide-ranging applications in fields like humanoid robotics, virtual assistants, game agents and so on. + +Previous research on human motion has explored various tasks, including motion generation [23, 32, 66, 92, 102, 116], motion captioning [20, 24, 32], motion prediction [32, 56, 111, 121], and motion composition [3]. Recent works in text-to-motion [67, 92, 102, 117] have involved pre-trained language models [16, 73] for motion generation. For instance, TEMOS [67] employs BERT [16] text embeddings in an end-to-end transformer architecture, while MDM [92] and MLD [102] both utilize text embeddings from CLIP [73] during the conditional diffusion process. On the other hand, MotionCLIP [91] and TMR [68] focus on modeling the coupled relationship between motion and text description, and MotionGPT [32] introduces a motion-language model that represents human motion and language in one unified vocabulary. However, these above methods treat all tasks as a one-turn conditioned generation, lacking contextual understanding and multi-turn continuous generation abilities. Therefore, we construct a Vision-Motion language model, integrating multi-turn conversations and continuous human motions. + +Two crucial challenges need to be addressed in this conversational motion generation. The first challenge is to contextually generate human motion in a continuous manner, resembling the way real humans move. The second challenge is the scarcity of text-motion paired datasets compared to datasets with pairs of image-language [11, 81], image-pose [37, 58, 61, 64] and video-motion [4, 5, 8, 30, 96]. Fortunately, both human motion and language are sequential and can be continuously "written". Building upon this observation, we employ the general vision-language instruction-tuning approach [50, 123] to enable conversational motion generation and question-answering through multi-modal instructions. By integrating image, motion, and language data and encoding them into tokens, the relationship between these three modalities becomes more evident. Therefore, with the advent of large vision-motion and vision-language data, Vision-Motion-language pre-training can enhance the performance of motion-related tasks. + +In this study, we introduce MotionChain, a comprehensive framework that integrates vision, motion, and language. MotionChain leverages large-scale vision-language data, vision-motion data, and pre-trained language models' strong language generation abilities to assist in motion-related generation tasks. To enable MotionChain to comprehend and generate human-like motions, we first train a motion-specific vector quantized variational autoencoder (VQ-VAE) model. This model constructs a "motion vocabulary" similar to the English word vocabulary and converts raw motion data into a sequence of motion tokens. To incorporate vision inputs into MotionChain, we then introduce a specialized vision tokenizer that connects a pre-trained vision encoder to the language model. This tokenizer converts image data into visual tokens within the language-motion "words" embedding space. These tokens are then processed by a pre-trained language model [14,74,93,94], which learns the relationship between image, motion and language. To enable conversational generation, we construct a multi-modal motion conversation dataset based on the existing text-motion dataset [23] and vision-motion dataset [5]. We then train the language model using our multimodal conversation dataset to learn the correlation between the three modalities. Extensive experiments demonstrate that MotionChain achieves state-of-the-art performance in multiple motion-related tasks. + +We summarize our contributions as follows: (1) We propose MotionChain, a unified vision-motion-language generative pre-trained model, which performs conversational generation tasks via multi-modal inputs with language models. (2) We introduce a motion composition technique, to generate 3D human motions following the temporal order of instructions. (3) We propose a multi-modal motion conversation benchmark, wherein MotionChain achieves competitive performance across diverse motion tasks. + +# 2 Related Work + +Human Motion Modeling. There have been numerous attempts to model the relationship between 3D human motion and multiple modalities including incomplete motion [56, 111, 121], action [3, 25, 40, 66, 99, 102], text [22-24, 32, 55, 67, 82, 92, 117-119], image [19, 21, 33, 114, 115] and video [5, 13, 19, 36, 75]. Text-to-motion is one of the most important motion generation tasks, due to the user-friendly and convenient language input. MDM [92], MotionDiffuse [117] and MLD [102] proposes a diffusion-based generative model [27, 77, 85] to generate motions conditioned on different inputs. TM2T [24] and T2M-GPT [116] investigate a generative framework based on VQ-VAE [76, 95] and generative transformer for motion generation. Motion completion task generates motion conditioning on partial motions, such as classical motion prediction [56, 111, 121] or motion in-between [92], which generates the intermediate motion while the first and last parts are fixed. TEACH [3] proposes a past-conditioned transformer model that generates motion from a sequence of actions autoregressively. Apart from motion generation, there is also work investigating other modalities of generation from motion. Two statistical models [90] and recurrent networks [70, 105] + +![](images/2ece0ff6c53ff2ed370c8c580373637a05cb859bee80aa4474ce9fa6307beb0f.jpg) +Fig. 2: Method overview: MotionChain consists of a motion tokenizer $\mathcal{V}_{\mathcal{M}}$ (Sec. 3.2), a vision tokenizer $\mathcal{V}_{\mathcal{I}}$ (r Sec. 3.2) and a vision-motion-aware language model (Sec. 3.3). By leveraging motion tokens generated by $\mathcal{V}_{\mathcal{M}}$ , alongside visual language token embeddings projected by vision tokenizer $\mathcal{V}_{\mathcal{I}}$ , and text tokens by text tokenizer, MotionChian achieves a unified learning paradigm for both motion and linguistic data. + +are learned in mapping motions to language. TM2T [24] proposed a new motion representation that compresses motions into a short sequence of discrete variables and then uses a neural translation network to build mappings between two modalities. In contrast to the above methods limited to only several tasks, MotionGPT [32] treats human motion as a foreign language and leverages language understanding and zero-shot transfer abilities of pre-trained language models. + +Character Control and Animation. Character control involves generating interactive motion sequences based on user instruction signals. One kind of approach [38, 59, 79] is to construct a graph representing transitions between motion clips and plan motion using graph search. Considering the limitations of these graph-based approaches in coarse discreteness, alternative methods like frame blending and concatenation [41], low-dimensional latent space learning [42], motion matching [15] proposed for embedding the task in the feature and [88] do the similar thing through hierarchical setup. Although the control signals for motion control and character animation are different from the instructions in text-to-motion tasks, we still recognize textual commands of conventional human motion generation as a boost for intuitive character control. + +Multi-Modal Language Models. In the field of computer vision, there has been a recent surge of interest in multi-modal models that can process text along with other modalities, including images [18, 29, 44, 104], audio [6, 80, 101], and 3D [12, 45, 54, 83, 108]. CLIP [73] is an example of such a model, which learns a semantic latent representation that connects images with corresponding language descriptions. While language models have achieved success in various tasks, the development of multi-modal language models capable of handling human motion is still limited. Existing works in computer vision can be broadly categorized into two classes. The first consists of end-to-end trained models explored separately for specific research topics. For example, tasks like vision-language navi + +![](images/5b5fc2e2f5bbf4400f66ff844044bc3d92ff303935cbe7943668403b496b8640.jpg) +Fig. 3: Data collection overview: Our initial step in collecting the motion reasoning data involves the utilization of human motion captions derived from an existing text-motion dataset. Subsequent to this, the text-motion retrieval model TMR [68] aids in the segmentation of motion pairs into categories based on the similarity between them. With the assistance of ChatGPT, we proceed to craft motion editing task data that correspond to these categorized similarity levels. Incorporating both motion reasoning and editing single-turn tasks, as well as the extensive 14 tasks delineated in [32], we construct a rich multi-modal multi-turn conversation dataset. + +gation [2, 7] and Habitat [89] require embodied AI agents to follow natural language instructions and take actions to accomplish goals in visual environments. InstructPix2Pix [7] in image translation enables agents to edit images based on human instructions. The second involves systems that coordinate various models using approaches like LangChain or LLMs [51, 62, 103]. Examples of such systems include Visual ChatGPT [100], X-GPT [125], and MMREACT [110]. While these methods focus on building instruction-following agents, we aim to develop an end-to-end trained multimodal model that can perform conversational motion generation tasks via multi-modal inputs with language models. + +# 3 Methods + +To leverage large language data, vision-language data, and vision-motion data for assisting motion-related tasks, we propose a motion-language-vision framework called MotionChain. The framework, as depicted in Fig. 2, consists of a multi-modal tokenizer that converts various types of data (text, image, and motion) into discrete tokens (Sec. 3.2), and a vision-motion-aware Language model that comprehends information from different modalities and generates corresponding answers based on input instructions (Sec. 3.3). Additionally, to simultaneously understand data from multiple modalities, we employ a multi-stage training strategy (Sec. 3.4) for the training of the multi-modal tokenizer and the motion-language-vision framework. + +We first introduced the multi-modal tokenizer, which comprises three branches for processing textual, image, and motion inputs. For textual inputs $w^{1:N} = \{w^i\}$ of length $N$ that describes a motion-related question or demand, we employ the SentencePiece model [39] used in previous works [14, 74, 93, 94], which has a vocabulary size of $K_{t}$ and is trained on a large number of language datasets. The motion branch consists of a motion encoder $\mathcal{E}_{\mathcal{M}}$ that encodes a motion sequence $m^{1:M} = \{x^i\}$ of $M$ frames into $L$ motion tokens $z^{1:L} = \{z^i\}$ , where $L = M / l$ and $l$ represents the temporal downsampling rate on motion frames. It also includes a motion decoder $\mathcal{D}_{\mathcal{M}}$ that can decode motion tokens back to human motion $\hat{m}^{1:M}$ . The vision branch processes the input image $X$ with a pre-trained CLIP visual encoder and a learnable linear projection that follows it, converted into language token embeddings $H_{q}$ . Given a textual sentence $w^{1:N}$ , a sequence of motion $m^{1:M}$ , and an image condition $X$ , all encoded as language tokens, our vision-motion-aware language model is designed to produce an answer comprising $L$ tokens, denoted as $\hat{x}^{1:L} = \{\hat{x}^i\}$ . These output tokens can represent either motion sequences $\hat{x}_{m}^{1:L}$ or textual descriptions $\hat{x}_{t}^{1:L}$ , which integrate both human motion $\hat{m}^{1:M}$ , and text $\hat{w}^{1:L}$ within the given context. + +# 3.1 Data Collection + +With the emergence of text-conditioned motion generation tasks, datasets like KIT [69], BABEL [71], HumanML3D [23] and the more recent Motion-X [49] have been developed. However, these datasets predominantly offer text labels as simple action phrases or captions. Building upon these foundations, MotionGPT [32] introduces an instruction-based motion-language dataset that encapsulates 14 core tasks, including motion prediction, translation, and editing, through thousands of instruction templates in a unified format. Despite this advancement, MotionGPT's data lack a deep engagement with the nuances of human motion analysis and are limited to single-turn generation tasks without incorporating contextual memory. Inspired by the recent success of GPT models across text-annotation tasks [17], image-annotation tasks [50], 3D-annotation tasks [28], we propose a data collection methodology integrates the capabilities of existing LLMs like ChatGPT [62], with the text-motion retrieval model TMR [68] to facilitate motion conversation data collection. In addition to the 14 motion-related tasks in MotionGPT [32], we introduce tasks centered around motion reasoning and motion editing, leveraging contextual insights for a deeper motion analysis. + +Utilizing ChatGPT [62], we initiate the collection of motion reasoning data using human motion captions from the text-motion dataset [23], starting with manually designed example queries that explore the contextual scenarios surrounding motions, possible preceding or succeeding actions, the subjects' roles, and the tools or equipment involved, etc. Following this, we employ TMR [68] for categorizing motions from the dataset into varying similarity levels. For medium-similarity motion pairs, we utilize ChatGPT [62] to generate motion editing directives that enable the transformation of one motion to another. For motions of + +high similarity, we manually devise tasks aimed at editing their lengths, further enriching the dataset's versatility and analytical scope. + +After the collection of single-turn generation tasks, we progress to develop multi-turn conversation data. This involves the deliberate association of initial motion generation tasks with a variety of follow-up tasks randomly chosen among motion translation, reasoning, editing, etc. Following [122] we construct our conversation data in a structured format, as depicted below: + +$X_{\mathrm{system - message}}$ + +USER: $X_{v}$ $\dot{X}_s^1$ ASSISTANT: $X_{a}^{1} < / s>$ + +USER: $X_{v} X_{s}^{2}$ ASSISTANT: $X_{a}^{2} < / \mathrm{s}>$ + +USER: $X_{v} X_{s}^{3}$ ASSISTANT: $X_{a}^{3} < / \mathrm{s}>$ ... + +Where $X_{v}$ is defined as the vision language token embeddings, processed via the visual tokenizer. $X_{s}^{i}$ and $X_{a}^{i}$ are used to denote the source inputs and target answers for each round $i$ , respectively. Both sets of tokens originate from the integrated motion-language vocabulary $V$ , which includes motions, texts, or a blend thereof. The dataset exhibits variability in the number of generation turns up to 10; for the sake of clarity, we present only three examples herein. MotionChain is trained to predict answers, incorporating a learning mechanism that determines whether to stop generation by outputting end of sentence flag $< / \mathrm{s} >$ based on the current instruction and all preceding questions and answers. In the computation of the loss, as defined in Eq. (5), only the green tokens are utilized. + +# 3.2 Multi-modal Tokenizer + +Motion tokenizer, denoted as $\mathcal{V}_{\mathcal{M}}$ , is based on the architecture of Vector Quantized Variational Autoencoders (VQ-VAE) utilized in previous studies [22,24,32, 84,95,98,106,109,116]. Once pre-trained, it can represent motion using discrete tokens, facilitating the integration of motion and language. The Motion tokenizer consists of a motion encoder $\mathcal{E}_{\mathcal{M}}$ and a motion decoder $\mathcal{D}_{\mathcal{M}}$ . Initially, the motion encoder $\mathcal{E}$ applies 1D convolutions to the motion features $m^{1:M}$ along the temporal dimension to obtain latent vectors $\hat{z}^{1:L} = \mathcal{E}_{\mathcal{M}}(m^{1:M})$ . Subsequently, the latent vectors $\hat{z}$ are quantized and transformed into a collection of codebook entries $z$ . The learnable codebook $Z = \{z_{i}\}_{i=1}^{K} \subset \mathbb{R}^{d}$ comprises $K$ latent embedding vectors, each with a dimension of $d$ . The quantization process $Q(\cdot)$ replaces each row vector $\hat{z}$ with its nearest codebook entry $z_{k}$ in $Z$ , which can be expressed as: + +$$ +z _ {i} = Q \left(\hat {z} ^ {i}\right) := \arg \min _ {z _ {k} \in Z} \| \hat {z} _ {i} - z _ {k} \| _ {2}. \tag {1} +$$ + +We assign $s^i$ as the index number of motion tokens $z^{1:L}$ , so motion tokens $z^{1:L}$ can be represented as a sequence of indices $s^{1:L} = \{s^i\}_{i=1}^L$ . The motion decoder $\mathcal{D}_{\mathcal{M}}$ can project $z^{1:L} = \{z^i\}_{i=1}^L$ back to the raw motion space, resulting in the motion $\hat{m}^{1:M}$ with $M$ frames. Following [22,24,32,98,116], we adopt three distinct loss functions when training the motion tokenizer: + +$$ +\mathcal {L} _ {\mathcal {V}} = \mathcal {L} _ {r} + \mathcal {L} _ {e} + \mathcal {L} _ {c} \tag {2} +$$ + +![](images/7a41c3c7f2ec84c74344d9d49bbf140e657db0b17849fef6493b0d086748291c.jpg) +(a) Independent + +![](images/65da011d9b7b6a97fa4383c8d2c5d774a27ccef12991c5a264f7d1c7b3fbefe4.jpg) +(b) Past-condition +Fig. 4: Motion Composition Variants: We illustrate the baselines for motion composition during multi-turn motion generation (a). independent decoding each turn (b). separate decoding conditioned on the last few tokens from the prior turn (c). decoding with joint motion tokens. Green tokens stand for image condition, blue tokens stand for textual instruction, and orange tokens stand for human motions. + +![](images/7e2d5e43eaa0c0fd055ee3e0f8fd642875832808805ab1728c09fdf3a0dd7f5f.jpg) +(c) Tokens-joint + +where $\mathcal{L}_r$ denotes reconstruction loss, $\mathcal{L}_e$ denotes the embedding loss, and $\mathcal{L}_c$ denotes commitment loss. + +During multi-turn motion generation, the motion continuity between turns is achieved through our motion decoder, which links the motion of the current turn with that of the preceding ones. Taking the composition of two motions as an example: we concatenate the past motion tokens, denoted as $z_{p}^{1:L_{p}}$ , with the tokens representing the current motion, $z_{c}^{1:L_{c}}$ . This concatenated sequence of tokens is subsequently decoded into a comprehensive set of continuous motion features, represented as $m_{\mathrm{whole}}^{1:M_{\mathrm{whole}}}$ , as depicted below: + +$$ +z _ {\text {w h o l e}} ^ {1: \left(L _ {p} + L _ {c}\right)} = \left[ z _ {p} ^ {1: L _ {p}}, z _ {c} ^ {1: L _ {c}} \right]. \tag {3} +$$ + +Similarly, this framework is adept at executing composition tasks involving an array of motions. The comparison results in Tab. 3 demonstrate that our motion tokenizer could effectively perform motion composition tasks. + +Visual Tokenizer accepts images $X_{I}$ as inputs. We employ the CLIP visual encoder that is pre-trained on image-text pairs to derive visual feature $Z_{I}$ . These features are then projected into language token embeddings $X_{v}$ via a linear layer like previous work [50], maintaining consistency in the dimensionality with the language model's word embedding space. + +# 3.3 Motion-aware Language Model + +Language models such as Llama [93, 94] and T5 [14, 74] employ the Sentence-Piece [39] model to encode textual inputs into WordPiece tokens, utilizing a $K_{t}$ word piece vocabulary. Unlike prior text-to-motion [24, 102, 116, 118] and motion-to-text [24] methods that process text and motion separately, we merge the text vocabulary $V_{t} = \{v_{t}^{i}\}_{i=1}^{K_{t}}$ with the motion vocabulary $V_{m} = \{v_{m}^{i}\}_{i=1}^{K_{m}}$ , maintaining the motion tokenizer's codebook $Z$ order and including special tokens for boundary demarcation. This creates a unified vocabulary $V = \{V_{t}, V_{m}\}$ , enabling the formulation of motion-centric tasks in a universal template, where inputs and + +outputs share the same vocabulary. For visual input, our visual tokenizer converts images into visual token embeddings $X_{v}$ , aligning with the language model [39, 62, 74, 93] token space for integrated representation. + +For single conditioned generation tasks, our input comprises a sequence of $N$ length tokens $X_{s} = \{x_{s}^{i}\}_{i = 1}^{N}$ , where $x_{s}\in \{V_{t},V_{m}\}$ representing either text, motion, or a combination thereof, drawn from the unified vocabularies. In cases involving image inputs, visual tokens $X_{v}$ are interspersed at the beginning of the source tokens sequence, forming $[X_v,X_s]$ . Subsequent interaction rounds generate target answer tokens $X_{a}$ . To facilitate iterative result generation and content retention, our framework generates multi-turn conversation data $(X_{v},X_{s}^{1},X_{a}^{1},X_{s}^{2},X_{a}^{2},\dots ,X_{s}^{T},X_{a}^{T})$ , with $T$ indicating the total turn count. Notably, visual tokens are consistently placed at the forefront of the initial turn's source tokens. The processing sequence is organized such that to predict target answer tokens autoregressively, as shown in Fig. 2. Source tokens are processed by the transformer to predict the next token's probability distribution, formulated as: + +$$ +p _ {\theta} \left(X _ {a} \mid X _ {v}, X _ {s}\right) = \prod_ {i} p _ {\theta} \left(x _ {a} ^ {i} \mid X _ {v}, X _ {s, < i}, X _ {a, < i}\right) \tag {4} +$$ + +with $\theta$ indicating trainable parameters, and $X_{s, < i}, X_{a, < i}$ representing the sequences of source and preceding target tokens. The training objective is maximizing the log-likelihood of distribution: + +$$ +\mathcal {L} _ {L M} = - \sum_ {i = 0} ^ {L _ {t} - 1} \log p _ {\theta} \left(x _ {a} ^ {i} \mid X _ {v}, X _ {s, < i}, X _ {a, < i}\right). \tag {5} +$$ + +By optimizing this objective, MotionChain captures the complex interrelations among images, motion, and text, facilitating accurate target "word" generation. + +During the inference phase, target tokens are recursively sampled from the model's predicted distribution $p_{\theta}\left(\hat{x}_a^i \mid X_v, X_s, \hat{X}_{a, < i}\right)$ , ceasing with the appearance of a special end token. This strategy facilitates a step-by-step target sequence generation, where each token's probability is conditioned on all previous turns' sources and targets and current source input. + +# 3.4 Training Strategy + +To facilitate the integration of image and motion comprehension within the language modeling context, we adopt a 3-stage training strategy. (1) The initial stage involves pre-training the motion tokenizer on a corpus of human motion data, in line with [32]. This process establishes the motion vocabulary $V_{m}$ , which serves as a foundation for encoding human motions as a series of discrete tokens. (2) Subsequently, the motion tokenizer remains frozen while we connect the visual tokenizer to the language model framework. This integration is supported by a suite of supervised objectives, including text-to-motion, motion-to-text, and image-based motion generation, aiming to learn the intricate relationships between images, motion, and language. (3) The final stage involves instruction + +tuning, and refines the model's capabilities through the application of prompt-based instructions. These instructions are framed within multi-turn conversation sequences, as detailed in Sec. 3.3, to expanded range of motion-related tasks. + +Training of Motion Tokenizer. The initial step involves training the motion tokenizer, guided by the loss objective in Equation 2. This stage enables the tokenizer to represent human motion sequences $\hat{x}^{1:L}$ as discrete motion tokens, a key step for merging motion data with textual information seamlessly. Once optimized, the motion tokenizer remains frozen. + +Motion-language Pre-training Stage. Leveraging recent developments in language modeling [14, 74, 93, 94, 122] pre-trained on natural language datasets and then fine-tuned with instruction-based phrasing [14, 62]. To augment the model's ability to discern relationships between images and human motions, we first pre-train our MotionChain using a mix of language, image, and motion datasets. Following the stage 1 training of the motion tokenizer, we have established a unified motion-language vocabulary $V = V_{t}, V_{m}$ , capable of representing motions in discrete token form. Moreover, we maintain the visual encoder's weights in the visual tokenizer as fixed, while the linear projection weight $W$ is jointly optimized with the language model. During this stage, the model undertakes three fundamental single-turn modality translation tasks: text-to-motion, motion-to-text, and image-conditioned motion generation, as outlined in Sec. 3.1. The primary objective is to maximize the likelihood of the model according to the loss function specified in Eq. (5), thereby letting the model understand the relationship between language, vision conditions, and motions. + +Instruction Tuning Stage. As described in Sec. 3.1, we construct a multimodal, multi-task, and multi-turn motion conversation dataset by augmenting existing text-to-motion [23] and human mesh reconstruction datasets [5] with targeted instructional prompts and leverage the capabilities of LLMs [62] and the text-motion retrieval model [68] for motion reasoning and editing tasks. The efficacy of instruction tuning, as evidenced across language models [14, 50, 62, 122], is well-established, yielding enhancements in model performance across a wide range of tasks. After instruction tuning, MotionChain can handle more motion-related tasks including the proficient handling of previously unseen tasks + +# 4 Experiments + +We evaluate the proposed MotionChain encompasses comprehensive comparisons across both one-turn motion-related tasks and multi-turn motion generation tasks. Firstly, we provide details of the dataset settings, evaluation criteria, and implementation details as specified in Sec. 4.1. Subsequently, comparative analyses are presented, focusing on the motion reasoning task (Sec. 4.2) and the temporal motion composition task (Sec. 4.3). In Sec. 4.4, we evaluate the choice of motion composition technique and different architectures of vision tokenizer. + +# 4.1 Experimental Setup + +Datasets. For one-turn motion reasoning tasks, the study employs our proposed multi-modal multi-turn conversation dataset upon HumanML3D [23] with 44,970 sequence-level textual descriptions for 14,616 motion sequences obtained from AMASS [57] and HumanAct12 [25]. The datasets are divided into training, testing, and validation sets with a ratio of $0.8:0.15:0.05$ . To evaluate the multi-turn motion generation task, we focus on BABEL [71] that provides textual descriptions for the motions in the AMASS [57] with annotated segments that overlap in each sequence, which allows evaluating generation of a sequence of motion or actions. We adopt the processed text labels by [3] and motion representation of HumanML3D [23] which combines joint velocities, positions, and rotations. Following [3] we consider pairs of actions for simplicity but Motion-Chain applies to a sequence of actions or motion of arbitrary length. For the image-conditioned motion generation task, we mainly focus on BEDLAM [5], a large synthetic dataset of realistic moving 3D humans containing more than 200 subjects and 380K frames video and motion pair. + +Evaluation Metrics are summarized as four parts. (1) Motion quality: We adopt Frechet Inception Distance (FID) as the primary metric, FID quantifies the divergence in feature distributions between generated and actual motion sequences. Utilizing feature extractors from prior studies [23,49,68], FID measures the distance of feature distributions between the generated and real motions. Following [5,32,33,65,102], we also adopt MPJPE, PA-MPJPE to measure global and local errors in millimeters and ACCL for acceleration errors, to evaluate the quality of the reconstructed motions. (2) Motion Diversity: Utilizing the Diversity (DIV) metric, we calculate variance across motion features to evaluate generation diversity. (3) Text matching: The precision of text-to-motion matches is quantified by the R Precision metric, based on the feature evaluator [23,49,68], and includes an analysis of Top 1/2/3 retrieval accuracy. The Multi-modal Distance (MM Dist) quantifies the semantic gap between motions and texts. (4) Linguistic quality: We follow [24] utilizing linguistic metrics from natural language studies, including BLUE [63], Rouge [48], Cider [97], and BertScore [120] to evaluate the quality of generated motion captions. More detailed benchmark information is provided in the supplementary materials. + +Implementation Details. We set the codebook of the motion tokenizer as $K \in \mathbb{R}^{512 \times 1024}$ for most experiments. The motion encoder, denoted as $\mathcal{E}_{\mathcal{M}}$ , integrates a temporal downsampling rate, $l = 4$ . Our vision tokenizer incorporates a frozen Vision Transformer (ViT-L/14) [73] as visual encoder for most experiments. Additionally, for comprehensive ablation studies, we explored the use of both a frozen vision encoder and a Q-former from BLIP-2 [43] as a vision tokenizer. We mainly utilize Flan-T5-base [14] as the underlying architecture for our language model. Moreover, all our models employ the AdamW [53] optimizer with $[\beta_1, \beta_2] = [0.9, 0.99]$ for training. The motion tokenizers are trained to utilize a $10^{-4}$ learning rate employing cosine annealing scheduler and a 256 mini-batch size. Our language models based on Flan-T5-base [14] have a $10^{-4}$ learning rate with cosine annealing scheduler and 16 mini-batch sizes in both the + +Table 1: Comparison of motion reasoning on the test set of our conversation dataset. Our proposed MotionChain is fine-tuned on motion reasoning tasks while other methods' results are generated by their pre-trained weight. $Length_{avg}$ represents the average words in generated answers to all questions. We adopt metrics commonly used in natural language processing tasks for evaluation. + +
MethodsParamsLengthavgBleu@1↑Bleu@4↑Rouge↑Cider↑BertScore↑
Flan-t5-base [14]250M8.344.641.7815.3215.933.45
Flan-t5-large [14]780M11.9512.184.8322.8115.0214.19
Flan-t5-xl [14]3B9.098.544.0124.8915.0318.34
Llama-2-7b [94]7B130.8411.123.6719.141.046.81
Vicuna-1.5-7b [122]7B71.4919.277.3925.755.4419.05
Vicuna-1.5-13b [122]13B84.7417.206.5324.187.7718.00
MotionChain (Ours)573M22.1737.9219.1938.0524.5332.24
+ +pre-train stage and the instruction tuning stage. The motion tokenizer undergoes 10000 epochs of training, while the language model undergoes 500 epochs during the pre-train stage and another 50 epochs during the instruction tuning stage. Most models are trained on 8 Tesla V100 GPUs. + +# 4.2 Comparisons on Motion Reasoning. + +In Sec. 3.1, we introduce a multi-modal motion conversation dataset, enriched with motion reasoning data facilitated by ChatGPT [62]. This task evaluates the model's reasoning capabilities with motion reasoning tasks, where a motion sequence or its corresponding textual descriptions serve as inputs. Our evaluation compares our MotionChain, which integrates motion perception, against contemporary Large Language Models (LLMs) that possess solely textual processing capabilities. The compared LLMs are assessed using their original pretrained weight. Results in Tab. 1, illustrate that MotionChain exhibits superior motion reasoning proficiency, benefiting from its integrated motion perception. + +# 4.3 Comparisons on Temporal Composition. + +The temporal motion composition task involves generating a continuous motion sequence from two actions in a time series. We conducted our experiments following the settings in TEACH [3] and used the Amass [57] subset BABEL [71] validation set. Additionally, we processed the motion in AMASS [57] into the format proposed by HumanML3D [23] and trained our MotionChain on the action-to-motion task. To compare with TEACH, we initially used an officially provided pre-trained model to sample motion on the validation set 20 times. Subsequently, we post-processed their motion into the HumanML3D format, represented in SMPL [52]. The performance of our MotionChain is summarized in Table 2. As evaluating generative models quantitatively is challenging, we also provide qualitative comparisons in the supplementary materials. + +Table 2: Comparison of temporal motion composition on Babel [71]. We evaluate the state-of-the-art motion temporal composition method Teach [3] under the $95\%$ confidence interval from 20 times running. (cf. Sec. 4.1 for notations.) + +
MethodsDiversityMPJPE↓PA-MPJPE↓ACCL↓
Real15.74±.149---
Teach [3]27.11±.159979.21±.215933.32±.25423.02±.018
MotionChain (Ours)43.25±.159276.05±6.7253.72±.5807.11±0.100
+ +Table 3: Evaluation of motion composition methods on HumanML3D [23]. Here Independent, Past-condition, and Tokens-joint stand for different motion composition variants during multi-turn motion conversation, as illustrated in Fig. 4. + +
MethodMPJPE↓PA-MPJPE↓ACCL↓Diversity
Independent350.79102.9711.406.47
Past-condition232.4646.156.186.01
Tokens-joint108.7718.852.265.56
+ +# 4.4 Ablation Studies + +MotionChain enables multi-modal motion conversation using two main techniques. The first technique involves generating a smooth sequence of motions by concatenating motion tokens which are then decoded back to motion by motion decoder $\mathcal{D}_{\mathcal{M}}$ . The second technique involves processing multi-modal visual input through a vision tokenizer, which consists of a frozen vision encoder and a trainable linear projection. To evaluate the effectiveness of these two designs, we compare them with other variants. For a more comprehensive analysis, detailed ablation studies can be found in the supplementary materials. + +Motion Composition Mechanism Apart from the jointly token concatenating mechanism, we also evaluate the performance of temporal motion composition through the other motion temporal composition variants Motion-cat: concatenating the motion in final motion level rather than token level. Experimental results in Tab. 3 show that jointly concatenating motion tokens achieved remarkable performance compared to the other variants. For further information regarding the implementation of the aforementioned vision tokenizer, please refer to the supplementary materials. + +Image Tokenizer Architecture. MotionChain connects the frozen vision encoder to the language model through a linear layer. However, previous vision-language [1,43] works also demonstrate the effectiveness of other kinds of visual-aligning modules. Here we consider the other two vision tokenizer variants: (a) inspired by [1,10,31], we introduce a perceiver module that incorporates a transformer receiving a predefined number of latent input queries. These queries cross-attend to the visual features, enabling effective information exchange. (b) We directly adopt the pre-trained Q-former from BLIP-2 [43] to align visual + +Table 4: Evaluation of vision tokenizer architecture on Bedlam [49]. We implement three different architectures, including Q-former, Perceiver, and Linear. We evaluate these results with the metrics in motion reconstruction. Additional information regarding the implementation is in the supplementary materials. (cf. Tab. 2 for notations.) + +
ArchitectureFirst-frameLast-frame
MPJPE ↓PA-MPJPE ↓MPJPE ↓PA-MPJPE ↓
Q-former195.4986.56134.7357.17
Perceiver185.6199.21134.8957.58
Linear144.3776.48133.7356.73
+ +inputs with the language model. We evaluate the different architectures under the single human image as the first frame condition and the last frame condition separately. Experimental results in Tab. 4 show that a lightweight linear projection is sufficient for comprehending the human pose from visual input. Additional details about the implementation of the above vision tokenizer can be found in the supplements. + +# 5 Conclusion and Limitation + +Limitation. As the trial to explore conversational human motion generation with visual language models, the proposed MotionChain still has limitations as follows. MotionChain utilizes indeterministic generative models, similar to other language models, but other traditional or neural motion controllers [86, 87] are mostly deterministic and sensitive to control signals. Besides, our method can only generate motion on articulated human bodies, excluding many other human parts such as faces [9, 34, 72] and hands [46, 46, 47, 47, 78]. Although we utilize vision, language, and motion as multimodal conditional inputs akin to human perception, MotionChain is still restricted to the collision signals for human-object and human-scene interactions [35, 82, 112]. + +Conclusion. We summarize the proposed MotionChain as a conversational human motion controller to generate continuous and long-term human motion through multimodal prompts. Compared to these one-turn motion generation methods [32, 92, 117], our MotionChain produces more contextually rich generation and can achieve the step-by-step process of human task execution for humanoid robotics and game agents. By leveraging large-scale language, vision-language, and vision-motion data to assist motion-related generation tasks, MotionChain thus comprehends each instruction in multi-turn conversation and generates human motions followed by these prompts. Extensive experiments validate the efficacy of MotionChain, demonstrating state-of-the-art performance in conversational motion generation, as well as more intuitive manners of controlling and interacting with virtual humans. + +# Acknowledgements + +This work is supported by National Natural Science Foundation of China (No. 62101137), National Key Research and Development Program of China (No. 2022ZD0160101), Shanghai Natural Science Foundation (No. 23ZR1402900), Shanghai Municipal Science and Technology Major Project (No.2021SHZDZX0103). The computations in this research were performed using the CFFF platform of Fudan University. + +# References + +1. Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., et al.: Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems 35, 23716-23736 (2022) +2. Anderson, P., Wu, Q., Teney, D., Bruce, J., Johnson, M., Sünderhauf, N., Reid, I., Gould, S., Van Den Hengel, A.: Vision-and-language navigation: Interpreting visually-grounded navigation instructions in real environments. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3674-3683 (2018) +3. Athanasiou, N., Petrovich, M., Black, M.J., Varol, G.: Teach: Temporal action compositions for 3d humans. In: International Conference on 3D Vision (3DV) (September 2022) +4. Bazavan, E.G., Zanfir, A., Zanfir, M., Freeman, W.T., Sukthankar, R., Sminchisescu, C.: Hspace: Synthetic parametric humans animated in complex environments. arXiv preprint arXiv:2112.12867 (2021) +5. Black, M.J., Patel, P., Tesch, J., Yang, J.: Bedlam: A synthetic dataset of bodies exhibiting detailed lifelike animated motion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8726-8737 (2023) +6. Borsos, Z., Marinier, R., Vincent, D., Kharitonov, E., Pietquin, O., Sharifi, M., Roblek, D., Teboul, O., Grangier, D., Tagliasacchi, M., et al.: Audiolm: a language modeling approach to audio generation. IEEE/ACM transactions on audio, speech, and language processing 31, 2523-2533 (2023) +7. Brooks, T., Holynski, A., Efros, A.A.: Instructpix2pix: Learning to follow image editing instructions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18392-18402 (2023) +8. Cai, Z., Zhang, M., Ren, J., Wei, C., Ren, D., Lin, Z., Zhao, H., Yang, L., Loy, C.C., Liu, Z.: Playing for 3d human recovery. arXiv preprint arXiv:2110.07588 (2021) +9. Cao, X., Chen, Z., Chen, A., Chen, X., Li, S., Yu, J.: Sparse photometric 3d face reconstruction guided by morphable models. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 4635-4644 (2018) +10. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: European conference on computer vision. pp. 213-229. Springer (2020) +11. Changpinyo, S., Sharma, P., Ding, N., Soricut, R.: Conceptual 12m: Pushing web-scale image-text pre-training to recognize long-tail visual concepts. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3558-3568 (2021) + +12. Chen, S., Chen, X., Zhang, C., Li, M., Yu, G., Fei, H., Zhu, H., Fan, J., Chen, T.: Ll3da: Visual interactive instruction tuning for omni-3d understanding reasoning and planning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 26428-26438 (2024) +13. Choudhury, R., Kitani, K.M., Jeni, L.A.: Tempo: Efficient multi-view pose estimation, tracking, and forecasting. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 14750-14760 (2023) +14. Chung, H.W., Hou, L., Longpre, S., Zoph, B., Tay, Y., Fedus, W., Li, E., Wang, X., Dehghani, M., Brahma, S., et al.: Scaling instruction-finetuned language models. arXiv preprint arXiv:2210.11416 (2022) +15. Clavet, S.: Motion matching and the road to next-gen animation. In: Proc. of GDC. vol. 2, p. 9 (2016) +16. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018) +17. Gilardi, F., Alizadeh, M., Kubli, M.: Chatgpt outperforms crowd-workers for text-annotation tasks. arXiv preprint arXiv:2303.15056 (2023) +18. Girdhar, R., El-Nouby, A., Liu, Z., Singh, M., Alwala, K.V., Joulin, A., Misra, I.: Imagebind: One embedding space to bind them all. arXiv preprint arXiv:2305.05665 (2023) +19. Goel, S., Pavlakos, G., Rajasegaran, J., Kanazawa, A., Malik, J.: Humans in 4d: Reconstructing and tracking humans with transformers. arXiv preprint arXiv:2305.20091 (2023) +20. Goutsu, Y., Inamura, T.: Linguistic descriptions of human motion with generative adversarial seq2seq learning. In: 2021 IEEE International Conference on Robotics and Automation (ICRA). pp. 4281-4287. IEEE (2021) +21. Guler, R.A., Kokkinos, I.: Holopose: Holistic 3d human reconstruction in-the-wild. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10884-10894 (2019) +22. Guo, C., Mu, Y., Javed, M.G., Wang, S., Cheng, L.: Momask: Generative masked modeling of 3d human motions. arXiv preprint arXiv:2312.00063 (2023) +23. Guo, C., Zou, S., Zuo, X., Wang, S., Ji, W., Li, X., Cheng, L.: Generating diverse and natural 3d human motions from text. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5152-5161 (June 2022) +24. Guo, C., Zuo, X., Wang, S., Cheng, L.: Tm2t: Stochastic and tokenized modeling for the reciprocal generation of 3d human motions and texts. In: ECCV (2022) +25. Guo, C., Zuo, X., Wang, S., Zou, S., Sun, Q., Deng, A., Gong, M., Cheng, L.: Action2motion: Conditioned generation of 3d human motions. In: Proceedings of the 28th ACM International Conference on Multimedia. pp. 2021-2029 (2020) +26. Guo, Z., Zhang, R., Zhu, X., Tang, Y., Ma, X., Han, J., Chen, K., Gao, P., Li, X., Li, H., et al.: Point-bind & point-llm: Aligning point cloud with multimodality for 3d understanding, generation, and instruction following. arXiv preprint arXiv:2309.00615 (2023) +27. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems 33, 6840-6851 (2020) +28. Hong, Y., Zhen, H., Chen, P., Zheng, S., Du, Y., Chen, Z., Gan, C.: 3d-lm: Injecting the 3d world into large language models. arXiv preprint arXiv:2307.12981 (2023) + +29. Huang, S., Dong, L., Wang, W., Hao, Y., Singhal, S., Ma, S., Lv, T., Cui, L., Mohammed, O.K., Liu, Q., et al.: Language is not all you need: Aligning perception with language models. arXiv preprint arXiv:2302.14045 (2023) +30. Ionescu, C., Papava, D., Olaru, V., Sminchisescu, C.: Human3. 6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. IEEE transactions on pattern analysis and machine intelligence 36(7), 1325-1339 (2013) +31. Jaegle, A., Gimeno, F., Brock, A., Vinyals, O., Zisserman, A., Carreira, J.: Perceiver: General perception with iterative attention. In: International conference on machine learning. pp. 4651-4664. PMLR (2021) +32. Jiang, B., Chen, X., Liu, W., Yu, J., Yu, G., Chen, T.: Motiongpt: Human motion as a foreign language. arXiv preprint arXiv:2306.14795 (2023) +33. Kanazawa, A., Black, M.J., Jacobs, D.W., Malik, J.: End-to-end recovery of human shape and pose. In: Computer Vision and Pattern Regognition (CVPR) (2018) +34. Karras, T., Aila, T., Laine, S., Herva, A., Lehtinen, J.: Audio-driven facial animation by joint end-to-end learning of pose and emotion. ACM Transactions on Graphics (TOG) 36(4), 1-12 (2017) +35. Karunratanakul, K., Preechakul, K., Suwajanakorn, S., Tang, S.: Guided motion diffusion for controllable human motion synthesis. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2151-2162 (2023) +36. Kocabas, M., Athanasiou, N., Black, M.J.: Vibe: Video inference for human body pose and shape estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2020) +37. Kocabas, M., Huang, C.H.P., Tesch, J., Müller, L., Hilliges, O., Black, M.J.: Spec: Seeing people in the wild with an estimated camera. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 11035-11045 (2021) +38. Kovar, L., Gleicher, M., Pighin, F.: Motion graphs. In: Seminal Graphics Papers: Pushing the Boundaries, Volume 2, pp. 723-732 (2023) +39. Kudo, T., Richardson, J.: Sentencepiece: A simple and language independent subword tokenizer and detokenizer for neural text processing. arXiv preprint arXiv:1808.06226 (2018) +40. Lee, T., Moon, G., Lee, K.M.: Multiact: Long-term 3d human motion generation from multiple action labels. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 1231-1239 (2023) +41. Lee, Y., Wampler, K., Bernstein, G., Popovic, J., Popovic, Z.: Motion fields for interactive character locomotion. In: ACM SIGGRAPH Asia 2010 papers, pp. 1-8 (2010) +42. Levine, S., Wang, J.M., Haraux, A., Popovic, Z., Koltun, V.: Continuous character control with low-dimensional embeddings. ACM Transactions on Graphics (TOG) 31(4), 1-10 (2012) +43. Li, J., Li, D., Savarese, S., Hoi, S.: Blip-2: Bootstrapping language-image pretraining with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597 (2023) +44. Li, J., Li, D., Xiong, C., Hoi, S.: Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In: International Conference on Machine Learning. pp. 12888-12900. PMLR (2022) +45. Li, M., Chen, X., Zhang, C., Chen, S., Zhu, H., Yin, F., Yu, G., Chen, T.: M3dbench: Let's instruct large models with multi-modal 3d prompts. arXiv preprint arXiv:2312.10763 (2023) + +46. Li, Y., Wu, M., Zhang, Y., Xu, L., Yu, J.: Piano: A parametric hand bone model from magnetic resonance imaging. arXiv preprint arXiv:2106.10893 (2021) +47. Li, Y., Zhang, L., Qiu, Z., Jiang, Y., Li, N., Ma, Y., Zhang, Y., Xu, L., Yu, J.: Nimble: a non-rigid hand model with bones and muscles. ACM Transactions on Graphics (TOG) 41(4), 1-16 (2022) +48. Lin, C.Y.: Rouge: A package for automatic evaluation of summaries. In: Text summarization branches out. pp. 74-81 (2004) +49. Lin, J., Zeng, A., Lu, S., Cai, Y., Zhang, R., Wang, H., Zhang, L.: Motion-x: A large-scale 3d expressive whole-body human motion dataset. arXiv preprint arXiv:2307.00818 (2023) +50. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. arXiv preprint arXiv:2304.08485 (2023) +51. Liu, X., Yan, H., Zhang, S., An, C., Qiu, X., Lin, D.: Scaling laws of rope-based extrapolation. arXiv preprint arXiv:2310.05209 (2023) +52. Loper, M., Mahmood, N., Romero, J., Pons-Moll, G., Black, M.J.: Smpl: A skinned multi-person linear model. ACM Trans. Graph. 34(6), 248:1-248:16 (Oct 2015) +53. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101 (2017) +54. Lu, C., Yin, F., Chen, X., Liu, W., Chen, T., Yu, G., Fan, J.: A large-scale outdoor multi-modal dataset and benchmark for novel view synthesis and implicit scene reconstruction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7557-7567 (2023) +55. Lu, S., Chen, L.H., Zeng, A., Lin, J., Zhang, R., Zhang, L., Shum, H.Y.: Humantomato: Text-aligned whole-body motion generation. arXiv preprint arXiv:2310.12978 (2023) +56. Ma, H., Li, J., Hosseini, R., Tomizuka, M., Choi, C.: Multi-objective diverse human motion prediction with knowledge distillation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8161-8171 (2022) +57. Mahmood, N., Ghorbani, N., Troje, N.F., Pons-Moll, G., Black, M.J.: Amass: Archive of motion capture as surface shapes. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (October 2019) +58. Mehta, D., Rhodin, H., Casas, D., Fua, P., Sotnychenko, O., Xu, W., Theobalt, C.: Monocular 3d human pose estimation in the wild using improved cnn supervision. In: International Conference on 3D Vision (3DV) (2017). https://doi.org/10.1109/3dv.2017.00064, http://gvv.mpi-inf.mpg.de/3dhp_dataset +59. Min, J., Chai, J.: Motion graphs++ a compact generative model for semantic motion analysis and synthesis. ACM Transactions on Graphics (TOG) 31(6), 1-12 (2012) +60. OpenAI: Gpt-4 technical report (2023) +61. Ordonez, V., Kulkarni, G., Berg, T.: Im2text: Describing images using 1 million captioned photographs. Advances in neural information processing systems 24 (2011) +62. Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., et al.: Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems 35, 27730-27744 (2022) +63. Papineni, K., Roukos, S., Ward, T., Zhu, W.J.: Bleu: a method for automatic evaluation of machine translation. In: Proceedings of the 40th annual meeting of the Association for Computational Linguistics. pp. 311-318 (2002) + +64. Patel, P., Huang, C.H.P., Tesch, J., Hoffmann, D.T., Tripathi, S., Black, M.J.: Agora: Avatars in geography optimized for regression analysis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13468-13478 (2021) +65. Pavlakos, G., Choutas, V., Ghorbani, N., Bolkart, T., Osman, A.A.A., Tzionas, D., Black, M.J.: Expressive body capture: 3d hands, face, and body from a single image. In: Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR) (2019) +66. Petrovich, M., Black, M.J., Varol, G.: Action-conditioned 3D human motion synthesis with transformer VAE. In: International Conference on Computer Vision (ICCV) (2021) +67. Petrovich, M., Black, M.J., Varol, G.: TEMOS: Generating diverse human motions from textual descriptions. In: European Conference on Computer Vision (ECCV) (2022) +68. Petrovich, M., Black, M.J., Varol, G.: Tmr: Text-to-motion retrieval using contrastive 3d human motion synthesis. arXiv preprint arXiv:2305.00976 (2023) +69. Plappert, M., Mandery, C., Asfour, T.: The kit motion-language dataset. Big Data 4(4), 236-252 (dec 2016). https://doi.org/10.1089/big.2016.0028, http://dx.doi.org/10.1089/big.2016.0028 +70. Plappert, M., Mandery, C., Asfour, T.: Learning a bidirectional mapping between human whole-body motion and natural language using deep recurrent neural networks. Robotics and Autonomous Systems 109, 13-26 (2018) +71. Punnakkal, A.R., Chandrasekaran, A., Athanasiou, N., Quiros-Ramirez, A., Black, M.J.: BABEL: Bodies, action and behavior with english labels. In: Proceedings IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR). pp. 722-731 (Jun 2021) +72. Qiu, Z., Li, Y., He, D., Zhang, Q., Zhang, L., Zhang, Y., Wang, J., Xu, L., Wang, X., Zhang, Y., et al.: Sculptor: Skeleton-consistent face creation using a learned parametric generator. ACM Transactions on Graphics (TOG) 41(6), 1-17 (2022) +73. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International Conference on Machine Learning. pp. 8748-8763. PMLR (2021) +74. Raffel, C., Shazeer, N., Roberts, A., Lee, K., Narang, S., Matena, M., Zhou, Y., Li, W., Liu, P.J.: Exploring the limits of transfer learning with a unified text-to-text transformer. The Journal of Machine Learning Research 21(1), 5485-5551 (2020) +75. Rajasegaran, J., Pavlakos, G., Kanazawa, A., Malik, J.: Tracking people by predicting 3d appearance, location and pose. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2740-2749 (2022) +76. Razavi, A., Van den Oord, A., Vinyals, O.: Generating diverse high-fidelity images with vq-vae-2. Advances in neural information processing systems 32 (2019) +77. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2022), https://github.com/CompVis/latent-diffusionhttps://arxiv.org/abs/2112.10752 +78. Romero, J., Tzionas, D., Black, M.J.: Embodied hands: Modeling and capturing hands and bodies together. arXiv preprint arXiv:2201.02610 (2022) +79. Rose, C., Cohen, M.F., Bodenheimer, B.: Verbs and adverbs: Multidimensional motion interpolation. IEEE Computer Graphics and Applications 18(5), 32-40 (1998) + +80. Rubenstein, P.K., Asawaroengchai, C., Nguyen, D.D., Bapna, A., Borsos, Z., Quitry, F.d.C., Chen, P., Badawy, D.E., Han, W., Kharitonov, E., et al.: Audiopalm: A large language model that can speak and listen. arXiv preprint arXiv:2306.12925 (2023) +81. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems 35, 25278-25294 (2022) +82. Shafir, Y., Tevet, G., Kapon, R., Bermano, A.H.: Human motion diffusion as a generative prior. arXiv preprint arXiv:2303.01418 (2023) +83. Siddiqui, Y., Alliegro, A., Artemov, A., Tommasi, T., Sirigatti, D., Rosov, V., Dai, A., Nießner, M.: Meshgpt: Generating triangle meshes with decoder-only transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19615-19625 (2024) +84. Siyao, L., Yu, W., Gu, T., Lin, C., Wang, Q., Qian, C., Loy, C.C., Liu, Z.: Bailando: 3d dance generation by actor-critic gpt with choreographic memory. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11050-11059 (2022) +85. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502 (2020) +86. Starke, S., Mason, I., Komura, T.: Deepphase: periodic autoencoders for learning motion phase manifolds. ACM Transactions on Graphics (TOG) 41(4), 1-13 (2022) +87. Starke, S., Zhang, H., Komura, T., Saito, J.: Neural state machine for character-scene interactions. ACM Trans. Graph. 38(6), 209-1 (2019) +88. Starke, S., Zhao, Y., Zinno, F., Komura, T.: Neural animation layering for synthesizing martial arts movements. ACM Transactions on Graphics (TOG) 40(4), 1-16 (2021) +89. Szot, A., Clegg, A., Undersander, E., Wijmans, E., Zhao, Y., Turner, J., Maestre, N., Mukadam, M., Chaplot, D.S., Maksymets, O., et al.: Habitat 2.0: Training home assistants to rearrange their habitat. Advances in Neural Information Processing Systems 34, 251-266 (2021) +90. Takano, W., Nakamura, Y.: Statistical mutual conversion between whole body motion primitives and linguistic sentences for human motions. The International Journal of Robotics Research 34(10), 1314-1328 (2015) +91. Tevet, G., Gordon, B., Hertz, A., Bermano, A.H., Cohen-Or, D.: Motionclip: Exposing human motion generation to clip space. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXII. pp. 358-374. Springer (2022) +92. Tevet, G., Raab, S., Gordon, B., Shafir, Y., Bermano, A.H., Cohen-Or, D.: Human motion diffusion model. arXiv preprint arXiv:2209.14916 (2022) +93. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., et al.: Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 (2023) +94. Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi, A., Babaei, Y., Bashlykov, N., Batra, S., Bhargava, P., Bhosale, S., et al.: Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288 (2023) +95. Van Den Oord, A., Vinyals, O., et al.: Neural discrete representation learning. Advances in neural information processing systems 30 (2017) + +96. Varol, G., Romero, J., Martin, X., Mahmood, N., Black, M., Laptev, I., Schmid, C.: Learning from synthetic humans. In: Computer Vision and Pattern Recognition (CVPR) (2017) +97. Vedantam, R., Lawrence Zitnick, C., Parikh, D.: Cider: Consensus-based image description evaluation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4566-4575 (2015) +98. Wang, C.: T2m-hifigpt: Generating high quality human motion from textual descriptions with residual discrete representations. arXiv preprint arXiv:2312.10628 (2023) +99. Wang, W., Zhe, X., Chen, H., Kang, D., Li, T., Chen, R., Bao, L.: Neural marionette: A transformer-based multi-action human motion synthesis system. arXiv preprint arXiv:2209.13204 (2022) +100. Wu, C., Yin, S., Qi, W., Wang, X., Tang, Z., Duan, N.: Visual chatgpt: Talking, drawing and editing with visual foundation models. arXiv preprint arXiv:2303.04671 (2023) +101. Wu, Y.C., Gebru, I.D., Markovic, D., Richard, A.: Audiodec: An open-source streaming high-fidelity neural audio codec. In: ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 1-5. IEEE (2023) +102. Xin, C., Jiang, B., Liu, W., Huang, Z., Fu, B., Chen, T., Yu, J., Yu, G.: Executing your commands via motion diffusion in latent space. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2023) +103. Xiong, W., Liu, J., Molybog, I., Zhang, H., Bhargava, P., Hou, R., Martin, L., Rungta, R., Sankararaman, K.A., Oguz, B., et al.: Effective long-context scaling of foundation models. arXiv preprint arXiv:2309.16039 (2023) +104. Xu, H., Ghosh, G., Huang, P.Y., Okhonko, D., Aghajanyan, A., Metze, F., Zettlemoyer, L., Feichtenhofer, C.: Videoclip: Contrastive pre-training for zero-shot video-text understanding. arXiv preprint arXiv:2109.14084 (2021) +105. Yamada, T., Matsunaga, H., Ogata, T.: Paired recurrent autoencoders for bidirectional translation between robot actions and linguistic descriptions. IEEE Robotics and Automation Letters 3(4), 3441-3448 (2018) +106. Yao, H., Song, Z., Zhou, Y., Ao, T., Chen, B., Liu, L.: Moconvq: Unified physics-based motion control via scalable discrete representations (2023) +107. Ye, Q., Xu, H., Xu, G., Ye, J., Yan, M., Zhou, Y., Wang, J., Hu, A., Shi, P., Shi, Y., et al.: mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178 (2023) +108. Yin, F., Chen, X., Zhang, C., Jiang, B., Zhao, Z., Fan, J., Yu, G., Li, T., Chen, T.: Shapeqpt: 3d shape generation with a unified multi-modal language model. arXiv preprint arXiv:2311.17618 (2023) +109. Yin, F., Liu, W., Huang, Z., Cheng, P., Chen, T., Yu, G.: Coordinates are not lonely-codebook prior helps implicit neural 3d representations. Advances in Neural Information Processing Systems 35, 12705-12717 (2022) +110. Yu, J., Xu, Y., Koh, J.Y., Luong, T., Baid, G., Wang, Z., Vasudevan, V., Ku, A., Yang, Y., Ayan, B.K., et al.: Scaling autoregressive models for content-rich text-to-image generation. arXiv preprint arXiv:2206.10789 (2022) +111. Yuan, Y., Kitani, K.: Dlow: Diversifying latent flows for diverse human motion prediction. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part IX 16. pp. 346-364. Springer (2020) + +112. Yuan, Y., Song, J., Iqbal, U., Vahdat, A., Kautz, J.: PhysDiff: Physics-guided human motion diffusion model. In: IEEE International Conference on Computer Vision (ICCV) (October 2023) +113. Zhang, H., Li, X., Bing, L.: Video-llama: An instruction-tuned audio-visual language model for video understanding. arXiv preprint arXiv:2306.02858 (2023) +114. Zhang, H., Cao, J., Lu, G., Ouyang, W., Sun, Z.: Danet: Decompose-and-aggregate network for 3d human shape and pose estimation. In: Proceedings of the 27th ACM International Conference on Multimedia. pp. 935-944 (2019) +115. Zhang, H., Tian, Y., Zhang, Y., Li, M., An, L., Sun, Z., Liu, Y.: Pymaf-x: Towards well-aligned full-body model regression from monocular images. IEEE Transactions on Pattern Analysis and Machine Intelligence (2023) +116. Zhang, J., Zhang, Y., Cun, X., Huang, S., Zhang, Y., Zhao, H., Lu, H., Shen, X.: T2m-gpt: Generating human motion from textual descriptions with discrete representations. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023) +117. Zhang, M., Cai, Z., Pan, L., Hong, F., Guo, X., Yang, L., Liu, Z.: Motiondiffuse: Text-driven human motion generation with diffusion model. arXiv preprint arXiv:2208.15001 (2022) +118. Zhang, M., Guo, X., Pan, L., Cai, Z., Hong, F., Li, H., Yang, L., Liu, Z.: Remodiffuse: Retrieval-augmented motion diffusion model. arXiv preprint arXiv:2304.01116 (2023) +119. Zhang, Q., Song, J., Huang, X., Chen, Y., Liu, M.Y.: Diffcollage: Parallel generation of large content with diffusion models. In: 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 10188-10198. IEEE (2023) +120. Zhang, T., Kishore, V., Wu, F., Weinberger, K.Q., Artzi, Y.: Bertscore: Evaluating text generation with bert. arXiv preprint arXiv:1904.09675 (2019) +121. Zhang, Y., Black, M.J., Tang, S.: We are more than our joints: Predicting how 3d bodies move. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3372-3382 (2021) +122. Zheng, L., Chiang, W.L., Sheng, Y., Zhuang, S., Wu, Z., Zhuang, Y., Lin, Z., Li, Z., Li, D., Xing, E.P., Zhang, H., Gonzalez, J.E., Stoica, I.: Judging lvm-as-a-judge with mt-bench and chatbot arena (2023) +123. Zhou, C., Liu, P., Xu, P., Iyer, S., Sun, J., Mao, Y., Ma, X., Efrat, A., Yu, P., Yu, L., et al.: Lima: Less is more for alignment. Advances in Neural Information Processing Systems 36 (2024) +124. Zhu, D., Chen, J., Shen, X., Li, X., Elhoseiny, M.: Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592 (2023) +125. Zou, X., Dou, Z.Y., Yang, J., Gan, Z., Li, L., Li, C., Dai, X., Behl, H., Wang, J., Yuan, L., et al.: Generalized decoding for pixel, image, and language. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15116-15127 (2023) \ No newline at end of file diff --git a/2024/MotionChain_ Conversational Motion Controllers via Multimodal Prompts/images.zip b/2024/MotionChain_ Conversational Motion Controllers via Multimodal Prompts/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f15319f0ca754aca202bece222c1a8c0b961164b --- /dev/null +++ b/2024/MotionChain_ Conversational Motion Controllers via Multimodal Prompts/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d2289156b41811066b796e363bf48047aeb41175f843d1d46a094deab14638d6 +size 301715 diff --git a/2024/MotionChain_ Conversational Motion Controllers via Multimodal Prompts/layout.json b/2024/MotionChain_ Conversational Motion Controllers via Multimodal Prompts/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..1b8ab3b887cd22ac961efd143cab543a269449b7 --- /dev/null +++ b/2024/MotionChain_ Conversational Motion Controllers via Multimodal Prompts/layout.json @@ -0,0 +1,11599 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 135, + 111, + 479, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 111, + 479, + 148 + ], + "spans": [ + { + "bbox": [ + 135, + 111, + 479, + 148 + ], + "type": "text", + "content": "MotionChain: Conversational Motion Controllers via Multimodal Prompts" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 133, + 167, + 480, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 167, + 480, + 194 + ], + "spans": [ + { + "bbox": [ + 133, + 167, + 480, + 194 + ], + "type": "text", + "content": "Biao Jiang" + }, + { + "bbox": [ + 133, + 167, + 480, + 194 + ], + "type": "inline_equation", + "content": "^{1,2\\star}" + }, + { + "bbox": [ + 133, + 167, + 480, + 194 + ], + "type": "text", + "content": ", Xin Chen" + }, + { + "bbox": [ + 133, + 167, + 480, + 194 + ], + "type": "inline_equation", + "content": "^{2\\star\\star}" + }, + { + "bbox": [ + 133, + 167, + 480, + 194 + ], + "type": "text", + "content": ", Chi Zhang" + }, + { + "bbox": [ + 133, + 167, + 480, + 194 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 133, + 167, + 480, + 194 + ], + "type": "text", + "content": ", Fukun Yin" + }, + { + "bbox": [ + 133, + 167, + 480, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 133, + 167, + 480, + 194 + ], + "type": "text", + "content": ", Zhuoyuan Li" + }, + { + "bbox": [ + 133, + 167, + 480, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 133, + 167, + 480, + 194 + ], + "type": "text", + "content": ", Gang Yu" + }, + { + "bbox": [ + 133, + 167, + 480, + 194 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 133, + 167, + 480, + 194 + ], + "type": "text", + "content": ", and Jiayuan Fan" + }, + { + "bbox": [ + 133, + 167, + 480, + 194 + ], + "type": "inline_equation", + "content": "^{1\\star\\star\\star}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 266, + 201, + 347, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 266, + 201, + 347, + 213 + ], + "spans": [ + { + "bbox": [ + 266, + 201, + 347, + 213 + ], + "type": "text", + "content": "1 Fudan University" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 286, + 213, + 328, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 213, + 328, + 223 + ], + "spans": [ + { + "bbox": [ + 286, + 213, + 328, + 223 + ], + "type": "text", + "content": "2 Tencent" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 254, + 455, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 254, + 455, + 474 + ], + "spans": [ + { + "bbox": [ + 160, + 254, + 455, + 474 + ], + "type": "text", + "content": "Abstract. Recent advancements in language models have demonstrated their adeptness in conducting multi-turn dialogues and retaining conversational context. However, this proficiency remains largely unexplored in other multimodal generative models, particularly in human motion models. By integrating multi-turn conversations in controlling continuous virtual human movements, generative human motion models can achieve an intuitive and step-by-step process of human task execution for humanoid robotics, game agents, or other embodied systems. In this work, we present MotionChain, a conversational human motion controller that generates continuous and long-term human motion through multimodal prompts. Specifically, MotionChain consists of multi-modal tokenizers that transform various data types such as text, image, and motion, into discrete tokens, coupled with a Vision-Motion-aware Language model. By leveraging large-scale language, vision-language, and vision-motion data to assist motion-related generation tasks, MotionChain thus comprehends each instruction in multi-turn conversation and generates human motions followed by these prompts. Extensive experiments validate the efficacy of MotionChain, demonstrating state-of-the-art performance in conversational motion generation, as well as more intuitive manners of controlling and interacting with virtual humans." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 160, + 484, + 416, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 484, + 416, + 495 + ], + "spans": [ + { + "bbox": [ + 160, + 484, + 416, + 495 + ], + "type": "text", + "content": "Keywords: 3D Motion " + }, + { + "bbox": [ + 160, + 484, + 416, + 495 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 484, + 416, + 495 + ], + "type": "text", + "content": " Motion Generation " + }, + { + "bbox": [ + 160, + 484, + 416, + 495 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 484, + 416, + 495 + ], + "type": "text", + "content": " Text-to-Motion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 516, + 230, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 516, + 230, + 529 + ], + "spans": [ + { + "bbox": [ + 132, + 516, + 230, + 529 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 540, + 482, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 540, + 482, + 625 + ], + "spans": [ + { + "bbox": [ + 130, + 540, + 482, + 625 + ], + "type": "text", + "content": "The success of large language models (LLMs) [60, 62, 93, 94, 122] has sparked significant interest in the development of multi-modal language models. These models aim to transfer instruction-following and zero-shot abilities to other modalities tasks, such as image-language models [1, 50, 107, 124], video-language models [1, 43, 44, 113], and 3D-language models [12, 26, 28, 108]. However, a comprehensive model that can perceive visual input and generate continuous motion through multi-turn conversations has not yet been developed. Such a" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 132, + 632, + 408, + 643 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 408, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 408, + 643 + ], + "type": "text", + "content": "* Work done while Biao Jiang was a Research Intern with Tencent." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 643, + 196, + 654 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 196, + 654 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 196, + 654 + ], + "type": "text", + "content": "\\*\\* Project lead." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 126, + 654, + 235, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 126, + 654, + 235, + 665 + ], + "spans": [ + { + "bbox": [ + 126, + 654, + 235, + 665 + ], + "type": "text", + "content": "\\*\\*\\* Corresponding author." + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 119, + 478, + 200 + ], + "blocks": [ + { + "bbox": [ + 136, + 119, + 478, + 200 + ], + "lines": [ + { + "bbox": [ + 136, + 119, + 478, + 200 + ], + "spans": [ + { + "bbox": [ + 136, + 119, + 478, + 200 + ], + "type": "image", + "image_path": "86a343de4f67d818efac918a57e0a0271f3ff81233309de547b2138df43313c3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 213, + 482, + 278 + ], + "lines": [ + { + "bbox": [ + 130, + 213, + 482, + 278 + ], + "spans": [ + { + "bbox": [ + 130, + 213, + 482, + 278 + ], + "type": "text", + "content": "Fig. 1: MotionChain can interpret instructions from multi-turn conversations and generate human motions or textual answers based on text, motion, or image inputs. We provide the conversation results in image-conditioned motion generation (1st column), motion reasoning (second column), motion editing (third column), and motion translation (third column), with each subsequent turn informed by all previous conversations. Left-to-right represents the temporal order." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 304, + 480, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 304, + 480, + 327 + ], + "spans": [ + { + "bbox": [ + 130, + 304, + 480, + 327 + ], + "type": "text", + "content": "multi-modal model would have wide-ranging applications in fields like humanoid robotics, virtual assistants, game agents and so on." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 329, + 482, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 329, + 482, + 508 + ], + "spans": [ + { + "bbox": [ + 130, + 329, + 482, + 508 + ], + "type": "text", + "content": "Previous research on human motion has explored various tasks, including motion generation [23, 32, 66, 92, 102, 116], motion captioning [20, 24, 32], motion prediction [32, 56, 111, 121], and motion composition [3]. Recent works in text-to-motion [67, 92, 102, 117] have involved pre-trained language models [16, 73] for motion generation. For instance, TEMOS [67] employs BERT [16] text embeddings in an end-to-end transformer architecture, while MDM [92] and MLD [102] both utilize text embeddings from CLIP [73] during the conditional diffusion process. On the other hand, MotionCLIP [91] and TMR [68] focus on modeling the coupled relationship between motion and text description, and MotionGPT [32] introduces a motion-language model that represents human motion and language in one unified vocabulary. However, these above methods treat all tasks as a one-turn conditioned generation, lacking contextual understanding and multi-turn continuous generation abilities. Therefore, we construct a Vision-Motion language model, integrating multi-turn conversations and continuous human motions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": "Two crucial challenges need to be addressed in this conversational motion generation. The first challenge is to contextually generate human motion in a continuous manner, resembling the way real humans move. The second challenge is the scarcity of text-motion paired datasets compared to datasets with pairs of image-language [11, 81], image-pose [37, 58, 61, 64] and video-motion [4, 5, 8, 30, 96]. Fortunately, both human motion and language are sequential and can be continuously \"written\". Building upon this observation, we employ the general vision-language instruction-tuning approach [50, 123] to enable conversational motion generation and question-answering through multi-modal instructions. By integrating image, motion, and language data and encoding them into tokens, the relationship between these three modalities becomes more evident. Therefore, with the advent of large vision-motion and vision-language data, Vision-Motion-language pre-training can enhance the performance of motion-related tasks." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 279, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 279, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 279, + 102 + ], + "type": "text", + "content": "Biao Jiang, Xin Chen et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 482, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 482, + 342 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 482, + 342 + ], + "type": "text", + "content": "In this study, we introduce MotionChain, a comprehensive framework that integrates vision, motion, and language. MotionChain leverages large-scale vision-language data, vision-motion data, and pre-trained language models' strong language generation abilities to assist in motion-related generation tasks. To enable MotionChain to comprehend and generate human-like motions, we first train a motion-specific vector quantized variational autoencoder (VQ-VAE) model. This model constructs a \"motion vocabulary\" similar to the English word vocabulary and converts raw motion data into a sequence of motion tokens. To incorporate vision inputs into MotionChain, we then introduce a specialized vision tokenizer that connects a pre-trained vision encoder to the language model. This tokenizer converts image data into visual tokens within the language-motion \"words\" embedding space. These tokens are then processed by a pre-trained language model [14,74,93,94], which learns the relationship between image, motion and language. To enable conversational generation, we construct a multi-modal motion conversation dataset based on the existing text-motion dataset [23] and vision-motion dataset [5]. We then train the language model using our multimodal conversation dataset to learn the correlation between the three modalities. Extensive experiments demonstrate that MotionChain achieves state-of-the-art performance in multiple motion-related tasks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 343, + 482, + 428 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 343, + 482, + 428 + ], + "spans": [ + { + "bbox": [ + 130, + 343, + 482, + 428 + ], + "type": "text", + "content": "We summarize our contributions as follows: (1) We propose MotionChain, a unified vision-motion-language generative pre-trained model, which performs conversational generation tasks via multi-modal inputs with language models. (2) We introduce a motion composition technique, to generate 3D human motions following the temporal order of instructions. (3) We propose a multi-modal motion conversation benchmark, wherein MotionChain achieves competitive performance across diverse motion tasks." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 447, + 238, + 460 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 447, + 238, + 460 + ], + "spans": [ + { + "bbox": [ + 132, + 447, + 238, + 460 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 473, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 473, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 473, + 482, + 668 + ], + "type": "text", + "content": "Human Motion Modeling. There have been numerous attempts to model the relationship between 3D human motion and multiple modalities including incomplete motion [56, 111, 121], action [3, 25, 40, 66, 99, 102], text [22-24, 32, 55, 67, 82, 92, 117-119], image [19, 21, 33, 114, 115] and video [5, 13, 19, 36, 75]. Text-to-motion is one of the most important motion generation tasks, due to the user-friendly and convenient language input. MDM [92], MotionDiffuse [117] and MLD [102] proposes a diffusion-based generative model [27, 77, 85] to generate motions conditioned on different inputs. TM2T [24] and T2M-GPT [116] investigate a generative framework based on VQ-VAE [76, 95] and generative transformer for motion generation. Motion completion task generates motion conditioning on partial motions, such as classical motion prediction [56, 111, 121] or motion in-between [92], which generates the intermediate motion while the first and last parts are fixed. TEACH [3] proposes a past-conditioned transformer model that generates motion from a sequence of actions autoregressively. Apart from motion generation, there is also work investigating other modalities of generation from motion. Two statistical models [90] and recurrent networks [70, 105]" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 392, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 392, + 91, + 447, + 101 + ], + "type": "text", + "content": "MotionChain" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 140, + 115, + 480, + 252 + ], + "blocks": [ + { + "bbox": [ + 140, + 115, + 480, + 252 + ], + "lines": [ + { + "bbox": [ + 140, + 115, + 480, + 252 + ], + "spans": [ + { + "bbox": [ + 140, + 115, + 480, + 252 + ], + "type": "image", + "image_path": "2ece0ff6c53ff2ed370c8c580373637a05cb859bee80aa4474ce9fa6307beb0f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 262, + 482, + 319 + ], + "lines": [ + { + "bbox": [ + 130, + 262, + 482, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 262, + 482, + 319 + ], + "type": "text", + "content": "Fig. 2: Method overview: MotionChain consists of a motion tokenizer " + }, + { + "bbox": [ + 130, + 262, + 482, + 319 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_{\\mathcal{M}}" + }, + { + "bbox": [ + 130, + 262, + 482, + 319 + ], + "type": "text", + "content": " (Sec. 3.2), a vision tokenizer " + }, + { + "bbox": [ + 130, + 262, + 482, + 319 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_{\\mathcal{I}}" + }, + { + "bbox": [ + 130, + 262, + 482, + 319 + ], + "type": "text", + "content": " (r Sec. 3.2) and a vision-motion-aware language model (Sec. 3.3). By leveraging motion tokens generated by " + }, + { + "bbox": [ + 130, + 262, + 482, + 319 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_{\\mathcal{M}}" + }, + { + "bbox": [ + 130, + 262, + 482, + 319 + ], + "type": "text", + "content": ", alongside visual language token embeddings projected by vision tokenizer " + }, + { + "bbox": [ + 130, + 262, + 482, + 319 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_{\\mathcal{I}}" + }, + { + "bbox": [ + 130, + 262, + 482, + 319 + ], + "type": "text", + "content": ", and text tokens by text tokenizer, MotionChian achieves a unified learning paradigm for both motion and linguistic data." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 342, + 480, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 342, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 130, + 342, + 480, + 413 + ], + "type": "text", + "content": "are learned in mapping motions to language. TM2T [24] proposed a new motion representation that compresses motions into a short sequence of discrete variables and then uses a neural translation network to build mappings between two modalities. In contrast to the above methods limited to only several tasks, MotionGPT [32] treats human motion as a foreign language and leverages language understanding and zero-shot transfer abilities of pre-trained language models." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 415, + 482, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 415, + 482, + 545 + ], + "spans": [ + { + "bbox": [ + 130, + 415, + 482, + 545 + ], + "type": "text", + "content": "Character Control and Animation. Character control involves generating interactive motion sequences based on user instruction signals. One kind of approach [38, 59, 79] is to construct a graph representing transitions between motion clips and plan motion using graph search. Considering the limitations of these graph-based approaches in coarse discreteness, alternative methods like frame blending and concatenation [41], low-dimensional latent space learning [42], motion matching [15] proposed for embedding the task in the feature and [88] do the similar thing through hierarchical setup. Although the control signals for motion control and character animation are different from the instructions in text-to-motion tasks, we still recognize textual commands of conventional human motion generation as a boost for intuitive character control." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "content": "Multi-Modal Language Models. In the field of computer vision, there has been a recent surge of interest in multi-modal models that can process text along with other modalities, including images [18, 29, 44, 104], audio [6, 80, 101], and 3D [12, 45, 54, 83, 108]. CLIP [73] is an example of such a model, which learns a semantic latent representation that connects images with corresponding language descriptions. While language models have achieved success in various tasks, the development of multi-modal language models capable of handling human motion is still limited. Existing works in computer vision can be broadly categorized into two classes. The first consists of end-to-end trained models explored separately for specific research topics. For example, tasks like vision-language navi" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 278, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 278, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 278, + 102 + ], + "type": "text", + "content": "Biao Jiang, Xin Chen et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 200, + 124, + 414, + 261 + ], + "blocks": [ + { + "bbox": [ + 200, + 124, + 414, + 261 + ], + "lines": [ + { + "bbox": [ + 200, + 124, + 414, + 261 + ], + "spans": [ + { + "bbox": [ + 200, + 124, + 414, + 261 + ], + "type": "image", + "image_path": "5b5fc2e2f5bbf4400f66ff844044bc3d92ff303935cbe7943668403b496b8640.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 282, + 482, + 370 + ], + "lines": [ + { + "bbox": [ + 130, + 282, + 482, + 370 + ], + "spans": [ + { + "bbox": [ + 130, + 282, + 482, + 370 + ], + "type": "text", + "content": "Fig. 3: Data collection overview: Our initial step in collecting the motion reasoning data involves the utilization of human motion captions derived from an existing text-motion dataset. Subsequent to this, the text-motion retrieval model TMR [68] aids in the segmentation of motion pairs into categories based on the similarity between them. With the assistance of ChatGPT, we proceed to craft motion editing task data that correspond to these categorized similarity levels. Incorporating both motion reasoning and editing single-turn tasks, as well as the extensive 14 tasks delineated in [32], we construct a rich multi-modal multi-turn conversation dataset." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 394, + 482, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 394, + 482, + 502 + ], + "spans": [ + { + "bbox": [ + 130, + 394, + 482, + 502 + ], + "type": "text", + "content": "gation [2, 7] and Habitat [89] require embodied AI agents to follow natural language instructions and take actions to accomplish goals in visual environments. InstructPix2Pix [7] in image translation enables agents to edit images based on human instructions. The second involves systems that coordinate various models using approaches like LangChain or LLMs [51, 62, 103]. Examples of such systems include Visual ChatGPT [100], X-GPT [125], and MMREACT [110]. While these methods focus on building instruction-following agents, we aim to develop an end-to-end trained multimodal model that can perform conversational motion generation tasks via multi-modal inputs with language models." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 520, + 207, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 520, + 207, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 520, + 207, + 533 + ], + "type": "text", + "content": "3 Methods" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 546, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 546, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 546, + 482, + 665 + ], + "type": "text", + "content": "To leverage large language data, vision-language data, and vision-motion data for assisting motion-related tasks, we propose a motion-language-vision framework called MotionChain. The framework, as depicted in Fig. 2, consists of a multi-modal tokenizer that converts various types of data (text, image, and motion) into discrete tokens (Sec. 3.2), and a vision-motion-aware Language model that comprehends information from different modalities and generates corresponding answers based on input instructions (Sec. 3.3). Additionally, to simultaneously understand data from multiple modalities, we employ a multi-stage training strategy (Sec. 3.4) for the training of the multi-modal tokenizer and the motion-language-vision framework." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 392, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 392, + 91, + 447, + 100 + ], + "type": "text", + "content": "MotionChain" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 319 + ], + "type": "text", + "content": "We first introduced the multi-modal tokenizer, which comprises three branches for processing textual, image, and motion inputs. For textual inputs " + }, + { + "bbox": [ + 130, + 116, + 482, + 319 + ], + "type": "inline_equation", + "content": "w^{1:N} = \\{w^i\\}" + }, + { + "bbox": [ + 130, + 116, + 482, + 319 + ], + "type": "text", + "content": " of length " + }, + { + "bbox": [ + 130, + 116, + 482, + 319 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 116, + 482, + 319 + ], + "type": "text", + "content": " that describes a motion-related question or demand, we employ the SentencePiece model [39] used in previous works [14, 74, 93, 94], which has a vocabulary size of " + }, + { + "bbox": [ + 130, + 116, + 482, + 319 + ], + "type": "inline_equation", + "content": "K_{t}" + }, + { + "bbox": [ + 130, + 116, + 482, + 319 + ], + "type": "text", + "content": " and is trained on a large number of language datasets. The motion branch consists of a motion encoder " + }, + { + "bbox": [ + 130, + 116, + 482, + 319 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathcal{M}}" + }, + { + "bbox": [ + 130, + 116, + 482, + 319 + ], + "type": "text", + "content": " that encodes a motion sequence " + }, + { + "bbox": [ + 130, + 116, + 482, + 319 + ], + "type": "inline_equation", + "content": "m^{1:M} = \\{x^i\\}" + }, + { + "bbox": [ + 130, + 116, + 482, + 319 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 130, + 116, + 482, + 319 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 116, + 482, + 319 + ], + "type": "text", + "content": " frames into " + }, + { + "bbox": [ + 130, + 116, + 482, + 319 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 116, + 482, + 319 + ], + "type": "text", + "content": " motion tokens " + }, + { + "bbox": [ + 130, + 116, + 482, + 319 + ], + "type": "inline_equation", + "content": "z^{1:L} = \\{z^i\\}" + }, + { + "bbox": [ + 130, + 116, + 482, + 319 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 116, + 482, + 319 + ], + "type": "inline_equation", + "content": "L = M / l" + }, + { + "bbox": [ + 130, + 116, + 482, + 319 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 482, + 319 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 130, + 116, + 482, + 319 + ], + "type": "text", + "content": " represents the temporal downsampling rate on motion frames. It also includes a motion decoder " + }, + { + "bbox": [ + 130, + 116, + 482, + 319 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathcal{M}}" + }, + { + "bbox": [ + 130, + 116, + 482, + 319 + ], + "type": "text", + "content": " that can decode motion tokens back to human motion " + }, + { + "bbox": [ + 130, + 116, + 482, + 319 + ], + "type": "inline_equation", + "content": "\\hat{m}^{1:M}" + }, + { + "bbox": [ + 130, + 116, + 482, + 319 + ], + "type": "text", + "content": ". The vision branch processes the input image " + }, + { + "bbox": [ + 130, + 116, + 482, + 319 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 130, + 116, + 482, + 319 + ], + "type": "text", + "content": " with a pre-trained CLIP visual encoder and a learnable linear projection that follows it, converted into language token embeddings " + }, + { + "bbox": [ + 130, + 116, + 482, + 319 + ], + "type": "inline_equation", + "content": "H_{q}" + }, + { + "bbox": [ + 130, + 116, + 482, + 319 + ], + "type": "text", + "content": ". Given a textual sentence " + }, + { + "bbox": [ + 130, + 116, + 482, + 319 + ], + "type": "inline_equation", + "content": "w^{1:N}" + }, + { + "bbox": [ + 130, + 116, + 482, + 319 + ], + "type": "text", + "content": ", a sequence of motion " + }, + { + "bbox": [ + 130, + 116, + 482, + 319 + ], + "type": "inline_equation", + "content": "m^{1:M}" + }, + { + "bbox": [ + 130, + 116, + 482, + 319 + ], + "type": "text", + "content": ", and an image condition " + }, + { + "bbox": [ + 130, + 116, + 482, + 319 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 130, + 116, + 482, + 319 + ], + "type": "text", + "content": ", all encoded as language tokens, our vision-motion-aware language model is designed to produce an answer comprising " + }, + { + "bbox": [ + 130, + 116, + 482, + 319 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 116, + 482, + 319 + ], + "type": "text", + "content": " tokens, denoted as " + }, + { + "bbox": [ + 130, + 116, + 482, + 319 + ], + "type": "inline_equation", + "content": "\\hat{x}^{1:L} = \\{\\hat{x}^i\\}" + }, + { + "bbox": [ + 130, + 116, + 482, + 319 + ], + "type": "text", + "content": ". These output tokens can represent either motion sequences " + }, + { + "bbox": [ + 130, + 116, + 482, + 319 + ], + "type": "inline_equation", + "content": "\\hat{x}_{m}^{1:L}" + }, + { + "bbox": [ + 130, + 116, + 482, + 319 + ], + "type": "text", + "content": " or textual descriptions " + }, + { + "bbox": [ + 130, + 116, + 482, + 319 + ], + "type": "inline_equation", + "content": "\\hat{x}_{t}^{1:L}" + }, + { + "bbox": [ + 130, + 116, + 482, + 319 + ], + "type": "text", + "content": ", which integrate both human motion " + }, + { + "bbox": [ + 130, + 116, + 482, + 319 + ], + "type": "inline_equation", + "content": "\\hat{m}^{1:M}" + }, + { + "bbox": [ + 130, + 116, + 482, + 319 + ], + "type": "text", + "content": ", and text " + }, + { + "bbox": [ + 130, + 116, + 482, + 319 + ], + "type": "inline_equation", + "content": "\\hat{w}^{1:L}" + }, + { + "bbox": [ + 130, + 116, + 482, + 319 + ], + "type": "text", + "content": " within the given context." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 341, + 240, + 352 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 341, + 240, + 352 + ], + "spans": [ + { + "bbox": [ + 132, + 341, + 240, + 352 + ], + "type": "text", + "content": "3.1 Data Collection" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 365, + 482, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 365, + 482, + 568 + ], + "spans": [ + { + "bbox": [ + 130, + 365, + 482, + 568 + ], + "type": "text", + "content": "With the emergence of text-conditioned motion generation tasks, datasets like KIT [69], BABEL [71], HumanML3D [23] and the more recent Motion-X [49] have been developed. However, these datasets predominantly offer text labels as simple action phrases or captions. Building upon these foundations, MotionGPT [32] introduces an instruction-based motion-language dataset that encapsulates 14 core tasks, including motion prediction, translation, and editing, through thousands of instruction templates in a unified format. Despite this advancement, MotionGPT's data lack a deep engagement with the nuances of human motion analysis and are limited to single-turn generation tasks without incorporating contextual memory. Inspired by the recent success of GPT models across text-annotation tasks [17], image-annotation tasks [50], 3D-annotation tasks [28], we propose a data collection methodology integrates the capabilities of existing LLMs like ChatGPT [62], with the text-motion retrieval model TMR [68] to facilitate motion conversation data collection. In addition to the 14 motion-related tasks in MotionGPT [32], we introduce tasks centered around motion reasoning and motion editing, leveraging contextual insights for a deeper motion analysis." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": "Utilizing ChatGPT [62], we initiate the collection of motion reasoning data using human motion captions from the text-motion dataset [23], starting with manually designed example queries that explore the contextual scenarios surrounding motions, possible preceding or succeeding actions, the subjects' roles, and the tools or equipment involved, etc. Following this, we employ TMR [68] for categorizing motions from the dataset into varying similarity levels. For medium-similarity motion pairs, we utilize ChatGPT [62] to generate motion editing directives that enable the transformation of one motion to another. For motions of" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 278, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 278, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 278, + 102 + ], + "type": "text", + "content": "Biao Jiang, Xin Chen et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "high similarity, we manually devise tasks aimed at editing their lengths, further enriching the dataset's versatility and analytical scope." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 480, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 480, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 480, + 200 + ], + "type": "text", + "content": "After the collection of single-turn generation tasks, we progress to develop multi-turn conversation data. This involves the deliberate association of initial motion generation tasks with a variety of follow-up tasks randomly chosen among motion translation, reasoning, editing, etc. Following [122] we construct our conversation data in a structured format, as depicted below:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 149, + 208, + 210, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 208, + 210, + 217 + ], + "spans": [ + { + "bbox": [ + 149, + 208, + 210, + 217 + ], + "type": "inline_equation", + "content": "X_{\\mathrm{system - message}}" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 149, + 217, + 308, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 217, + 308, + 228 + ], + "spans": [ + { + "bbox": [ + 149, + 217, + 308, + 228 + ], + "type": "text", + "content": "USER: " + }, + { + "bbox": [ + 149, + 217, + 308, + 228 + ], + "type": "inline_equation", + "content": "X_{v}" + }, + { + "bbox": [ + 149, + 217, + 308, + 228 + ], + "type": "inline_equation", + "content": "\\dot{X}_s^1" + }, + { + "bbox": [ + 149, + 217, + 308, + 228 + ], + "type": "text", + "content": " ASSISTANT: " + }, + { + "bbox": [ + 149, + 217, + 308, + 228 + ], + "type": "inline_equation", + "content": "X_{a}^{1} < / s>" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 150, + 228, + 308, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 150, + 228, + 308, + 239 + ], + "spans": [ + { + "bbox": [ + 150, + 228, + 308, + 239 + ], + "type": "text", + "content": "USER: " + }, + { + "bbox": [ + 150, + 228, + 308, + 239 + ], + "type": "inline_equation", + "content": "X_{v} X_{s}^{2}" + }, + { + "bbox": [ + 150, + 228, + 308, + 239 + ], + "type": "text", + "content": " ASSISTANT: " + }, + { + "bbox": [ + 150, + 228, + 308, + 239 + ], + "type": "inline_equation", + "content": "X_{a}^{2} < / \\mathrm{s}>" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 150, + 239, + 319, + 250 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 150, + 239, + 319, + 250 + ], + "spans": [ + { + "bbox": [ + 150, + 239, + 319, + 250 + ], + "type": "text", + "content": "USER: " + }, + { + "bbox": [ + 150, + 239, + 319, + 250 + ], + "type": "inline_equation", + "content": "X_{v} X_{s}^{3}" + }, + { + "bbox": [ + 150, + 239, + 319, + 250 + ], + "type": "text", + "content": " ASSISTANT: " + }, + { + "bbox": [ + 150, + 239, + 319, + 250 + ], + "type": "inline_equation", + "content": "X_{a}^{3} < / \\mathrm{s}>" + }, + { + "bbox": [ + 150, + 239, + 319, + 250 + ], + "type": "text", + "content": " ..." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 258, + 479, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 258, + 479, + 378 + ], + "spans": [ + { + "bbox": [ + 130, + 258, + 479, + 378 + ], + "type": "text", + "content": "Where " + }, + { + "bbox": [ + 130, + 258, + 479, + 378 + ], + "type": "inline_equation", + "content": "X_{v}" + }, + { + "bbox": [ + 130, + 258, + 479, + 378 + ], + "type": "text", + "content": " is defined as the vision language token embeddings, processed via the visual tokenizer. " + }, + { + "bbox": [ + 130, + 258, + 479, + 378 + ], + "type": "inline_equation", + "content": "X_{s}^{i}" + }, + { + "bbox": [ + 130, + 258, + 479, + 378 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 258, + 479, + 378 + ], + "type": "inline_equation", + "content": "X_{a}^{i}" + }, + { + "bbox": [ + 130, + 258, + 479, + 378 + ], + "type": "text", + "content": " are used to denote the source inputs and target answers for each round " + }, + { + "bbox": [ + 130, + 258, + 479, + 378 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 258, + 479, + 378 + ], + "type": "text", + "content": ", respectively. Both sets of tokens originate from the integrated motion-language vocabulary " + }, + { + "bbox": [ + 130, + 258, + 479, + 378 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 130, + 258, + 479, + 378 + ], + "type": "text", + "content": ", which includes motions, texts, or a blend thereof. The dataset exhibits variability in the number of generation turns up to 10; for the sake of clarity, we present only three examples herein. MotionChain is trained to predict answers, incorporating a learning mechanism that determines whether to stop generation by outputting end of sentence flag " + }, + { + "bbox": [ + 130, + 258, + 479, + 378 + ], + "type": "inline_equation", + "content": "< / \\mathrm{s} >" + }, + { + "bbox": [ + 130, + 258, + 479, + 378 + ], + "type": "text", + "content": " based on the current instruction and all preceding questions and answers. In the computation of the loss, as defined in Eq. (5), only the green tokens are utilized." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 394, + 276, + 406 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 394, + 276, + 406 + ], + "spans": [ + { + "bbox": [ + 131, + 394, + 276, + 406 + ], + "type": "text", + "content": "3.2 Multi-modal Tokenizer" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 413, + 480, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 413, + 480, + 556 + ], + "spans": [ + { + "bbox": [ + 130, + 413, + 480, + 556 + ], + "type": "text", + "content": "Motion tokenizer, denoted as " + }, + { + "bbox": [ + 130, + 413, + 480, + 556 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_{\\mathcal{M}}" + }, + { + "bbox": [ + 130, + 413, + 480, + 556 + ], + "type": "text", + "content": ", is based on the architecture of Vector Quantized Variational Autoencoders (VQ-VAE) utilized in previous studies [22,24,32, 84,95,98,106,109,116]. Once pre-trained, it can represent motion using discrete tokens, facilitating the integration of motion and language. The Motion tokenizer consists of a motion encoder " + }, + { + "bbox": [ + 130, + 413, + 480, + 556 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathcal{M}}" + }, + { + "bbox": [ + 130, + 413, + 480, + 556 + ], + "type": "text", + "content": " and a motion decoder " + }, + { + "bbox": [ + 130, + 413, + 480, + 556 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathcal{M}}" + }, + { + "bbox": [ + 130, + 413, + 480, + 556 + ], + "type": "text", + "content": ". Initially, the motion encoder " + }, + { + "bbox": [ + 130, + 413, + 480, + 556 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 130, + 413, + 480, + 556 + ], + "type": "text", + "content": " applies 1D convolutions to the motion features " + }, + { + "bbox": [ + 130, + 413, + 480, + 556 + ], + "type": "inline_equation", + "content": "m^{1:M}" + }, + { + "bbox": [ + 130, + 413, + 480, + 556 + ], + "type": "text", + "content": " along the temporal dimension to obtain latent vectors " + }, + { + "bbox": [ + 130, + 413, + 480, + 556 + ], + "type": "inline_equation", + "content": "\\hat{z}^{1:L} = \\mathcal{E}_{\\mathcal{M}}(m^{1:M})" + }, + { + "bbox": [ + 130, + 413, + 480, + 556 + ], + "type": "text", + "content": ". Subsequently, the latent vectors " + }, + { + "bbox": [ + 130, + 413, + 480, + 556 + ], + "type": "inline_equation", + "content": "\\hat{z}" + }, + { + "bbox": [ + 130, + 413, + 480, + 556 + ], + "type": "text", + "content": " are quantized and transformed into a collection of codebook entries " + }, + { + "bbox": [ + 130, + 413, + 480, + 556 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 130, + 413, + 480, + 556 + ], + "type": "text", + "content": ". The learnable codebook " + }, + { + "bbox": [ + 130, + 413, + 480, + 556 + ], + "type": "inline_equation", + "content": "Z = \\{z_{i}\\}_{i=1}^{K} \\subset \\mathbb{R}^{d}" + }, + { + "bbox": [ + 130, + 413, + 480, + 556 + ], + "type": "text", + "content": " comprises " + }, + { + "bbox": [ + 130, + 413, + 480, + 556 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 413, + 480, + 556 + ], + "type": "text", + "content": " latent embedding vectors, each with a dimension of " + }, + { + "bbox": [ + 130, + 413, + 480, + 556 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 130, + 413, + 480, + 556 + ], + "type": "text", + "content": ". The quantization process " + }, + { + "bbox": [ + 130, + 413, + 480, + 556 + ], + "type": "inline_equation", + "content": "Q(\\cdot)" + }, + { + "bbox": [ + 130, + 413, + 480, + 556 + ], + "type": "text", + "content": " replaces each row vector " + }, + { + "bbox": [ + 130, + 413, + 480, + 556 + ], + "type": "inline_equation", + "content": "\\hat{z}" + }, + { + "bbox": [ + 130, + 413, + 480, + 556 + ], + "type": "text", + "content": " with its nearest codebook entry " + }, + { + "bbox": [ + 130, + 413, + 480, + 556 + ], + "type": "inline_equation", + "content": "z_{k}" + }, + { + "bbox": [ + 130, + 413, + 480, + 556 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 130, + 413, + 480, + 556 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 130, + 413, + 480, + 556 + ], + "type": "text", + "content": ", which can be expressed as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 225, + 569, + 480, + 582 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 569, + 480, + 582 + ], + "spans": [ + { + "bbox": [ + 225, + 569, + 480, + 582 + ], + "type": "interline_equation", + "content": "z _ {i} = Q \\left(\\hat {z} ^ {i}\\right) := \\arg \\min _ {z _ {k} \\in Z} \\| \\hat {z} _ {i} - z _ {k} \\| _ {2}. \\tag {1}", + "image_path": "2bfabfe0570fff650ecea9cccaf268c0d98441afc4cf3f766fee6037e096bfd3.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 585, + 479, + 645 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 585, + 479, + 645 + ], + "spans": [ + { + "bbox": [ + 130, + 585, + 479, + 645 + ], + "type": "text", + "content": "We assign " + }, + { + "bbox": [ + 130, + 585, + 479, + 645 + ], + "type": "inline_equation", + "content": "s^i" + }, + { + "bbox": [ + 130, + 585, + 479, + 645 + ], + "type": "text", + "content": " as the index number of motion tokens " + }, + { + "bbox": [ + 130, + 585, + 479, + 645 + ], + "type": "inline_equation", + "content": "z^{1:L}" + }, + { + "bbox": [ + 130, + 585, + 479, + 645 + ], + "type": "text", + "content": ", so motion tokens " + }, + { + "bbox": [ + 130, + 585, + 479, + 645 + ], + "type": "inline_equation", + "content": "z^{1:L}" + }, + { + "bbox": [ + 130, + 585, + 479, + 645 + ], + "type": "text", + "content": " can be represented as a sequence of indices " + }, + { + "bbox": [ + 130, + 585, + 479, + 645 + ], + "type": "inline_equation", + "content": "s^{1:L} = \\{s^i\\}_{i=1}^L" + }, + { + "bbox": [ + 130, + 585, + 479, + 645 + ], + "type": "text", + "content": ". The motion decoder " + }, + { + "bbox": [ + 130, + 585, + 479, + 645 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathcal{M}}" + }, + { + "bbox": [ + 130, + 585, + 479, + 645 + ], + "type": "text", + "content": " can project " + }, + { + "bbox": [ + 130, + 585, + 479, + 645 + ], + "type": "inline_equation", + "content": "z^{1:L} = \\{z^i\\}_{i=1}^L" + }, + { + "bbox": [ + 130, + 585, + 479, + 645 + ], + "type": "text", + "content": " back to the raw motion space, resulting in the motion " + }, + { + "bbox": [ + 130, + 585, + 479, + 645 + ], + "type": "inline_equation", + "content": "\\hat{m}^{1:M}" + }, + { + "bbox": [ + 130, + 585, + 479, + 645 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 130, + 585, + 479, + 645 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 585, + 479, + 645 + ], + "type": "text", + "content": " frames. Following [22,24,32,98,116], we adopt three distinct loss functions when training the motion tokenizer:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 264, + 654, + 480, + 666 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 264, + 654, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 264, + 654, + 480, + 666 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathcal {V}} = \\mathcal {L} _ {r} + \\mathcal {L} _ {e} + \\mathcal {L} _ {c} \\tag {2}", + "image_path": "fb654a4dfbe61dde831931614187bc5f14e3a1704995adb753d1d96fc6a0a0ba.jpg" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 392, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 392, + 91, + 447, + 101 + ], + "type": "text", + "content": "MotionChain" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 140, + 112, + 245, + 186 + ], + "blocks": [ + { + "bbox": [ + 140, + 112, + 245, + 186 + ], + "lines": [ + { + "bbox": [ + 140, + 112, + 245, + 186 + ], + "spans": [ + { + "bbox": [ + 140, + 112, + 245, + 186 + ], + "type": "image", + "image_path": "7a41c3c7f2ec84c74344d9d49bbf140e657db0b17849fef6493b0d086748291c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 162, + 190, + 221, + 199 + ], + "lines": [ + { + "bbox": [ + 162, + 190, + 221, + 199 + ], + "spans": [ + { + "bbox": [ + 162, + 190, + 221, + 199 + ], + "type": "text", + "content": "(a) Independent" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 255, + 112, + 359, + 186 + ], + "blocks": [ + { + "bbox": [ + 255, + 112, + 359, + 186 + ], + "lines": [ + { + "bbox": [ + 255, + 112, + 359, + 186 + ], + "spans": [ + { + "bbox": [ + 255, + 112, + 359, + 186 + ], + "type": "image", + "image_path": "65da011d9b7b6a97fa4383c8d2c5d774a27ccef12991c5a264f7d1c7b3fbefe4.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 273, + 190, + 340, + 199 + ], + "lines": [ + { + "bbox": [ + 273, + 190, + 340, + 199 + ], + "spans": [ + { + "bbox": [ + 273, + 190, + 340, + 199 + ], + "type": "text", + "content": "(b) Past-condition" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 208, + 481, + 264 + ], + "lines": [ + { + "bbox": [ + 130, + 208, + 481, + 264 + ], + "spans": [ + { + "bbox": [ + 130, + 208, + 481, + 264 + ], + "type": "text", + "content": "Fig. 4: Motion Composition Variants: We illustrate the baselines for motion composition during multi-turn motion generation (a). independent decoding each turn (b). separate decoding conditioned on the last few tokens from the prior turn (c). decoding with joint motion tokens. Green tokens stand for image condition, blue tokens stand for textual instruction, and orange tokens stand for human motions." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 369, + 112, + 475, + 186 + ], + "blocks": [ + { + "bbox": [ + 369, + 112, + 475, + 186 + ], + "lines": [ + { + "bbox": [ + 369, + 112, + 475, + 186 + ], + "spans": [ + { + "bbox": [ + 369, + 112, + 475, + 186 + ], + "type": "image", + "image_path": "7e2d5e43eaa0c0fd055ee3e0f8fd642875832808805ab1728c09fdf3a0dd7f5f.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 393, + 190, + 452, + 199 + ], + "lines": [ + { + "bbox": [ + 393, + 190, + 452, + 199 + ], + "spans": [ + { + "bbox": [ + 393, + 190, + 452, + 199 + ], + "type": "text", + "content": "(c) Tokens-joint" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 288, + 480, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 288, + 480, + 310 + ], + "spans": [ + { + "bbox": [ + 130, + 288, + 480, + 310 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 288, + 480, + 310 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_r" + }, + { + "bbox": [ + 130, + 288, + 480, + 310 + ], + "type": "text", + "content": " denotes reconstruction loss, " + }, + { + "bbox": [ + 130, + 288, + 480, + 310 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_e" + }, + { + "bbox": [ + 130, + 288, + 480, + 310 + ], + "type": "text", + "content": " denotes the embedding loss, and " + }, + { + "bbox": [ + 130, + 288, + 480, + 310 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_c" + }, + { + "bbox": [ + 130, + 288, + 480, + 310 + ], + "type": "text", + "content": " denotes commitment loss." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 312, + 482, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 312, + 482, + 399 + ], + "spans": [ + { + "bbox": [ + 130, + 312, + 482, + 399 + ], + "type": "text", + "content": "During multi-turn motion generation, the motion continuity between turns is achieved through our motion decoder, which links the motion of the current turn with that of the preceding ones. Taking the composition of two motions as an example: we concatenate the past motion tokens, denoted as " + }, + { + "bbox": [ + 130, + 312, + 482, + 399 + ], + "type": "inline_equation", + "content": "z_{p}^{1:L_{p}}" + }, + { + "bbox": [ + 130, + 312, + 482, + 399 + ], + "type": "text", + "content": ", with the tokens representing the current motion, " + }, + { + "bbox": [ + 130, + 312, + 482, + 399 + ], + "type": "inline_equation", + "content": "z_{c}^{1:L_{c}}" + }, + { + "bbox": [ + 130, + 312, + 482, + 399 + ], + "type": "text", + "content": ". This concatenated sequence of tokens is subsequently decoded into a comprehensive set of continuous motion features, represented as " + }, + { + "bbox": [ + 130, + 312, + 482, + 399 + ], + "type": "inline_equation", + "content": "m_{\\mathrm{whole}}^{1:M_{\\mathrm{whole}}}" + }, + { + "bbox": [ + 130, + 312, + 482, + 399 + ], + "type": "text", + "content": ", as depicted below:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 249, + 407, + 481, + 425 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 249, + 407, + 481, + 425 + ], + "spans": [ + { + "bbox": [ + 249, + 407, + 481, + 425 + ], + "type": "interline_equation", + "content": "z _ {\\text {w h o l e}} ^ {1: \\left(L _ {p} + L _ {c}\\right)} = \\left[ z _ {p} ^ {1: L _ {p}}, z _ {c} ^ {1: L _ {c}} \\right]. \\tag {3}", + "image_path": "53c340e13702bca503716cd46d259208bb224b5c3c934cede4ae99038ea2c99e.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 433, + 480, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 433, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 130, + 433, + 480, + 468 + ], + "type": "text", + "content": "Similarly, this framework is adept at executing composition tasks involving an array of motions. The comparison results in Tab. 3 demonstrate that our motion tokenizer could effectively perform motion composition tasks." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 469, + 481, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 469, + 481, + 529 + ], + "spans": [ + { + "bbox": [ + 130, + 469, + 481, + 529 + ], + "type": "text", + "content": "Visual Tokenizer accepts images " + }, + { + "bbox": [ + 130, + 469, + 481, + 529 + ], + "type": "inline_equation", + "content": "X_{I}" + }, + { + "bbox": [ + 130, + 469, + 481, + 529 + ], + "type": "text", + "content": " as inputs. We employ the CLIP visual encoder that is pre-trained on image-text pairs to derive visual feature " + }, + { + "bbox": [ + 130, + 469, + 481, + 529 + ], + "type": "inline_equation", + "content": "Z_{I}" + }, + { + "bbox": [ + 130, + 469, + 481, + 529 + ], + "type": "text", + "content": ". These features are then projected into language token embeddings " + }, + { + "bbox": [ + 130, + 469, + 481, + 529 + ], + "type": "inline_equation", + "content": "X_{v}" + }, + { + "bbox": [ + 130, + 469, + 481, + 529 + ], + "type": "text", + "content": " via a linear layer like previous work [50], maintaining consistency in the dimensionality with the language model's word embedding space." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 131, + 548, + 317, + 560 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 548, + 317, + 560 + ], + "spans": [ + { + "bbox": [ + 131, + 548, + 317, + 560 + ], + "type": "text", + "content": "3.3 Motion-aware Language Model" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "content": "Language models such as Llama [93, 94] and T5 [14, 74] employ the Sentence-Piece [39] model to encode textual inputs into WordPiece tokens, utilizing a " + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "inline_equation", + "content": "K_{t}" + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "content": " word piece vocabulary. Unlike prior text-to-motion [24, 102, 116, 118] and motion-to-text [24] methods that process text and motion separately, we merge the text vocabulary " + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "inline_equation", + "content": "V_{t} = \\{v_{t}^{i}\\}_{i=1}^{K_{t}}" + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "content": " with the motion vocabulary " + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "inline_equation", + "content": "V_{m} = \\{v_{m}^{i}\\}_{i=1}^{K_{m}}" + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "content": ", maintaining the motion tokenizer's codebook " + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "content": " order and including special tokens for boundary demarcation. This creates a unified vocabulary " + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "inline_equation", + "content": "V = \\{V_{t}, V_{m}\\}" + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "content": ", enabling the formulation of motion-centric tasks in a universal template, where inputs and" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 278, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 278, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 278, + 102 + ], + "type": "text", + "content": "Biao Jiang, Xin Chen et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "outputs share the same vocabulary. For visual input, our visual tokenizer converts images into visual token embeddings " + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "inline_equation", + "content": "X_{v}" + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": ", aligning with the language model [39, 62, 74, 93] token space for integrated representation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 480, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 480, + 305 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 480, + 305 + ], + "type": "text", + "content": "For single conditioned generation tasks, our input comprises a sequence of " + }, + { + "bbox": [ + 130, + 152, + 480, + 305 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 152, + 480, + 305 + ], + "type": "text", + "content": " length tokens " + }, + { + "bbox": [ + 130, + 152, + 480, + 305 + ], + "type": "inline_equation", + "content": "X_{s} = \\{x_{s}^{i}\\}_{i = 1}^{N}" + }, + { + "bbox": [ + 130, + 152, + 480, + 305 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 152, + 480, + 305 + ], + "type": "inline_equation", + "content": "x_{s}\\in \\{V_{t},V_{m}\\}" + }, + { + "bbox": [ + 130, + 152, + 480, + 305 + ], + "type": "text", + "content": " representing either text, motion, or a combination thereof, drawn from the unified vocabularies. In cases involving image inputs, visual tokens " + }, + { + "bbox": [ + 130, + 152, + 480, + 305 + ], + "type": "inline_equation", + "content": "X_{v}" + }, + { + "bbox": [ + 130, + 152, + 480, + 305 + ], + "type": "text", + "content": " are interspersed at the beginning of the source tokens sequence, forming " + }, + { + "bbox": [ + 130, + 152, + 480, + 305 + ], + "type": "inline_equation", + "content": "[X_v,X_s]" + }, + { + "bbox": [ + 130, + 152, + 480, + 305 + ], + "type": "text", + "content": ". Subsequent interaction rounds generate target answer tokens " + }, + { + "bbox": [ + 130, + 152, + 480, + 305 + ], + "type": "inline_equation", + "content": "X_{a}" + }, + { + "bbox": [ + 130, + 152, + 480, + 305 + ], + "type": "text", + "content": ". To facilitate iterative result generation and content retention, our framework generates multi-turn conversation data " + }, + { + "bbox": [ + 130, + 152, + 480, + 305 + ], + "type": "inline_equation", + "content": "(X_{v},X_{s}^{1},X_{a}^{1},X_{s}^{2},X_{a}^{2},\\dots ,X_{s}^{T},X_{a}^{T})" + }, + { + "bbox": [ + 130, + 152, + 480, + 305 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 130, + 152, + 480, + 305 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 152, + 480, + 305 + ], + "type": "text", + "content": " indicating the total turn count. Notably, visual tokens are consistently placed at the forefront of the initial turn's source tokens. The processing sequence is organized such that to predict target answer tokens autoregressively, as shown in Fig. 2. Source tokens are processed by the transformer to predict the next token's probability distribution, formulated as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 203, + 306, + 480, + 329 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 203, + 306, + 480, + 329 + ], + "spans": [ + { + "bbox": [ + 203, + 306, + 480, + 329 + ], + "type": "interline_equation", + "content": "p _ {\\theta} \\left(X _ {a} \\mid X _ {v}, X _ {s}\\right) = \\prod_ {i} p _ {\\theta} \\left(x _ {a} ^ {i} \\mid X _ {v}, X _ {s, < i}, X _ {a, < i}\\right) \\tag {4}", + "image_path": "0ccd421979a8738cbe6f9198cacd872fc72e41684bdbfe3b5e55fc229b3551fb.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 334, + 479, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 334, + 479, + 370 + ], + "spans": [ + { + "bbox": [ + 130, + 334, + 479, + 370 + ], + "type": "text", + "content": "with " + }, + { + "bbox": [ + 130, + 334, + 479, + 370 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 130, + 334, + 479, + 370 + ], + "type": "text", + "content": " indicating trainable parameters, and " + }, + { + "bbox": [ + 130, + 334, + 479, + 370 + ], + "type": "inline_equation", + "content": "X_{s, < i}, X_{a, < i}" + }, + { + "bbox": [ + 130, + 334, + 479, + 370 + ], + "type": "text", + "content": " representing the sequences of source and preceding target tokens. The training objective is maximizing the log-likelihood of distribution:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 210, + 380, + 480, + 411 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 210, + 380, + 480, + 411 + ], + "spans": [ + { + "bbox": [ + 210, + 380, + 480, + 411 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {L M} = - \\sum_ {i = 0} ^ {L _ {t} - 1} \\log p _ {\\theta} \\left(x _ {a} ^ {i} \\mid X _ {v}, X _ {s, < i}, X _ {a, < i}\\right). \\tag {5}", + "image_path": "ed9c4f3cd7f9944107150a494f2e47dc8e9268743c2f37c6456a5425b7a81c85.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 419, + 479, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 419, + 479, + 443 + ], + "spans": [ + { + "bbox": [ + 130, + 419, + 479, + 443 + ], + "type": "text", + "content": "By optimizing this objective, MotionChain captures the complex interrelations among images, motion, and text, facilitating accurate target \"word\" generation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 444, + 480, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 444, + 480, + 508 + ], + "spans": [ + { + "bbox": [ + 130, + 444, + 480, + 508 + ], + "type": "text", + "content": "During the inference phase, target tokens are recursively sampled from the model's predicted distribution " + }, + { + "bbox": [ + 130, + 444, + 480, + 508 + ], + "type": "inline_equation", + "content": "p_{\\theta}\\left(\\hat{x}_a^i \\mid X_v, X_s, \\hat{X}_{a, < i}\\right)" + }, + { + "bbox": [ + 130, + 444, + 480, + 508 + ], + "type": "text", + "content": ", ceasing with the appearance of a special end token. This strategy facilitates a step-by-step target sequence generation, where each token's probability is conditioned on all previous turns' sources and targets and current source input." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 525, + 250, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 525, + 250, + 538 + ], + "spans": [ + { + "bbox": [ + 131, + 525, + 250, + 538 + ], + "type": "text", + "content": "3.4 Training Strategy" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 546, + 480, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 546, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 546, + 480, + 666 + ], + "type": "text", + "content": "To facilitate the integration of image and motion comprehension within the language modeling context, we adopt a 3-stage training strategy. (1) The initial stage involves pre-training the motion tokenizer on a corpus of human motion data, in line with [32]. This process establishes the motion vocabulary " + }, + { + "bbox": [ + 130, + 546, + 480, + 666 + ], + "type": "inline_equation", + "content": "V_{m}" + }, + { + "bbox": [ + 130, + 546, + 480, + 666 + ], + "type": "text", + "content": ", which serves as a foundation for encoding human motions as a series of discrete tokens. (2) Subsequently, the motion tokenizer remains frozen while we connect the visual tokenizer to the language model framework. This integration is supported by a suite of supervised objectives, including text-to-motion, motion-to-text, and image-based motion generation, aiming to learn the intricate relationships between images, motion, and language. (3) The final stage involves instruction" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 392, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 392, + 91, + 447, + 101 + ], + "type": "text", + "content": "MotionChain" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "tuning, and refines the model's capabilities through the application of prompt-based instructions. These instructions are framed within multi-turn conversation sequences, as detailed in Sec. 3.3, to expanded range of motion-related tasks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 155, + 482, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 155, + 482, + 215 + ], + "spans": [ + { + "bbox": [ + 130, + 155, + 482, + 215 + ], + "type": "text", + "content": "Training of Motion Tokenizer. The initial step involves training the motion tokenizer, guided by the loss objective in Equation 2. This stage enables the tokenizer to represent human motion sequences " + }, + { + "bbox": [ + 130, + 155, + 482, + 215 + ], + "type": "inline_equation", + "content": "\\hat{x}^{1:L}" + }, + { + "bbox": [ + 130, + 155, + 482, + 215 + ], + "type": "text", + "content": " as discrete motion tokens, a key step for merging motion data with textual information seamlessly. Once optimized, the motion tokenizer remains frozen." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 218, + 482, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 218, + 482, + 397 + ], + "spans": [ + { + "bbox": [ + 130, + 218, + 482, + 397 + ], + "type": "text", + "content": "Motion-language Pre-training Stage. Leveraging recent developments in language modeling [14, 74, 93, 94, 122] pre-trained on natural language datasets and then fine-tuned with instruction-based phrasing [14, 62]. To augment the model's ability to discern relationships between images and human motions, we first pre-train our MotionChain using a mix of language, image, and motion datasets. Following the stage 1 training of the motion tokenizer, we have established a unified motion-language vocabulary " + }, + { + "bbox": [ + 130, + 218, + 482, + 397 + ], + "type": "inline_equation", + "content": "V = V_{t}, V_{m}" + }, + { + "bbox": [ + 130, + 218, + 482, + 397 + ], + "type": "text", + "content": ", capable of representing motions in discrete token form. Moreover, we maintain the visual encoder's weights in the visual tokenizer as fixed, while the linear projection weight " + }, + { + "bbox": [ + 130, + 218, + 482, + 397 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 130, + 218, + 482, + 397 + ], + "type": "text", + "content": " is jointly optimized with the language model. During this stage, the model undertakes three fundamental single-turn modality translation tasks: text-to-motion, motion-to-text, and image-conditioned motion generation, as outlined in Sec. 3.1. The primary objective is to maximize the likelihood of the model according to the loss function specified in Eq. (5), thereby letting the model understand the relationship between language, vision conditions, and motions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 400, + 482, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 400, + 482, + 509 + ], + "spans": [ + { + "bbox": [ + 130, + 400, + 482, + 509 + ], + "type": "text", + "content": "Instruction Tuning Stage. As described in Sec. 3.1, we construct a multimodal, multi-task, and multi-turn motion conversation dataset by augmenting existing text-to-motion [23] and human mesh reconstruction datasets [5] with targeted instructional prompts and leverage the capabilities of LLMs [62] and the text-motion retrieval model [68] for motion reasoning and editing tasks. The efficacy of instruction tuning, as evidenced across language models [14, 50, 62, 122], is well-established, yielding enhancements in model performance across a wide range of tasks. After instruction tuning, MotionChain can handle more motion-related tasks including the proficient handling of previously unseen tasks" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 541, + 230, + 555 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 541, + 230, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 541, + 230, + 555 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "We evaluate the proposed MotionChain encompasses comprehensive comparisons across both one-turn motion-related tasks and multi-turn motion generation tasks. Firstly, we provide details of the dataset settings, evaluation criteria, and implementation details as specified in Sec. 4.1. Subsequently, comparative analyses are presented, focusing on the motion reasoning task (Sec. 4.2) and the temporal motion composition task (Sec. 4.3). In Sec. 4.4, we evaluate the choice of motion composition technique and different architectures of vision tokenizer." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 278, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 278, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 278, + 102 + ], + "type": "text", + "content": "Biao Jiang, Xin Chen et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 261, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 261, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 261, + 129 + ], + "type": "text", + "content": "4.1 Experimental Setup" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 138, + 482, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 138, + 482, + 317 + ], + "spans": [ + { + "bbox": [ + 130, + 138, + 482, + 317 + ], + "type": "text", + "content": "Datasets. For one-turn motion reasoning tasks, the study employs our proposed multi-modal multi-turn conversation dataset upon HumanML3D [23] with 44,970 sequence-level textual descriptions for 14,616 motion sequences obtained from AMASS [57] and HumanAct12 [25]. The datasets are divided into training, testing, and validation sets with a ratio of " + }, + { + "bbox": [ + 130, + 138, + 482, + 317 + ], + "type": "inline_equation", + "content": "0.8:0.15:0.05" + }, + { + "bbox": [ + 130, + 138, + 482, + 317 + ], + "type": "text", + "content": ". To evaluate the multi-turn motion generation task, we focus on BABEL [71] that provides textual descriptions for the motions in the AMASS [57] with annotated segments that overlap in each sequence, which allows evaluating generation of a sequence of motion or actions. We adopt the processed text labels by [3] and motion representation of HumanML3D [23] which combines joint velocities, positions, and rotations. Following [3] we consider pairs of actions for simplicity but Motion-Chain applies to a sequence of actions or motion of arbitrary length. For the image-conditioned motion generation task, we mainly focus on BEDLAM [5], a large synthetic dataset of realistic moving 3D humans containing more than 200 subjects and 380K frames video and motion pair." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 318, + 482, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 318, + 482, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 318, + 482, + 521 + ], + "type": "text", + "content": "Evaluation Metrics are summarized as four parts. (1) Motion quality: We adopt Frechet Inception Distance (FID) as the primary metric, FID quantifies the divergence in feature distributions between generated and actual motion sequences. Utilizing feature extractors from prior studies [23,49,68], FID measures the distance of feature distributions between the generated and real motions. Following [5,32,33,65,102], we also adopt MPJPE, PA-MPJPE to measure global and local errors in millimeters and ACCL for acceleration errors, to evaluate the quality of the reconstructed motions. (2) Motion Diversity: Utilizing the Diversity (DIV) metric, we calculate variance across motion features to evaluate generation diversity. (3) Text matching: The precision of text-to-motion matches is quantified by the R Precision metric, based on the feature evaluator [23,49,68], and includes an analysis of Top 1/2/3 retrieval accuracy. The Multi-modal Distance (MM Dist) quantifies the semantic gap between motions and texts. (4) Linguistic quality: We follow [24] utilizing linguistic metrics from natural language studies, including BLUE [63], Rouge [48], Cider [97], and BertScore [120] to evaluate the quality of generated motion captions. More detailed benchmark information is provided in the supplementary materials." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 522, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 522, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 522, + 482, + 665 + ], + "type": "text", + "content": "Implementation Details. We set the codebook of the motion tokenizer as " + }, + { + "bbox": [ + 130, + 522, + 482, + 665 + ], + "type": "inline_equation", + "content": "K \\in \\mathbb{R}^{512 \\times 1024}" + }, + { + "bbox": [ + 130, + 522, + 482, + 665 + ], + "type": "text", + "content": " for most experiments. The motion encoder, denoted as " + }, + { + "bbox": [ + 130, + 522, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathcal{M}}" + }, + { + "bbox": [ + 130, + 522, + 482, + 665 + ], + "type": "text", + "content": ", integrates a temporal downsampling rate, " + }, + { + "bbox": [ + 130, + 522, + 482, + 665 + ], + "type": "inline_equation", + "content": "l = 4" + }, + { + "bbox": [ + 130, + 522, + 482, + 665 + ], + "type": "text", + "content": ". Our vision tokenizer incorporates a frozen Vision Transformer (ViT-L/14) [73] as visual encoder for most experiments. Additionally, for comprehensive ablation studies, we explored the use of both a frozen vision encoder and a Q-former from BLIP-2 [43] as a vision tokenizer. We mainly utilize Flan-T5-base [14] as the underlying architecture for our language model. Moreover, all our models employ the AdamW [53] optimizer with " + }, + { + "bbox": [ + 130, + 522, + 482, + 665 + ], + "type": "inline_equation", + "content": "[\\beta_1, \\beta_2] = [0.9, 0.99]" + }, + { + "bbox": [ + 130, + 522, + 482, + 665 + ], + "type": "text", + "content": " for training. The motion tokenizers are trained to utilize a " + }, + { + "bbox": [ + 130, + 522, + 482, + 665 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 130, + 522, + 482, + 665 + ], + "type": "text", + "content": " learning rate employing cosine annealing scheduler and a 256 mini-batch size. Our language models based on Flan-T5-base [14] have a " + }, + { + "bbox": [ + 130, + 522, + 482, + 665 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 130, + 522, + 482, + 665 + ], + "type": "text", + "content": " learning rate with cosine annealing scheduler and 16 mini-batch sizes in both the" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 392, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 392, + 91, + 447, + 101 + ], + "type": "text", + "content": "MotionChain" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 179, + 481, + 286 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 481, + 170 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 481, + 170 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 481, + 170 + ], + "type": "text", + "content": "Table 1: Comparison of motion reasoning on the test set of our conversation dataset. Our proposed MotionChain is fine-tuned on motion reasoning tasks while other methods' results are generated by their pre-trained weight. " + }, + { + "bbox": [ + 132, + 114, + 481, + 170 + ], + "type": "inline_equation", + "content": "Length_{avg}" + }, + { + "bbox": [ + 132, + 114, + 481, + 170 + ], + "type": "text", + "content": " represents the average words in generated answers to all questions. We adopt metrics commonly used in natural language processing tasks for evaluation." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 179, + 481, + 286 + ], + "lines": [ + { + "bbox": [ + 135, + 179, + 481, + 286 + ], + "spans": [ + { + "bbox": [ + 135, + 179, + 481, + 286 + ], + "type": "table", + "html": "
MethodsParamsLengthavgBleu@1↑Bleu@4↑Rouge↑Cider↑BertScore↑
Flan-t5-base [14]250M8.344.641.7815.3215.933.45
Flan-t5-large [14]780M11.9512.184.8322.8115.0214.19
Flan-t5-xl [14]3B9.098.544.0124.8915.0318.34
Llama-2-7b [94]7B130.8411.123.6719.141.046.81
Vicuna-1.5-7b [122]7B71.4919.277.3925.755.4419.05
Vicuna-1.5-13b [122]13B84.7417.206.5324.187.7718.00
MotionChain (Ours)573M22.1737.9219.1938.0524.5332.24
", + "image_path": "a3ebed76ae76380f8af1425574d735e86c2456539d427e0eba7e7a08c28ca116.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 307, + 481, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 307, + 481, + 355 + ], + "spans": [ + { + "bbox": [ + 132, + 307, + 481, + 355 + ], + "type": "text", + "content": "pre-train stage and the instruction tuning stage. The motion tokenizer undergoes 10000 epochs of training, while the language model undergoes 500 epochs during the pre-train stage and another 50 epochs during the instruction tuning stage. Most models are trained on 8 Tesla V100 GPUs." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 371, + 339, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 371, + 339, + 384 + ], + "spans": [ + { + "bbox": [ + 133, + 371, + 339, + 384 + ], + "type": "text", + "content": "4.2 Comparisons on Motion Reasoning." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 390, + 481, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 390, + 481, + 498 + ], + "spans": [ + { + "bbox": [ + 132, + 390, + 481, + 498 + ], + "type": "text", + "content": "In Sec. 3.1, we introduce a multi-modal motion conversation dataset, enriched with motion reasoning data facilitated by ChatGPT [62]. This task evaluates the model's reasoning capabilities with motion reasoning tasks, where a motion sequence or its corresponding textual descriptions serve as inputs. Our evaluation compares our MotionChain, which integrates motion perception, against contemporary Large Language Models (LLMs) that possess solely textual processing capabilities. The compared LLMs are assessed using their original pretrained weight. Results in Tab. 1, illustrate that MotionChain exhibits superior motion reasoning proficiency, benefiting from its integrated motion perception." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 514, + 362, + 527 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 514, + 362, + 527 + ], + "spans": [ + { + "bbox": [ + 133, + 514, + 362, + 527 + ], + "type": "text", + "content": "4.3 Comparisons on Temporal Composition." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 533, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 533, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 533, + 481, + 665 + ], + "type": "text", + "content": "The temporal motion composition task involves generating a continuous motion sequence from two actions in a time series. We conducted our experiments following the settings in TEACH [3] and used the Amass [57] subset BABEL [71] validation set. Additionally, we processed the motion in AMASS [57] into the format proposed by HumanML3D [23] and trained our MotionChain on the action-to-motion task. To compare with TEACH, we initially used an officially provided pre-trained model to sample motion on the validation set 20 times. Subsequently, we post-processed their motion into the HumanML3D format, represented in SMPL [52]. The performance of our MotionChain is summarized in Table 2. As evaluating generative models quantitatively is challenging, we also provide qualitative comparisons in the supplementary materials." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 278, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 278, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 278, + 102 + ], + "type": "text", + "content": "Biao Jiang, Xin Chen et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 165, + 157, + 449, + 220 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 149 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 149 + ], + "type": "text", + "content": "Table 2: Comparison of temporal motion composition on Babel [71]. We evaluate the state-of-the-art motion temporal composition method Teach [3] under the " + }, + { + "bbox": [ + 130, + 114, + 482, + 149 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 130, + 114, + 482, + 149 + ], + "type": "text", + "content": " confidence interval from 20 times running. (cf. Sec. 4.1 for notations.)" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 165, + 157, + 449, + 220 + ], + "lines": [ + { + "bbox": [ + 165, + 157, + 449, + 220 + ], + "spans": [ + { + "bbox": [ + 165, + 157, + 449, + 220 + ], + "type": "table", + "html": "
MethodsDiversityMPJPE↓PA-MPJPE↓ACCL↓
Real15.74±.149---
Teach [3]27.11±.159979.21±.215933.32±.25423.02±.018
MotionChain (Ours)43.25±.159276.05±6.7253.72±.5807.11±0.100
", + "image_path": "bd4c4bd18243fa1ca18d6361086ab82baf9ef9bea4279cd54c24af155416cb04.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 192, + 274, + 422, + 332 + ], + "blocks": [ + { + "bbox": [ + 130, + 231, + 482, + 266 + ], + "lines": [ + { + "bbox": [ + 130, + 231, + 482, + 266 + ], + "spans": [ + { + "bbox": [ + 130, + 231, + 482, + 266 + ], + "type": "text", + "content": "Table 3: Evaluation of motion composition methods on HumanML3D [23]. Here Independent, Past-condition, and Tokens-joint stand for different motion composition variants during multi-turn motion conversation, as illustrated in Fig. 4." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 192, + 274, + 422, + 332 + ], + "lines": [ + { + "bbox": [ + 192, + 274, + 422, + 332 + ], + "spans": [ + { + "bbox": [ + 192, + 274, + 422, + 332 + ], + "type": "table", + "html": "
MethodMPJPE↓PA-MPJPE↓ACCL↓Diversity
Independent350.79102.9711.406.47
Past-condition232.4646.156.186.01
Tokens-joint108.7718.852.265.56
", + "image_path": "f6c80ae33c1d635aad20d36ec9842fb327502938d709b845b4ab7e49179f44d2.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 354, + 246, + 366 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 354, + 246, + 366 + ], + "spans": [ + { + "bbox": [ + 132, + 354, + 246, + 366 + ], + "type": "text", + "content": "4.4 Ablation Studies" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 377, + 482, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 377, + 482, + 472 + ], + "spans": [ + { + "bbox": [ + 130, + 377, + 482, + 472 + ], + "type": "text", + "content": "MotionChain enables multi-modal motion conversation using two main techniques. The first technique involves generating a smooth sequence of motions by concatenating motion tokens which are then decoded back to motion by motion decoder " + }, + { + "bbox": [ + 130, + 377, + 482, + 472 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathcal{M}}" + }, + { + "bbox": [ + 130, + 377, + 482, + 472 + ], + "type": "text", + "content": ". The second technique involves processing multi-modal visual input through a vision tokenizer, which consists of a frozen vision encoder and a trainable linear projection. To evaluate the effectiveness of these two designs, we compare them with other variants. For a more comprehensive analysis, detailed ablation studies can be found in the supplementary materials." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 474, + 482, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 482, + 568 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 482, + 568 + ], + "type": "text", + "content": "Motion Composition Mechanism Apart from the jointly token concatenating mechanism, we also evaluate the performance of temporal motion composition through the other motion temporal composition variants Motion-cat: concatenating the motion in final motion level rather than token level. Experimental results in Tab. 3 show that jointly concatenating motion tokens achieved remarkable performance compared to the other variants. For further information regarding the implementation of the aforementioned vision tokenizer, please refer to the supplementary materials." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "Image Tokenizer Architecture. MotionChain connects the frozen vision encoder to the language model through a linear layer. However, previous vision-language [1,43] works also demonstrate the effectiveness of other kinds of visual-aligning modules. Here we consider the other two vision tokenizer variants: (a) inspired by [1,10,31], we introduce a perceiver module that incorporates a transformer receiving a predefined number of latent input queries. These queries cross-attend to the visual features, enabling effective information exchange. (b) We directly adopt the pre-trained Q-former from BLIP-2 [43] to align visual" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 392, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 392, + 91, + 447, + 101 + ], + "type": "text", + "content": "MotionChain" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 179, + 168, + 434, + 243 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 160 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 160 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 160 + ], + "type": "text", + "content": "Table 4: Evaluation of vision tokenizer architecture on Bedlam [49]. We implement three different architectures, including Q-former, Perceiver, and Linear. We evaluate these results with the metrics in motion reconstruction. Additional information regarding the implementation is in the supplementary materials. (cf. Tab. 2 for notations.)" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 179, + 168, + 434, + 243 + ], + "lines": [ + { + "bbox": [ + 179, + 168, + 434, + 243 + ], + "spans": [ + { + "bbox": [ + 179, + 168, + 434, + 243 + ], + "type": "table", + "html": "
ArchitectureFirst-frameLast-frame
MPJPE ↓PA-MPJPE ↓MPJPE ↓PA-MPJPE ↓
Q-former195.4986.56134.7357.17
Perceiver185.6199.21134.8957.58
Linear144.3776.48133.7356.73
", + "image_path": "20efcd51b2d5d3e8c826b95b148cfe69c873d871d2168248ffc94b3fc89d8732.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 268, + 482, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 268, + 482, + 341 + ], + "spans": [ + { + "bbox": [ + 130, + 268, + 482, + 341 + ], + "type": "text", + "content": "inputs with the language model. We evaluate the different architectures under the single human image as the first frame condition and the last frame condition separately. Experimental results in Tab. 4 show that a lightweight linear projection is sufficient for comprehending the human pose from visual input. Additional details about the implementation of the above vision tokenizer can be found in the supplements." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 366, + 312, + 380 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 366, + 312, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 366, + 312, + 380 + ], + "type": "text", + "content": "5 Conclusion and Limitation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 400, + 482, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 400, + 482, + 520 + ], + "spans": [ + { + "bbox": [ + 130, + 400, + 482, + 520 + ], + "type": "text", + "content": "Limitation. As the trial to explore conversational human motion generation with visual language models, the proposed MotionChain still has limitations as follows. MotionChain utilizes indeterministic generative models, similar to other language models, but other traditional or neural motion controllers [86, 87] are mostly deterministic and sensitive to control signals. Besides, our method can only generate motion on articulated human bodies, excluding many other human parts such as faces [9, 34, 72] and hands [46, 46, 47, 47, 78]. Although we utilize vision, language, and motion as multimodal conditional inputs akin to human perception, MotionChain is still restricted to the collision signals for human-object and human-scene interactions [35, 82, 112]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": "Conclusion. We summarize the proposed MotionChain as a conversational human motion controller to generate continuous and long-term human motion through multimodal prompts. Compared to these one-turn motion generation methods [32, 92, 117], our MotionChain produces more contextually rich generation and can achieve the step-by-step process of human task execution for humanoid robotics and game agents. By leveraging large-scale language, vision-language, and vision-motion data to assist motion-related generation tasks, MotionChain thus comprehends each instruction in multi-turn conversation and generates human motions followed by these prompts. Extensive experiments validate the efficacy of MotionChain, demonstrating state-of-the-art performance in conversational motion generation, as well as more intuitive manners of controlling and interacting with virtual humans." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 278, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 278, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 278, + 102 + ], + "type": "text", + "content": "Biao Jiang, Xin Chen et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 138, + 494, + 210 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 138, + 494, + 210 + ], + "spans": [ + { + "bbox": [ + 130, + 138, + 494, + 210 + ], + "type": "text", + "content": "This work is supported by National Natural Science Foundation of China (No. 62101137), National Key Research and Development Program of China (No. 2022ZD0160101), Shanghai Natural Science Foundation (No. 23ZR1402900), Shanghai Municipal Science and Technology Major Project (No.2021SHZDZX0103). The computations in this research were performed using the CFFF platform of Fudan University." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 228, + 197, + 240 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 228, + 197, + 240 + ], + "spans": [ + { + "bbox": [ + 132, + 228, + 197, + 240 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 251, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 141, + 251, + 481, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 251, + 481, + 295 + ], + "spans": [ + { + "bbox": [ + 141, + 251, + 481, + 295 + ], + "type": "text", + "content": "1. Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., et al.: Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems 35, 23716-23736 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 141, + 295, + 481, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 295, + 481, + 350 + ], + "spans": [ + { + "bbox": [ + 141, + 295, + 481, + 350 + ], + "type": "text", + "content": "2. Anderson, P., Wu, Q., Teney, D., Bruce, J., Johnson, M., Sünderhauf, N., Reid, I., Gould, S., Van Den Hengel, A.: Vision-and-language navigation: Interpreting visually-grounded navigation instructions in real environments. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3674-3683 (2018)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 141, + 350, + 481, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 350, + 481, + 382 + ], + "spans": [ + { + "bbox": [ + 141, + 350, + 481, + 382 + ], + "type": "text", + "content": "3. Athanasiou, N., Petrovich, M., Black, M.J., Varol, G.: Teach: Temporal action compositions for 3d humans. In: International Conference on 3D Vision (3DV) (September 2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 141, + 383, + 481, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 383, + 481, + 415 + ], + "spans": [ + { + "bbox": [ + 141, + 383, + 481, + 415 + ], + "type": "text", + "content": "4. Bazavan, E.G., Zanfir, A., Zanfir, M., Freeman, W.T., Sukthankar, R., Sminchisescu, C.: Hspace: Synthetic parametric humans animated in complex environments. arXiv preprint arXiv:2112.12867 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 415, + 481, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 415, + 481, + 449 + ], + "spans": [ + { + "bbox": [ + 141, + 415, + 481, + 449 + ], + "type": "text", + "content": "5. Black, M.J., Patel, P., Tesch, J., Yang, J.: Bedlam: A synthetic dataset of bodies exhibiting detailed lifelike animated motion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8726-8737 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 141, + 449, + 481, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 449, + 481, + 491 + ], + "spans": [ + { + "bbox": [ + 141, + 449, + 481, + 491 + ], + "type": "text", + "content": "6. Borsos, Z., Marinier, R., Vincent, D., Kharitonov, E., Pietquin, O., Sharifi, M., Roblek, D., Teboul, O., Grangier, D., Tagliasacchi, M., et al.: Audiolm: a language modeling approach to audio generation. IEEE/ACM transactions on audio, speech, and language processing 31, 2523-2533 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 141, + 491, + 481, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 491, + 481, + 524 + ], + "spans": [ + { + "bbox": [ + 141, + 491, + 481, + 524 + ], + "type": "text", + "content": "7. Brooks, T., Holynski, A., Efros, A.A.: Instructpix2pix: Learning to follow image editing instructions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18392-18402 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 141, + 524, + 481, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 524, + 481, + 556 + ], + "spans": [ + { + "bbox": [ + 141, + 524, + 481, + 556 + ], + "type": "text", + "content": "8. Cai, Z., Zhang, M., Ren, J., Wei, C., Ren, D., Lin, Z., Zhao, H., Yang, L., Loy, C.C., Liu, Z.: Playing for 3d human recovery. arXiv preprint arXiv:2110.07588 (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 141, + 556, + 481, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 556, + 481, + 589 + ], + "spans": [ + { + "bbox": [ + 141, + 556, + 481, + 589 + ], + "type": "text", + "content": "9. Cao, X., Chen, Z., Chen, A., Chen, X., Li, S., Yu, J.: Sparse photometric 3d face reconstruction guided by morphable models. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 4635-4644 (2018)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 589, + 481, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 589, + 481, + 622 + ], + "spans": [ + { + "bbox": [ + 138, + 589, + 481, + 622 + ], + "type": "text", + "content": "10. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: European conference on computer vision. pp. 213-229. Springer (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 622, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 622, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 622, + 481, + 665 + ], + "type": "text", + "content": "11. Changpinyo, S., Sharma, P., Ding, N., Soricut, R.: Conceptual 12m: Pushing web-scale image-text pre-training to recognize long-tail visual concepts. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3558-3568 (2021)" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 392, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 392, + 91, + 447, + 101 + ], + "type": "text", + "content": "MotionChain" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 138, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 138, + 116, + 481, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 116, + 481, + 161 + ], + "spans": [ + { + "bbox": [ + 138, + 116, + 481, + 161 + ], + "type": "text", + "content": "12. Chen, S., Chen, X., Zhang, C., Li, M., Yu, G., Fei, H., Zhu, H., Fan, J., Chen, T.: Ll3da: Visual interactive instruction tuning for omni-3d understanding reasoning and planning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 26428-26438 (2024)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 162, + 481, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 162, + 481, + 195 + ], + "spans": [ + { + "bbox": [ + 138, + 162, + 481, + 195 + ], + "type": "text", + "content": "13. Choudhury, R., Kitani, K.M., Jeni, L.A.: Tempo: Efficient multi-view pose estimation, tracking, and forecasting. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 14750-14760 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 196, + 481, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 196, + 481, + 228 + ], + "spans": [ + { + "bbox": [ + 138, + 196, + 481, + 228 + ], + "type": "text", + "content": "14. Chung, H.W., Hou, L., Longpre, S., Zoph, B., Tay, Y., Fedus, W., Li, E., Wang, X., Dehghani, M., Brahma, S., et al.: Scaling instruction-finetuned language models. arXiv preprint arXiv:2210.11416 (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 228, + 481, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 228, + 481, + 251 + ], + "spans": [ + { + "bbox": [ + 138, + 228, + 481, + 251 + ], + "type": "text", + "content": "15. Clavet, S.: Motion matching and the road to next-gen animation. In: Proc. of GDC. vol. 2, p. 9 (2016)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 251, + 481, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 251, + 481, + 284 + ], + "spans": [ + { + "bbox": [ + 138, + 251, + 481, + 284 + ], + "type": "text", + "content": "16. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 285, + 481, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 285, + 481, + 307 + ], + "spans": [ + { + "bbox": [ + 138, + 285, + 481, + 307 + ], + "type": "text", + "content": "17. Gilardi, F., Alizadeh, M., Kubli, M.: Chatgpt outperforms crowd-workers for text-annotation tasks. arXiv preprint arXiv:2303.15056 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 308, + 481, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 308, + 481, + 340 + ], + "spans": [ + { + "bbox": [ + 138, + 308, + 481, + 340 + ], + "type": "text", + "content": "18. Girdhar, R., El-Nouby, A., Liu, Z., Singh, M., Alwala, K.V., Joulin, A., Misra, I.: Imagebind: One embedding space to bind them all. arXiv preprint arXiv:2305.05665 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 341, + 481, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 341, + 481, + 373 + ], + "spans": [ + { + "bbox": [ + 138, + 341, + 481, + 373 + ], + "type": "text", + "content": "19. Goel, S., Pavlakos, G., Rajasegaran, J., Kanazawa, A., Malik, J.: Humans in 4d: Reconstructing and tracking humans with transformers. arXiv preprint arXiv:2305.20091 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 375, + 481, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 375, + 481, + 407 + ], + "spans": [ + { + "bbox": [ + 138, + 375, + 481, + 407 + ], + "type": "text", + "content": "20. Goutsu, Y., Inamura, T.: Linguistic descriptions of human motion with generative adversarial seq2seq learning. In: 2021 IEEE International Conference on Robotics and Automation (ICRA). pp. 4281-4287. IEEE (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 408, + 481, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 408, + 481, + 441 + ], + "spans": [ + { + "bbox": [ + 138, + 408, + 481, + 441 + ], + "type": "text", + "content": "21. Guler, R.A., Kokkinos, I.: Holopose: Holistic 3d human reconstruction in-the-wild. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10884-10894 (2019)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 442, + 481, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 442, + 481, + 464 + ], + "spans": [ + { + "bbox": [ + 138, + 442, + 481, + 464 + ], + "type": "text", + "content": "22. Guo, C., Mu, Y., Javed, M.G., Wang, S., Cheng, L.: Momask: Generative masked modeling of 3d human motions. arXiv preprint arXiv:2312.00063 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 464, + 481, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 464, + 481, + 508 + ], + "spans": [ + { + "bbox": [ + 138, + 464, + 481, + 508 + ], + "type": "text", + "content": "23. Guo, C., Zou, S., Zuo, X., Wang, S., Ji, W., Li, X., Cheng, L.: Generating diverse and natural 3d human motions from text. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5152-5161 (June 2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 509, + 481, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 509, + 481, + 531 + ], + "spans": [ + { + "bbox": [ + 138, + 509, + 481, + 531 + ], + "type": "text", + "content": "24. Guo, C., Zuo, X., Wang, S., Cheng, L.: Tm2t: Stochastic and tokenized modeling for the reciprocal generation of 3d human motions and texts. In: ECCV (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 532, + 481, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 532, + 481, + 564 + ], + "spans": [ + { + "bbox": [ + 138, + 532, + 481, + 564 + ], + "type": "text", + "content": "25. Guo, C., Zuo, X., Wang, S., Zou, S., Sun, Q., Deng, A., Gong, M., Cheng, L.: Action2motion: Conditioned generation of 3d human motions. In: Proceedings of the 28th ACM International Conference on Multimedia. pp. 2021-2029 (2020)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 565, + 481, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 565, + 481, + 609 + ], + "spans": [ + { + "bbox": [ + 138, + 565, + 481, + 609 + ], + "type": "text", + "content": "26. Guo, Z., Zhang, R., Zhu, X., Tang, Y., Ma, X., Han, J., Chen, K., Gao, P., Li, X., Li, H., et al.: Point-bind & point-llm: Aligning point cloud with multimodality for 3d understanding, generation, and instruction following. arXiv preprint arXiv:2309.00615 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 609, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 609, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 609, + 481, + 632 + ], + "type": "text", + "content": "27. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems 33, 6840-6851 (2020)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 632, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 632, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 632, + 481, + 665 + ], + "type": "text", + "content": "28. Hong, Y., Zhen, H., Chen, P., Zheng, S., Du, Y., Chen, Z., Gan, C.: 3d-lm: Injecting the 3d world into large language models. arXiv preprint arXiv:2307.12981 (2023)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 278, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 278, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 278, + 101 + ], + "type": "text", + "content": "Biao Jiang, Xin Chen et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 138, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 138, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 138, + 116, + 480, + 149 + ], + "type": "text", + "content": "29. Huang, S., Dong, L., Wang, W., Hao, Y., Singhal, S., Ma, S., Lv, T., Cui, L., Mohammed, O.K., Liu, Q., et al.: Language is not all you need: Aligning perception with language models. arXiv preprint arXiv:2302.14045 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 149, + 481, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 149, + 481, + 192 + ], + "spans": [ + { + "bbox": [ + 138, + 149, + 481, + 192 + ], + "type": "text", + "content": "30. Ionescu, C., Papava, D., Olaru, V., Sminchisescu, C.: Human3. 6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. IEEE transactions on pattern analysis and machine intelligence 36(7), 1325-1339 (2013)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 192, + 480, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 192, + 480, + 225 + ], + "spans": [ + { + "bbox": [ + 138, + 192, + 480, + 225 + ], + "type": "text", + "content": "31. Jaegle, A., Gimeno, F., Brock, A., Vinyals, O., Zisserman, A., Carreira, J.: Perceiver: General perception with iterative attention. In: International conference on machine learning. pp. 4651-4664. PMLR (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 225, + 480, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 225, + 480, + 246 + ], + "spans": [ + { + "bbox": [ + 138, + 225, + 480, + 246 + ], + "type": "text", + "content": "32. Jiang, B., Chen, X., Liu, W., Yu, J., Yu, G., Chen, T.: Motiongpt: Human motion as a foreign language. arXiv preprint arXiv:2306.14795 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 246, + 480, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 246, + 480, + 278 + ], + "spans": [ + { + "bbox": [ + 138, + 246, + 480, + 278 + ], + "type": "text", + "content": "33. Kanazawa, A., Black, M.J., Jacobs, D.W., Malik, J.: End-to-end recovery of human shape and pose. In: Computer Vision and Pattern Regognition (CVPR) (2018)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 278, + 480, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 278, + 480, + 311 + ], + "spans": [ + { + "bbox": [ + 138, + 278, + 480, + 311 + ], + "type": "text", + "content": "34. Karras, T., Aila, T., Laine, S., Herva, A., Lehtinen, J.: Audio-driven facial animation by joint end-to-end learning of pose and emotion. ACM Transactions on Graphics (TOG) 36(4), 1-12 (2017)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 311, + 480, + 343 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 311, + 480, + 343 + ], + "spans": [ + { + "bbox": [ + 138, + 311, + 480, + 343 + ], + "type": "text", + "content": "35. Karunratanakul, K., Preechakul, K., Suwajanakorn, S., Tang, S.: Guided motion diffusion for controllable human motion synthesis. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2151-2162 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 343, + 480, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 343, + 480, + 374 + ], + "spans": [ + { + "bbox": [ + 138, + 343, + 480, + 374 + ], + "type": "text", + "content": "36. Kocabas, M., Athanasiou, N., Black, M.J.: Vibe: Video inference for human body pose and shape estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 374, + 480, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 374, + 480, + 418 + ], + "spans": [ + { + "bbox": [ + 138, + 374, + 480, + 418 + ], + "type": "text", + "content": "37. Kocabas, M., Huang, C.H.P., Tesch, J., Müller, L., Hilliges, O., Black, M.J.: Spec: Seeing people in the wild with an estimated camera. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 11035-11045 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 418, + 480, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 418, + 480, + 439 + ], + "spans": [ + { + "bbox": [ + 138, + 418, + 480, + 439 + ], + "type": "text", + "content": "38. Kovar, L., Gleicher, M., Pighin, F.: Motion graphs. In: Seminal Graphics Papers: Pushing the Boundaries, Volume 2, pp. 723-732 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 439, + 480, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 439, + 480, + 472 + ], + "spans": [ + { + "bbox": [ + 138, + 439, + 480, + 472 + ], + "type": "text", + "content": "39. Kudo, T., Richardson, J.: Sentencepiece: A simple and language independent subword tokenizer and detokenizer for neural text processing. arXiv preprint arXiv:1808.06226 (2018)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 472, + 480, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 472, + 480, + 504 + ], + "spans": [ + { + "bbox": [ + 138, + 472, + 480, + 504 + ], + "type": "text", + "content": "40. Lee, T., Moon, G., Lee, K.M.: Multiact: Long-term 3d human motion generation from multiple action labels. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 1231-1239 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 504, + 480, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 504, + 480, + 536 + ], + "spans": [ + { + "bbox": [ + 138, + 504, + 480, + 536 + ], + "type": "text", + "content": "41. Lee, Y., Wampler, K., Bernstein, G., Popovic, J., Popovic, Z.: Motion fields for interactive character locomotion. In: ACM SIGGRAPH Asia 2010 papers, pp. 1-8 (2010)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 536, + 480, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 536, + 480, + 567 + ], + "spans": [ + { + "bbox": [ + 138, + 536, + 480, + 567 + ], + "type": "text", + "content": "42. Levine, S., Wang, J.M., Haraux, A., Popovic, Z., Koltun, V.: Continuous character control with low-dimensional embeddings. ACM Transactions on Graphics (TOG) 31(4), 1-10 (2012)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 568, + 480, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 568, + 480, + 601 + ], + "spans": [ + { + "bbox": [ + 138, + 568, + 480, + 601 + ], + "type": "text", + "content": "43. Li, J., Li, D., Savarese, S., Hoi, S.: Blip-2: Bootstrapping language-image pretraining with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 601, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 601, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 601, + 480, + 632 + ], + "type": "text", + "content": "44. Li, J., Li, D., Xiong, C., Hoi, S.: Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In: International Conference on Machine Learning. pp. 12888-12900. PMLR (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "type": "text", + "content": "45. Li, M., Chen, X., Zhang, C., Chen, S., Zhu, H., Yin, F., Yu, G., Chen, T.: M3dbench: Let's instruct large models with multi-modal 3d prompts. arXiv preprint arXiv:2312.10763 (2023)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 392, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 392, + 91, + 447, + 100 + ], + "type": "text", + "content": "MotionChain" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 138, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 138, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 138, + 116, + 480, + 138 + ], + "type": "text", + "content": "46. Li, Y., Wu, M., Zhang, Y., Xu, L., Yu, J.: Piano: A parametric hand bone model from magnetic resonance imaging. arXiv preprint arXiv:2106.10893 (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 138, + 480, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 138, + 480, + 171 + ], + "spans": [ + { + "bbox": [ + 138, + 138, + 480, + 171 + ], + "type": "text", + "content": "47. Li, Y., Zhang, L., Qiu, Z., Jiang, Y., Li, N., Ma, Y., Zhang, Y., Xu, L., Yu, J.: Nimble: a non-rigid hand model with bones and muscles. ACM Transactions on Graphics (TOG) 41(4), 1-16 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 171, + 480, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 171, + 480, + 192 + ], + "spans": [ + { + "bbox": [ + 138, + 171, + 480, + 192 + ], + "type": "text", + "content": "48. Lin, C.Y.: Rouge: A package for automatic evaluation of summaries. In: Text summarization branches out. pp. 74-81 (2004)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 192, + 480, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 192, + 480, + 224 + ], + "spans": [ + { + "bbox": [ + 138, + 192, + 480, + 224 + ], + "type": "text", + "content": "49. Lin, J., Zeng, A., Lu, S., Cai, Y., Zhang, R., Wang, H., Zhang, L.: Motion-x: A large-scale 3d expressive whole-body human motion dataset. arXiv preprint arXiv:2307.00818 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 224, + 480, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 224, + 480, + 245 + ], + "spans": [ + { + "bbox": [ + 138, + 224, + 480, + 245 + ], + "type": "text", + "content": "50. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. arXiv preprint arXiv:2304.08485 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 245, + 480, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 245, + 480, + 266 + ], + "spans": [ + { + "bbox": [ + 138, + 245, + 480, + 266 + ], + "type": "text", + "content": "51. Liu, X., Yan, H., Zhang, S., An, C., Qiu, X., Lin, D.: Scaling laws of rope-based extrapolation. arXiv preprint arXiv:2310.05209 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 266, + 480, + 299 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 266, + 480, + 299 + ], + "spans": [ + { + "bbox": [ + 138, + 266, + 480, + 299 + ], + "type": "text", + "content": "52. Loper, M., Mahmood, N., Romero, J., Pons-Moll, G., Black, M.J.: Smpl: A skinned multi-person linear model. ACM Trans. Graph. 34(6), 248:1-248:16 (Oct 2015)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 299, + 480, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 299, + 480, + 319 + ], + "spans": [ + { + "bbox": [ + 138, + 299, + 480, + 319 + ], + "type": "text", + "content": "53. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101 (2017)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 319, + 480, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 319, + 480, + 363 + ], + "spans": [ + { + "bbox": [ + 138, + 319, + 480, + 363 + ], + "type": "text", + "content": "54. Lu, C., Yin, F., Chen, X., Liu, W., Chen, T., Yu, G., Fan, J.: A large-scale outdoor multi-modal dataset and benchmark for novel view synthesis and implicit scene reconstruction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7557-7567 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 363, + 480, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 363, + 480, + 396 + ], + "spans": [ + { + "bbox": [ + 138, + 363, + 480, + 396 + ], + "type": "text", + "content": "55. Lu, S., Chen, L.H., Zeng, A., Lin, J., Zhang, R., Zhang, L., Shum, H.Y.: Humantomato: Text-aligned whole-body motion generation. arXiv preprint arXiv:2310.12978 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 396, + 480, + 438 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 396, + 480, + 438 + ], + "spans": [ + { + "bbox": [ + 138, + 396, + 480, + 438 + ], + "type": "text", + "content": "56. Ma, H., Li, J., Hosseini, R., Tomizuka, M., Choi, C.: Multi-objective diverse human motion prediction with knowledge distillation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8161-8171 (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 438, + 480, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 438, + 480, + 471 + ], + "spans": [ + { + "bbox": [ + 138, + 438, + 480, + 471 + ], + "type": "text", + "content": "57. Mahmood, N., Ghorbani, N., Troje, N.F., Pons-Moll, G., Black, M.J.: Amass: Archive of motion capture as surface shapes. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (October 2019)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 471, + 480, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 471, + 480, + 514 + ], + "spans": [ + { + "bbox": [ + 138, + 471, + 480, + 514 + ], + "type": "text", + "content": "58. Mehta, D., Rhodin, H., Casas, D., Fua, P., Sotnychenko, O., Xu, W., Theobalt, C.: Monocular 3d human pose estimation in the wild using improved cnn supervision. In: International Conference on 3D Vision (3DV) (2017). https://doi.org/10.1109/3dv.2017.00064, http://gvv.mpi-inf.mpg.de/3dhp_dataset" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 514, + 480, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 514, + 480, + 547 + ], + "spans": [ + { + "bbox": [ + 138, + 514, + 480, + 547 + ], + "type": "text", + "content": "59. Min, J., Chai, J.: Motion graphs++ a compact generative model for semantic motion analysis and synthesis. ACM Transactions on Graphics (TOG) 31(6), 1-12 (2012)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 547, + 312, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 547, + 312, + 557 + ], + "spans": [ + { + "bbox": [ + 138, + 547, + 312, + 557 + ], + "type": "text", + "content": "60. OpenAI: Gpt-4 technical report (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 557, + 480, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 557, + 480, + 589 + ], + "spans": [ + { + "bbox": [ + 138, + 557, + 480, + 589 + ], + "type": "text", + "content": "61. Ordonez, V., Kulkarni, G., Berg, T.: Im2text: Describing images using 1 million captioned photographs. Advances in neural information processing systems 24 (2011)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 589, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 589, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 589, + 480, + 632 + ], + "type": "text", + "content": "62. Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., et al.: Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems 35, 27730-27744 (2022)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "type": "text", + "content": "63. Papineni, K., Roukos, S., Ward, T., Zhu, W.J.: Bleu: a method for automatic evaluation of machine translation. In: Proceedings of the 40th annual meeting of the Association for Computational Linguistics. pp. 311-318 (2002)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 278, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 278, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 278, + 101 + ], + "type": "text", + "content": "Biao Jiang, Xin Chen et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 138, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 138, + 116, + 480, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 116, + 480, + 160 + ], + "spans": [ + { + "bbox": [ + 138, + 116, + 480, + 160 + ], + "type": "text", + "content": "64. Patel, P., Huang, C.H.P., Tesch, J., Hoffmann, D.T., Tripathi, S., Black, M.J.: Agora: Avatars in geography optimized for regression analysis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13468-13478 (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 160, + 480, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 160, + 480, + 204 + ], + "spans": [ + { + "bbox": [ + 138, + 160, + 480, + 204 + ], + "type": "text", + "content": "65. Pavlakos, G., Choutas, V., Ghorbani, N., Bolkart, T., Osman, A.A.A., Tzionas, D., Black, M.J.: Expressive body capture: 3d hands, face, and body from a single image. In: Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR) (2019)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 204, + 480, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 204, + 480, + 236 + ], + "spans": [ + { + "bbox": [ + 138, + 204, + 480, + 236 + ], + "type": "text", + "content": "66. Petrovich, M., Black, M.J., Varol, G.: Action-conditioned 3D human motion synthesis with transformer VAE. In: International Conference on Computer Vision (ICCV) (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 236, + 480, + 268 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 236, + 480, + 268 + ], + "spans": [ + { + "bbox": [ + 138, + 236, + 480, + 268 + ], + "type": "text", + "content": "67. Petrovich, M., Black, M.J., Varol, G.: TEMOS: Generating diverse human motions from textual descriptions. In: European Conference on Computer Vision (ECCV) (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 268, + 480, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 268, + 480, + 289 + ], + "spans": [ + { + "bbox": [ + 138, + 268, + 480, + 289 + ], + "type": "text", + "content": "68. Petrovich, M., Black, M.J., Varol, G.: Tmr: Text-to-motion retrieval using contrastive 3d human motion synthesis. arXiv preprint arXiv:2305.00976 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 289, + 480, + 321 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 289, + 480, + 321 + ], + "spans": [ + { + "bbox": [ + 138, + 289, + 480, + 321 + ], + "type": "text", + "content": "69. Plappert, M., Mandery, C., Asfour, T.: The kit motion-language dataset. Big Data 4(4), 236-252 (dec 2016). https://doi.org/10.1089/big.2016.0028, http://dx.doi.org/10.1089/big.2016.0028" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 321, + 480, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 321, + 480, + 354 + ], + "spans": [ + { + "bbox": [ + 138, + 321, + 480, + 354 + ], + "type": "text", + "content": "70. Plappert, M., Mandery, C., Asfour, T.: Learning a bidirectional mapping between human whole-body motion and natural language using deep recurrent neural networks. Robotics and Autonomous Systems 109, 13-26 (2018)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 354, + 480, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 354, + 480, + 396 + ], + "spans": [ + { + "bbox": [ + 138, + 354, + 480, + 396 + ], + "type": "text", + "content": "71. Punnakkal, A.R., Chandrasekaran, A., Athanasiou, N., Quiros-Ramirez, A., Black, M.J.: BABEL: Bodies, action and behavior with english labels. In: Proceedings IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR). pp. 722-731 (Jun 2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 396, + 480, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 396, + 480, + 429 + ], + "spans": [ + { + "bbox": [ + 138, + 396, + 480, + 429 + ], + "type": "text", + "content": "72. Qiu, Z., Li, Y., He, D., Zhang, Q., Zhang, L., Zhang, Y., Wang, J., Xu, L., Wang, X., Zhang, Y., et al.: Sculptor: Skeleton-consistent face creation using a learned parametric generator. ACM Transactions on Graphics (TOG) 41(6), 1-17 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 429, + 480, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 429, + 480, + 472 + ], + "spans": [ + { + "bbox": [ + 138, + 429, + 480, + 472 + ], + "type": "text", + "content": "73. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International Conference on Machine Learning. pp. 8748-8763. PMLR (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 472, + 480, + 515 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 472, + 480, + 515 + ], + "spans": [ + { + "bbox": [ + 138, + 472, + 480, + 515 + ], + "type": "text", + "content": "74. Raffel, C., Shazeer, N., Roberts, A., Lee, K., Narang, S., Matena, M., Zhou, Y., Li, W., Liu, P.J.: Exploring the limits of transfer learning with a unified text-to-text transformer. The Journal of Machine Learning Research 21(1), 5485-5551 (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 515, + 480, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 515, + 480, + 548 + ], + "spans": [ + { + "bbox": [ + 138, + 515, + 480, + 548 + ], + "type": "text", + "content": "75. Rajasegaran, J., Pavlakos, G., Kanazawa, A., Malik, J.: Tracking people by predicting 3d appearance, location and pose. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2740-2749 (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 548, + 480, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 548, + 480, + 568 + ], + "spans": [ + { + "bbox": [ + 138, + 548, + 480, + 568 + ], + "type": "text", + "content": "76. Razavi, A., Van den Oord, A., Vinyals, O.: Generating diverse high-fidelity images with vq-vae-2. Advances in neural information processing systems 32 (2019)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 568, + 480, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 568, + 480, + 611 + ], + "spans": [ + { + "bbox": [ + 138, + 568, + 480, + 611 + ], + "type": "text", + "content": "77. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2022), https://github.com/CompVis/latent-diffusionhttps://arxiv.org/abs/2112.10752" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 611, + 480, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 611, + 480, + 633 + ], + "spans": [ + { + "bbox": [ + 138, + 611, + 480, + 633 + ], + "type": "text", + "content": "78. Romero, J., Tzionas, D., Black, M.J.: Embodied hands: Modeling and capturing hands and bodies together. arXiv preprint arXiv:2201.02610 (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "type": "text", + "content": "79. Rose, C., Cohen, M.F., Bodenheimer, B.: Verbs and adverbs: Multidimensional motion interpolation. IEEE Computer Graphics and Applications 18(5), 32-40 (1998)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 392, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 392, + 91, + 447, + 101 + ], + "type": "text", + "content": "MotionChain" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + }, + { + "para_blocks": [ + { + "bbox": [ + 138, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 138, + 116, + 480, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 116, + 480, + 160 + ], + "spans": [ + { + "bbox": [ + 138, + 116, + 480, + 160 + ], + "type": "text", + "content": "80. Rubenstein, P.K., Asawaroengchai, C., Nguyen, D.D., Bapna, A., Borsos, Z., Quitry, F.d.C., Chen, P., Badawy, D.E., Han, W., Kharitonov, E., et al.: Audiopalm: A large language model that can speak and listen. arXiv preprint arXiv:2306.12925 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 161, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 161, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 138, + 161, + 480, + 205 + ], + "type": "text", + "content": "81. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems 35, 25278-25294 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 206, + 480, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 206, + 480, + 228 + ], + "spans": [ + { + "bbox": [ + 138, + 206, + 480, + 228 + ], + "type": "text", + "content": "82. Shafir, Y., Tevet, G., Kapon, R., Bermano, A.H.: Human motion diffusion as a generative prior. arXiv preprint arXiv:2303.01418 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 228, + 480, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 228, + 480, + 272 + ], + "spans": [ + { + "bbox": [ + 138, + 228, + 480, + 272 + ], + "type": "text", + "content": "83. Siddiqui, Y., Alliegro, A., Artemov, A., Tommasi, T., Sirigatti, D., Rosov, V., Dai, A., Nießner, M.: Meshgpt: Generating triangle meshes with decoder-only transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19615-19625 (2024)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 274, + 480, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 274, + 480, + 317 + ], + "spans": [ + { + "bbox": [ + 138, + 274, + 480, + 317 + ], + "type": "text", + "content": "84. Siyao, L., Yu, W., Gu, T., Lin, C., Wang, Q., Qian, C., Loy, C.C., Liu, Z.: Bailando: 3d dance generation by actor-critic gpt with choreographic memory. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11050-11059 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 318, + 480, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 318, + 480, + 340 + ], + "spans": [ + { + "bbox": [ + 138, + 318, + 480, + 340 + ], + "type": "text", + "content": "85. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502 (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 341, + 480, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 341, + 480, + 373 + ], + "spans": [ + { + "bbox": [ + 138, + 341, + 480, + 373 + ], + "type": "text", + "content": "86. Starke, S., Mason, I., Komura, T.: Deepphase: periodic autoencoders for learning motion phase manifolds. ACM Transactions on Graphics (TOG) 41(4), 1-13 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 374, + 480, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 374, + 480, + 396 + ], + "spans": [ + { + "bbox": [ + 138, + 374, + 480, + 396 + ], + "type": "text", + "content": "87. Starke, S., Zhang, H., Komura, T., Saito, J.: Neural state machine for character-scene interactions. ACM Trans. Graph. 38(6), 209-1 (2019)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 397, + 480, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 397, + 480, + 430 + ], + "spans": [ + { + "bbox": [ + 138, + 397, + 480, + 430 + ], + "type": "text", + "content": "88. Starke, S., Zhao, Y., Zinno, F., Komura, T.: Neural animation layering for synthesizing martial arts movements. ACM Transactions on Graphics (TOG) 40(4), 1-16 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 430, + 480, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 430, + 480, + 474 + ], + "spans": [ + { + "bbox": [ + 138, + 430, + 480, + 474 + ], + "type": "text", + "content": "89. Szot, A., Clegg, A., Undersander, E., Wijmans, E., Zhao, Y., Turner, J., Maestre, N., Mukadam, M., Chaplot, D.S., Maksymets, O., et al.: Habitat 2.0: Training home assistants to rearrange their habitat. Advances in Neural Information Processing Systems 34, 251-266 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 475, + 480, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 475, + 480, + 507 + ], + "spans": [ + { + "bbox": [ + 138, + 475, + 480, + 507 + ], + "type": "text", + "content": "90. Takano, W., Nakamura, Y.: Statistical mutual conversion between whole body motion primitives and linguistic sentences for human motions. The International Journal of Robotics Research 34(10), 1314-1328 (2015)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 508, + 480, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 508, + 480, + 552 + ], + "spans": [ + { + "bbox": [ + 138, + 508, + 480, + 552 + ], + "type": "text", + "content": "91. Tevet, G., Gordon, B., Hertz, A., Bermano, A.H., Cohen-Or, D.: Motionclip: Exposing human motion generation to clip space. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXII. pp. 358-374. Springer (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 553, + 480, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 553, + 480, + 575 + ], + "spans": [ + { + "bbox": [ + 138, + 553, + 480, + 575 + ], + "type": "text", + "content": "92. Tevet, G., Raab, S., Gordon, B., Shafir, Y., Bermano, A.H., Cohen-Or, D.: Human motion diffusion model. arXiv preprint arXiv:2209.14916 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 576, + 480, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 576, + 480, + 609 + ], + "spans": [ + { + "bbox": [ + 138, + 576, + 480, + 609 + ], + "type": "text", + "content": "93. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., et al.: Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 609, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 609, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 138, + 609, + 480, + 643 + ], + "type": "text", + "content": "94. Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi, A., Babaei, Y., Bashlykov, N., Batra, S., Bhargava, P., Bhosale, S., et al.: Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "text", + "content": "95. Van Den Oord, A., Vinyals, O., et al.: Neural discrete representation learning. Advances in neural information processing systems 30 (2017)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "20" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 278, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 278, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 278, + 101 + ], + "type": "text", + "content": "Biao Jiang, Xin Chen et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 19 + }, + { + "para_blocks": [ + { + "bbox": [ + 134, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 136, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 136, + 116, + 480, + 149 + ], + "type": "text", + "content": "96. Varol, G., Romero, J., Martin, X., Mahmood, N., Black, M., Laptev, I., Schmid, C.: Learning from synthetic humans. In: Computer Vision and Pattern Recognition (CVPR) (2017)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 150, + 480, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 150, + 480, + 183 + ], + "spans": [ + { + "bbox": [ + 138, + 150, + 480, + 183 + ], + "type": "text", + "content": "97. Vedantam, R., Lawrence Zitnick, C., Parikh, D.: Cider: Consensus-based image description evaluation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4566-4575 (2015)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 183, + 480, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 183, + 480, + 215 + ], + "spans": [ + { + "bbox": [ + 138, + 183, + 480, + 215 + ], + "type": "text", + "content": "98. Wang, C.: T2m-hifigpt: Generating high quality human motion from textual descriptions with residual discrete representations. arXiv preprint arXiv:2312.10628 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 216, + 480, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 216, + 480, + 248 + ], + "spans": [ + { + "bbox": [ + 138, + 216, + 480, + 248 + ], + "type": "text", + "content": "99. Wang, W., Zhe, X., Chen, H., Kang, D., Li, T., Chen, R., Bao, L.: Neural marionette: A transformer-based multi-action human motion synthesis system. arXiv preprint arXiv:2209.13204 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 134, + 249, + 480, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 249, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 134, + 249, + 480, + 281 + ], + "type": "text", + "content": "100. Wu, C., Yin, S., Qi, W., Wang, X., Tang, Z., Duan, N.: Visual chatgpt: Talking, drawing and editing with visual foundation models. arXiv preprint arXiv:2303.04671 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 134, + 281, + 480, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 281, + 480, + 326 + ], + "spans": [ + { + "bbox": [ + 134, + 281, + 480, + 326 + ], + "type": "text", + "content": "101. Wu, Y.C., Gebru, I.D., Markovic, D., Richard, A.: Audiodec: An open-source streaming high-fidelity neural audio codec. In: ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 1-5. IEEE (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 134, + 326, + 480, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 326, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 134, + 326, + 480, + 369 + ], + "type": "text", + "content": "102. Xin, C., Jiang, B., Liu, W., Huang, Z., Fu, B., Chen, T., Yu, J., Yu, G.: Executing your commands via motion diffusion in latent space. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 134, + 369, + 480, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 369, + 480, + 402 + ], + "spans": [ + { + "bbox": [ + 134, + 369, + 480, + 402 + ], + "type": "text", + "content": "103. Xiong, W., Liu, J., Molybog, I., Zhang, H., Bhargava, P., Hou, R., Martin, L., Rungta, R., Sankararaman, K.A., Oguz, B., et al.: Effective long-context scaling of foundation models. arXiv preprint arXiv:2309.16039 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 134, + 402, + 480, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 402, + 480, + 435 + ], + "spans": [ + { + "bbox": [ + 134, + 402, + 480, + 435 + ], + "type": "text", + "content": "104. Xu, H., Ghosh, G., Huang, P.Y., Okhonko, D., Aghajanyan, A., Metze, F., Zettlemoyer, L., Feichtenhofer, C.: Videoclip: Contrastive pre-training for zero-shot video-text understanding. arXiv preprint arXiv:2109.14084 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 134, + 435, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 435, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 134, + 435, + 480, + 468 + ], + "type": "text", + "content": "105. Yamada, T., Matsunaga, H., Ogata, T.: Paired recurrent autoencoders for bidirectional translation between robot actions and linguistic descriptions. IEEE Robotics and Automation Letters 3(4), 3441-3448 (2018)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 134, + 468, + 480, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 468, + 480, + 490 + ], + "spans": [ + { + "bbox": [ + 134, + 468, + 480, + 490 + ], + "type": "text", + "content": "106. Yao, H., Song, Z., Zhou, Y., Ao, T., Chen, B., Liu, L.: Moconvq: Unified physics-based motion control via scalable discrete representations (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 134, + 491, + 480, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 491, + 480, + 522 + ], + "spans": [ + { + "bbox": [ + 134, + 491, + 480, + 522 + ], + "type": "text", + "content": "107. Ye, Q., Xu, H., Xu, G., Ye, J., Yan, M., Zhou, Y., Wang, J., Hu, A., Shi, P., Shi, Y., et al.: mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 522, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 522, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 134, + 522, + 480, + 555 + ], + "type": "text", + "content": "108. Yin, F., Chen, X., Zhang, C., Jiang, B., Zhao, Z., Fan, J., Yu, G., Li, T., Chen, T.: Shapeqpt: 3d shape generation with a unified multi-modal language model. arXiv preprint arXiv:2311.17618 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 555, + 480, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 555, + 480, + 588 + ], + "spans": [ + { + "bbox": [ + 134, + 555, + 480, + 588 + ], + "type": "text", + "content": "109. Yin, F., Liu, W., Huang, Z., Cheng, P., Chen, T., Yu, G.: Coordinates are not lonely-codebook prior helps implicit neural 3d representations. Advances in Neural Information Processing Systems 35, 12705-12717 (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 589, + 480, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 589, + 480, + 621 + ], + "spans": [ + { + "bbox": [ + 134, + 589, + 480, + 621 + ], + "type": "text", + "content": "110. Yu, J., Xu, Y., Koh, J.Y., Luong, T., Baid, G., Wang, Z., Vasudevan, V., Ku, A., Yang, Y., Ayan, B.K., et al.: Scaling autoregressive models for content-rich text-to-image generation. arXiv preprint arXiv:2206.10789 (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 134, + 621, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 621, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 621, + 480, + 665 + ], + "type": "text", + "content": "111. Yuan, Y., Kitani, K.: Dlow: Diversifying latent flows for diverse human motion prediction. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part IX 16. pp. 346-364. Springer (2020)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 392, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 392, + 91, + 447, + 101 + ], + "type": "text", + "content": "MotionChain" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "21" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 20 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 590 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "text", + "content": "112. Yuan, Y., Song, J., Iqbal, U., Vahdat, A., Kautz, J.: PhysDiff: Physics-guided human motion diffusion model. In: IEEE International Conference on Computer Vision (ICCV) (October 2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 480, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 480, + 172 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 480, + 172 + ], + "type": "text", + "content": "113. Zhang, H., Li, X., Bing, L.: Video-llama: An instruction-tuned audio-visual language model for video understanding. arXiv preprint arXiv:2306.02858 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 172, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 172, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 133, + 172, + 480, + 205 + ], + "type": "text", + "content": "114. Zhang, H., Cao, J., Lu, G., Ouyang, W., Sun, Z.: Danet: Decompose-and-aggregate network for 3d human shape and pose estimation. In: Proceedings of the 27th ACM International Conference on Multimedia. pp. 935-944 (2019)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 205, + 480, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 205, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 134, + 205, + 480, + 237 + ], + "type": "text", + "content": "115. Zhang, H., Tian, Y., Zhang, Y., Li, M., An, L., Sun, Z., Liu, Y.: Pymaf-x: Towards well-aligned full-body model regression from monocular images. IEEE Transactions on Pattern Analysis and Machine Intelligence (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 134, + 237, + 480, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 237, + 480, + 282 + ], + "spans": [ + { + "bbox": [ + 134, + 237, + 480, + 282 + ], + "type": "text", + "content": "116. Zhang, J., Zhang, Y., Cun, X., Huang, S., Zhang, Y., Zhao, H., Lu, H., Shen, X.: T2m-gpt: Generating human motion from textual descriptions with discrete representations. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 134, + 282, + 480, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 282, + 480, + 315 + ], + "spans": [ + { + "bbox": [ + 134, + 282, + 480, + 315 + ], + "type": "text", + "content": "117. Zhang, M., Cai, Z., Pan, L., Hong, F., Guo, X., Yang, L., Liu, Z.: Motiondiffuse: Text-driven human motion generation with diffusion model. arXiv preprint arXiv:2208.15001 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 134, + 315, + 480, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 315, + 480, + 347 + ], + "spans": [ + { + "bbox": [ + 134, + 315, + 480, + 347 + ], + "type": "text", + "content": "118. Zhang, M., Guo, X., Pan, L., Cai, Z., Hong, F., Li, H., Yang, L., Liu, Z.: Remodiffuse: Retrieval-augmented motion diffusion model. arXiv preprint arXiv:2304.01116 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 134, + 347, + 480, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 347, + 480, + 392 + ], + "spans": [ + { + "bbox": [ + 134, + 347, + 480, + 392 + ], + "type": "text", + "content": "119. Zhang, Q., Song, J., Huang, X., Chen, Y., Liu, M.Y.: Diffcollage: Parallel generation of large content with diffusion models. In: 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 10188-10198. IEEE (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 134, + 391, + 480, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 391, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 134, + 391, + 480, + 413 + ], + "type": "text", + "content": "120. Zhang, T., Kishore, V., Wu, F., Weinberger, K.Q., Artzi, Y.: Bertscore: Evaluating text generation with bert. arXiv preprint arXiv:1904.09675 (2019)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 134, + 413, + 480, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 413, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 134, + 413, + 480, + 445 + ], + "type": "text", + "content": "121. Zhang, Y., Black, M.J., Tang, S.: We are more than our joints: Predicting how 3d bodies move. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3372-3382 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 134, + 446, + 480, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 446, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 134, + 446, + 480, + 479 + ], + "type": "text", + "content": "122. Zheng, L., Chiang, W.L., Sheng, Y., Zhuang, S., Wu, Z., Zhuang, Y., Lin, Z., Li, Z., Li, D., Xing, E.P., Zhang, H., Gonzalez, J.E., Stoica, I.: Judging lvm-as-a-judge with mt-bench and chatbot arena (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 134, + 479, + 480, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 479, + 480, + 512 + ], + "spans": [ + { + "bbox": [ + 134, + 479, + 480, + 512 + ], + "type": "text", + "content": "123. Zhou, C., Liu, P., Xu, P., Iyer, S., Sun, J., Mao, Y., Ma, X., Efrat, A., Yu, P., Yu, L., et al.: Lima: Less is more for alignment. Advances in Neural Information Processing Systems 36 (2024)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 512, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 512, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 134, + 512, + 480, + 544 + ], + "type": "text", + "content": "124. Zhu, D., Chen, J., Shen, X., Li, X., Elhoseiny, M.: Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 544, + 480, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 544, + 480, + 590 + ], + "spans": [ + { + "bbox": [ + 134, + 544, + 480, + 590 + ], + "type": "text", + "content": "125. Zou, X., Dou, Z.Y., Yang, J., Gan, Z., Li, L., Li, C., Dai, X., Behl, H., Wang, J., Yuan, L., et al.: Generalized decoding for pixel, image, and language. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15116-15127 (2023)" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "22" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 278, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 278, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 278, + 101 + ], + "type": "text", + "content": "Biao Jiang, Xin Chen et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 21 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/MotionDirector_ Motion Customization of Text-to-Video Diffusion Models/8325ffaa-f0c0-4f28-929f-63aa2f3d7c76_content_list.json b/2024/MotionDirector_ Motion Customization of Text-to-Video Diffusion Models/8325ffaa-f0c0-4f28-929f-63aa2f3d7c76_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..56f84b147f2e40111e2fe5e3961adc25c70940c4 --- /dev/null +++ b/2024/MotionDirector_ Motion Customization of Text-to-Video Diffusion Models/8325ffaa-f0c0-4f28-929f-63aa2f3d7c76_content_list.json @@ -0,0 +1,1636 @@ +[ + { + "type": "text", + "text": "MotionDirector: Motion Customization of Text-to-Video Diffusion Models", + "text_level": 1, + "bbox": [ + 261, + 140, + 741, + 183 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Rui Zhao $^{1}$ , Yuchao Gu $^{1}$ , Jay Zhangjie Wu $^{1}$ , David Junhao Zhang $^{1}$ , Jia-Wei Liu $^{1}$ , Weijia Wu $^{1}$ , Jussi Keppo $^{2}$ , and Mike Zheng Shou $^{1}$", + "bbox": [ + 228, + 210, + 772, + 243 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Show Lab, $^{2}$ National University of Singapore, Singapore", + "bbox": [ + 308, + 253, + 692, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Large-scale pre-trained diffusion models have exhibited remarkable capabilities in diverse video generations. Given a set of video clips of the same motion concept, the task of Motion Customization is to adapt existing text-to-video diffusion models to generate videos with this motion. Adaptation methods have been developed for customizing appearance like subject or style, yet under-explored for motion. It is straightforward to extend mainstream adaption methods for motion customization, including full model tuning and Low-Rank Adaptions (LoRAs). However, the motion concept learned by these methods is often coupled with the limited appearances in the training videos, making it difficult to generalize the customized motion to other appearances. To overcome this challenge, we propose MotionDirector, with a dual-path LoRAs architecture to decouple the learning of appearance and motion. Further, we design a novel appearance-debiased temporal loss to mitigate the influence of appearance on the temporal training objective. Experimental results show the proposed method can generate videos of diverse appearances for the customized motions. Our method also supports various downstream applications, such as the mixing of different videos with their appearance and motion respectively, and animating a single image with customized motions. The project website is at: MotionDirector.", + "bbox": [ + 261, + 305, + 738, + 580 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: motion customization $\\cdot$ text-to-video generation $\\cdot$ diffusion models", + "bbox": [ + 261, + 595, + 740, + 622 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 217, + 648, + 375, + 666 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Text-to-video diffusion models [17,21,50] are approaching generating high-quality diverse videos given text instructions. The open-sourcing of foundational text-to-video models [55,60] pre-trained on large-scale data has sparked enthusiasm for video generation in both the community and academia. Users can create videos that are either realistic or imaginative simply by providing text prompts. While foundation models generate diverse videos from the same text, adapting them to generate more specific content can better accommodate the preferences of users. Similar to the customization of text-to-image foundation models [45], tuning video foundation models to generate videos of a certain concept of appearance,", + "bbox": [ + 212, + 680, + 787, + 816 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding Author.", + "bbox": [ + 217, + 824, + 388, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3582cef6c28440356b1b66594b0965f76d24ef0669ceb60bef4b3bee14c7130f.jpg", + "image_caption": [ + "Fig. 1: Motion customization of the text-to-video diffusion." + ], + "image_footnote": [], + "bbox": [ + 220, + 145, + 782, + 340 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "like subject or style, has also been explored [17]. Compared with images, videos consist of not just appearances but also motion dynamics. Users may desire to create videos with specific motions, such as a car moving forward and then turning left under a predefined camera perspective, as illustrated on the right side in Fig. 1. However, customizing the motions in the text-to-video generation is still under-explored.", + "bbox": [ + 212, + 398, + 784, + 489 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The task of Motion Customization is formulated as follows: given a set of reference videos representing a specific motion concept, the objective is to customize pre-trained foundational models to generate videos that accurately capture and replicate the depicted motion. The customized motion concept can be challenging motions that are hard for the pre-trained model to generate, novel motions not present in the pre-training dataset, and specific motions that are difficult to express in text. Examples of these are respectively showcased in the first, second, and third columns of Fig 1. In contrast, previous works on appearance customization adapt the foundation models to generate samples with desired appearance, like subject or style, given reference videos or images representing such appearance [17,45]. It is straightforward to use previous adaption methods for motion customization. For example, on the given reference videos, fine-tuning the weights of foundation models [45], parameter-efficient tuning additional layers [70], or training Low-Rank Adaptions (LoRAs) [26] injected in the layers of foundation models. However, customizing diffusion models to generate desired motions without harming their appearance diversity is challenging because the motion and appearance are coupled with each other at the step-by-step denoising stage. Directly deploying previous adaption methods to learn motions makes the models fit the limited appearances seen in the reference videos, posing challenges in generalizing the learned motions to various appearances. Recent works on controllable text-to-video generations [11, 17, 63] generate videos controlled by signals representing pre-defined motions. However, the control signals, such as depth maps or edges, impose constraints on the shapes of subjects and back", + "bbox": [ + 212, + 493, + 787, + 843 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Zhao et al.", + "bbox": [ + 271, + 114, + 346, + 126 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/0b4a830b7c8818495a38fd991e4e79d94464b1abd952a55177059708a12cb370.jpg", + "image_caption": [ + "Fig. 2: (Row 1) Take two videos to train the proposed MotionDirector, respectively. (Row 2) MotionDirector can generalize the learned motions to diverse appearances. (Row 3) MotionDirector can mix the learned motion and appearance from different videos to generate new videos. (Row 4) MotionDirector can animate a single image with learned motions." + ], + "image_footnote": [], + "bbox": [ + 218, + 143, + 781, + 345 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "grounds, thus influencing the appearance of generated videos in a coupled way. Besides, these methods accept only one sequence of control signals to generate one video, which may not be suitable for users seeking certain motion types without strict spatial constraints, such as the example of lifting weights in Fig. 1.", + "bbox": [ + 212, + 460, + 784, + 521 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To achieve motion customization of text-to-video diffusion models while preserving appearance diversity, we propose the MotionDirector, which tunes the foundation models to learn the appearance and motions in the given single or multiple reference videos in a decoupled way. MotionDirector tunes the models with low-rank adaptions (LoRAs) while keeping their pre-trained parameters fixed to retain the learned generation knowledge. Specifically, the MotionDirector employs a dual-path architecture, as shown in Fig. 3. For each video, a spatial path consists of a foundation model with trainable spatial LoRAs injected into its spatial transformer layers. These spatial LoRAs are trained on a single frame randomly sampled per training step to capture the appearance characteristics of the input videos. The temporal path, on the other hand, is a replica of the foundation model that shares the spatial LoRAs with the spatial path to fit the appearance of the corresponding input video. Additionally, the temporal transformers in this path are equipped with temporal LoRAs, which are trained on multiple frames of input videos to capture the underlying motion patterns. To further enhance the learning of motions, we propose an appearance-debiased temporal loss to mitigate the influence of appearance on the temporal training objective.", + "bbox": [ + 212, + 522, + 787, + 792 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Only deploying the trained temporal LoRAs enables the foundation model to generate videos of the learned motions with diverse appearances, as shown in the second row of Fig 2. The decoupled paradigm further makes an interesting", + "bbox": [ + 212, + 794, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "MotionDirector", + "bbox": [ + 625, + 114, + 730, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "kind of video generation feasible, which is the mix of the appearance from one video with the motion from another video, called the mix of videos, as shown in the third row of Fig 2. The key to this success lies in that MotionDirector can decouple the appearance and motion of videos and then combine them from various source videos. It is achieved by injecting spatial LoRAs trained on one video and temporal LoRAs trained on another video into the foundation model. Besides, the learned motions can be deployed to animate images, as images can be treated as appearance providers, as shown in the last row of Fig 2.", + "bbox": [ + 212, + 146, + 787, + 267 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We conducted experiments on two benchmarks with 86 different motions and over 600 text prompts to test proposed methods, baselines, and comparison methods. The results show our method can be applied to different diffusion-based foundation models and achieve motion customization of various motion concepts. Compared with other methods, our method avoids fitting the limited appearance of reference videos, and can generalize the learned motions to diverse appearances.", + "bbox": [ + 212, + 268, + 787, + 372 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our contributions are summarized as follows:", + "bbox": [ + 238, + 375, + 568, + 388 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce and define the task of Motion Customization. The challenges lie in generalizing the customized motions to various appearances.", + "- We propose the MotionDirector with a dual-path architecture and a novel appearance-debiased temporal training objective, to decouple the learning of appearance and motion.", + "- Experiments on two benchmarks demonstrate that MotionDirector can customize various base models to generate diverse videos with desired motion concepts, and outperforms controllable generation methods and tuning-based methods." + ], + "bbox": [ + 225, + 401, + 782, + 534 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 565, + 388, + 580 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.1 Text-to-Video Generation.", + "text_level": 1, + "bbox": [ + 215, + 599, + 478, + 614 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To achieve high-quality video generation, various methods have been developed, such as Generative Adversarial Networks (GANs) [2, 47, 49, 56, 57, 59], autoregressive models [12, 25, 33, 54, 75] and implicit neural representations [51, 80]. Diffusion-based models [3, 10, 13, 18, 22-24, 27, 34, 35, 40, 41, 50, 58, 60, 62, 65, 71, 78, 79, 81, 84, 86] are also approaching high-quality generation by training conditional 3D U-Nets [44] to denoise from randomly sampled sequences of Gaussian noises [74]. Recent foundation models [4, 17, 21, 36, 50, 64, 82] are pre-trained on large-scale image and video datasets [1, 8, 48], to learn powerful generation ability. Some works turn text-to-image foundation models to text-to-video generation by manipulation on cross-frame attention or training additional temporal layers, like Tune-A-Video [71], Text2Video-Zero [30], and AnimiteDiff [15]. The recently open-sourced foundation models [55, 60] have ignited enthusiasm among users to generate realistic or imaginative videos, and further make it possible for users to customize and build their own private models.", + "bbox": [ + 212, + 628, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Zhao et al.", + "bbox": [ + 271, + 114, + 344, + 126 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Generation Model Customization.", + "text_level": 1, + "bbox": [ + 215, + 146, + 542, + 161 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Customizing the pre-trained large foundation models can fit the preferences of users better while maintaining powerful generation knowledge without training from scratch. Previous customization methods for text-to-image diffusion models [7,14,32,45,52,68] aim to generate certain subjects or styles, given a set of example images. Dreambooth [45] or LoRA [26] can be simply applied to customizing video foundation models to generate videos with certain subjects or styles, given a set of reference video clips or images. The recently proposed VideoCrafter [19] has explored this, which we categorize as appearance customization. In addition to appearances, videos are also characterized by the motion dynamics of subjects and camera movements across frames. After we proposed the task of motion customization, some concurrent works [28, 39, 43, 66, 67, 72, 76] emerged in this field. DreamVideo [67] trains several adapters in U-Net to customize both the subject and motion in video generation. Lamp [72] learns a motion pattern from a set of videos to utilize image diffusion models to generate videos.", + "bbox": [ + 212, + 175, + 787, + 388 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "2.3 Controllable Video Generation.", + "text_level": 1, + "bbox": [ + 215, + 412, + 519, + 426 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Controllable generation aims to ensure the generation results align with the given explicit control signals, such as depth maps, human pose, optical flows, etc. [5, 16, 29, 37, 38, 42, 61, 73, 77, 83, 85]. For the controllable text-to-video generation methods, i.e. the VideoCrafter [17], VideoComposer [63], Control-A-Video [6], they train additional branches that take condition signals to align the generated videos with them. Unlike the human poses for specifically controlling the generation of human bodies, the general control singles, such as depth maps, are typically extracted from reference videos and are coupled with both appearance and motion. This results in the generation results being influenced by both the appearance and motion in reference videos. Applying these methods directly in motion customization is challenging when it comes to generalizing the desired motions to diverse appearances.", + "bbox": [ + 212, + 441, + 787, + 625 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Methodology", + "text_level": 1, + "bbox": [ + 215, + 651, + 380, + 669 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Preliminaries", + "text_level": 1, + "bbox": [ + 215, + 686, + 372, + 700 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Video Diffusion Model. Video diffusion models train a 3D U-Net to denoise from a randomly sampled sequence of Gaussian noises to generate videos, guided by text prompts. The 3D U-net basically consists of down-sample, middle, and up-sample blocks. Each block has several convolution layers, spatial transformers, and temporal transformers as shown in Fig 3. The 3D U-Net $\\epsilon_{\\theta}$ and a text encoder $\\tau_{\\theta}$ are jointly optimized by the noise-prediction loss, as detailed in [9]:", + "bbox": [ + 212, + 715, + 785, + 808 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathbb {E} _ {z _ {0}, y, \\epsilon \\sim \\mathcal {N} (0, I), t \\sim \\mathcal {U} (0, T)} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} \\left(z _ {t}, t, \\tau_ {\\theta} (y)\\right) \\| _ {2} ^ {2} \\right], \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 321, + 824, + 785, + 842 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "MotionDirector", + "bbox": [ + 625, + 114, + 730, + 127 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/73f3d477644768419b0a8c81f48f3ad3fb9d92540b2eeecc7f57ad3c9ce486fe.jpg", + "image_caption": [ + "Fig. 3: The dual-path architecture of the proposed method. All pre-trained weights of the base diffusion model remain fixed. In the spatial path, the spatial transformers are injected with trainable spatial LoRAs as shown on the right side. In the temporal path, the spatial transformers are injected with spatial LoRAs sharing weights with those ones in the spatial path, and the temporal transformers are injected with trainable temporal LoRAs." + ], + "image_footnote": [], + "bbox": [ + 218, + 145, + 784, + 306 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $z_0$ is the latent code of the training videos, $y$ is the text prompt, $\\epsilon$ is the Gaussian noise added to the latent code, and $t$ is the time step. As discussed in [9], the noised latent code $z_t$ is determined as:", + "bbox": [ + 212, + 431, + 787, + 479 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nz _ {t} = \\sqrt {\\bar {\\alpha} _ {t}} z _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon , \\bar {\\alpha} _ {t} = \\prod_ {i = 1} ^ {t} \\alpha_ {t}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 367, + 489, + 785, + 532 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\alpha_{t}$ is a hyper-parameter controlling the noise strength.", + "bbox": [ + 212, + 541, + 655, + 556 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Low-Rank Adaption. Low-rank adaption (LoRA) [26] was proposed to adapt the pre-trained large language models to downstream tasks. Recently it has been applied in text-to-image generation and text-to-video generation tasks to achieve appearance customization [19, 46]. LoRA employs a low-rank factorization technique to update the weight matrix $W$ as", + "bbox": [ + 212, + 556, + 787, + 632 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nW = W _ {0} + \\Delta W = W _ {0} + B A, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 393, + 645, + 785, + 662 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $W_0 \\in \\mathbb{R}^{d \\times k}$ represents the original weights of the pre-trained model, $B \\in \\mathbb{R}^{d \\times r}$ and $A \\in \\mathbb{R}^{r \\times k}$ represent the low-rank factors, where $r$ is much smaller than original dimensions $d$ and $k$ . LoRA requires smaller computing sources than fine-tuning the weights of the entire network like DreamBooth [45], and it is convenient to spread and deploy as a plug-and-play plugin for pre-trained models.", + "bbox": [ + 212, + 672, + 787, + 762 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In the following sections, we introduce the proposed MotionDirector, containing the dual-path Low-Rank Adaptions (in Sec. 3.2) for learning the appearance and motion in reference videos in a decoupled manner, and appearance-debiased temporal loss (in Sec. 3.3)) to further mitigate the influence of appearance on the temporal training objective.", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Zhao et al.", + "bbox": [ + 271, + 114, + 346, + 127 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/bf379e347ad1d2416a527089af9d55c14e67a0e0b6f113aaef00e343a5ff23d3.jpg", + "image_caption": [ + "Fig. 4: Comparison of videos generated by the base model, the model fine-tuned in the coupled manner, the model fine-tuned without the proposed Appearance-Debiased Temporal Loss (AD-Loss), and our full-version MotionDirector method." + ], + "image_footnote": [], + "bbox": [ + 218, + 145, + 784, + 318 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.2 Dual-Path Low-Rank Adaptions", + "text_level": 1, + "bbox": [ + 215, + 402, + 526, + 419 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "At each time-step $t$ , the 3D U-Net takes in the latent code $z_{t} \\in \\mathbb{R}^{b \\times f \\times w \\times h \\times c}$ and the conditional input $y$ (e.g., text), where $b, f, w, h, c$ represents the size of the batch, frame, width, height, and channel dimensions, respectively. The spatial transformers apply spatial self-attention along the spatial dimensions $w, h$ to improve the correlation between pixels, and then leverage the cross-attention between the latent code and the conditional input $y$ to improve textual alignment. The temporal transformers apply temporal self-attention along the frame dimension $f$ to improve the temporal consistency between frames. However, spatial and temporal information in the latent code gradually become coupled with each other during the step-by-step denoising stage. Attempting to directly learn and fit the motions in reference videos will inevitably lead to fitting their limited appearances. As shown in the third row \"Coupled Tuned\" of Fig. 4. From the results on the left side, we can observe that the generated \"panda\" is coupled with the appearance of \"human\" in the reference videos and the generated background is also coupled with that of reference videos. Besides, the results on the right side show that the coupled tuned model can not generate the \"Eiffel Tower\" because it is overfitted on the limited backgrounds in the reference videos.", + "bbox": [ + 212, + 430, + 787, + 686 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To address this problem, we propose to tune the spatial and temporal transformers in a dual-path way to learn the earn the appearance and motion in reference videos, respectively, as shown in Fig. 3. Specifically, for the spatial path, we inject LoRAs into spatial transformers to learn the appearance of training data, and for the temporal path, we inject LoRAs into temporal transformers to learn the motion in videos. The training of spatial and temporal LoRAs are as follows.", + "bbox": [ + 212, + 688, + 787, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Spatial Training. For the spatial path, we inject unique spatial LoRAs into the spatial transformers for each training video while keeping the weights of pre-trained 3D U-Net fixed. To maintain the learned strong and diverse textual", + "bbox": [ + 212, + 794, + 787, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "MotionDirector", + "bbox": [ + 625, + 114, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "alignment ability, we do not inject LoRAs into cross-attention layers of spatial transformers, since their weights influence the correlations between the pixels and text prompts. On the other hand, we inject LoRAs into spatial self-attention layers and feed-forward layers to update the correlations in spatial dimensions to enable the model to reconstruct the appearance of training data. For each training step, the spatial LoRAs are trained on a single frame randomly sampled from the training video to fit its appearance while ignoring its motion, based on spatial loss, which is reformulated as", + "bbox": [ + 212, + 146, + 787, + 268 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {s p a t i a l}} = \\mathbb {E} _ {z _ {0}, y, \\epsilon , t, i \\sim \\mathcal {U} (0, F)} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} \\left(z _ {t, i}, t, \\tau_ {\\theta} (y)\\right) \\| _ {2} ^ {2} \\right], \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 318, + 282, + 785, + 301 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\mathbf{F}$ is the number of frames of the training data and the $z_{t,i}$ is the sampled frame from the latent code $z_{t}$ .", + "bbox": [ + 212, + 306, + 785, + 335 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Temporal Training. For the temporal path, we inject the temporal LoRAs into self-attention and feed-forward layers of temporal transformers to update the correlations along the frame dimension. Besides, the spatial transformers are injected with LoRAs sharing the same weights learned from the spatial path, to force the trainable temporal LoRAs to ignore the appearance of the training data. The temporal LoRAs could be simply trained on all frames of training data based on the temporal loss $\\mathcal{L}_{org-temp}$ , formulated in the same way as equation (1).", + "bbox": [ + 212, + 338, + 787, + 459 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.3 Appearance-Debiased Temporal Loss", + "text_level": 1, + "bbox": [ + 214, + 482, + 566, + 498 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The videos generated by models trained with original temporal loss are shown in the fourth row of Fig. 4. We can observe that the motions are not learned well. To explore the underlying reasons, we visualized the denoising process of several videos.", + "bbox": [ + 212, + 508, + 785, + 568 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As illustrated in Fig. 5, when considering the latent codes of each frame $z_{t,i_{i = 1}^{F}}$ as a set of points in the latent space, we found that motion primarily impacts the underlying dependencies between these point sets, whereas the distances between different sets of points are more influenced by appearance. The original temporal training objective does not take this into account. The learning of motion dynamics, embedded within the internal connectivity structure of latent codes, remains biased by positional differences of latent codes. These biases are primarily influenced by the appearance of training samples and pose a challenge to learning motions efficiently.", + "bbox": [ + 212, + 569, + 787, + 705 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To further decouple the motion from appearance, we proposed to eliminate the appearance bias among the noises and predicted noises, and calculate the appearance-debiased temporal loss on them. The debiasing of each noise $\\epsilon_{i} \\in \\{\\epsilon_{i}\\}_{i=1}^{F}$ is as follows,", + "bbox": [ + 214, + 705, + 787, + 767 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\phi (\\epsilon_ {i}) = \\sqrt {\\beta^ {2} + 1} \\epsilon_ {i} - \\beta \\epsilon_ {a n c h o r}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 393, + 779, + 785, + 797 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\beta$ is the strength factor controlling the debiasing strength and a higher value leads to stronger debiasing effects. The $\\epsilon_{anchor}$ is the anchor among the", + "bbox": [ + 214, + 809, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Zhao et al.", + "bbox": [ + 271, + 114, + 344, + 127 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6dae295d058d052c7b45fb02d1b226b792d3ff94cbf9b58a37a7ac22d83c6106.jpg", + "image_caption": [ + "Fig. 5: (a) Four example videos (the same as the videos in the first and fourth rows of Fig. 2) and their relationships in terms of motion and appearance. (b) We inverse the four videos based on the video diffusion model and visualize the denoising process. Each point corresponds to a latent code $z_{t,i,j}$ at time step $t$ of $i$ -th frame of $j$ -th video, reduced to 2 dimensions through PCA. (c) Take latent codes at time step 0 for example, the ones of the same video are connected in order of frames. We find that the internal connectivity structure between latent codes is more influenced by motion, while the distance between sets of latent codes is primarily affected by the difference in appearance. (d) Through the proposed appearance-debiased operation, as formulated in Equation 5, the latent codes eliminate distance bias introduced by appearance, while preserving the internal connectivity structure that represents motion dynamics." + ], + "image_footnote": [], + "bbox": [ + 222, + 143, + 782, + 311 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "frames from the same training data. If $\\beta$ equals 1, it indicates that we attribute equal importance to both appearance and motion in the video. The detailed mathematical derivation of this equation is provided in the supplementary materials. In practice, we find that simply setting $\\beta = 1$ and randomly sampling $\\epsilon_{i} \\in \\{\\epsilon_{i}\\}_{i=1}^{F}$ as the anchor can achieve good performance. The appearance-debiased temporal loss (AD-Loss) is reformulated as", + "bbox": [ + 212, + 508, + 787, + 599 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {a d - t e m p} = \\mathbb {E} _ {z _ {0}, y, \\epsilon , t} \\left[ \\| \\phi (\\epsilon) - \\phi \\left(\\epsilon_ {\\theta} \\left(z _ {t}, t, \\tau_ {\\theta} (y)\\right)\\right) \\| _ {2} ^ {2} \\right]. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 321, + 609, + 785, + 630 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "For temporal LoRAs, the loss function is the combination of temporal loss and appearance-debiased temporal loss as follows,", + "bbox": [ + 212, + 640, + 787, + 671 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t e m p o r a l}} = \\mathcal {L} _ {\\text {o r g - t e m p}} + \\mathcal {L} _ {\\text {a d - t e m p}}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 379, + 684, + 785, + 702 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "3.4 Inference of MotionDirector", + "text_level": 1, + "bbox": [ + 214, + 722, + 495, + 736 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In the inference stage, we inject the trained temporal LoRAs into the pre-trained video diffusion model to enable it to generate diverse videos with the learned motion from the training data. If the training data is a single video, the learned motion will be a specific motion, such as an object first moving forward and then turning to the left. If the training data is a set of videos, the learned motion will be the motion concept provided by them, like lifting weights or playing golf.", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "MotionDirector", + "bbox": [ + 625, + 114, + 730, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/4fc96f5fcd42560b73670b90cbcd8cab094aacc4b6cdd5589afa8609e2d9f8de.jpg", + "image_caption": [ + "Fig. 6: Qualitative comparison results of motion customization on multiple videos." + ], + "image_footnote": [], + "bbox": [ + 222, + 143, + 785, + 351 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/42bc7282cc3173da84a5af5c6410bcfa8c727724e993d605b037fdee4b5d6822.jpg", + "table_caption": [ + "Tab. 1: Automatic and human evaluations results of motion customization on multiple videos." + ], + "table_footnote": [], + "table_body": "
Automatic EvaluationsHuman Evaluations
Appearance Diversity (↑)Temporal Consistency (↑)Pick Score (↑)Appearance DiversityTemporal ConsistencyMotion Fidelity
Tune-A-Video28.2292.4520.20v.s. Base Model (ModelScope)25.00 v.s. 75.0025.00 v.s. 75.0040.00 v.s. 60.00
v.s. Base Model (ZeroScope)44.00 v.s. 56.0016.67 v.s. 83.3353.33 v.s. 46.67
ModelScopeBase Model28.5592.5420.33v.s. Base Model (ModelScope)23.08 v.s. 76.9240.00 v.s. 60.0052.00 v.s. 48.00
Coupled Tuned25.66 (-2.89)90.6619.85
w/o AD-Loss28.32 (-0.23)91.1720.34v.s. Base Model (ModelScope)53.12 v.s. 46.8849.84 v.s. 50.1662.45 v.s. 37.55
ours28.66 (+0.11)92.3620.59v.s. Base Model (ModelScope)54.84 v.s. 45.1656.00 v.s. 44.0075.00 v.s. 25.00
ZeroScopeBase Model28.4092.9420.76v.s. Base Model (ZeroScope)37.81 v.s. 62.1941.67 v.s. 58.3354.55 v.s. 45.45
Coupled Tuned25.52 (-2.88)90.6719.99
w/o AD-Loss28.61 (+0.21)91.3720.56v.s. Base Model (ZeroScope)50.10 v.s. 49.9048.00 v.s. 52.0058.33 v.s. 41.67
ours28.94 (+0.54)92.6720.80v.s. Base Model (ZeroScope)52.94 v.s. 47.0655.00 v.s. 45.0076.47 v.s. 23.53
", + "bbox": [ + 218, + 470, + 784, + 575 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "The motion concepts can be ones preferred by users or ones that lie in the long-tailed distribution that can not be synthesized well by pre-trained models. Since appearance and motion are decoupled by our method, the spatial LoRAs can also be used to influence the appearance of generated videos, as shown in Fig. 2. Users can flexibly adjust the influence strength of learned appearance and motion on the generation according to their preferences by simply setting the strength of LoRAs as $W = W_{0} + \\gamma \\Delta W$ , where $\\gamma$ is called the LoRA scale, and $\\Delta W$ is the learned weights. For instance, a small scale of temporal LoRAs leads to the generated results weakly expressing the learned motion concepts, while a significantly large scale leads to discordant generation. More experimental results showing the effect of LoRA scales are provided in the supplementary materials. To ensure comparative fairness, we set the scale of temporal LoRAs to 1 in all experiments in Sec. 4.", + "bbox": [ + 212, + 643, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Zhao et al.", + "bbox": [ + 271, + 114, + 346, + 126 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 143, + 375, + 162 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.1 Motion customization on multiple videos", + "text_level": 1, + "bbox": [ + 215, + 183, + 599, + 198 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Dataset. We conduct experiments on the adapted UCF Sports Action data set [53], which includes 95 videos of 12 different human motions, like playing golf, lifting weights, etc. For each type of motion, we label one original text prompt describing the motion, such as \"a person is playing golf, side view\". For these motions, we set 72 different text prompts in total as input to generate videos using comparison methods, such as \"a monkey is playing golf, side view\".", + "bbox": [ + 212, + 214, + 787, + 305 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Comparison Methods. We compare the proposed method with three baselines and the video generation method Tune-A-Video [70] that can be adapted to this task. Tune-A-Video was initially proposed for training temporal modules on a single video to learn its motion information, while here we adapt it to train on multiple videos. The baseline methods are compared with the proposed method on two different foundational text-to-video diffusion models, i.e. the ModelScope [60] and the ZeroScope [55]. We employ three baseline methods: the first is directly applying the vanilla foundation models, the second is tuning the foundation models with LoRAs in a coupled manner, and the third is the proposed dual-path method excluding the appearance-debiased temporal loss.", + "bbox": [ + 212, + 306, + 787, + 457 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Qualitative Results As shown in Fig. 6, taking a set of videos with motions of playing golf as training data, the Tune-A-Video fails to generate diverse appearances with the learned motions, like a monkey playing golf. To compare the baseline methods and proposed method fairly, we feed the same initial Gaussian noise to these methods to generate videos. The pre-trained foundation model, ZeroScope, correctly generates the appearance but lacks the realistic motion that swings a golf club, as those desired motions in the reference videos. The coupled tuned model could generate the desired motion but the learned motion is coupled with too much appearance information causing the generated subject in the video to be more like a human rather than a monkey. The last two rows show that the proposed dual-path LoRAs can avoid hurting the appearance generation and the proposed appearance-debiased temporal loss enhances the learning of desired motion better. We could draw a similar conclusion from the second example showing the motion of riding a panda.", + "bbox": [ + 212, + 458, + 787, + 671 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Quantitative Results. We evaluate the methods with automatic evaluations and human evaluations, and the results are shown in Table. 1.", + "bbox": [ + 214, + 671, + 785, + 700 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Automatic Metrics. Following the LOVEU-TGVE competition [69], the appearance diversity is computing the average CLIP score [20] between the diverse text prompts and all frames of the generated videos, the temporal consistency is the average CLIP score between frames, and the Pick Score is the average PickScore [31] between all frames of output videos.", + "bbox": [ + 214, + 702, + 785, + 777 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "In Table 1, the automatic evaluation of temporal consistency is calculated as the CLIP score between frames. This may result in videos that are nearly static and without desired motion receiving high scores. Thus, we further apply human evaluation to evaluate generated videos more comprehensively and accurately.", + "bbox": [ + 214, + 779, + 785, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "MotionDirector", + "bbox": [ + 625, + 114, + 730, + 127 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Human Preference. On the Amazon MTurk $^{1}$ , each generated video is evaluated by 5 human raters in terms of appearance diversity, temporal consistency, and motion fidelity, which evaluate whether the generated motion is similar to the references. To simplify the comparison for raters, they are asked to compare the results pairwise and select their preferred one, where the videos are shuffled and their source methods are anonymous. In Table. 1, the pairwise numbers \"p1 v.s. p2\" means $p_1\\%$ results of the first method are preferred while $p_2\\%$ results of the second method are preferred.", + "bbox": [ + 212, + 146, + 787, + 267 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "The evaluation results show that coupled tuning will destroy the appearance diversity of pre-trained models, while our method will preserve it and achieve the highest motion fidelity.", + "bbox": [ + 212, + 268, + 787, + 313 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.2 Motion customization on a single video", + "text_level": 1, + "bbox": [ + 214, + 337, + 584, + 353 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Dataset. We conduct the comparison experiments on the open-sourced benchmark released by the LOVEU-TGVE competition at CVPR 2023 [69]. The dataset comprises 76 videos, each originally associated with 4 editing text prompts. Additionally, we introduced 3 more prompts with significant changes.", + "bbox": [ + 212, + 364, + 795, + 425 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Comparison Methods. We compare the proposed method with SOTA controllable generation methods, the VideoCrafter [17], VideoComposer [63], and Control-A-Video [6], and the tuning-based method Tune-A-Video [70]. To ensure a fair comparison, we use the depth control mode of controllable generation methods, which is available in all of them.", + "bbox": [ + 212, + 426, + 787, + 501 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/58d5620fac8497f224f5113feb5b0cf49363aa06b9f5242c48e71b4084ce4a60.jpg", + "image_caption": [ + "Fig. 7: Qualitative comparison results of motion customization on a single video." + ], + "image_footnote": [], + "bbox": [ + 220, + 527, + 784, + 767 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Zhao et al.", + "bbox": [ + 271, + 114, + 346, + 127 + ], + "page_idx": 11 + }, + { + "type": "page_footnote", + "text": "1 https://requester.mturk.com/", + "bbox": [ + 218, + 823, + 433, + 840 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/ddcc2d2cef987346087f3c37c5b04bff2b611475f6ab8253c1eb89be73aa5c00.jpg", + "table_caption": [ + "Tab. 2: Automatic and human evaluations results of motion customization on a single video." + ], + "table_footnote": [], + "table_body": "
Automatic EvaluationsHuman Evaluations
Text Alignment (↑)Appearance Diversity (↑)Temporal Consistency (↑)Pick Score (↑)Text AlignmentAppearance DiversityTemporal ConsistencyMotion Fidelity
VideoComposer27.6627.0392.2220.26ours v.s. VideoComposer54.55 v.s. 45.4572.83 v.s. 27.1761.57 v.s. 38.4361.24 v.s. 38.76
Control-a-Video26.5425.3592.6319.75ours v.s. Control-A-Video68.00 v.s. 32.0078.43 v.s. 21.5771.28 v.s. 29.7256.47 v.s. 43.53
VideoCrafter28.0327.6992.2620.12ours v.s. VideoCrafter52.72 v.s. 47.2871.11 v.s. 28.8960.22 v.s. 39.7860.00 v.s. 40.00
Tune-a-Video25.6425.9592.4220.09ours v.s. Tune-A-Video67.86 v.s. 32.1469.14 v.s. 30.8671.67 v.s. 28.3356.52 v.s. 43.48
ours27.8228.4893.0020.74
", + "bbox": [ + 218, + 176, + 785, + 241 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/f07b22b517f3924135187f8be2ea53bada783e21e4b60984dad8df35282af69c.jpg", + "image_caption": [ + "Fig. 8: Videos generated by MotionDirector. Their appearances are customized according to the reference images shown on the top-left side. Their motions are customized based on the reference videos shown on the top-right side." + ], + "image_footnote": [], + "bbox": [ + 217, + 255, + 784, + 410 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Qualitative and Quantitative Results. As shown in Fig. 7, comparison methods fail to generalize the desired motions to diverse appearances, like the ears of bears and the Arc de Triomphe. Similar to the depth, using the edge and optical flow as control signals also introduces interference from the appearance into the generated videos. In Table. 2, we refer to the alignment between the generated videos and the original 4 editing text prompts as text alignment, and the alignment with the 3 new text prompts with significant changes as appearance diversity. The results show that our method outperforms other methods by a large margin when generalizing the motions to diverse appearances, and achieves competitive motion fidelity.", + "bbox": [ + 212, + 494, + 787, + 647 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.3 Customizing Both Appearance and Motion", + "text_level": 1, + "bbox": [ + 215, + 672, + 617, + 690 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Since MotionDirector learns the appearance and motion in a decoupled manner, it can also reorganize and combine the appearances and motions from different sources. For example, as shown in Fig. 2, MotionDirector can combine the appearance from one video with the motion from another video to generate a new video that exhibits similar appearance and motion characteristics to each of the input videos, respectively. Here we provide more results in Fig. 8. The base model is injected with the spatial LoRAs trained on the given images of \"Terracotta Warrior\" and injected with different temporal LoRAs trained on different sets of videos of different motion concepts, such as \"riding horse\".", + "bbox": [ + 212, + 704, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "MotionDirector", + "bbox": [ + 625, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.4 Efficiency Performance", + "text_level": 1, + "bbox": [ + 217, + 146, + 450, + 161 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "The lightweight LoRAs enable our method to tune the foundation models efficiently. Taking the foundation model ZeroScope for example, it has over 1.8 billion pre-trained parameters. Each set of trainable spatial and temporal LoRAs only adds 9 million and 12 million parameters, respectively. Requiring 14 GB VRAM, MotionDirector takes 20 minutes to converge on multiple reference videos, and 8 minutes for a single reference video, competitive to the 10 minutes required by Tune-A-Video [70]. After training, MotionDirector can generate diverse videos with learned motions. In the inference stage, MotionDirector does not introduce significant increases in computational costs or inference time.", + "bbox": [ + 218, + 171, + 785, + 306 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Limitations and Future Works", + "text_level": 1, + "bbox": [ + 218, + 330, + 544, + 345 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Despite the MotionDirector can learn the motions of one or two subjects in the reference videos, it is still hard to learn complex motions of multiple subjects, such as a group of boys playing soccer. Previous appearance customization methods suffer similar problems when generating multiple customized subjects [14]. A possible solution is to further decouple the motions of different subjects in the latent space and learn them separately.", + "bbox": [ + 218, + 362, + 785, + 453 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 218, + 476, + 356, + 492 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We introduce and formulate the task of Motion Customization, which is adapting the pre-trained foundation text-to-video diffusion models to generate videos with desired motions. The challenge of this task is generalizing the customized motions to various appearances. To overcome this challenge, we propose the MotionDirector with a dual-path architecture and a novel appearance-debiased temporal training objective to decouple the learning of appearance and motion. Experimental results show that MotionDirector can learn motion concepts from a limited number of reference videos, and generalize them to diverse appearances. The automatic and human evaluations on two benchmarks demonstrate the MontionDirector outperforms other methods in terms of appearance diversity and motion fidelity.", + "bbox": [ + 218, + 507, + 785, + 674 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 218, + 696, + 401, + 715 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "This research is supported by National Research Foundation, Singapore and A\\*STAR, under its RIE2020 Industry Alignment Fund - Industry Collaboration Projects (IAF-ICP) grant call (Grant No. I2001E0059) - SIA-NUS Digital Aviation Corp Lab. Mike Shou is supported by the Ministry of Education, Singapore, under the Academic Research Fund Tier 1 (FY2023).", + "bbox": [ + 218, + 729, + 785, + 805 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Zhao et al.", + "bbox": [ + 271, + 114, + 344, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 143, + 323, + 160 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Bain, M., Nagrani, A., Varol, G., Zisserman, A.: Frozen in time: A joint video and image encoder for end-to-end retrieval. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1728-1738 (2021)", + "2. Balaji, Y., Min, M.R., Bai, B., Chellappa, R., Graf, H.P.: Conditional gan with discriminative filter generation for text-to-video synthesis. In: IJCAI. vol. 1, p. 2 (2019)", + "3. Blattmann, A., Rombach, R., Ling, H., Dockhorn, T., Kim, S.W., Fidler, S., Kreis, K.: Align your latents: High-resolution video synthesis with latent diffusion models. In: CVPR (2023)", + "4. Blattmann, A., Rombach, R., Ling, H., Dockhorn, T., Kim, S.W., Fidler, S., Kreis, K.: Align your latents: High-resolution video synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22563-22575 (2023)", + "5. Chen, T.S., Lin, C.H., Tseng, H.Y., Lin, T.Y., Yang, M.H.: Motion-conditioned diffusion model for controllable video synthesis. arXiv:2304.14404 (2023)", + "6. Chen, W., Wu, J., Xie, P., Wu, H., Li, J., Xia, X., Xiao, X., Lin, L.: Control-a-video: Controllable text-to-video generation with diffusion models. arXiv preprint arXiv:2305.13840 (2023)", + "7. Chen, X., Huang, L., Liu, Y., Shen, Y., Zhao, D., Zhao, H.: Anydoor: Zero-shot object-level image customization. arXiv preprint arXiv:2307.09481 (2023)", + "8. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 248-255. IEEE (2009)", + "9. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems 34, 8780-8794 (2021)", + "0. Duan, Z., You, L., Wang, C., Chen, C., Wu, Z., Qian, W., Huang, J., Chao, F., Ji, R.: Diffsynth: Latent in-iteration deflickering for realistic video synthesis. arXiv:2308.03463 (2023)", + "1. Esser, P., Chiu, J., Atighehchian, P., Granskog, J., Germanidis, A.: Structure and content-guided video synthesis with diffusion models. arXiv preprint arXiv:2302.03011 (2023)", + "2. Ge, S., Hayes, T., Yang, H., Yin, X., Pang, G., Jacobs, D., Huang, J.B., Parikh, D.: Long video generation with time-agnostic vqgan and time-sensitive transformer. arXiv preprint arXiv:2204.03638 (2022)", + "3. Ge, S., Nah, S., Liu, G., Poon, T., Tao, A., Catanzaro, B., Jacobs, D., Huang, J.B., Liu, M.Y., Balaji, Y.: Preserve your own correlation: A noise prior for video diffusion models. arXiv:2305.10474 (2023)", + "4. Gu, Y., Wang, X., Wu, J.Z., Shi, Y., Chen, Y., Fan, Z., Xiao, W., Zhao, R., Chang, S., Wu, W., et al.: Mix-of-show: Decentralized low-rank adaptation for multi-concept customization of diffusion models. arXiv preprint arXiv:2305.18292 (2023)", + "5. Guo, Y., Yang, C., Rao, A., Wang, Y., Qiao, Y., Lin, D., Dai, B.: Animatediff: Animate your personalized text-to-image diffusion models without specific tuning. arXiv preprint arXiv:2307.04725 (2023)", + "6. He, Y., Xia, M., Chen, H., Cun, X., Gong, Y., Xing, J., Zhang, Y., Wang, X., Weng, C., Shan, Y., et al.: animate-a-story: Storytelling with retrieval-augmented video generation. arXiv:2307.06940 (2023)" + ], + "bbox": [ + 225, + 178, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "MotionDirector", + "bbox": [ + 625, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "17. He, Y., Yang, T., Zhang, Y., Shan, Y., Chen, Q.: Latent video diffusion models for high-fidelity long video generation (2022)", + "18. He, Y., Yang, T., Zhang, Y., Shan, Y., Chen, Q.: Latent video diffusion models for high-fidelity video generation with arbitrary lengths. arXiv:2211.13221 (2022)", + "19. He, Y., Yang, T., Zhang, Y., Shan, Y., Chen, Q.: Videocrafter: A toolkit for text-to-video generation and editing. https://github.com/AILab-CVC/VideoCrafter (2023)", + "20. Hessel, J., Holtzman, A., Forbes, M., Bras, R.L., Choi, Y.: Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718 (2021)", + "21. Ho, J., Chan, W., Sahara, C., Whang, J., Gao, R., Gritsenko, A., Kingma, D.P., Poole, B., Norouzi, M., Fleet, D.J., et al.: Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303 (2022)", + "22. Ho, J., Chan, W., Sahara, C., Whang, J., Gao, R., Gritsenko, A., Kingma, D.P., Poole, B., Norouzi, M., Fleet, D.J., et al.: Imagen video: High definition video generation with diffusion models. arXiv:2210.02303 (2022)", + "23. Ho, J., Salimans, T., Gritsenko, A., Chan, W., Norouzi, M., Fleet, D.J.: Video diffusion models. In: NeurIPS (2022)", + "24. Hong, S., Seo, J., Hong, S., Shin, H., Kim, S.: Large language models are frame-level directors for zero-shot text-to-video generation. arXiv:2305.14330 (2023)", + "25. Hong, W., Ding, M., Zheng, W., Liu, X., Tang, J.: Cogvideo: Large-scale pretraining for text-to-video generation via transformers. arXiv preprint arXiv:2205.15868 (2022)", + "26. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685 (2021)", + "27. Huang, H., Feng, Y., Shi, C., Xu, L., Yu, J., Yang, S.: Free-bloom: Zero-shot text-to-video generator with lvm director and ldm animator. In: NeurIPS (2023)", + "28. Jeong, H., Park, G.Y., Ye, J.C.: Vmc: Video motion customization using temporal attention adaption for text-to-video diffusion models. arXiv preprint arXiv:2312.00845 (2023)", + "29. Karras, J., Holynski, A., Wang, T.C., Kemelmacher-Shlizerman, I.: Dreampose: Fashion image-to-video synthesis via stable diffusion. arXiv:2304.06025 (2023)", + "30. Khachatryan, L., Movsisyan, A., Tadevosyan, V., Henschel, R., Wang, Z., Navasardyan, S., Shi, H.: Text2video-zero: Text-to-image diffusion models are zero-shot video generators. In: ICCV (2023)", + "31. Kirstain, Y., Polyak, A., Singer, U., Matiana, S., Penna, J., Levy, O.: Pick-a-pic: An open dataset of user preferences for text-to-image generation. arXiv preprint arXiv:2305.01569 (2023)", + "32. Kumari, N., Zhang, B., Zhang, R., Shechtman, E., Zhu, J.Y.: Multi-concept customization of text-to-image diffusion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1931-1941 (2023)", + "33. Le Moing, G., Ponce, J., Schmid, C.: Ccvs: context-aware controllable video synthesis. Advances in Neural Information Processing Systems 34, 14042-14055 (2021)", + "34. Li, X., Chu, W., Wu, Y., Yuan, W., Liu, F., Zhang, Q., Li, F., Feng, H., Ding, E., Wang, J.: Videogen: A reference-guided latent diffusion approach for high definition text-to-video generation. arXiv preprint arXiv:2309.00398 (2023)", + "35. Lian, L., Shi, B., Yala, A., Darrell, T., Li, B.: Llm-grounded video diffusion models. arXiv:2309.17444 (2023)", + "36. Luo, Z., Chen, D., Zhang, Y., Huang, Y., Wang, L., Shen, Y., Zhao, D., Zhou, J., Tan, T.: Videofusion: Decomposed diffusion models for high-quality video gen" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Zhao et al.", + "bbox": [ + 271, + 114, + 344, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "eration. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2023)", + "37. Ma, Y., He, Y., Cun, X., Wang, X., Shan, Y., Li, X., Chen, Q.: Follow your pose: Pose-guided text-to-video generation using pose-free videos. arXiv preprint arXiv:2304.01186 (2023)", + "38. Ma, Y., He, Y., Cun, X., Wang, X., Shan, Y., Li, X., Chen, Q.: Follow your pose: Pose-guided text-to-video generation using pose-free videos. arXiv:2304.01186 (2023)", + "39. Materzynska, J., Sivic, J., Shechtman, E., Torralba, A., Zhang, R., Russell, B.: Customizing motion in text-to-video diffusion models. arXiv preprint arXiv:2312.04966 (2023)", + "40. Mei, K., Patel, V.: Vidm: Video implicit diffusion models. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 9117-9125 (2023)", + "41. Ni, H., Shi, C., Li, K., Huang, S.X., Min, M.R.: Conditional image-to-video generation with latent flow diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18444-18455 (2023)", + "42. Qin, B., Ye, W., Yu, Q., Tang, S., Zhuang, Y.: Dancing avatar: Pose and text-guided human motion videos synthesis with image diffusion model. arXiv:2308.07749 (2023)", + "43. Ren, Y., Zhou, Y., Yang, J., Shi, J., Liu, D., Liu, F., Kwon, M., Shrivastava, A.: Customize-a-video: One-shot motion customization of text-to-video diffusion models. arXiv preprint arXiv:2402.14780 (2024)", + "44. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: MICCAI (2015)", + "45. Ruiz, N., Li, Y., Jampani, V., Pritch, Y., Rubinstein, M., Aberman, K.: Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22500-22510 (2023)", + "46. Ryu, S.: Low-rank adaptation for fast text-to-image diffusion fine-tuning. https://github.com/cloneofsimo/lora (2023)", + "47. Saito, M., Matsumoto, E., Saito, S.: Temporal generative adversarial nets with singular value clipping. In: Proceedings of the IEEE international conference on computer vision. pp. 2830-2839 (2017)", + "48. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems 35, 25278-25294 (2022)", + "49. Shen, X., Li, X., Elhoseiny, M.: Mostgan-v: Video generation with temporal motion styles. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5652-5661 (2023)", + "50. Singer, U., Polyak, A., Hayes, T., Yin, X., An, J., Zhang, S., Hu, Q., Yang, H., Ashual, O., Gafni, O., et al.: Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792 (2022)", + "51. Skorokhodov, I., Tulyakov, S., Elhoseiny, M.: Stylegan-v: A continuous video generator with the price, image quality and perks of stylegan2. arXiv preprint arXiv:2112.14683 (2021)", + "52. Smith, J.S., Hsu, Y.C., Zhang, L., Hua, T., Kira, Z., Shen, Y., Jin, H.: Continual diffusion: Continual customization of text-to-image diffusion with c-lora. arXiv preprint arXiv:2304.06027 (2023)", + "53. Soomro, K., Zamir, A.R.: Action recognition in realistic sports videos. In: Computer vision in sports, pp. 181-208. Springer (2015)" + ], + "bbox": [ + 215, + 146, + 785, + 840 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "MotionDirector", + "bbox": [ + 625, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "54. Srivastava, N., Mansimov, E., Salakhudinov, R.: Unsupervised learning of video representations using lstms. In: International conference on machine learning. pp. 843-852. PMLR (2015)", + "55. Sterling, S.: Zeroscope. https://huggingface.co/cerspense/zeroscope_v2_576w (2023)", + "56. Tian, Y., Ren, J., Chai, M., Olszewski, K., Peng, X., Metaxas, D.N., Tulyakov, S.: A good image generator is what you need for high-resolution video synthesis. In: International Conference on Learning Representations (2020)", + "57. Tulyakov, S., Liu, M.Y., Yang, X., Kautz, J.: Mocogan: Decomposing motion and content for video generation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1526-1535 (2018)", + "58. Voleti, V., Jolicoeur-Martineau, A., Pal, C.: Mcvd-masked conditional video diffusion for prediction, generation, and interpolation. Advances in Neural Information Processing Systems 35, 23371-23385 (2022)", + "59. Vondrick, C., Pirsiavash, H., Torralba, A.: Generating videos with scene dynamics. Advances in neural information processing systems 29 (2016)", + "60. Wang, J., Yuan, H., Chen, D., Zhang, Y., Wang, X., Zhang, S.: Modelscope text-to-video technical report. arXiv preprint arXiv:2308.06571 (2023)", + "61. Wang, T., Li, L., Lin, K., Lin, C.C., Yang, Z., Zhang, H., Liu, Z., Wang, L.: Disco: Disentangled control for referring human dance generation in real world. arXiv:2307.00040 (2023)", + "62. Wang, W., Yang, H., Tuo, Z., He, H., Zhu, J., Fu, J., Liu, J.: Videofactory: Swap attention in spatiotemporal diffusions for text-to-video generation. arXiv:2305.10874 (2023)", + "63. Wang, X., Yuan, H., Zhang, S., Chen, D., Wang, J., Zhang, Y., Shen, Y., Zhao, D., Zhou, J.: Videocomposer: Compositional video synthesis with motion controllability. arXiv preprint arXiv:2306.02018 (2023)", + "64. Wang, Y., Chen, X., Ma, X., Zhou, S., Huang, Z., Wang, Y., Yang, C., He, Y., Yu, J., der Yang, P., Guo, Y., Wu, T., Si, C., Jiang, Y., Chen, C., Loy, C.C., Dai, B., Lin, D., Qiao, Y., Liu, Z.: Lavie: High-quality video generation with cascaded latent diffusion models (2023), https://api-semanticscholar.org/CorpusID: 262823915", + "65. Wang, Y., Chen, X., Ma, X., Zhou, S., Huang, Z., Wang, Y., Yang, C., He, Y., Yu, J., Yang, P., et al.: Lavie: High-quality video generation with cascaded latent diffusion models. arXiv:2309.15103 (2023)", + "66. Wang, Z., Yuan, Z., Wang, X., Chen, T., Xia, M., Luo, P., Shan, Y.: Motionctrl: A unified and flexible motion controller for video generation. arXiv preprint arXiv:2312.03641 (2023)", + "67. Wei, Y., Zhang, S., Qing, Z., Yuan, H., Liu, Z., Liu, Y., Zhang, Y., Zhou, J., Shan, H.: Dreamvideo: Composing your dream videos with customized subject and motion. arXiv preprint arXiv:2312.04433 (2023)", + "68. Wei, Y., Zhang, Y., Ji, Z., Bai, J., Zhang, L., Zuo, W.: Elite: Encoding visual concepts into textual embeddings for customized text-to-image generation. arXiv preprint arXiv:2302.13848 (2023)", + "69. Wu, J.Z., Gao, D., Bai, J., Shou, M., Li, X., Dong, Z., Singh, A., Keutzer, K., Iandola, F.: The text-guided video editing benchmark at loveu 2023. https:// sites.google.com/view/loveucvpr23/track4 (2023)", + "70. Wu, J.Z., Ge, Y., Wang, X., Lei, W., Gu, Y., Hsu, W., Shan, Y., Qie, X., Shou, M.Z.: Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. arXiv preprint arXiv:2212.11565 (2022)" + ], + "bbox": [ + 215, + 147, + 784, + 840 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Zhao et al.", + "bbox": [ + 271, + 114, + 344, + 126 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "71. Wu, J.Z., Ge, Y., Wang, X., Lei, W., Gu, Y., Hsu, W., Shan, Y., Qie, X., Shou, M.Z.: Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In: ICCV (2023)", + "72. Wu, R., Chen, L., Yang, T., Guo, C., Li, C., Zhang, X.: Lamp: Learn a motion pattern for few-shot-based video generation. arXiv preprint arXiv:2310.10769 (2023)", + "73. Xing, J., Xia, M., Liu, Y., Zhang, Y., Zhang, Y., He, Y., Liu, H., Chen, H., Cun, X., Wang, X., et al.: Make-your-video: Customized video generation using textual and structural guidance. arXiv:2306.00943 (2023)", + "74. Xing, Z., Feng, Q., Chen, H., Dai, Q., Hu, H., Xu, H., Wu, Z., Jiang, Y.G.: A survey on video diffusion models. arXiv preprint arXiv:2310.10647 (2023)", + "75. Yan, W., Zhang, Y., Abbeel, P., Srinivas, A.: Videogpt: Video generation using vq-vae and transformers. arXiv preprint arXiv:2104.10157 (2021)", + "76. Yang, S., Hou, L., Huang, H., Ma, C., Wan, P., Zhang, D., Chen, X., Liao, J.: Direct-a-video: Customized video generation with user-directed camera movement and object motion. arXiv preprint arXiv:2402.03162 (2024)", + "77. Yin, S., Wu, C., Liang, J., Shi, J., Li, H., Ming, G., Duan, N.: Dragnuwa: Fine-grained control in video generation by integrating text, image, and trajectory. arXiv:2308.08089 (2023)", + "78. Yin, S., Wu, C., Yang, H., Wang, J., Wang, X., Ni, M., Yang, Z., Li, L., Liu, S., Yang, F., et al.: Nuwa-xl: Diffusion over diffusion for extremely long video generation. arXiv:2303.12346 (2023)", + "79. Yu, S., Sohn, K., Kim, S., Shin, J.: Video probabilistic diffusion models in projected latent space. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18456-18466 (2023)", + "80. Yu, S., Tack, J., Mo, S., Kim, H., Kim, J., Ha, J.W., Shin, J.: Generating videos with dynamics-aware implicit generative adversarial networks. In: International Conference on Learning Representations (2021)", + "81. Zhang, D.J., Wu, J.Z., Liu, J.W., Zhao, R., Ran, L., Gu, Y., Gao, D., Shou, M.Z.: Show-1: Marrying pixel and latent diffusion models for text-to-video generation. arXiv:2309.15818 (2023)", + "82. Zhang, D.J., Wu, J.Z., Liu, J.W., Zhao, R., Ran, L., Gu, Y., Gao, D., Shou, M.Z.: Show-1: Marrying pixel and latent diffusion models for text-to-video generation (2023)", + "83. Zhang, L., Agrawala, M.: Adding conditional control to text-to-image diffusion models. arXiv preprint arXiv:2302.05543 (2023)", + "84. Zhao, R., Gu, Y., Wu, J.Z., Zhang, D.J., Liu, J., Wu, W., Keppo, J., Shou, M.Z.: Motiondirector: Motion customization of text-to-video diffusion models. arXiv preprint arXiv:2310.08465 (2023)", + "85. Zhao, S., Chen, D., Chen, Y.C., Bao, J., Hao, S., Yuan, L., Wong, K.Y.K.: Uncontrolnet: All-in-one control to text-to-image diffusion models. arXiv preprint arXiv:2305.16322 (2023)", + "86. Zhou, D., Wang, W., Yan, H., Lv, W., Zhu, Y., Feng, J.: Magicvideo: Efficient video generation with latent diffusion models. arXiv:2211.11018 (2022)" + ], + "bbox": [ + 212, + 146, + 787, + 743 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "MotionDirector", + "bbox": [ + 625, + 114, + 730, + 126 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 18 + } +] \ No newline at end of file diff --git a/2024/MotionDirector_ Motion Customization of Text-to-Video Diffusion Models/8325ffaa-f0c0-4f28-929f-63aa2f3d7c76_model.json b/2024/MotionDirector_ Motion Customization of Text-to-Video Diffusion Models/8325ffaa-f0c0-4f28-929f-63aa2f3d7c76_model.json new file mode 100644 index 0000000000000000000000000000000000000000..ed721352ec044e62b28ee6fc778b946044e05b90 --- /dev/null +++ b/2024/MotionDirector_ Motion Customization of Text-to-Video Diffusion Models/8325ffaa-f0c0-4f28-929f-63aa2f3d7c76_model.json @@ -0,0 +1,2603 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.263, + 0.141, + 0.743, + 0.184 + ], + "angle": 0, + "content": "MotionDirector: Motion Customization of Text-to-Video Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.23, + 0.212, + 0.773, + 0.244 + ], + "angle": 0, + "content": "Rui Zhao\\(^{1}\\), Yuchao Gu\\(^{1}\\), Jay Zhangjie Wu\\(^{1}\\), David Junhao Zhang\\(^{1}\\), Jia-Wei Liu\\(^{1}\\), Weijia Wu\\(^{1}\\), Jussi Keppo\\(^{2}\\), and Mike Zheng Shou\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.31, + 0.254, + 0.694, + 0.27 + ], + "angle": 0, + "content": "\\(^{1}\\)Show Lab, \\(^{2}\\)National University of Singapore, Singapore" + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.306, + 0.74, + 0.582 + ], + "angle": 0, + "content": "Abstract. Large-scale pre-trained diffusion models have exhibited remarkable capabilities in diverse video generations. Given a set of video clips of the same motion concept, the task of Motion Customization is to adapt existing text-to-video diffusion models to generate videos with this motion. Adaptation methods have been developed for customizing appearance like subject or style, yet under-explored for motion. It is straightforward to extend mainstream adaption methods for motion customization, including full model tuning and Low-Rank Adaptions (LoRAs). However, the motion concept learned by these methods is often coupled with the limited appearances in the training videos, making it difficult to generalize the customized motion to other appearances. To overcome this challenge, we propose MotionDirector, with a dual-path LoRAs architecture to decouple the learning of appearance and motion. Further, we design a novel appearance-debiased temporal loss to mitigate the influence of appearance on the temporal training objective. Experimental results show the proposed method can generate videos of diverse appearances for the customized motions. Our method also supports various downstream applications, such as the mixing of different videos with their appearance and motion respectively, and animating a single image with customized motions. The project website is at: MotionDirector." + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.597, + 0.741, + 0.623 + ], + "angle": 0, + "content": "Keywords: motion customization \\(\\cdot\\) text-to-video generation \\(\\cdot\\) diffusion models" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.65, + 0.377, + 0.667 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.681, + 0.788, + 0.818 + ], + "angle": 0, + "content": "Text-to-video diffusion models [17,21,50] are approaching generating high-quality diverse videos given text instructions. The open-sourcing of foundational text-to-video models [55,60] pre-trained on large-scale data has sparked enthusiasm for video generation in both the community and academia. Users can create videos that are either realistic or imaginative simply by providing text prompts. While foundation models generate diverse videos from the same text, adapting them to generate more specific content can better accommodate the preferences of users. Similar to the customization of text-to-image foundation models [45], tuning video foundation models to generate videos of a certain concept of appearance," + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.39, + 0.84 + ], + "angle": 0, + "content": "* Corresponding Author." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.116, + 0.347, + 0.127 + ], + "angle": 0, + "content": "Zhao et al." + }, + { + "type": "image", + "bbox": [ + 0.222, + 0.146, + 0.784, + 0.341 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.302, + 0.35, + 0.7, + 0.365 + ], + "angle": 0, + "content": "Fig. 1: Motion customization of the text-to-video diffusion." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.399, + 0.785, + 0.49 + ], + "angle": 0, + "content": "like subject or style, has also been explored [17]. Compared with images, videos consist of not just appearances but also motion dynamics. Users may desire to create videos with specific motions, such as a car moving forward and then turning left under a predefined camera perspective, as illustrated on the right side in Fig. 1. However, customizing the motions in the text-to-video generation is still under-explored." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.494, + 0.788, + 0.844 + ], + "angle": 0, + "content": "The task of Motion Customization is formulated as follows: given a set of reference videos representing a specific motion concept, the objective is to customize pre-trained foundational models to generate videos that accurately capture and replicate the depicted motion. The customized motion concept can be challenging motions that are hard for the pre-trained model to generate, novel motions not present in the pre-training dataset, and specific motions that are difficult to express in text. Examples of these are respectively showcased in the first, second, and third columns of Fig 1. In contrast, previous works on appearance customization adapt the foundation models to generate samples with desired appearance, like subject or style, given reference videos or images representing such appearance [17,45]. It is straightforward to use previous adaption methods for motion customization. For example, on the given reference videos, fine-tuning the weights of foundation models [45], parameter-efficient tuning additional layers [70], or training Low-Rank Adaptions (LoRAs) [26] injected in the layers of foundation models. However, customizing diffusion models to generate desired motions without harming their appearance diversity is challenging because the motion and appearance are coupled with each other at the step-by-step denoising stage. Directly deploying previous adaption methods to learn motions makes the models fit the limited appearances seen in the reference videos, posing challenges in generalizing the learned motions to various appearances. Recent works on controllable text-to-video generations [11, 17, 63] generate videos controlled by signals representing pre-defined motions. However, the control signals, such as depth maps or edges, impose constraints on the shapes of subjects and back" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.626, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "MotionDirector" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.145, + 0.782, + 0.347 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.361, + 0.788, + 0.43 + ], + "angle": 0, + "content": "Fig. 2: (Row 1) Take two videos to train the proposed MotionDirector, respectively. (Row 2) MotionDirector can generalize the learned motions to diverse appearances. (Row 3) MotionDirector can mix the learned motion and appearance from different videos to generate new videos. (Row 4) MotionDirector can animate a single image with learned motions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.462, + 0.785, + 0.522 + ], + "angle": 0, + "content": "grounds, thus influencing the appearance of generated videos in a coupled way. Besides, these methods accept only one sequence of control signals to generate one video, which may not be suitable for users seeking certain motion types without strict spatial constraints, such as the example of lifting weights in Fig. 1." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.523, + 0.788, + 0.794 + ], + "angle": 0, + "content": "To achieve motion customization of text-to-video diffusion models while preserving appearance diversity, we propose the MotionDirector, which tunes the foundation models to learn the appearance and motions in the given single or multiple reference videos in a decoupled way. MotionDirector tunes the models with low-rank adaptions (LoRAs) while keeping their pre-trained parameters fixed to retain the learned generation knowledge. Specifically, the MotionDirector employs a dual-path architecture, as shown in Fig. 3. For each video, a spatial path consists of a foundation model with trainable spatial LoRAs injected into its spatial transformer layers. These spatial LoRAs are trained on a single frame randomly sampled per training step to capture the appearance characteristics of the input videos. The temporal path, on the other hand, is a replica of the foundation model that shares the spatial LoRAs with the spatial path to fit the appearance of the corresponding input video. Additionally, the temporal transformers in this path are equipped with temporal LoRAs, which are trained on multiple frames of input videos to capture the underlying motion patterns. To further enhance the learning of motions, we propose an appearance-debiased temporal loss to mitigate the influence of appearance on the temporal training objective." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Only deploying the trained temporal LoRAs enables the foundation model to generate videos of the learned motions with diverse appearances, as shown in the second row of Fig 2. The decoupled paradigm further makes an interesting" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.346, + 0.127 + ], + "angle": 0, + "content": "Zhao et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.268 + ], + "angle": 0, + "content": "kind of video generation feasible, which is the mix of the appearance from one video with the motion from another video, called the mix of videos, as shown in the third row of Fig 2. The key to this success lies in that MotionDirector can decouple the appearance and motion of videos and then combine them from various source videos. It is achieved by injecting spatial LoRAs trained on one video and temporal LoRAs trained on another video into the foundation model. Besides, the learned motions can be deployed to animate images, as images can be treated as appearance providers, as shown in the last row of Fig 2." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.269, + 0.788, + 0.373 + ], + "angle": 0, + "content": "We conducted experiments on two benchmarks with 86 different motions and over 600 text prompts to test proposed methods, baselines, and comparison methods. The results show our method can be applied to different diffusion-based foundation models and achieve motion customization of various motion concepts. Compared with other methods, our method avoids fitting the limited appearance of reference videos, and can generalize the learned motions to diverse appearances." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.375, + 0.57, + 0.389 + ], + "angle": 0, + "content": "Our contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.402, + 0.784, + 0.431 + ], + "angle": 0, + "content": "- We introduce and define the task of Motion Customization. The challenges lie in generalizing the customized motions to various appearances." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.433, + 0.784, + 0.475 + ], + "angle": 0, + "content": "- We propose the MotionDirector with a dual-path architecture and a novel appearance-debiased temporal training objective, to decouple the learning of appearance and motion." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.479, + 0.784, + 0.535 + ], + "angle": 0, + "content": "- Experiments on two benchmarks demonstrate that MotionDirector can customize various base models to generate diverse videos with desired motion concepts, and outperforms controllable generation methods and tuning-based methods." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.402, + 0.784, + 0.535 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.566, + 0.389, + 0.582 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.601, + 0.48, + 0.615 + ], + "angle": 0, + "content": "2.1 Text-to-Video Generation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.789, + 0.841 + ], + "angle": 0, + "content": "To achieve high-quality video generation, various methods have been developed, such as Generative Adversarial Networks (GANs) [2, 47, 49, 56, 57, 59], autoregressive models [12, 25, 33, 54, 75] and implicit neural representations [51, 80]. Diffusion-based models [3, 10, 13, 18, 22-24, 27, 34, 35, 40, 41, 50, 58, 60, 62, 65, 71, 78, 79, 81, 84, 86] are also approaching high-quality generation by training conditional 3D U-Nets [44] to denoise from randomly sampled sequences of Gaussian noises [74]. Recent foundation models [4, 17, 21, 36, 50, 64, 82] are pre-trained on large-scale image and video datasets [1, 8, 48], to learn powerful generation ability. Some works turn text-to-image foundation models to text-to-video generation by manipulation on cross-frame attention or training additional temporal layers, like Tune-A-Video [71], Text2Video-Zero [30], and AnimiteDiff [15]. The recently open-sourced foundation models [55, 60] have ignited enthusiasm among users to generate realistic or imaginative videos, and further make it possible for users to customize and build their own private models." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.626, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "MotionDirector" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.544, + 0.162 + ], + "angle": 0, + "content": "2.2 Generation Model Customization." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.176, + 0.788, + 0.389 + ], + "angle": 0, + "content": "Customizing the pre-trained large foundation models can fit the preferences of users better while maintaining powerful generation knowledge without training from scratch. Previous customization methods for text-to-image diffusion models [7,14,32,45,52,68] aim to generate certain subjects or styles, given a set of example images. Dreambooth [45] or LoRA [26] can be simply applied to customizing video foundation models to generate videos with certain subjects or styles, given a set of reference video clips or images. The recently proposed VideoCrafter [19] has explored this, which we categorize as appearance customization. In addition to appearances, videos are also characterized by the motion dynamics of subjects and camera movements across frames. After we proposed the task of motion customization, some concurrent works [28, 39, 43, 66, 67, 72, 76] emerged in this field. DreamVideo [67] trains several adapters in U-Net to customize both the subject and motion in video generation. Lamp [72] learns a motion pattern from a set of videos to utilize image diffusion models to generate videos." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.414, + 0.52, + 0.428 + ], + "angle": 0, + "content": "2.3 Controllable Video Generation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.443, + 0.788, + 0.625 + ], + "angle": 0, + "content": "Controllable generation aims to ensure the generation results align with the given explicit control signals, such as depth maps, human pose, optical flows, etc. [5, 16, 29, 37, 38, 42, 61, 73, 77, 83, 85]. For the controllable text-to-video generation methods, i.e. the VideoCrafter [17], VideoComposer [63], Control-A-Video [6], they train additional branches that take condition signals to align the generated videos with them. Unlike the human poses for specifically controlling the generation of human bodies, the general control singles, such as depth maps, are typically extracted from reference videos and are coupled with both appearance and motion. This results in the generation results being influenced by both the appearance and motion in reference videos. Applying these methods directly in motion customization is challenging when it comes to generalizing the desired motions to diverse appearances." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.652, + 0.381, + 0.67 + ], + "angle": 0, + "content": "3 Methodology" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.687, + 0.374, + 0.702 + ], + "angle": 0, + "content": "3.1 Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.716, + 0.787, + 0.809 + ], + "angle": 0, + "content": "Video Diffusion Model. Video diffusion models train a 3D U-Net to denoise from a randomly sampled sequence of Gaussian noises to generate videos, guided by text prompts. The 3D U-net basically consists of down-sample, middle, and up-sample blocks. Each block has several convolution layers, spatial transformers, and temporal transformers as shown in Fig 3. The 3D U-Net \\(\\epsilon_{\\theta}\\) and a text encoder \\(\\tau_{\\theta}\\) are jointly optimized by the noise-prediction loss, as detailed in [9]:" + }, + { + "type": "equation", + "bbox": [ + 0.323, + 0.825, + 0.786, + 0.843 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathbb {E} _ {z _ {0}, y, \\epsilon \\sim \\mathcal {N} (0, I), t \\sim \\mathcal {U} (0, T)} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} \\left(z _ {t}, t, \\tau_ {\\theta} (y)\\right) \\| _ {2} ^ {2} \\right], \\tag {1}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.347, + 0.128 + ], + "angle": 0, + "content": "Zhao et al." + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.146, + 0.785, + 0.307 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.318, + 0.788, + 0.403 + ], + "angle": 0, + "content": "Fig. 3: The dual-path architecture of the proposed method. All pre-trained weights of the base diffusion model remain fixed. In the spatial path, the spatial transformers are injected with trainable spatial LoRAs as shown on the right side. In the temporal path, the spatial transformers are injected with spatial LoRAs sharing weights with those ones in the spatial path, and the temporal transformers are injected with trainable temporal LoRAs." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.433, + 0.788, + 0.48 + ], + "angle": 0, + "content": "where \\( z_0 \\) is the latent code of the training videos, \\( y \\) is the text prompt, \\( \\epsilon \\) is the Gaussian noise added to the latent code, and \\( t \\) is the time step. As discussed in [9], the noised latent code \\( z_t \\) is determined as:" + }, + { + "type": "equation", + "bbox": [ + 0.368, + 0.491, + 0.786, + 0.533 + ], + "angle": 0, + "content": "\\[\nz _ {t} = \\sqrt {\\bar {\\alpha} _ {t}} z _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon , \\bar {\\alpha} _ {t} = \\prod_ {i = 1} ^ {t} \\alpha_ {t}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.542, + 0.656, + 0.557 + ], + "angle": 0, + "content": "where \\(\\alpha_{t}\\) is a hyper-parameter controlling the noise strength." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.557, + 0.788, + 0.633 + ], + "angle": 0, + "content": "Low-Rank Adaption. Low-rank adaption (LoRA) [26] was proposed to adapt the pre-trained large language models to downstream tasks. Recently it has been applied in text-to-image generation and text-to-video generation tasks to achieve appearance customization [19, 46]. LoRA employs a low-rank factorization technique to update the weight matrix \\( W \\) as" + }, + { + "type": "equation", + "bbox": [ + 0.395, + 0.646, + 0.786, + 0.663 + ], + "angle": 0, + "content": "\\[\nW = W _ {0} + \\Delta W = W _ {0} + B A, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.673, + 0.788, + 0.763 + ], + "angle": 0, + "content": "where \\( W_0 \\in \\mathbb{R}^{d \\times k} \\) represents the original weights of the pre-trained model, \\( B \\in \\mathbb{R}^{d \\times r} \\) and \\( A \\in \\mathbb{R}^{r \\times k} \\) represent the low-rank factors, where \\( r \\) is much smaller than original dimensions \\( d \\) and \\( k \\). LoRA requires smaller computing sources than fine-tuning the weights of the entire network like DreamBooth [45], and it is convenient to spread and deploy as a plug-and-play plugin for pre-trained models." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In the following sections, we introduce the proposed MotionDirector, containing the dual-path Low-Rank Adaptions (in Sec. 3.2) for learning the appearance and motion in reference videos in a decoupled manner, and appearance-debiased temporal loss (in Sec. 3.3)) to further mitigate the influence of appearance on the temporal training objective." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.626, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "MotionDirector" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.146, + 0.785, + 0.319 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.331, + 0.788, + 0.375 + ], + "angle": 0, + "content": "Fig. 4: Comparison of videos generated by the base model, the model fine-tuned in the coupled manner, the model fine-tuned without the proposed Appearance-Debiased Temporal Loss (AD-Loss), and our full-version MotionDirector method." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.404, + 0.527, + 0.42 + ], + "angle": 0, + "content": "3.2 Dual-Path Low-Rank Adaptions" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.431, + 0.789, + 0.688 + ], + "angle": 0, + "content": "At each time-step \\( t \\), the 3D U-Net takes in the latent code \\( z_{t} \\in \\mathbb{R}^{b \\times f \\times w \\times h \\times c} \\) and the conditional input \\( y \\) (e.g., text), where \\( b, f, w, h, c \\) represents the size of the batch, frame, width, height, and channel dimensions, respectively. The spatial transformers apply spatial self-attention along the spatial dimensions \\( w, h \\) to improve the correlation between pixels, and then leverage the cross-attention between the latent code and the conditional input \\( y \\) to improve textual alignment. The temporal transformers apply temporal self-attention along the frame dimension \\( f \\) to improve the temporal consistency between frames. However, spatial and temporal information in the latent code gradually become coupled with each other during the step-by-step denoising stage. Attempting to directly learn and fit the motions in reference videos will inevitably lead to fitting their limited appearances. As shown in the third row \"Coupled Tuned\" of Fig. 4. From the results on the left side, we can observe that the generated \"panda\" is coupled with the appearance of \"human\" in the reference videos and the generated background is also coupled with that of reference videos. Besides, the results on the right side show that the coupled tuned model can not generate the \"Eiffel Tower\" because it is overfitted on the limited backgrounds in the reference videos." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.793 + ], + "angle": 0, + "content": "To address this problem, we propose to tune the spatial and temporal transformers in a dual-path way to learn the earn the appearance and motion in reference videos, respectively, as shown in Fig. 3. Specifically, for the spatial path, we inject LoRAs into spatial transformers to learn the appearance of training data, and for the temporal path, we inject LoRAs into temporal transformers to learn the motion in videos. The training of spatial and temporal LoRAs are as follows." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Spatial Training. For the spatial path, we inject unique spatial LoRAs into the spatial transformers for each training video while keeping the weights of pre-trained 3D U-Net fixed. To maintain the learned strong and diverse textual" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.346, + 0.128 + ], + "angle": 0, + "content": "Zhao et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.269 + ], + "angle": 0, + "content": "alignment ability, we do not inject LoRAs into cross-attention layers of spatial transformers, since their weights influence the correlations between the pixels and text prompts. On the other hand, we inject LoRAs into spatial self-attention layers and feed-forward layers to update the correlations in spatial dimensions to enable the model to reconstruct the appearance of training data. For each training step, the spatial LoRAs are trained on a single frame randomly sampled from the training video to fit its appearance while ignoring its motion, based on spatial loss, which is reformulated as" + }, + { + "type": "equation", + "bbox": [ + 0.32, + 0.283, + 0.786, + 0.302 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {s p a t i a l}} = \\mathbb {E} _ {z _ {0}, y, \\epsilon , t, i \\sim \\mathcal {U} (0, F)} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} \\left(z _ {t, i}, t, \\tau_ {\\theta} (y)\\right) \\| _ {2} ^ {2} \\right], \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.308, + 0.786, + 0.337 + ], + "angle": 0, + "content": "where \\( \\mathbf{F} \\) is the number of frames of the training data and the \\( z_{t,i} \\) is the sampled frame from the latent code \\( z_{t} \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.339, + 0.788, + 0.46 + ], + "angle": 0, + "content": "Temporal Training. For the temporal path, we inject the temporal LoRAs into self-attention and feed-forward layers of temporal transformers to update the correlations along the frame dimension. Besides, the spatial transformers are injected with LoRAs sharing the same weights learned from the spatial path, to force the trainable temporal LoRAs to ignore the appearance of the training data. The temporal LoRAs could be simply trained on all frames of training data based on the temporal loss \\(\\mathcal{L}_{org-temp}\\), formulated in the same way as equation (1)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.483, + 0.567, + 0.499 + ], + "angle": 0, + "content": "3.3 Appearance-Debiased Temporal Loss" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.51, + 0.787, + 0.569 + ], + "angle": 0, + "content": "The videos generated by models trained with original temporal loss are shown in the fourth row of Fig. 4. We can observe that the motions are not learned well. To explore the underlying reasons, we visualized the denoising process of several videos." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.57, + 0.788, + 0.706 + ], + "angle": 0, + "content": "As illustrated in Fig. 5, when considering the latent codes of each frame \\( z_{t,i_{i = 1}^{F}} \\) as a set of points in the latent space, we found that motion primarily impacts the underlying dependencies between these point sets, whereas the distances between different sets of points are more influenced by appearance. The original temporal training objective does not take this into account. The learning of motion dynamics, embedded within the internal connectivity structure of latent codes, remains biased by positional differences of latent codes. These biases are primarily influenced by the appearance of training samples and pose a challenge to learning motions efficiently." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.707, + 0.788, + 0.768 + ], + "angle": 0, + "content": "To further decouple the motion from appearance, we proposed to eliminate the appearance bias among the noises and predicted noises, and calculate the appearance-debiased temporal loss on them. The debiasing of each noise \\(\\epsilon_{i} \\in \\{\\epsilon_{i}\\}_{i=1}^{F}\\) is as follows," + }, + { + "type": "equation", + "bbox": [ + 0.395, + 0.78, + 0.786, + 0.799 + ], + "angle": 0, + "content": "\\[\n\\phi (\\epsilon_ {i}) = \\sqrt {\\beta^ {2} + 1} \\epsilon_ {i} - \\beta \\epsilon_ {a n c h o r}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.788, + 0.841 + ], + "angle": 0, + "content": "where \\(\\beta\\) is the strength factor controlling the debiasing strength and a higher value leads to stronger debiasing effects. The \\(\\epsilon_{anchor}\\) is the anchor among the" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.626, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "MotionDirector" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.145, + 0.784, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.326, + 0.788, + 0.48 + ], + "angle": 0, + "content": "Fig. 5: (a) Four example videos (the same as the videos in the first and fourth rows of Fig. 2) and their relationships in terms of motion and appearance. (b) We inverse the four videos based on the video diffusion model and visualize the denoising process. Each point corresponds to a latent code \\( z_{t,i,j} \\) at time step \\( t \\) of \\( i \\)-th frame of \\( j \\)-th video, reduced to 2 dimensions through PCA. (c) Take latent codes at time step 0 for example, the ones of the same video are connected in order of frames. We find that the internal connectivity structure between latent codes is more influenced by motion, while the distance between sets of latent codes is primarily affected by the difference in appearance. (d) Through the proposed appearance-debiased operation, as formulated in Equation 5, the latent codes eliminate distance bias introduced by appearance, while preserving the internal connectivity structure that represents motion dynamics." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.509, + 0.788, + 0.601 + ], + "angle": 0, + "content": "frames from the same training data. If \\(\\beta\\) equals 1, it indicates that we attribute equal importance to both appearance and motion in the video. The detailed mathematical derivation of this equation is provided in the supplementary materials. In practice, we find that simply setting \\(\\beta = 1\\) and randomly sampling \\(\\epsilon_{i} \\in \\{\\epsilon_{i}\\}_{i=1}^{F}\\) as the anchor can achieve good performance. The appearance-debiased temporal loss (AD-Loss) is reformulated as" + }, + { + "type": "equation", + "bbox": [ + 0.323, + 0.611, + 0.787, + 0.631 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {a d - t e m p} = \\mathbb {E} _ {z _ {0}, y, \\epsilon , t} \\left[ \\| \\phi (\\epsilon) - \\phi \\left(\\epsilon_ {\\theta} \\left(z _ {t}, t, \\tau_ {\\theta} (y)\\right)\\right) \\| _ {2} ^ {2} \\right]. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.641, + 0.788, + 0.672 + ], + "angle": 0, + "content": "For temporal LoRAs, the loss function is the combination of temporal loss and appearance-debiased temporal loss as follows," + }, + { + "type": "equation", + "bbox": [ + 0.38, + 0.685, + 0.787, + 0.703 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t e m p o r a l}} = \\mathcal {L} _ {\\text {o r g - t e m p}} + \\mathcal {L} _ {\\text {a d - t e m p}}. \\tag {7}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.723, + 0.496, + 0.737 + ], + "angle": 0, + "content": "3.4 Inference of MotionDirector" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.842 + ], + "angle": 0, + "content": "In the inference stage, we inject the trained temporal LoRAs into the pre-trained video diffusion model to enable it to generate diverse videos with the learned motion from the training data. If the training data is a single video, the learned motion will be a specific motion, such as an object first moving forward and then turning to the left. If the training data is a set of videos, the learned motion will be the motion concept provided by them, like lifting weights or playing golf." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.116, + 0.347, + 0.127 + ], + "angle": 0, + "content": "Zhao et al." + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.145, + 0.787, + 0.352 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.225, + 0.363, + 0.777, + 0.379 + ], + "angle": 0, + "content": "Fig. 6: Qualitative comparison results of motion customization on multiple videos." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.431, + 0.787, + 0.459 + ], + "angle": 0, + "content": "Tab. 1: Automatic and human evaluations results of motion customization on multiple videos." + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.472, + 0.785, + 0.577 + ], + "angle": 0, + "content": "
Automatic EvaluationsHuman Evaluations
Appearance Diversity (↑)Temporal Consistency (↑)Pick Score (↑)Appearance DiversityTemporal ConsistencyMotion Fidelity
Tune-A-Video28.2292.4520.20v.s. Base Model (ModelScope)25.00 v.s. 75.0025.00 v.s. 75.0040.00 v.s. 60.00
v.s. Base Model (ZeroScope)44.00 v.s. 56.0016.67 v.s. 83.3353.33 v.s. 46.67
ModelScopeBase Model28.5592.5420.33v.s. Base Model (ModelScope)23.08 v.s. 76.9240.00 v.s. 60.0052.00 v.s. 48.00
Coupled Tuned25.66 (-2.89)90.6619.85
w/o AD-Loss28.32 (-0.23)91.1720.34v.s. Base Model (ModelScope)53.12 v.s. 46.8849.84 v.s. 50.1662.45 v.s. 37.55
ours28.66 (+0.11)92.3620.59v.s. Base Model (ModelScope)54.84 v.s. 45.1656.00 v.s. 44.0075.00 v.s. 25.00
ZeroScopeBase Model28.4092.9420.76v.s. Base Model (ZeroScope)37.81 v.s. 62.1941.67 v.s. 58.3354.55 v.s. 45.45
Coupled Tuned25.52 (-2.88)90.6719.99
w/o AD-Loss28.61 (+0.21)91.3720.56v.s. Base Model (ZeroScope)50.10 v.s. 49.9048.00 v.s. 52.0058.33 v.s. 41.67
ours28.94 (+0.54)92.6720.80v.s. Base Model (ZeroScope)52.94 v.s. 47.0655.00 v.s. 45.0076.47 v.s. 23.53
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.788, + 0.841 + ], + "angle": 0, + "content": "The motion concepts can be ones preferred by users or ones that lie in the long-tailed distribution that can not be synthesized well by pre-trained models. Since appearance and motion are decoupled by our method, the spatial LoRAs can also be used to influence the appearance of generated videos, as shown in Fig. 2. Users can flexibly adjust the influence strength of learned appearance and motion on the generation according to their preferences by simply setting the strength of LoRAs as \\( W = W_{0} + \\gamma \\Delta W \\), where \\( \\gamma \\) is called the LoRA scale, and \\( \\Delta W \\) is the learned weights. For instance, a small scale of temporal LoRAs leads to the generated results weakly expressing the learned motion concepts, while a significantly large scale leads to discordant generation. More experimental results showing the effect of LoRA scales are provided in the supplementary materials. To ensure comparative fairness, we set the scale of temporal LoRAs to 1 in all experiments in Sec. 4." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.626, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "MotionDirector" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.145, + 0.377, + 0.163 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.184, + 0.6, + 0.199 + ], + "angle": 0, + "content": "4.1 Motion customization on multiple videos" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.215, + 0.788, + 0.306 + ], + "angle": 0, + "content": "Dataset. We conduct experiments on the adapted UCF Sports Action data set [53], which includes 95 videos of 12 different human motions, like playing golf, lifting weights, etc. For each type of motion, we label one original text prompt describing the motion, such as \"a person is playing golf, side view\". For these motions, we set 72 different text prompts in total as input to generate videos using comparison methods, such as \"a monkey is playing golf, side view\"." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.307, + 0.788, + 0.458 + ], + "angle": 0, + "content": "Comparison Methods. We compare the proposed method with three baselines and the video generation method Tune-A-Video [70] that can be adapted to this task. Tune-A-Video was initially proposed for training temporal modules on a single video to learn its motion information, while here we adapt it to train on multiple videos. The baseline methods are compared with the proposed method on two different foundational text-to-video diffusion models, i.e. the ModelScope [60] and the ZeroScope [55]. We employ three baseline methods: the first is directly applying the vanilla foundation models, the second is tuning the foundation models with LoRAs in a coupled manner, and the third is the proposed dual-path method excluding the appearance-debiased temporal loss." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.459, + 0.788, + 0.672 + ], + "angle": 0, + "content": "Qualitative Results As shown in Fig. 6, taking a set of videos with motions of playing golf as training data, the Tune-A-Video fails to generate diverse appearances with the learned motions, like a monkey playing golf. To compare the baseline methods and proposed method fairly, we feed the same initial Gaussian noise to these methods to generate videos. The pre-trained foundation model, ZeroScope, correctly generates the appearance but lacks the realistic motion that swings a golf club, as those desired motions in the reference videos. The coupled tuned model could generate the desired motion but the learned motion is coupled with too much appearance information causing the generated subject in the video to be more like a human rather than a monkey. The last two rows show that the proposed dual-path LoRAs can avoid hurting the appearance generation and the proposed appearance-debiased temporal loss enhances the learning of desired motion better. We could draw a similar conclusion from the second example showing the motion of riding a panda." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.672, + 0.787, + 0.701 + ], + "angle": 0, + "content": "Quantitative Results. We evaluate the methods with automatic evaluations and human evaluations, and the results are shown in Table. 1." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.703, + 0.787, + 0.779 + ], + "angle": 0, + "content": "Automatic Metrics. Following the LOVEU-TGVE competition [69], the appearance diversity is computing the average CLIP score [20] between the diverse text prompts and all frames of the generated videos, the temporal consistency is the average CLIP score between frames, and the Pick Score is the average PickScore [31] between all frames of output videos." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.78, + 0.787, + 0.841 + ], + "angle": 0, + "content": "In Table 1, the automatic evaluation of temporal consistency is calculated as the CLIP score between frames. This may result in videos that are nearly static and without desired motion receiving high scores. Thus, we further apply human evaluation to evaluate generated videos more comprehensively and accurately." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.347, + 0.128 + ], + "angle": 0, + "content": "Zhao et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.268 + ], + "angle": 0, + "content": "Human Preference. On the Amazon MTurk \\(^{1}\\), each generated video is evaluated by 5 human raters in terms of appearance diversity, temporal consistency, and motion fidelity, which evaluate whether the generated motion is similar to the references. To simplify the comparison for raters, they are asked to compare the results pairwise and select their preferred one, where the videos are shuffled and their source methods are anonymous. In Table. 1, the pairwise numbers \"p1 v.s. p2\" means \\(p_1\\%\\) results of the first method are preferred while \\(p_2\\%\\) results of the second method are preferred." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.269, + 0.788, + 0.314 + ], + "angle": 0, + "content": "The evaluation results show that coupled tuning will destroy the appearance diversity of pre-trained models, while our method will preserve it and achieve the highest motion fidelity." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.338, + 0.586, + 0.354 + ], + "angle": 0, + "content": "4.2 Motion customization on a single video" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.366, + 0.796, + 0.426 + ], + "angle": 0, + "content": "Dataset. We conduct the comparison experiments on the open-sourced benchmark released by the LOVEU-TGVE competition at CVPR 2023 [69]. The dataset comprises 76 videos, each originally associated with 4 editing text prompts. Additionally, we introduced 3 more prompts with significant changes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.427, + 0.788, + 0.502 + ], + "angle": 0, + "content": "Comparison Methods. We compare the proposed method with SOTA controllable generation methods, the VideoCrafter [17], VideoComposer [63], and Control-A-Video [6], and the tuning-based method Tune-A-Video [70]. To ensure a fair comparison, we use the depth control mode of controllable generation methods, which is available in all of them." + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.528, + 0.785, + 0.768 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.23, + 0.771, + 0.773, + 0.786 + ], + "angle": 0, + "content": "Fig. 7: Qualitative comparison results of motion customization on a single video." + }, + { + "type": "page_footnote", + "bbox": [ + 0.22, + 0.824, + 0.434, + 0.841 + ], + "angle": 0, + "content": "1 https://requester.mturk.com/" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.626, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "MotionDirector" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.785, + 0.172 + ], + "angle": 0, + "content": "Tab. 2: Automatic and human evaluations results of motion customization on a single video." + }, + { + "type": "table", + "bbox": [ + 0.219, + 0.178, + 0.787, + 0.242 + ], + "angle": 0, + "content": "
Automatic EvaluationsHuman Evaluations
Text Alignment (↑)Appearance Diversity (↑)Temporal Consistency (↑)Pick Score (↑)Text AlignmentAppearance DiversityTemporal ConsistencyMotion Fidelity
VideoComposer27.6627.0392.2220.26ours v.s. VideoComposer54.55 v.s. 45.4572.83 v.s. 27.1761.57 v.s. 38.4361.24 v.s. 38.76
Control-a-Video26.5425.3592.6319.75ours v.s. Control-A-Video68.00 v.s. 32.0078.43 v.s. 21.5771.28 v.s. 29.7256.47 v.s. 43.53
VideoCrafter28.0327.6992.2620.12ours v.s. VideoCrafter52.72 v.s. 47.2871.11 v.s. 28.8960.22 v.s. 39.7860.00 v.s. 40.00
Tune-a-Video25.6425.9592.4220.09ours v.s. Tune-A-Video67.86 v.s. 32.1469.14 v.s. 30.8671.67 v.s. 28.3356.52 v.s. 43.48
ours27.8228.4893.0020.74
" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.256, + 0.785, + 0.411 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.421, + 0.785, + 0.464 + ], + "angle": 0, + "content": "Fig. 8: Videos generated by MotionDirector. Their appearances are customized according to the reference images shown on the top-left side. Their motions are customized based on the reference videos shown on the top-right side." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.495, + 0.788, + 0.648 + ], + "angle": 0, + "content": "Qualitative and Quantitative Results. As shown in Fig. 7, comparison methods fail to generalize the desired motions to diverse appearances, like the ears of bears and the Arc de Triomphe. Similar to the depth, using the edge and optical flow as control signals also introduces interference from the appearance into the generated videos. In Table. 2, we refer to the alignment between the generated videos and the original 4 editing text prompts as text alignment, and the alignment with the 3 new text prompts with significant changes as appearance diversity. The results show that our method outperforms other methods by a large margin when generalizing the motions to diverse appearances, and achieves competitive motion fidelity." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.674, + 0.618, + 0.691 + ], + "angle": 0, + "content": "4.3 Customizing Both Appearance and Motion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Since MotionDirector learns the appearance and motion in a decoupled manner, it can also reorganize and combine the appearances and motions from different sources. For example, as shown in Fig. 2, MotionDirector can combine the appearance from one video with the motion from another video to generate a new video that exhibits similar appearance and motion characteristics to each of the input videos, respectively. Here we provide more results in Fig. 8. The base model is injected with the spatial LoRAs trained on the given images of \"Terracotta Warrior\" and injected with different temporal LoRAs trained on different sets of videos of different motion concepts, such as \"riding horse\"." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.116, + 0.346, + 0.127 + ], + "angle": 0, + "content": "Zhao et al." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.147, + 0.451, + 0.162 + ], + "angle": 0, + "content": "4.4 Efficiency Performance" + }, + { + "type": "text", + "bbox": [ + 0.219, + 0.172, + 0.787, + 0.308 + ], + "angle": 0, + "content": "The lightweight LoRAs enable our method to tune the foundation models efficiently. Taking the foundation model ZeroScope for example, it has over 1.8 billion pre-trained parameters. Each set of trainable spatial and temporal LoRAs only adds 9 million and 12 million parameters, respectively. Requiring 14 GB VRAM, MotionDirector takes 20 minutes to converge on multiple reference videos, and 8 minutes for a single reference video, competitive to the 10 minutes required by Tune-A-Video [70]. After training, MotionDirector can generate diverse videos with learned motions. In the inference stage, MotionDirector does not introduce significant increases in computational costs or inference time." + }, + { + "type": "title", + "bbox": [ + 0.219, + 0.331, + 0.545, + 0.347 + ], + "angle": 0, + "content": "5 Limitations and Future Works" + }, + { + "type": "text", + "bbox": [ + 0.219, + 0.363, + 0.787, + 0.454 + ], + "angle": 0, + "content": "Despite the MotionDirector can learn the motions of one or two subjects in the reference videos, it is still hard to learn complex motions of multiple subjects, such as a group of boys playing soccer. Previous appearance customization methods suffer similar problems when generating multiple customized subjects [14]. A possible solution is to further decouple the motions of different subjects in the latent space and learn them separately." + }, + { + "type": "title", + "bbox": [ + 0.219, + 0.477, + 0.357, + 0.493 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.219, + 0.508, + 0.787, + 0.675 + ], + "angle": 0, + "content": "We introduce and formulate the task of Motion Customization, which is adapting the pre-trained foundation text-to-video diffusion models to generate videos with desired motions. The challenge of this task is generalizing the customized motions to various appearances. To overcome this challenge, we propose the MotionDirector with a dual-path architecture and a novel appearance-debiased temporal training objective to decouple the learning of appearance and motion. Experimental results show that MotionDirector can learn motion concepts from a limited number of reference videos, and generalize them to diverse appearances. The automatic and human evaluations on two benchmarks demonstrate the MontionDirector outperforms other methods in terms of appearance diversity and motion fidelity." + }, + { + "type": "title", + "bbox": [ + 0.219, + 0.698, + 0.402, + 0.716 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.219, + 0.73, + 0.787, + 0.806 + ], + "angle": 0, + "content": "This research is supported by National Research Foundation, Singapore and A\\*STAR, under its RIE2020 Industry Alignment Fund - Industry Collaboration Projects (IAF-ICP) grant call (Grant No. I2001E0059) - SIA-NUS Digital Aviation Corp Lab. Mike Shou is supported by the Ministry of Education, Singapore, under the Academic Research Fund Tier 1 (FY2023)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.626, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "MotionDirector" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.324, + 0.161 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.179, + 0.786, + 0.222 + ], + "angle": 0, + "content": "1. Bain, M., Nagrani, A., Varol, G., Zisserman, A.: Frozen in time: A joint video and image encoder for end-to-end retrieval. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1728-1738 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.223, + 0.786, + 0.264 + ], + "angle": 0, + "content": "2. Balaji, Y., Min, M.R., Bai, B., Chellappa, R., Graf, H.P.: Conditional gan with discriminative filter generation for text-to-video synthesis. In: IJCAI. vol. 1, p. 2 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.265, + 0.786, + 0.305 + ], + "angle": 0, + "content": "3. Blattmann, A., Rombach, R., Ling, H., Dockhorn, T., Kim, S.W., Fidler, S., Kreis, K.: Align your latents: High-resolution video synthesis with latent diffusion models. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.307, + 0.786, + 0.362 + ], + "angle": 0, + "content": "4. Blattmann, A., Rombach, R., Ling, H., Dockhorn, T., Kim, S.W., Fidler, S., Kreis, K.: Align your latents: High-resolution video synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22563-22575 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.363, + 0.786, + 0.39 + ], + "angle": 0, + "content": "5. Chen, T.S., Lin, C.H., Tseng, H.Y., Lin, T.Y., Yang, M.H.: Motion-conditioned diffusion model for controllable video synthesis. arXiv:2304.14404 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.392, + 0.786, + 0.433 + ], + "angle": 0, + "content": "6. Chen, W., Wu, J., Xie, P., Wu, H., Li, J., Xia, X., Xiao, X., Lin, L.: Control-a-video: Controllable text-to-video generation with diffusion models. arXiv preprint arXiv:2305.13840 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.434, + 0.786, + 0.461 + ], + "angle": 0, + "content": "7. Chen, X., Huang, L., Liu, Y., Shen, Y., Zhao, D., Zhao, H.: Anydoor: Zero-shot object-level image customization. arXiv preprint arXiv:2307.09481 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.462, + 0.786, + 0.503 + ], + "angle": 0, + "content": "8. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 248-255. IEEE (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.504, + 0.786, + 0.532 + ], + "angle": 0, + "content": "9. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems 34, 8780-8794 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.533, + 0.786, + 0.573 + ], + "angle": 0, + "content": "0. Duan, Z., You, L., Wang, C., Chen, C., Wu, Z., Qian, W., Huang, J., Chao, F., Ji, R.: Diffsynth: Latent in-iteration deflickering for realistic video synthesis. arXiv:2308.03463 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.574, + 0.786, + 0.615 + ], + "angle": 0, + "content": "1. Esser, P., Chiu, J., Atighehchian, P., Granskog, J., Germanidis, A.: Structure and content-guided video synthesis with diffusion models. arXiv preprint arXiv:2302.03011 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.616, + 0.786, + 0.657 + ], + "angle": 0, + "content": "2. Ge, S., Hayes, T., Yang, H., Yin, X., Pang, G., Jacobs, D., Huang, J.B., Parikh, D.: Long video generation with time-agnostic vqgan and time-sensitive transformer. arXiv preprint arXiv:2204.03638 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.659, + 0.786, + 0.699 + ], + "angle": 0, + "content": "3. Ge, S., Nah, S., Liu, G., Poon, T., Tao, A., Catanzaro, B., Jacobs, D., Huang, J.B., Liu, M.Y., Balaji, Y.: Preserve your own correlation: A noise prior for video diffusion models. arXiv:2305.10474 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.701, + 0.786, + 0.755 + ], + "angle": 0, + "content": "4. Gu, Y., Wang, X., Wu, J.Z., Shi, Y., Chen, Y., Fan, Z., Xiao, W., Zhao, R., Chang, S., Wu, W., et al.: Mix-of-show: Decentralized low-rank adaptation for multi-concept customization of diffusion models. arXiv preprint arXiv:2305.18292 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.757, + 0.786, + 0.798 + ], + "angle": 0, + "content": "5. Guo, Y., Yang, C., Rao, A., Wang, Y., Qiao, Y., Lin, D., Dai, B.: Animatediff: Animate your personalized text-to-image diffusion models without specific tuning. arXiv preprint arXiv:2307.04725 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.799, + 0.786, + 0.84 + ], + "angle": 0, + "content": "6. He, Y., Xia, M., Chen, H., Cun, X., Gong, Y., Xing, J., Zhang, Y., Wang, X., Weng, C., Shan, Y., et al.: animate-a-story: Storytelling with retrieval-augmented video generation. arXiv:2307.06940 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.179, + 0.786, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.346, + 0.127 + ], + "angle": 0, + "content": "Zhao et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "17. He, Y., Yang, T., Zhang, Y., Shan, Y., Chen, Q.: Latent video diffusion models for high-fidelity long video generation (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.176, + 0.785, + 0.203 + ], + "angle": 0, + "content": "18. He, Y., Yang, T., Zhang, Y., Shan, Y., Chen, Q.: Latent video diffusion models for high-fidelity video generation with arbitrary lengths. arXiv:2211.13221 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.204, + 0.785, + 0.245 + ], + "angle": 0, + "content": "19. He, Y., Yang, T., Zhang, Y., Shan, Y., Chen, Q.: Videocrafter: A toolkit for text-to-video generation and editing. https://github.com/AILab-CVC/VideoCrafter (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.246, + 0.785, + 0.273 + ], + "angle": 0, + "content": "20. Hessel, J., Holtzman, A., Forbes, M., Bras, R.L., Choi, Y.: Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.785, + 0.314 + ], + "angle": 0, + "content": "21. Ho, J., Chan, W., Sahara, C., Whang, J., Gao, R., Gritsenko, A., Kingma, D.P., Poole, B., Norouzi, M., Fleet, D.J., et al.: Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.785, + 0.356 + ], + "angle": 0, + "content": "22. Ho, J., Chan, W., Sahara, C., Whang, J., Gao, R., Gritsenko, A., Kingma, D.P., Poole, B., Norouzi, M., Fleet, D.J., et al.: Imagen video: High definition video generation with diffusion models. arXiv:2210.02303 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.356, + 0.785, + 0.383 + ], + "angle": 0, + "content": "23. Ho, J., Salimans, T., Gritsenko, A., Chan, W., Norouzi, M., Fleet, D.J.: Video diffusion models. In: NeurIPS (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.785, + 0.411 + ], + "angle": 0, + "content": "24. Hong, S., Seo, J., Hong, S., Shin, H., Kim, S.: Large language models are frame-level directors for zero-shot text-to-video generation. arXiv:2305.14330 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.412, + 0.785, + 0.451 + ], + "angle": 0, + "content": "25. Hong, W., Ding, M., Zheng, W., Liu, X., Tang, J.: Cogvideo: Large-scale pretraining for text-to-video generation via transformers. arXiv preprint arXiv:2205.15868 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.452, + 0.785, + 0.493 + ], + "angle": 0, + "content": "26. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.494, + 0.785, + 0.522 + ], + "angle": 0, + "content": "27. Huang, H., Feng, Y., Shi, C., Xu, L., Yu, J., Yang, S.: Free-bloom: Zero-shot text-to-video generator with lvm director and ldm animator. In: NeurIPS (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.523, + 0.785, + 0.563 + ], + "angle": 0, + "content": "28. Jeong, H., Park, G.Y., Ye, J.C.: Vmc: Video motion customization using temporal attention adaption for text-to-video diffusion models. arXiv preprint arXiv:2312.00845 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.564, + 0.785, + 0.591 + ], + "angle": 0, + "content": "29. Karras, J., Holynski, A., Wang, T.C., Kemelmacher-Shlizerman, I.: Dreampose: Fashion image-to-video synthesis via stable diffusion. arXiv:2304.06025 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.592, + 0.785, + 0.632 + ], + "angle": 0, + "content": "30. Khachatryan, L., Movsisyan, A., Tadevosyan, V., Henschel, R., Wang, Z., Navasardyan, S., Shi, H.: Text2video-zero: Text-to-image diffusion models are zero-shot video generators. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.633, + 0.785, + 0.673 + ], + "angle": 0, + "content": "31. Kirstain, Y., Polyak, A., Singer, U., Matiana, S., Penna, J., Levy, O.: Pick-a-pic: An open dataset of user preferences for text-to-image generation. arXiv preprint arXiv:2305.01569 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.674, + 0.785, + 0.715 + ], + "angle": 0, + "content": "32. Kumari, N., Zhang, B., Zhang, R., Shechtman, E., Zhu, J.Y.: Multi-concept customization of text-to-image diffusion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1931-1941 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.716, + 0.785, + 0.743 + ], + "angle": 0, + "content": "33. Le Moing, G., Ponce, J., Schmid, C.: Ccvs: context-aware controllable video synthesis. Advances in Neural Information Processing Systems 34, 14042-14055 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.744, + 0.785, + 0.785 + ], + "angle": 0, + "content": "34. Li, X., Chu, W., Wu, Y., Yuan, W., Liu, F., Zhang, Q., Li, F., Feng, H., Ding, E., Wang, J.: Videogen: A reference-guided latent diffusion approach for high definition text-to-video generation. arXiv preprint arXiv:2309.00398 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.785, + 0.785, + 0.812 + ], + "angle": 0, + "content": "35. Lian, L., Shi, B., Yala, A., Darrell, T., Li, B.: Llm-grounded video diffusion models. arXiv:2309.17444 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "36. Luo, Z., Chen, D., Zhang, Y., Huang, Y., Wang, L., Shen, Y., Zhao, D., Zhou, J., Tan, T.: Videofusion: Decomposed diffusion models for high-quality video gen" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.627, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "MotionDirector" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.241, + 0.147, + 0.786, + 0.175 + ], + "angle": 0, + "content": "eration. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.176, + 0.787, + 0.215 + ], + "angle": 0, + "content": "37. Ma, Y., He, Y., Cun, X., Wang, X., Shan, Y., Li, X., Chen, Q.: Follow your pose: Pose-guided text-to-video generation using pose-free videos. arXiv preprint arXiv:2304.01186 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.216, + 0.786, + 0.257 + ], + "angle": 0, + "content": "38. Ma, Y., He, Y., Cun, X., Wang, X., Shan, Y., Li, X., Chen, Q.: Follow your pose: Pose-guided text-to-video generation using pose-free videos. arXiv:2304.01186 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.257, + 0.787, + 0.298 + ], + "angle": 0, + "content": "39. Materzynska, J., Sivic, J., Shechtman, E., Torralba, A., Zhang, R., Russell, B.: Customizing motion in text-to-video diffusion models. arXiv preprint arXiv:2312.04966 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.298, + 0.787, + 0.325 + ], + "angle": 0, + "content": "40. Mei, K., Patel, V.: Vidm: Video implicit diffusion models. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 9117-9125 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.325, + 0.787, + 0.365 + ], + "angle": 0, + "content": "41. Ni, H., Shi, C., Li, K., Huang, S.X., Min, M.R.: Conditional image-to-video generation with latent flow diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18444-18455 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.365, + 0.787, + 0.405 + ], + "angle": 0, + "content": "42. Qin, B., Ye, W., Yu, Q., Tang, S., Zhuang, Y.: Dancing avatar: Pose and text-guided human motion videos synthesis with image diffusion model. arXiv:2308.07749 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.405, + 0.787, + 0.446 + ], + "angle": 0, + "content": "43. Ren, Y., Zhou, Y., Yang, J., Shi, J., Liu, D., Liu, F., Kwon, M., Shrivastava, A.: Customize-a-video: One-shot motion customization of text-to-video diffusion models. arXiv preprint arXiv:2402.14780 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.446, + 0.787, + 0.473 + ], + "angle": 0, + "content": "44. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: MICCAI (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.473, + 0.787, + 0.528 + ], + "angle": 0, + "content": "45. Ruiz, N., Li, Y., Jampani, V., Pritch, Y., Rubinstein, M., Aberman, K.: Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22500-22510 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.528, + 0.787, + 0.555 + ], + "angle": 0, + "content": "46. Ryu, S.: Low-rank adaptation for fast text-to-image diffusion fine-tuning. https://github.com/cloneofsimo/lora (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.555, + 0.787, + 0.595 + ], + "angle": 0, + "content": "47. Saito, M., Matsumoto, E., Saito, S.: Temporal generative adversarial nets with singular value clipping. In: Proceedings of the IEEE international conference on computer vision. pp. 2830-2839 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.595, + 0.787, + 0.65 + ], + "angle": 0, + "content": "48. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems 35, 25278-25294 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.65, + 0.787, + 0.691 + ], + "angle": 0, + "content": "49. Shen, X., Li, X., Elhoseiny, M.: Mostgan-v: Video generation with temporal motion styles. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5652-5661 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.691, + 0.787, + 0.732 + ], + "angle": 0, + "content": "50. Singer, U., Polyak, A., Hayes, T., Yin, X., An, J., Zhang, S., Hu, Q., Yang, H., Ashual, O., Gafni, O., et al.: Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.732, + 0.787, + 0.772 + ], + "angle": 0, + "content": "51. Skorokhodov, I., Tulyakov, S., Elhoseiny, M.: Stylegan-v: A continuous video generator with the price, image quality and perks of stylegan2. arXiv preprint arXiv:2112.14683 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.772, + 0.787, + 0.813 + ], + "angle": 0, + "content": "52. Smith, J.S., Hsu, Y.C., Zhang, L., Hua, T., Kira, Z., Shen, Y., Jin, H.: Continual diffusion: Continual customization of text-to-image diffusion with c-lora. arXiv preprint arXiv:2304.06027 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.787, + 0.841 + ], + "angle": 0, + "content": "53. Soomro, K., Zamir, A.R.: Action recognition in realistic sports videos. In: Computer vision in sports, pp. 181-208. Springer (2015)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.346, + 0.127 + ], + "angle": 0, + "content": "Zhao et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.189 + ], + "angle": 0, + "content": "54. Srivastava, N., Mansimov, E., Salakhudinov, R.: Unsupervised learning of video representations using lstms. In: International conference on machine learning. pp. 843-852. PMLR (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.785, + 0.218 + ], + "angle": 0, + "content": "55. Sterling, S.: Zeroscope. https://huggingface.co/cerspense/zeroscope_v2_576w (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.218, + 0.785, + 0.259 + ], + "angle": 0, + "content": "56. Tian, Y., Ren, J., Chai, M., Olszewski, K., Peng, X., Metaxas, D.N., Tulyakov, S.: A good image generator is what you need for high-resolution video synthesis. In: International Conference on Learning Representations (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.259, + 0.785, + 0.3 + ], + "angle": 0, + "content": "57. Tulyakov, S., Liu, M.Y., Yang, X., Kautz, J.: Mocogan: Decomposing motion and content for video generation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1526-1535 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.3, + 0.785, + 0.342 + ], + "angle": 0, + "content": "58. Voleti, V., Jolicoeur-Martineau, A., Pal, C.: Mcvd-masked conditional video diffusion for prediction, generation, and interpolation. Advances in Neural Information Processing Systems 35, 23371-23385 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.342, + 0.785, + 0.37 + ], + "angle": 0, + "content": "59. Vondrick, C., Pirsiavash, H., Torralba, A.: Generating videos with scene dynamics. Advances in neural information processing systems 29 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.37, + 0.785, + 0.398 + ], + "angle": 0, + "content": "60. Wang, J., Yuan, H., Chen, D., Zhang, Y., Wang, X., Zhang, S.: Modelscope text-to-video technical report. arXiv preprint arXiv:2308.06571 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.785, + 0.438 + ], + "angle": 0, + "content": "61. Wang, T., Li, L., Lin, K., Lin, C.C., Yang, Z., Zhang, H., Liu, Z., Wang, L.: Disco: Disentangled control for referring human dance generation in real world. arXiv:2307.00040 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.785, + 0.481 + ], + "angle": 0, + "content": "62. Wang, W., Yang, H., Tuo, Z., He, H., Zhu, J., Fu, J., Liu, J.: Videofactory: Swap attention in spatiotemporal diffusions for text-to-video generation. arXiv:2305.10874 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.481, + 0.785, + 0.522 + ], + "angle": 0, + "content": "63. Wang, X., Yuan, H., Zhang, S., Chen, D., Wang, J., Zhang, Y., Shen, Y., Zhao, D., Zhou, J.: Videocomposer: Compositional video synthesis with motion controllability. arXiv preprint arXiv:2306.02018 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.522, + 0.785, + 0.589 + ], + "angle": 0, + "content": "64. Wang, Y., Chen, X., Ma, X., Zhou, S., Huang, Z., Wang, Y., Yang, C., He, Y., Yu, J., der Yang, P., Guo, Y., Wu, T., Si, C., Jiang, Y., Chen, C., Loy, C.C., Dai, B., Lin, D., Qiao, Y., Liu, Z.: Lavie: High-quality video generation with cascaded latent diffusion models (2023), https://api-semanticscholar.org/CorpusID: 262823915" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.591, + 0.785, + 0.632 + ], + "angle": 0, + "content": "65. Wang, Y., Chen, X., Ma, X., Zhou, S., Huang, Z., Wang, Y., Yang, C., He, Y., Yu, J., Yang, P., et al.: Lavie: High-quality video generation with cascaded latent diffusion models. arXiv:2309.15103 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.632, + 0.785, + 0.674 + ], + "angle": 0, + "content": "66. Wang, Z., Yuan, Z., Wang, X., Chen, T., Xia, M., Luo, P., Shan, Y.: Motionctrl: A unified and flexible motion controller for video generation. arXiv preprint arXiv:2312.03641 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.674, + 0.785, + 0.715 + ], + "angle": 0, + "content": "67. Wei, Y., Zhang, S., Qing, Z., Yuan, H., Liu, Z., Liu, Y., Zhang, Y., Zhou, J., Shan, H.: Dreamvideo: Composing your dream videos with customized subject and motion. arXiv preprint arXiv:2312.04433 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.715, + 0.785, + 0.758 + ], + "angle": 0, + "content": "68. Wei, Y., Zhang, Y., Ji, Z., Bai, J., Zhang, L., Zuo, W.: Elite: Encoding visual concepts into textual embeddings for customized text-to-image generation. arXiv preprint arXiv:2302.13848 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.785, + 0.799 + ], + "angle": 0, + "content": "69. Wu, J.Z., Gao, D., Bai, J., Shou, M., Li, X., Dong, Z., Singh, A., Keutzer, K., Iandola, F.: The text-guided video editing benchmark at loveu 2023. https:// sites.google.com/view/loveucvpr23/track4 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.841 + ], + "angle": 0, + "content": "70. Wu, J.Z., Ge, Y., Wang, X., Lei, W., Gu, Y., Hsu, W., Shan, Y., Qie, X., Shou, M.Z.: Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. arXiv preprint arXiv:2212.11565 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.627, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "MotionDirector" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "71. Wu, J.Z., Ge, Y., Wang, X., Lei, W., Gu, Y., Hsu, W., Shan, Y., Qie, X., Shou, M.Z.: Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.787, + 0.218 + ], + "angle": 0, + "content": "72. Wu, R., Chen, L., Yang, T., Guo, C., Li, C., Zhang, X.: Lamp: Learn a motion pattern for few-shot-based video generation. arXiv preprint arXiv:2310.10769 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.218, + 0.787, + 0.259 + ], + "angle": 0, + "content": "73. Xing, J., Xia, M., Liu, Y., Zhang, Y., Zhang, Y., He, Y., Liu, H., Chen, H., Cun, X., Wang, X., et al.: Make-your-video: Customized video generation using textual and structural guidance. arXiv:2306.00943 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.259, + 0.787, + 0.286 + ], + "angle": 0, + "content": "74. Xing, Z., Feng, Q., Chen, H., Dai, Q., Hu, H., Xu, H., Wu, Z., Jiang, Y.G.: A survey on video diffusion models. arXiv preprint arXiv:2310.10647 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.286, + 0.787, + 0.314 + ], + "angle": 0, + "content": "75. Yan, W., Zhang, Y., Abbeel, P., Srinivas, A.: Videogpt: Video generation using vq-vae and transformers. arXiv preprint arXiv:2104.10157 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.314, + 0.787, + 0.357 + ], + "angle": 0, + "content": "76. Yang, S., Hou, L., Huang, H., Ma, C., Wan, P., Zhang, D., Chen, X., Liao, J.: Direct-a-video: Customized video generation with user-directed camera movement and object motion. arXiv preprint arXiv:2402.03162 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.356, + 0.787, + 0.397 + ], + "angle": 0, + "content": "77. Yin, S., Wu, C., Liang, J., Shi, J., Li, H., Ming, G., Duan, N.: Dragnuwa: Fine-grained control in video generation by integrating text, image, and trajectory. arXiv:2308.08089 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.397, + 0.787, + 0.439 + ], + "angle": 0, + "content": "78. Yin, S., Wu, C., Yang, H., Wang, J., Wang, X., Ni, M., Yang, Z., Li, L., Liu, S., Yang, F., et al.: Nuwa-xl: Diffusion over diffusion for extremely long video generation. arXiv:2303.12346 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.439, + 0.787, + 0.481 + ], + "angle": 0, + "content": "79. Yu, S., Sohn, K., Kim, S., Shin, J.: Video probabilistic diffusion models in projected latent space. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18456-18466 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.481, + 0.787, + 0.522 + ], + "angle": 0, + "content": "80. Yu, S., Tack, J., Mo, S., Kim, H., Kim, J., Ha, J.W., Shin, J.: Generating videos with dynamics-aware implicit generative adversarial networks. In: International Conference on Learning Representations (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.522, + 0.787, + 0.563 + ], + "angle": 0, + "content": "81. Zhang, D.J., Wu, J.Z., Liu, J.W., Zhao, R., Ran, L., Gu, Y., Gao, D., Shou, M.Z.: Show-1: Marrying pixel and latent diffusion models for text-to-video generation. arXiv:2309.15818 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.563, + 0.787, + 0.605 + ], + "angle": 0, + "content": "82. Zhang, D.J., Wu, J.Z., Liu, J.W., Zhao, R., Ran, L., Gu, Y., Gao, D., Shou, M.Z.: Show-1: Marrying pixel and latent diffusion models for text-to-video generation (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.605, + 0.787, + 0.633 + ], + "angle": 0, + "content": "83. Zhang, L., Agrawala, M.: Adding conditional control to text-to-image diffusion models. arXiv preprint arXiv:2302.05543 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.633, + 0.787, + 0.675 + ], + "angle": 0, + "content": "84. Zhao, R., Gu, Y., Wu, J.Z., Zhang, D.J., Liu, J., Wu, W., Keppo, J., Shou, M.Z.: Motiondirector: Motion customization of text-to-video diffusion models. arXiv preprint arXiv:2310.08465 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.674, + 0.787, + 0.715 + ], + "angle": 0, + "content": "85. Zhao, S., Chen, D., Chen, Y.C., Bao, J., Hao, S., Yuan, L., Wong, K.Y.K.: Uncontrolnet: All-in-one control to text-to-image diffusion models. arXiv preprint arXiv:2305.16322 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.715, + 0.787, + 0.744 + ], + "angle": 0, + "content": "86. Zhou, D., Wang, W., Yan, H., Lv, W., Zhu, Y., Feng, J.: Magicvideo: Efficient video generation with latent diffusion models. arXiv:2211.11018 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.744 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/MotionDirector_ Motion Customization of Text-to-Video Diffusion Models/8325ffaa-f0c0-4f28-929f-63aa2f3d7c76_origin.pdf b/2024/MotionDirector_ Motion Customization of Text-to-Video Diffusion Models/8325ffaa-f0c0-4f28-929f-63aa2f3d7c76_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..4bc021e21a2d4f8def82d2288f9cd50c8372dade --- /dev/null +++ b/2024/MotionDirector_ Motion Customization of Text-to-Video Diffusion Models/8325ffaa-f0c0-4f28-929f-63aa2f3d7c76_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:afc0b6e893896c5268f9702fabfae79f9f2dc07862ff7aad8ff7e7e3628da390 +size 3004549 diff --git a/2024/MotionDirector_ Motion Customization of Text-to-Video Diffusion Models/full.md b/2024/MotionDirector_ Motion Customization of Text-to-Video Diffusion Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..57250233c0ecca4cd1c17e054ca6d4d1dcd91151 --- /dev/null +++ b/2024/MotionDirector_ Motion Customization of Text-to-Video Diffusion Models/full.md @@ -0,0 +1,299 @@ +# MotionDirector: Motion Customization of Text-to-Video Diffusion Models + +Rui Zhao $^{1}$ , Yuchao Gu $^{1}$ , Jay Zhangjie Wu $^{1}$ , David Junhao Zhang $^{1}$ , Jia-Wei Liu $^{1}$ , Weijia Wu $^{1}$ , Jussi Keppo $^{2}$ , and Mike Zheng Shou $^{1}$ + +$^{1}$ Show Lab, $^{2}$ National University of Singapore, Singapore + +Abstract. Large-scale pre-trained diffusion models have exhibited remarkable capabilities in diverse video generations. Given a set of video clips of the same motion concept, the task of Motion Customization is to adapt existing text-to-video diffusion models to generate videos with this motion. Adaptation methods have been developed for customizing appearance like subject or style, yet under-explored for motion. It is straightforward to extend mainstream adaption methods for motion customization, including full model tuning and Low-Rank Adaptions (LoRAs). However, the motion concept learned by these methods is often coupled with the limited appearances in the training videos, making it difficult to generalize the customized motion to other appearances. To overcome this challenge, we propose MotionDirector, with a dual-path LoRAs architecture to decouple the learning of appearance and motion. Further, we design a novel appearance-debiased temporal loss to mitigate the influence of appearance on the temporal training objective. Experimental results show the proposed method can generate videos of diverse appearances for the customized motions. Our method also supports various downstream applications, such as the mixing of different videos with their appearance and motion respectively, and animating a single image with customized motions. The project website is at: MotionDirector. + +Keywords: motion customization $\cdot$ text-to-video generation $\cdot$ diffusion models + +# 1 Introduction + +Text-to-video diffusion models [17,21,50] are approaching generating high-quality diverse videos given text instructions. The open-sourcing of foundational text-to-video models [55,60] pre-trained on large-scale data has sparked enthusiasm for video generation in both the community and academia. Users can create videos that are either realistic or imaginative simply by providing text prompts. While foundation models generate diverse videos from the same text, adapting them to generate more specific content can better accommodate the preferences of users. Similar to the customization of text-to-image foundation models [45], tuning video foundation models to generate videos of a certain concept of appearance, + +![](images/3582cef6c28440356b1b66594b0965f76d24ef0669ceb60bef4b3bee14c7130f.jpg) +Fig. 1: Motion customization of the text-to-video diffusion. + +like subject or style, has also been explored [17]. Compared with images, videos consist of not just appearances but also motion dynamics. Users may desire to create videos with specific motions, such as a car moving forward and then turning left under a predefined camera perspective, as illustrated on the right side in Fig. 1. However, customizing the motions in the text-to-video generation is still under-explored. + +The task of Motion Customization is formulated as follows: given a set of reference videos representing a specific motion concept, the objective is to customize pre-trained foundational models to generate videos that accurately capture and replicate the depicted motion. The customized motion concept can be challenging motions that are hard for the pre-trained model to generate, novel motions not present in the pre-training dataset, and specific motions that are difficult to express in text. Examples of these are respectively showcased in the first, second, and third columns of Fig 1. In contrast, previous works on appearance customization adapt the foundation models to generate samples with desired appearance, like subject or style, given reference videos or images representing such appearance [17,45]. It is straightforward to use previous adaption methods for motion customization. For example, on the given reference videos, fine-tuning the weights of foundation models [45], parameter-efficient tuning additional layers [70], or training Low-Rank Adaptions (LoRAs) [26] injected in the layers of foundation models. However, customizing diffusion models to generate desired motions without harming their appearance diversity is challenging because the motion and appearance are coupled with each other at the step-by-step denoising stage. Directly deploying previous adaption methods to learn motions makes the models fit the limited appearances seen in the reference videos, posing challenges in generalizing the learned motions to various appearances. Recent works on controllable text-to-video generations [11, 17, 63] generate videos controlled by signals representing pre-defined motions. However, the control signals, such as depth maps or edges, impose constraints on the shapes of subjects and back + +![](images/0b4a830b7c8818495a38fd991e4e79d94464b1abd952a55177059708a12cb370.jpg) +Fig. 2: (Row 1) Take two videos to train the proposed MotionDirector, respectively. (Row 2) MotionDirector can generalize the learned motions to diverse appearances. (Row 3) MotionDirector can mix the learned motion and appearance from different videos to generate new videos. (Row 4) MotionDirector can animate a single image with learned motions. + +grounds, thus influencing the appearance of generated videos in a coupled way. Besides, these methods accept only one sequence of control signals to generate one video, which may not be suitable for users seeking certain motion types without strict spatial constraints, such as the example of lifting weights in Fig. 1. + +To achieve motion customization of text-to-video diffusion models while preserving appearance diversity, we propose the MotionDirector, which tunes the foundation models to learn the appearance and motions in the given single or multiple reference videos in a decoupled way. MotionDirector tunes the models with low-rank adaptions (LoRAs) while keeping their pre-trained parameters fixed to retain the learned generation knowledge. Specifically, the MotionDirector employs a dual-path architecture, as shown in Fig. 3. For each video, a spatial path consists of a foundation model with trainable spatial LoRAs injected into its spatial transformer layers. These spatial LoRAs are trained on a single frame randomly sampled per training step to capture the appearance characteristics of the input videos. The temporal path, on the other hand, is a replica of the foundation model that shares the spatial LoRAs with the spatial path to fit the appearance of the corresponding input video. Additionally, the temporal transformers in this path are equipped with temporal LoRAs, which are trained on multiple frames of input videos to capture the underlying motion patterns. To further enhance the learning of motions, we propose an appearance-debiased temporal loss to mitigate the influence of appearance on the temporal training objective. + +Only deploying the trained temporal LoRAs enables the foundation model to generate videos of the learned motions with diverse appearances, as shown in the second row of Fig 2. The decoupled paradigm further makes an interesting + +kind of video generation feasible, which is the mix of the appearance from one video with the motion from another video, called the mix of videos, as shown in the third row of Fig 2. The key to this success lies in that MotionDirector can decouple the appearance and motion of videos and then combine them from various source videos. It is achieved by injecting spatial LoRAs trained on one video and temporal LoRAs trained on another video into the foundation model. Besides, the learned motions can be deployed to animate images, as images can be treated as appearance providers, as shown in the last row of Fig 2. + +We conducted experiments on two benchmarks with 86 different motions and over 600 text prompts to test proposed methods, baselines, and comparison methods. The results show our method can be applied to different diffusion-based foundation models and achieve motion customization of various motion concepts. Compared with other methods, our method avoids fitting the limited appearance of reference videos, and can generalize the learned motions to diverse appearances. + +Our contributions are summarized as follows: + +- We introduce and define the task of Motion Customization. The challenges lie in generalizing the customized motions to various appearances. +- We propose the MotionDirector with a dual-path architecture and a novel appearance-debiased temporal training objective, to decouple the learning of appearance and motion. +- Experiments on two benchmarks demonstrate that MotionDirector can customize various base models to generate diverse videos with desired motion concepts, and outperforms controllable generation methods and tuning-based methods. + +# 2 Related Work + +# 2.1 Text-to-Video Generation. + +To achieve high-quality video generation, various methods have been developed, such as Generative Adversarial Networks (GANs) [2, 47, 49, 56, 57, 59], autoregressive models [12, 25, 33, 54, 75] and implicit neural representations [51, 80]. Diffusion-based models [3, 10, 13, 18, 22-24, 27, 34, 35, 40, 41, 50, 58, 60, 62, 65, 71, 78, 79, 81, 84, 86] are also approaching high-quality generation by training conditional 3D U-Nets [44] to denoise from randomly sampled sequences of Gaussian noises [74]. Recent foundation models [4, 17, 21, 36, 50, 64, 82] are pre-trained on large-scale image and video datasets [1, 8, 48], to learn powerful generation ability. Some works turn text-to-image foundation models to text-to-video generation by manipulation on cross-frame attention or training additional temporal layers, like Tune-A-Video [71], Text2Video-Zero [30], and AnimiteDiff [15]. The recently open-sourced foundation models [55, 60] have ignited enthusiasm among users to generate realistic or imaginative videos, and further make it possible for users to customize and build their own private models. + +# 2.2 Generation Model Customization. + +Customizing the pre-trained large foundation models can fit the preferences of users better while maintaining powerful generation knowledge without training from scratch. Previous customization methods for text-to-image diffusion models [7,14,32,45,52,68] aim to generate certain subjects or styles, given a set of example images. Dreambooth [45] or LoRA [26] can be simply applied to customizing video foundation models to generate videos with certain subjects or styles, given a set of reference video clips or images. The recently proposed VideoCrafter [19] has explored this, which we categorize as appearance customization. In addition to appearances, videos are also characterized by the motion dynamics of subjects and camera movements across frames. After we proposed the task of motion customization, some concurrent works [28, 39, 43, 66, 67, 72, 76] emerged in this field. DreamVideo [67] trains several adapters in U-Net to customize both the subject and motion in video generation. Lamp [72] learns a motion pattern from a set of videos to utilize image diffusion models to generate videos. + +# 2.3 Controllable Video Generation. + +Controllable generation aims to ensure the generation results align with the given explicit control signals, such as depth maps, human pose, optical flows, etc. [5, 16, 29, 37, 38, 42, 61, 73, 77, 83, 85]. For the controllable text-to-video generation methods, i.e. the VideoCrafter [17], VideoComposer [63], Control-A-Video [6], they train additional branches that take condition signals to align the generated videos with them. Unlike the human poses for specifically controlling the generation of human bodies, the general control singles, such as depth maps, are typically extracted from reference videos and are coupled with both appearance and motion. This results in the generation results being influenced by both the appearance and motion in reference videos. Applying these methods directly in motion customization is challenging when it comes to generalizing the desired motions to diverse appearances. + +# 3 Methodology + +# 3.1 Preliminaries + +Video Diffusion Model. Video diffusion models train a 3D U-Net to denoise from a randomly sampled sequence of Gaussian noises to generate videos, guided by text prompts. The 3D U-net basically consists of down-sample, middle, and up-sample blocks. Each block has several convolution layers, spatial transformers, and temporal transformers as shown in Fig 3. The 3D U-Net $\epsilon_{\theta}$ and a text encoder $\tau_{\theta}$ are jointly optimized by the noise-prediction loss, as detailed in [9]: + +$$ +\mathcal {L} = \mathbb {E} _ {z _ {0}, y, \epsilon \sim \mathcal {N} (0, I), t \sim \mathcal {U} (0, T)} \left[ \| \epsilon - \epsilon_ {\theta} \left(z _ {t}, t, \tau_ {\theta} (y)\right) \| _ {2} ^ {2} \right], \tag {1} +$$ + +![](images/73f3d477644768419b0a8c81f48f3ad3fb9d92540b2eeecc7f57ad3c9ce486fe.jpg) +Fig. 3: The dual-path architecture of the proposed method. All pre-trained weights of the base diffusion model remain fixed. In the spatial path, the spatial transformers are injected with trainable spatial LoRAs as shown on the right side. In the temporal path, the spatial transformers are injected with spatial LoRAs sharing weights with those ones in the spatial path, and the temporal transformers are injected with trainable temporal LoRAs. + +where $z_0$ is the latent code of the training videos, $y$ is the text prompt, $\epsilon$ is the Gaussian noise added to the latent code, and $t$ is the time step. As discussed in [9], the noised latent code $z_t$ is determined as: + +$$ +z _ {t} = \sqrt {\bar {\alpha} _ {t}} z _ {0} + \sqrt {1 - \bar {\alpha} _ {t}} \epsilon , \bar {\alpha} _ {t} = \prod_ {i = 1} ^ {t} \alpha_ {t}, \tag {2} +$$ + +where $\alpha_{t}$ is a hyper-parameter controlling the noise strength. + +Low-Rank Adaption. Low-rank adaption (LoRA) [26] was proposed to adapt the pre-trained large language models to downstream tasks. Recently it has been applied in text-to-image generation and text-to-video generation tasks to achieve appearance customization [19, 46]. LoRA employs a low-rank factorization technique to update the weight matrix $W$ as + +$$ +W = W _ {0} + \Delta W = W _ {0} + B A, \tag {3} +$$ + +where $W_0 \in \mathbb{R}^{d \times k}$ represents the original weights of the pre-trained model, $B \in \mathbb{R}^{d \times r}$ and $A \in \mathbb{R}^{r \times k}$ represent the low-rank factors, where $r$ is much smaller than original dimensions $d$ and $k$ . LoRA requires smaller computing sources than fine-tuning the weights of the entire network like DreamBooth [45], and it is convenient to spread and deploy as a plug-and-play plugin for pre-trained models. + +In the following sections, we introduce the proposed MotionDirector, containing the dual-path Low-Rank Adaptions (in Sec. 3.2) for learning the appearance and motion in reference videos in a decoupled manner, and appearance-debiased temporal loss (in Sec. 3.3)) to further mitigate the influence of appearance on the temporal training objective. + +![](images/bf379e347ad1d2416a527089af9d55c14e67a0e0b6f113aaef00e343a5ff23d3.jpg) +Fig. 4: Comparison of videos generated by the base model, the model fine-tuned in the coupled manner, the model fine-tuned without the proposed Appearance-Debiased Temporal Loss (AD-Loss), and our full-version MotionDirector method. + +# 3.2 Dual-Path Low-Rank Adaptions + +At each time-step $t$ , the 3D U-Net takes in the latent code $z_{t} \in \mathbb{R}^{b \times f \times w \times h \times c}$ and the conditional input $y$ (e.g., text), where $b, f, w, h, c$ represents the size of the batch, frame, width, height, and channel dimensions, respectively. The spatial transformers apply spatial self-attention along the spatial dimensions $w, h$ to improve the correlation between pixels, and then leverage the cross-attention between the latent code and the conditional input $y$ to improve textual alignment. The temporal transformers apply temporal self-attention along the frame dimension $f$ to improve the temporal consistency between frames. However, spatial and temporal information in the latent code gradually become coupled with each other during the step-by-step denoising stage. Attempting to directly learn and fit the motions in reference videos will inevitably lead to fitting their limited appearances. As shown in the third row "Coupled Tuned" of Fig. 4. From the results on the left side, we can observe that the generated "panda" is coupled with the appearance of "human" in the reference videos and the generated background is also coupled with that of reference videos. Besides, the results on the right side show that the coupled tuned model can not generate the "Eiffel Tower" because it is overfitted on the limited backgrounds in the reference videos. + +To address this problem, we propose to tune the spatial and temporal transformers in a dual-path way to learn the earn the appearance and motion in reference videos, respectively, as shown in Fig. 3. Specifically, for the spatial path, we inject LoRAs into spatial transformers to learn the appearance of training data, and for the temporal path, we inject LoRAs into temporal transformers to learn the motion in videos. The training of spatial and temporal LoRAs are as follows. + +Spatial Training. For the spatial path, we inject unique spatial LoRAs into the spatial transformers for each training video while keeping the weights of pre-trained 3D U-Net fixed. To maintain the learned strong and diverse textual + +alignment ability, we do not inject LoRAs into cross-attention layers of spatial transformers, since their weights influence the correlations between the pixels and text prompts. On the other hand, we inject LoRAs into spatial self-attention layers and feed-forward layers to update the correlations in spatial dimensions to enable the model to reconstruct the appearance of training data. For each training step, the spatial LoRAs are trained on a single frame randomly sampled from the training video to fit its appearance while ignoring its motion, based on spatial loss, which is reformulated as + +$$ +\mathcal {L} _ {\text {s p a t i a l}} = \mathbb {E} _ {z _ {0}, y, \epsilon , t, i \sim \mathcal {U} (0, F)} \left[ \| \epsilon - \epsilon_ {\theta} \left(z _ {t, i}, t, \tau_ {\theta} (y)\right) \| _ {2} ^ {2} \right], \tag {4} +$$ + +where $\mathbf{F}$ is the number of frames of the training data and the $z_{t,i}$ is the sampled frame from the latent code $z_{t}$ . + +Temporal Training. For the temporal path, we inject the temporal LoRAs into self-attention and feed-forward layers of temporal transformers to update the correlations along the frame dimension. Besides, the spatial transformers are injected with LoRAs sharing the same weights learned from the spatial path, to force the trainable temporal LoRAs to ignore the appearance of the training data. The temporal LoRAs could be simply trained on all frames of training data based on the temporal loss $\mathcal{L}_{org-temp}$ , formulated in the same way as equation (1). + +# 3.3 Appearance-Debiased Temporal Loss + +The videos generated by models trained with original temporal loss are shown in the fourth row of Fig. 4. We can observe that the motions are not learned well. To explore the underlying reasons, we visualized the denoising process of several videos. + +As illustrated in Fig. 5, when considering the latent codes of each frame $z_{t,i_{i = 1}^{F}}$ as a set of points in the latent space, we found that motion primarily impacts the underlying dependencies between these point sets, whereas the distances between different sets of points are more influenced by appearance. The original temporal training objective does not take this into account. The learning of motion dynamics, embedded within the internal connectivity structure of latent codes, remains biased by positional differences of latent codes. These biases are primarily influenced by the appearance of training samples and pose a challenge to learning motions efficiently. + +To further decouple the motion from appearance, we proposed to eliminate the appearance bias among the noises and predicted noises, and calculate the appearance-debiased temporal loss on them. The debiasing of each noise $\epsilon_{i} \in \{\epsilon_{i}\}_{i=1}^{F}$ is as follows, + +$$ +\phi (\epsilon_ {i}) = \sqrt {\beta^ {2} + 1} \epsilon_ {i} - \beta \epsilon_ {a n c h o r}, \tag {5} +$$ + +where $\beta$ is the strength factor controlling the debiasing strength and a higher value leads to stronger debiasing effects. The $\epsilon_{anchor}$ is the anchor among the + +![](images/6dae295d058d052c7b45fb02d1b226b792d3ff94cbf9b58a37a7ac22d83c6106.jpg) +Fig. 5: (a) Four example videos (the same as the videos in the first and fourth rows of Fig. 2) and their relationships in terms of motion and appearance. (b) We inverse the four videos based on the video diffusion model and visualize the denoising process. Each point corresponds to a latent code $z_{t,i,j}$ at time step $t$ of $i$ -th frame of $j$ -th video, reduced to 2 dimensions through PCA. (c) Take latent codes at time step 0 for example, the ones of the same video are connected in order of frames. We find that the internal connectivity structure between latent codes is more influenced by motion, while the distance between sets of latent codes is primarily affected by the difference in appearance. (d) Through the proposed appearance-debiased operation, as formulated in Equation 5, the latent codes eliminate distance bias introduced by appearance, while preserving the internal connectivity structure that represents motion dynamics. + +frames from the same training data. If $\beta$ equals 1, it indicates that we attribute equal importance to both appearance and motion in the video. The detailed mathematical derivation of this equation is provided in the supplementary materials. In practice, we find that simply setting $\beta = 1$ and randomly sampling $\epsilon_{i} \in \{\epsilon_{i}\}_{i=1}^{F}$ as the anchor can achieve good performance. The appearance-debiased temporal loss (AD-Loss) is reformulated as + +$$ +\mathcal {L} _ {a d - t e m p} = \mathbb {E} _ {z _ {0}, y, \epsilon , t} \left[ \| \phi (\epsilon) - \phi \left(\epsilon_ {\theta} \left(z _ {t}, t, \tau_ {\theta} (y)\right)\right) \| _ {2} ^ {2} \right]. \tag {6} +$$ + +For temporal LoRAs, the loss function is the combination of temporal loss and appearance-debiased temporal loss as follows, + +$$ +\mathcal {L} _ {\text {t e m p o r a l}} = \mathcal {L} _ {\text {o r g - t e m p}} + \mathcal {L} _ {\text {a d - t e m p}}. \tag {7} +$$ + +# 3.4 Inference of MotionDirector + +In the inference stage, we inject the trained temporal LoRAs into the pre-trained video diffusion model to enable it to generate diverse videos with the learned motion from the training data. If the training data is a single video, the learned motion will be a specific motion, such as an object first moving forward and then turning to the left. If the training data is a set of videos, the learned motion will be the motion concept provided by them, like lifting weights or playing golf. + +![](images/4fc96f5fcd42560b73670b90cbcd8cab094aacc4b6cdd5589afa8609e2d9f8de.jpg) +Fig. 6: Qualitative comparison results of motion customization on multiple videos. + +Tab. 1: Automatic and human evaluations results of motion customization on multiple videos. + +
Automatic EvaluationsHuman Evaluations
Appearance Diversity (↑)Temporal Consistency (↑)Pick Score (↑)Appearance DiversityTemporal ConsistencyMotion Fidelity
Tune-A-Video28.2292.4520.20v.s. Base Model (ModelScope)25.00 v.s. 75.0025.00 v.s. 75.0040.00 v.s. 60.00
v.s. Base Model (ZeroScope)44.00 v.s. 56.0016.67 v.s. 83.3353.33 v.s. 46.67
ModelScopeBase Model28.5592.5420.33v.s. Base Model (ModelScope)23.08 v.s. 76.9240.00 v.s. 60.0052.00 v.s. 48.00
Coupled Tuned25.66 (-2.89)90.6619.85
w/o AD-Loss28.32 (-0.23)91.1720.34v.s. Base Model (ModelScope)53.12 v.s. 46.8849.84 v.s. 50.1662.45 v.s. 37.55
ours28.66 (+0.11)92.3620.59v.s. Base Model (ModelScope)54.84 v.s. 45.1656.00 v.s. 44.0075.00 v.s. 25.00
ZeroScopeBase Model28.4092.9420.76v.s. Base Model (ZeroScope)37.81 v.s. 62.1941.67 v.s. 58.3354.55 v.s. 45.45
Coupled Tuned25.52 (-2.88)90.6719.99
w/o AD-Loss28.61 (+0.21)91.3720.56v.s. Base Model (ZeroScope)50.10 v.s. 49.9048.00 v.s. 52.0058.33 v.s. 41.67
ours28.94 (+0.54)92.6720.80v.s. Base Model (ZeroScope)52.94 v.s. 47.0655.00 v.s. 45.0076.47 v.s. 23.53
+ +The motion concepts can be ones preferred by users or ones that lie in the long-tailed distribution that can not be synthesized well by pre-trained models. Since appearance and motion are decoupled by our method, the spatial LoRAs can also be used to influence the appearance of generated videos, as shown in Fig. 2. Users can flexibly adjust the influence strength of learned appearance and motion on the generation according to their preferences by simply setting the strength of LoRAs as $W = W_{0} + \gamma \Delta W$ , where $\gamma$ is called the LoRA scale, and $\Delta W$ is the learned weights. For instance, a small scale of temporal LoRAs leads to the generated results weakly expressing the learned motion concepts, while a significantly large scale leads to discordant generation. More experimental results showing the effect of LoRA scales are provided in the supplementary materials. To ensure comparative fairness, we set the scale of temporal LoRAs to 1 in all experiments in Sec. 4. + +# 4 Experiments + +# 4.1 Motion customization on multiple videos + +Dataset. We conduct experiments on the adapted UCF Sports Action data set [53], which includes 95 videos of 12 different human motions, like playing golf, lifting weights, etc. For each type of motion, we label one original text prompt describing the motion, such as "a person is playing golf, side view". For these motions, we set 72 different text prompts in total as input to generate videos using comparison methods, such as "a monkey is playing golf, side view". + +Comparison Methods. We compare the proposed method with three baselines and the video generation method Tune-A-Video [70] that can be adapted to this task. Tune-A-Video was initially proposed for training temporal modules on a single video to learn its motion information, while here we adapt it to train on multiple videos. The baseline methods are compared with the proposed method on two different foundational text-to-video diffusion models, i.e. the ModelScope [60] and the ZeroScope [55]. We employ three baseline methods: the first is directly applying the vanilla foundation models, the second is tuning the foundation models with LoRAs in a coupled manner, and the third is the proposed dual-path method excluding the appearance-debiased temporal loss. + +Qualitative Results As shown in Fig. 6, taking a set of videos with motions of playing golf as training data, the Tune-A-Video fails to generate diverse appearances with the learned motions, like a monkey playing golf. To compare the baseline methods and proposed method fairly, we feed the same initial Gaussian noise to these methods to generate videos. The pre-trained foundation model, ZeroScope, correctly generates the appearance but lacks the realistic motion that swings a golf club, as those desired motions in the reference videos. The coupled tuned model could generate the desired motion but the learned motion is coupled with too much appearance information causing the generated subject in the video to be more like a human rather than a monkey. The last two rows show that the proposed dual-path LoRAs can avoid hurting the appearance generation and the proposed appearance-debiased temporal loss enhances the learning of desired motion better. We could draw a similar conclusion from the second example showing the motion of riding a panda. + +Quantitative Results. We evaluate the methods with automatic evaluations and human evaluations, and the results are shown in Table. 1. + +Automatic Metrics. Following the LOVEU-TGVE competition [69], the appearance diversity is computing the average CLIP score [20] between the diverse text prompts and all frames of the generated videos, the temporal consistency is the average CLIP score between frames, and the Pick Score is the average PickScore [31] between all frames of output videos. + +In Table 1, the automatic evaluation of temporal consistency is calculated as the CLIP score between frames. This may result in videos that are nearly static and without desired motion receiving high scores. Thus, we further apply human evaluation to evaluate generated videos more comprehensively and accurately. + +Human Preference. On the Amazon MTurk $^{1}$ , each generated video is evaluated by 5 human raters in terms of appearance diversity, temporal consistency, and motion fidelity, which evaluate whether the generated motion is similar to the references. To simplify the comparison for raters, they are asked to compare the results pairwise and select their preferred one, where the videos are shuffled and their source methods are anonymous. In Table. 1, the pairwise numbers "p1 v.s. p2" means $p_1\%$ results of the first method are preferred while $p_2\%$ results of the second method are preferred. + +The evaluation results show that coupled tuning will destroy the appearance diversity of pre-trained models, while our method will preserve it and achieve the highest motion fidelity. + +# 4.2 Motion customization on a single video + +Dataset. We conduct the comparison experiments on the open-sourced benchmark released by the LOVEU-TGVE competition at CVPR 2023 [69]. The dataset comprises 76 videos, each originally associated with 4 editing text prompts. Additionally, we introduced 3 more prompts with significant changes. + +Comparison Methods. We compare the proposed method with SOTA controllable generation methods, the VideoCrafter [17], VideoComposer [63], and Control-A-Video [6], and the tuning-based method Tune-A-Video [70]. To ensure a fair comparison, we use the depth control mode of controllable generation methods, which is available in all of them. + +![](images/58d5620fac8497f224f5113feb5b0cf49363aa06b9f5242c48e71b4084ce4a60.jpg) +Fig. 7: Qualitative comparison results of motion customization on a single video. + +Tab. 2: Automatic and human evaluations results of motion customization on a single video. + +
Automatic EvaluationsHuman Evaluations
Text Alignment (↑)Appearance Diversity (↑)Temporal Consistency (↑)Pick Score (↑)Text AlignmentAppearance DiversityTemporal ConsistencyMotion Fidelity
VideoComposer27.6627.0392.2220.26ours v.s. VideoComposer54.55 v.s. 45.4572.83 v.s. 27.1761.57 v.s. 38.4361.24 v.s. 38.76
Control-a-Video26.5425.3592.6319.75ours v.s. Control-A-Video68.00 v.s. 32.0078.43 v.s. 21.5771.28 v.s. 29.7256.47 v.s. 43.53
VideoCrafter28.0327.6992.2620.12ours v.s. VideoCrafter52.72 v.s. 47.2871.11 v.s. 28.8960.22 v.s. 39.7860.00 v.s. 40.00
Tune-a-Video25.6425.9592.4220.09ours v.s. Tune-A-Video67.86 v.s. 32.1469.14 v.s. 30.8671.67 v.s. 28.3356.52 v.s. 43.48
ours27.8228.4893.0020.74
+ +![](images/f07b22b517f3924135187f8be2ea53bada783e21e4b60984dad8df35282af69c.jpg) +Fig. 8: Videos generated by MotionDirector. Their appearances are customized according to the reference images shown on the top-left side. Their motions are customized based on the reference videos shown on the top-right side. + +Qualitative and Quantitative Results. As shown in Fig. 7, comparison methods fail to generalize the desired motions to diverse appearances, like the ears of bears and the Arc de Triomphe. Similar to the depth, using the edge and optical flow as control signals also introduces interference from the appearance into the generated videos. In Table. 2, we refer to the alignment between the generated videos and the original 4 editing text prompts as text alignment, and the alignment with the 3 new text prompts with significant changes as appearance diversity. The results show that our method outperforms other methods by a large margin when generalizing the motions to diverse appearances, and achieves competitive motion fidelity. + +# 4.3 Customizing Both Appearance and Motion + +Since MotionDirector learns the appearance and motion in a decoupled manner, it can also reorganize and combine the appearances and motions from different sources. For example, as shown in Fig. 2, MotionDirector can combine the appearance from one video with the motion from another video to generate a new video that exhibits similar appearance and motion characteristics to each of the input videos, respectively. Here we provide more results in Fig. 8. The base model is injected with the spatial LoRAs trained on the given images of "Terracotta Warrior" and injected with different temporal LoRAs trained on different sets of videos of different motion concepts, such as "riding horse". + +# 4.4 Efficiency Performance + +The lightweight LoRAs enable our method to tune the foundation models efficiently. Taking the foundation model ZeroScope for example, it has over 1.8 billion pre-trained parameters. Each set of trainable spatial and temporal LoRAs only adds 9 million and 12 million parameters, respectively. Requiring 14 GB VRAM, MotionDirector takes 20 minutes to converge on multiple reference videos, and 8 minutes for a single reference video, competitive to the 10 minutes required by Tune-A-Video [70]. After training, MotionDirector can generate diverse videos with learned motions. In the inference stage, MotionDirector does not introduce significant increases in computational costs or inference time. + +# 5 Limitations and Future Works + +Despite the MotionDirector can learn the motions of one or two subjects in the reference videos, it is still hard to learn complex motions of multiple subjects, such as a group of boys playing soccer. Previous appearance customization methods suffer similar problems when generating multiple customized subjects [14]. A possible solution is to further decouple the motions of different subjects in the latent space and learn them separately. + +# 6 Conclusion + +We introduce and formulate the task of Motion Customization, which is adapting the pre-trained foundation text-to-video diffusion models to generate videos with desired motions. The challenge of this task is generalizing the customized motions to various appearances. To overcome this challenge, we propose the MotionDirector with a dual-path architecture and a novel appearance-debiased temporal training objective to decouple the learning of appearance and motion. Experimental results show that MotionDirector can learn motion concepts from a limited number of reference videos, and generalize them to diverse appearances. The automatic and human evaluations on two benchmarks demonstrate the MontionDirector outperforms other methods in terms of appearance diversity and motion fidelity. + +# Acknowledgements + +This research is supported by National Research Foundation, Singapore and A\*STAR, under its RIE2020 Industry Alignment Fund - Industry Collaboration Projects (IAF-ICP) grant call (Grant No. I2001E0059) - SIA-NUS Digital Aviation Corp Lab. Mike Shou is supported by the Ministry of Education, Singapore, under the Academic Research Fund Tier 1 (FY2023). + +# References + +1. Bain, M., Nagrani, A., Varol, G., Zisserman, A.: Frozen in time: A joint video and image encoder for end-to-end retrieval. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1728-1738 (2021) +2. Balaji, Y., Min, M.R., Bai, B., Chellappa, R., Graf, H.P.: Conditional gan with discriminative filter generation for text-to-video synthesis. In: IJCAI. vol. 1, p. 2 (2019) +3. Blattmann, A., Rombach, R., Ling, H., Dockhorn, T., Kim, S.W., Fidler, S., Kreis, K.: Align your latents: High-resolution video synthesis with latent diffusion models. In: CVPR (2023) +4. Blattmann, A., Rombach, R., Ling, H., Dockhorn, T., Kim, S.W., Fidler, S., Kreis, K.: Align your latents: High-resolution video synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22563-22575 (2023) +5. Chen, T.S., Lin, C.H., Tseng, H.Y., Lin, T.Y., Yang, M.H.: Motion-conditioned diffusion model for controllable video synthesis. arXiv:2304.14404 (2023) +6. Chen, W., Wu, J., Xie, P., Wu, H., Li, J., Xia, X., Xiao, X., Lin, L.: Control-a-video: Controllable text-to-video generation with diffusion models. arXiv preprint arXiv:2305.13840 (2023) +7. Chen, X., Huang, L., Liu, Y., Shen, Y., Zhao, D., Zhao, H.: Anydoor: Zero-shot object-level image customization. arXiv preprint arXiv:2307.09481 (2023) +8. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 248-255. IEEE (2009) +9. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems 34, 8780-8794 (2021) +0. Duan, Z., You, L., Wang, C., Chen, C., Wu, Z., Qian, W., Huang, J., Chao, F., Ji, R.: Diffsynth: Latent in-iteration deflickering for realistic video synthesis. arXiv:2308.03463 (2023) +1. Esser, P., Chiu, J., Atighehchian, P., Granskog, J., Germanidis, A.: Structure and content-guided video synthesis with diffusion models. arXiv preprint arXiv:2302.03011 (2023) +2. Ge, S., Hayes, T., Yang, H., Yin, X., Pang, G., Jacobs, D., Huang, J.B., Parikh, D.: Long video generation with time-agnostic vqgan and time-sensitive transformer. arXiv preprint arXiv:2204.03638 (2022) +3. Ge, S., Nah, S., Liu, G., Poon, T., Tao, A., Catanzaro, B., Jacobs, D., Huang, J.B., Liu, M.Y., Balaji, Y.: Preserve your own correlation: A noise prior for video diffusion models. arXiv:2305.10474 (2023) +4. Gu, Y., Wang, X., Wu, J.Z., Shi, Y., Chen, Y., Fan, Z., Xiao, W., Zhao, R., Chang, S., Wu, W., et al.: Mix-of-show: Decentralized low-rank adaptation for multi-concept customization of diffusion models. arXiv preprint arXiv:2305.18292 (2023) +5. Guo, Y., Yang, C., Rao, A., Wang, Y., Qiao, Y., Lin, D., Dai, B.: Animatediff: Animate your personalized text-to-image diffusion models without specific tuning. arXiv preprint arXiv:2307.04725 (2023) +6. He, Y., Xia, M., Chen, H., Cun, X., Gong, Y., Xing, J., Zhang, Y., Wang, X., Weng, C., Shan, Y., et al.: animate-a-story: Storytelling with retrieval-augmented video generation. arXiv:2307.06940 (2023) + +17. He, Y., Yang, T., Zhang, Y., Shan, Y., Chen, Q.: Latent video diffusion models for high-fidelity long video generation (2022) +18. He, Y., Yang, T., Zhang, Y., Shan, Y., Chen, Q.: Latent video diffusion models for high-fidelity video generation with arbitrary lengths. arXiv:2211.13221 (2022) +19. He, Y., Yang, T., Zhang, Y., Shan, Y., Chen, Q.: Videocrafter: A toolkit for text-to-video generation and editing. https://github.com/AILab-CVC/VideoCrafter (2023) +20. Hessel, J., Holtzman, A., Forbes, M., Bras, R.L., Choi, Y.: Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718 (2021) +21. Ho, J., Chan, W., Sahara, C., Whang, J., Gao, R., Gritsenko, A., Kingma, D.P., Poole, B., Norouzi, M., Fleet, D.J., et al.: Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303 (2022) +22. Ho, J., Chan, W., Sahara, C., Whang, J., Gao, R., Gritsenko, A., Kingma, D.P., Poole, B., Norouzi, M., Fleet, D.J., et al.: Imagen video: High definition video generation with diffusion models. arXiv:2210.02303 (2022) +23. Ho, J., Salimans, T., Gritsenko, A., Chan, W., Norouzi, M., Fleet, D.J.: Video diffusion models. In: NeurIPS (2022) +24. Hong, S., Seo, J., Hong, S., Shin, H., Kim, S.: Large language models are frame-level directors for zero-shot text-to-video generation. arXiv:2305.14330 (2023) +25. Hong, W., Ding, M., Zheng, W., Liu, X., Tang, J.: Cogvideo: Large-scale pretraining for text-to-video generation via transformers. arXiv preprint arXiv:2205.15868 (2022) +26. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685 (2021) +27. Huang, H., Feng, Y., Shi, C., Xu, L., Yu, J., Yang, S.: Free-bloom: Zero-shot text-to-video generator with lvm director and ldm animator. In: NeurIPS (2023) +28. Jeong, H., Park, G.Y., Ye, J.C.: Vmc: Video motion customization using temporal attention adaption for text-to-video diffusion models. arXiv preprint arXiv:2312.00845 (2023) +29. Karras, J., Holynski, A., Wang, T.C., Kemelmacher-Shlizerman, I.: Dreampose: Fashion image-to-video synthesis via stable diffusion. arXiv:2304.06025 (2023) +30. Khachatryan, L., Movsisyan, A., Tadevosyan, V., Henschel, R., Wang, Z., Navasardyan, S., Shi, H.: Text2video-zero: Text-to-image diffusion models are zero-shot video generators. In: ICCV (2023) +31. Kirstain, Y., Polyak, A., Singer, U., Matiana, S., Penna, J., Levy, O.: Pick-a-pic: An open dataset of user preferences for text-to-image generation. arXiv preprint arXiv:2305.01569 (2023) +32. Kumari, N., Zhang, B., Zhang, R., Shechtman, E., Zhu, J.Y.: Multi-concept customization of text-to-image diffusion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1931-1941 (2023) +33. Le Moing, G., Ponce, J., Schmid, C.: Ccvs: context-aware controllable video synthesis. Advances in Neural Information Processing Systems 34, 14042-14055 (2021) +34. Li, X., Chu, W., Wu, Y., Yuan, W., Liu, F., Zhang, Q., Li, F., Feng, H., Ding, E., Wang, J.: Videogen: A reference-guided latent diffusion approach for high definition text-to-video generation. arXiv preprint arXiv:2309.00398 (2023) +35. Lian, L., Shi, B., Yala, A., Darrell, T., Li, B.: Llm-grounded video diffusion models. arXiv:2309.17444 (2023) +36. Luo, Z., Chen, D., Zhang, Y., Huang, Y., Wang, L., Shen, Y., Zhao, D., Zhou, J., Tan, T.: Videofusion: Decomposed diffusion models for high-quality video gen + +eration. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2023) +37. Ma, Y., He, Y., Cun, X., Wang, X., Shan, Y., Li, X., Chen, Q.: Follow your pose: Pose-guided text-to-video generation using pose-free videos. arXiv preprint arXiv:2304.01186 (2023) +38. Ma, Y., He, Y., Cun, X., Wang, X., Shan, Y., Li, X., Chen, Q.: Follow your pose: Pose-guided text-to-video generation using pose-free videos. arXiv:2304.01186 (2023) +39. Materzynska, J., Sivic, J., Shechtman, E., Torralba, A., Zhang, R., Russell, B.: Customizing motion in text-to-video diffusion models. arXiv preprint arXiv:2312.04966 (2023) +40. Mei, K., Patel, V.: Vidm: Video implicit diffusion models. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 9117-9125 (2023) +41. Ni, H., Shi, C., Li, K., Huang, S.X., Min, M.R.: Conditional image-to-video generation with latent flow diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18444-18455 (2023) +42. Qin, B., Ye, W., Yu, Q., Tang, S., Zhuang, Y.: Dancing avatar: Pose and text-guided human motion videos synthesis with image diffusion model. arXiv:2308.07749 (2023) +43. Ren, Y., Zhou, Y., Yang, J., Shi, J., Liu, D., Liu, F., Kwon, M., Shrivastava, A.: Customize-a-video: One-shot motion customization of text-to-video diffusion models. arXiv preprint arXiv:2402.14780 (2024) +44. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: MICCAI (2015) +45. Ruiz, N., Li, Y., Jampani, V., Pritch, Y., Rubinstein, M., Aberman, K.: Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22500-22510 (2023) +46. Ryu, S.: Low-rank adaptation for fast text-to-image diffusion fine-tuning. https://github.com/cloneofsimo/lora (2023) +47. Saito, M., Matsumoto, E., Saito, S.: Temporal generative adversarial nets with singular value clipping. In: Proceedings of the IEEE international conference on computer vision. pp. 2830-2839 (2017) +48. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems 35, 25278-25294 (2022) +49. Shen, X., Li, X., Elhoseiny, M.: Mostgan-v: Video generation with temporal motion styles. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5652-5661 (2023) +50. Singer, U., Polyak, A., Hayes, T., Yin, X., An, J., Zhang, S., Hu, Q., Yang, H., Ashual, O., Gafni, O., et al.: Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792 (2022) +51. Skorokhodov, I., Tulyakov, S., Elhoseiny, M.: Stylegan-v: A continuous video generator with the price, image quality and perks of stylegan2. arXiv preprint arXiv:2112.14683 (2021) +52. Smith, J.S., Hsu, Y.C., Zhang, L., Hua, T., Kira, Z., Shen, Y., Jin, H.: Continual diffusion: Continual customization of text-to-image diffusion with c-lora. arXiv preprint arXiv:2304.06027 (2023) +53. Soomro, K., Zamir, A.R.: Action recognition in realistic sports videos. In: Computer vision in sports, pp. 181-208. Springer (2015) + +54. Srivastava, N., Mansimov, E., Salakhudinov, R.: Unsupervised learning of video representations using lstms. In: International conference on machine learning. pp. 843-852. PMLR (2015) +55. Sterling, S.: Zeroscope. https://huggingface.co/cerspense/zeroscope_v2_576w (2023) +56. Tian, Y., Ren, J., Chai, M., Olszewski, K., Peng, X., Metaxas, D.N., Tulyakov, S.: A good image generator is what you need for high-resolution video synthesis. In: International Conference on Learning Representations (2020) +57. Tulyakov, S., Liu, M.Y., Yang, X., Kautz, J.: Mocogan: Decomposing motion and content for video generation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1526-1535 (2018) +58. Voleti, V., Jolicoeur-Martineau, A., Pal, C.: Mcvd-masked conditional video diffusion for prediction, generation, and interpolation. Advances in Neural Information Processing Systems 35, 23371-23385 (2022) +59. Vondrick, C., Pirsiavash, H., Torralba, A.: Generating videos with scene dynamics. Advances in neural information processing systems 29 (2016) +60. Wang, J., Yuan, H., Chen, D., Zhang, Y., Wang, X., Zhang, S.: Modelscope text-to-video technical report. arXiv preprint arXiv:2308.06571 (2023) +61. Wang, T., Li, L., Lin, K., Lin, C.C., Yang, Z., Zhang, H., Liu, Z., Wang, L.: Disco: Disentangled control for referring human dance generation in real world. arXiv:2307.00040 (2023) +62. Wang, W., Yang, H., Tuo, Z., He, H., Zhu, J., Fu, J., Liu, J.: Videofactory: Swap attention in spatiotemporal diffusions for text-to-video generation. arXiv:2305.10874 (2023) +63. Wang, X., Yuan, H., Zhang, S., Chen, D., Wang, J., Zhang, Y., Shen, Y., Zhao, D., Zhou, J.: Videocomposer: Compositional video synthesis with motion controllability. arXiv preprint arXiv:2306.02018 (2023) +64. Wang, Y., Chen, X., Ma, X., Zhou, S., Huang, Z., Wang, Y., Yang, C., He, Y., Yu, J., der Yang, P., Guo, Y., Wu, T., Si, C., Jiang, Y., Chen, C., Loy, C.C., Dai, B., Lin, D., Qiao, Y., Liu, Z.: Lavie: High-quality video generation with cascaded latent diffusion models (2023), https://api-semanticscholar.org/CorpusID: 262823915 +65. Wang, Y., Chen, X., Ma, X., Zhou, S., Huang, Z., Wang, Y., Yang, C., He, Y., Yu, J., Yang, P., et al.: Lavie: High-quality video generation with cascaded latent diffusion models. arXiv:2309.15103 (2023) +66. Wang, Z., Yuan, Z., Wang, X., Chen, T., Xia, M., Luo, P., Shan, Y.: Motionctrl: A unified and flexible motion controller for video generation. arXiv preprint arXiv:2312.03641 (2023) +67. Wei, Y., Zhang, S., Qing, Z., Yuan, H., Liu, Z., Liu, Y., Zhang, Y., Zhou, J., Shan, H.: Dreamvideo: Composing your dream videos with customized subject and motion. arXiv preprint arXiv:2312.04433 (2023) +68. Wei, Y., Zhang, Y., Ji, Z., Bai, J., Zhang, L., Zuo, W.: Elite: Encoding visual concepts into textual embeddings for customized text-to-image generation. arXiv preprint arXiv:2302.13848 (2023) +69. Wu, J.Z., Gao, D., Bai, J., Shou, M., Li, X., Dong, Z., Singh, A., Keutzer, K., Iandola, F.: The text-guided video editing benchmark at loveu 2023. https:// sites.google.com/view/loveucvpr23/track4 (2023) +70. Wu, J.Z., Ge, Y., Wang, X., Lei, W., Gu, Y., Hsu, W., Shan, Y., Qie, X., Shou, M.Z.: Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. arXiv preprint arXiv:2212.11565 (2022) + +71. Wu, J.Z., Ge, Y., Wang, X., Lei, W., Gu, Y., Hsu, W., Shan, Y., Qie, X., Shou, M.Z.: Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In: ICCV (2023) +72. Wu, R., Chen, L., Yang, T., Guo, C., Li, C., Zhang, X.: Lamp: Learn a motion pattern for few-shot-based video generation. arXiv preprint arXiv:2310.10769 (2023) +73. Xing, J., Xia, M., Liu, Y., Zhang, Y., Zhang, Y., He, Y., Liu, H., Chen, H., Cun, X., Wang, X., et al.: Make-your-video: Customized video generation using textual and structural guidance. arXiv:2306.00943 (2023) +74. Xing, Z., Feng, Q., Chen, H., Dai, Q., Hu, H., Xu, H., Wu, Z., Jiang, Y.G.: A survey on video diffusion models. arXiv preprint arXiv:2310.10647 (2023) +75. Yan, W., Zhang, Y., Abbeel, P., Srinivas, A.: Videogpt: Video generation using vq-vae and transformers. arXiv preprint arXiv:2104.10157 (2021) +76. Yang, S., Hou, L., Huang, H., Ma, C., Wan, P., Zhang, D., Chen, X., Liao, J.: Direct-a-video: Customized video generation with user-directed camera movement and object motion. arXiv preprint arXiv:2402.03162 (2024) +77. Yin, S., Wu, C., Liang, J., Shi, J., Li, H., Ming, G., Duan, N.: Dragnuwa: Fine-grained control in video generation by integrating text, image, and trajectory. arXiv:2308.08089 (2023) +78. Yin, S., Wu, C., Yang, H., Wang, J., Wang, X., Ni, M., Yang, Z., Li, L., Liu, S., Yang, F., et al.: Nuwa-xl: Diffusion over diffusion for extremely long video generation. arXiv:2303.12346 (2023) +79. Yu, S., Sohn, K., Kim, S., Shin, J.: Video probabilistic diffusion models in projected latent space. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18456-18466 (2023) +80. Yu, S., Tack, J., Mo, S., Kim, H., Kim, J., Ha, J.W., Shin, J.: Generating videos with dynamics-aware implicit generative adversarial networks. In: International Conference on Learning Representations (2021) +81. Zhang, D.J., Wu, J.Z., Liu, J.W., Zhao, R., Ran, L., Gu, Y., Gao, D., Shou, M.Z.: Show-1: Marrying pixel and latent diffusion models for text-to-video generation. arXiv:2309.15818 (2023) +82. Zhang, D.J., Wu, J.Z., Liu, J.W., Zhao, R., Ran, L., Gu, Y., Gao, D., Shou, M.Z.: Show-1: Marrying pixel and latent diffusion models for text-to-video generation (2023) +83. Zhang, L., Agrawala, M.: Adding conditional control to text-to-image diffusion models. arXiv preprint arXiv:2302.05543 (2023) +84. Zhao, R., Gu, Y., Wu, J.Z., Zhang, D.J., Liu, J., Wu, W., Keppo, J., Shou, M.Z.: Motiondirector: Motion customization of text-to-video diffusion models. arXiv preprint arXiv:2310.08465 (2023) +85. Zhao, S., Chen, D., Chen, Y.C., Bao, J., Hao, S., Yuan, L., Wong, K.Y.K.: Uncontrolnet: All-in-one control to text-to-image diffusion models. arXiv preprint arXiv:2305.16322 (2023) +86. Zhou, D., Wang, W., Yan, H., Lv, W., Zhu, Y., Feng, J.: Magicvideo: Efficient video generation with latent diffusion models. arXiv:2211.11018 (2022) \ No newline at end of file diff --git a/2024/MotionDirector_ Motion Customization of Text-to-Video Diffusion Models/images.zip b/2024/MotionDirector_ Motion Customization of Text-to-Video Diffusion Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f72b8391fd60844526b083b7e912c20baa8cb6ba --- /dev/null +++ b/2024/MotionDirector_ Motion Customization of Text-to-Video Diffusion Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cdb07d97973199cdfffd6940bb26722564ceb0682cae6c27ab3fb9818e784a1b +size 830670 diff --git a/2024/MotionDirector_ Motion Customization of Text-to-Video Diffusion Models/layout.json b/2024/MotionDirector_ Motion Customization of Text-to-Video Diffusion Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..a7d430dc85e116afa70fc444af85ae7809421e2b --- /dev/null +++ b/2024/MotionDirector_ Motion Customization of Text-to-Video Diffusion Models/layout.json @@ -0,0 +1,8999 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 160, + 111, + 454, + 145 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 111, + 454, + 145 + ], + "spans": [ + { + "bbox": [ + 160, + 111, + 454, + 145 + ], + "type": "text", + "content": "MotionDirector: Motion Customization of Text-to-Video Diffusion Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 140, + 167, + 473, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 167, + 473, + 193 + ], + "spans": [ + { + "bbox": [ + 140, + 167, + 473, + 193 + ], + "type": "text", + "content": "Rui Zhao" + }, + { + "bbox": [ + 140, + 167, + 473, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 140, + 167, + 473, + 193 + ], + "type": "text", + "content": ", Yuchao Gu" + }, + { + "bbox": [ + 140, + 167, + 473, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 140, + 167, + 473, + 193 + ], + "type": "text", + "content": ", Jay Zhangjie Wu" + }, + { + "bbox": [ + 140, + 167, + 473, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 140, + 167, + 473, + 193 + ], + "type": "text", + "content": ", David Junhao Zhang" + }, + { + "bbox": [ + 140, + 167, + 473, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 140, + 167, + 473, + 193 + ], + "type": "text", + "content": ", Jia-Wei Liu" + }, + { + "bbox": [ + 140, + 167, + 473, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 140, + 167, + 473, + 193 + ], + "type": "text", + "content": ", Weijia Wu" + }, + { + "bbox": [ + 140, + 167, + 473, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 140, + 167, + 473, + 193 + ], + "type": "text", + "content": ", Jussi Keppo" + }, + { + "bbox": [ + 140, + 167, + 473, + 193 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 140, + 167, + 473, + 193 + ], + "type": "text", + "content": ", and Mike Zheng Shou" + }, + { + "bbox": [ + 140, + 167, + 473, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 189, + 201, + 424, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 189, + 201, + 424, + 213 + ], + "spans": [ + { + "bbox": [ + 189, + 201, + 424, + 213 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 189, + 201, + 424, + 213 + ], + "type": "text", + "content": "Show Lab, " + }, + { + "bbox": [ + 189, + 201, + 424, + 213 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 189, + 201, + 424, + 213 + ], + "type": "text", + "content": "National University of Singapore, Singapore" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 242, + 452, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 242, + 452, + 460 + ], + "spans": [ + { + "bbox": [ + 160, + 242, + 452, + 460 + ], + "type": "text", + "content": "Abstract. Large-scale pre-trained diffusion models have exhibited remarkable capabilities in diverse video generations. Given a set of video clips of the same motion concept, the task of Motion Customization is to adapt existing text-to-video diffusion models to generate videos with this motion. Adaptation methods have been developed for customizing appearance like subject or style, yet under-explored for motion. It is straightforward to extend mainstream adaption methods for motion customization, including full model tuning and Low-Rank Adaptions (LoRAs). However, the motion concept learned by these methods is often coupled with the limited appearances in the training videos, making it difficult to generalize the customized motion to other appearances. To overcome this challenge, we propose MotionDirector, with a dual-path LoRAs architecture to decouple the learning of appearance and motion. Further, we design a novel appearance-debiased temporal loss to mitigate the influence of appearance on the temporal training objective. Experimental results show the proposed method can generate videos of diverse appearances for the customized motions. Our method also supports various downstream applications, such as the mixing of different videos with their appearance and motion respectively, and animating a single image with customized motions. The project website is at: MotionDirector." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 472, + 453, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 472, + 453, + 493 + ], + "spans": [ + { + "bbox": [ + 160, + 472, + 453, + 493 + ], + "type": "text", + "content": "Keywords: motion customization " + }, + { + "bbox": [ + 160, + 472, + 453, + 493 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 472, + 453, + 493 + ], + "type": "text", + "content": " text-to-video generation " + }, + { + "bbox": [ + 160, + 472, + 453, + 493 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 472, + 453, + 493 + ], + "type": "text", + "content": " diffusion models" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 514, + 230, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 514, + 230, + 528 + ], + "spans": [ + { + "bbox": [ + 133, + 514, + 230, + 528 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 539, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 539, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 539, + 482, + 647 + ], + "type": "text", + "content": "Text-to-video diffusion models [17,21,50] are approaching generating high-quality diverse videos given text instructions. The open-sourcing of foundational text-to-video models [55,60] pre-trained on large-scale data has sparked enthusiasm for video generation in both the community and academia. Users can create videos that are either realistic or imaginative simply by providing text prompts. While foundation models generate diverse videos from the same text, adapting them to generate more specific content can better accommodate the preferences of users. Similar to the customization of text-to-image foundation models [45], tuning video foundation models to generate videos of a certain concept of appearance," + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 653, + 238, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 238, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 238, + 665 + ], + "type": "text", + "content": "* Corresponding Author." + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 115, + 479, + 270 + ], + "blocks": [ + { + "bbox": [ + 135, + 115, + 479, + 270 + ], + "lines": [ + { + "bbox": [ + 135, + 115, + 479, + 270 + ], + "spans": [ + { + "bbox": [ + 135, + 115, + 479, + 270 + ], + "type": "image", + "image_path": "3582cef6c28440356b1b66594b0965f76d24ef0669ceb60bef4b3bee14c7130f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 184, + 277, + 428, + 289 + ], + "lines": [ + { + "bbox": [ + 184, + 277, + 428, + 289 + ], + "spans": [ + { + "bbox": [ + 184, + 277, + 428, + 289 + ], + "type": "text", + "content": "Fig. 1: Motion customization of the text-to-video diffusion." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 316, + 480, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 316, + 480, + 388 + ], + "spans": [ + { + "bbox": [ + 130, + 316, + 480, + 388 + ], + "type": "text", + "content": "like subject or style, has also been explored [17]. Compared with images, videos consist of not just appearances but also motion dynamics. Users may desire to create videos with specific motions, such as a car moving forward and then turning left under a predefined camera perspective, as illustrated on the right side in Fig. 1. However, customizing the motions in the text-to-video generation is still under-explored." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 391, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 391, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 391, + 482, + 668 + ], + "type": "text", + "content": "The task of Motion Customization is formulated as follows: given a set of reference videos representing a specific motion concept, the objective is to customize pre-trained foundational models to generate videos that accurately capture and replicate the depicted motion. The customized motion concept can be challenging motions that are hard for the pre-trained model to generate, novel motions not present in the pre-training dataset, and specific motions that are difficult to express in text. Examples of these are respectively showcased in the first, second, and third columns of Fig 1. In contrast, previous works on appearance customization adapt the foundation models to generate samples with desired appearance, like subject or style, given reference videos or images representing such appearance [17,45]. It is straightforward to use previous adaption methods for motion customization. For example, on the given reference videos, fine-tuning the weights of foundation models [45], parameter-efficient tuning additional layers [70], or training Low-Rank Adaptions (LoRAs) [26] injected in the layers of foundation models. However, customizing diffusion models to generate desired motions without harming their appearance diversity is challenging because the motion and appearance are coupled with each other at the step-by-step denoising stage. Directly deploying previous adaption methods to learn motions makes the models fit the limited appearances seen in the reference videos, posing challenges in generalizing the learned motions to various appearances. Recent works on controllable text-to-video generations [11, 17, 63] generate videos controlled by signals representing pre-defined motions. However, the control signals, such as depth maps or edges, impose constraints on the shapes of subjects and back" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "text", + "content": "Zhao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 114, + 478, + 274 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 478, + 274 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 478, + 274 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 478, + 274 + ], + "type": "image", + "image_path": "0b4a830b7c8818495a38fd991e4e79d94464b1abd952a55177059708a12cb370.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 285, + 482, + 340 + ], + "lines": [ + { + "bbox": [ + 130, + 285, + 482, + 340 + ], + "spans": [ + { + "bbox": [ + 130, + 285, + 482, + 340 + ], + "type": "text", + "content": "Fig. 2: (Row 1) Take two videos to train the proposed MotionDirector, respectively. (Row 2) MotionDirector can generalize the learned motions to diverse appearances. (Row 3) MotionDirector can mix the learned motion and appearance from different videos to generate new videos. (Row 4) MotionDirector can animate a single image with learned motions." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 365, + 480, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 365, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 130, + 365, + 480, + 413 + ], + "type": "text", + "content": "grounds, thus influencing the appearance of generated videos in a coupled way. Besides, these methods accept only one sequence of control signals to generate one video, which may not be suitable for users seeking certain motion types without strict spatial constraints, such as the example of lifting weights in Fig. 1." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 414, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 414, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 414, + 482, + 628 + ], + "type": "text", + "content": "To achieve motion customization of text-to-video diffusion models while preserving appearance diversity, we propose the MotionDirector, which tunes the foundation models to learn the appearance and motions in the given single or multiple reference videos in a decoupled way. MotionDirector tunes the models with low-rank adaptions (LoRAs) while keeping their pre-trained parameters fixed to retain the learned generation knowledge. Specifically, the MotionDirector employs a dual-path architecture, as shown in Fig. 3. For each video, a spatial path consists of a foundation model with trainable spatial LoRAs injected into its spatial transformer layers. These spatial LoRAs are trained on a single frame randomly sampled per training step to capture the appearance characteristics of the input videos. The temporal path, on the other hand, is a replica of the foundation model that shares the spatial LoRAs with the spatial path to fit the appearance of the corresponding input video. Additionally, the temporal transformers in this path are equipped with temporal LoRAs, which are trained on multiple frames of input videos to capture the underlying motion patterns. To further enhance the learning of motions, we propose an appearance-debiased temporal loss to mitigate the influence of appearance on the temporal training objective." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": "Only deploying the trained temporal LoRAs enables the foundation model to generate videos of the learned motions with diverse appearances, as shown in the second row of Fig 2. The decoupled paradigm further makes an interesting" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 383, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 383, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 383, + 91, + 447, + 100 + ], + "type": "text", + "content": "MotionDirector" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "content": "kind of video generation feasible, which is the mix of the appearance from one video with the motion from another video, called the mix of videos, as shown in the third row of Fig 2. The key to this success lies in that MotionDirector can decouple the appearance and motion of videos and then combine them from various source videos. It is achieved by injecting spatial LoRAs trained on one video and temporal LoRAs trained on another video into the foundation model. Besides, the learned motions can be deployed to animate images, as images can be treated as appearance providers, as shown in the last row of Fig 2." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 213, + 482, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 213, + 482, + 295 + ], + "spans": [ + { + "bbox": [ + 130, + 213, + 482, + 295 + ], + "type": "text", + "content": "We conducted experiments on two benchmarks with 86 different motions and over 600 text prompts to test proposed methods, baselines, and comparison methods. The results show our method can be applied to different diffusion-based foundation models and achieve motion customization of various motion concepts. Compared with other methods, our method avoids fitting the limited appearance of reference videos, and can generalize the learned motions to diverse appearances." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 146, + 297, + 348, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 297, + 348, + 308 + ], + "spans": [ + { + "bbox": [ + 146, + 297, + 348, + 308 + ], + "type": "text", + "content": "Our contributions are summarized as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 318, + 479, + 423 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 138, + 318, + 479, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 318, + 479, + 341 + ], + "spans": [ + { + "bbox": [ + 138, + 318, + 479, + 341 + ], + "type": "text", + "content": "- We introduce and define the task of Motion Customization. The challenges lie in generalizing the customized motions to various appearances." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 342, + 479, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 342, + 479, + 376 + ], + "spans": [ + { + "bbox": [ + 138, + 342, + 479, + 376 + ], + "type": "text", + "content": "- We propose the MotionDirector with a dual-path architecture and a novel appearance-debiased temporal training objective, to decouple the learning of appearance and motion." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 379, + 479, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 379, + 479, + 423 + ], + "spans": [ + { + "bbox": [ + 138, + 379, + 479, + 423 + ], + "type": "text", + "content": "- Experiments on two benchmarks demonstrate that MotionDirector can customize various base models to generate diverse videos with desired motion concepts, and outperforms controllable generation methods and tuning-based methods." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 448, + 238, + 460 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 448, + 238, + 460 + ], + "spans": [ + { + "bbox": [ + 132, + 448, + 238, + 460 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 475, + 293, + 487 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 475, + 293, + 487 + ], + "spans": [ + { + "bbox": [ + 132, + 475, + 293, + 487 + ], + "type": "text", + "content": "2.1 Text-to-Video Generation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": "To achieve high-quality video generation, various methods have been developed, such as Generative Adversarial Networks (GANs) [2, 47, 49, 56, 57, 59], autoregressive models [12, 25, 33, 54, 75] and implicit neural representations [51, 80]. Diffusion-based models [3, 10, 13, 18, 22-24, 27, 34, 35, 40, 41, 50, 58, 60, 62, 65, 71, 78, 79, 81, 84, 86] are also approaching high-quality generation by training conditional 3D U-Nets [44] to denoise from randomly sampled sequences of Gaussian noises [74]. Recent foundation models [4, 17, 21, 36, 50, 64, 82] are pre-trained on large-scale image and video datasets [1, 8, 48], to learn powerful generation ability. Some works turn text-to-image foundation models to text-to-video generation by manipulation on cross-frame attention or training additional temporal layers, like Tune-A-Video [71], Text2Video-Zero [30], and AnimiteDiff [15]. The recently open-sourced foundation models [55, 60] have ignited enthusiasm among users to generate realistic or imaginative videos, and further make it possible for users to customize and build their own private models." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 211, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 211, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 211, + 100 + ], + "type": "text", + "content": "Zhao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 332, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 332, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 332, + 128 + ], + "type": "text", + "content": "2.2 Generation Model Customization." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 139, + 482, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 139, + 482, + 308 + ], + "spans": [ + { + "bbox": [ + 130, + 139, + 482, + 308 + ], + "type": "text", + "content": "Customizing the pre-trained large foundation models can fit the preferences of users better while maintaining powerful generation knowledge without training from scratch. Previous customization methods for text-to-image diffusion models [7,14,32,45,52,68] aim to generate certain subjects or styles, given a set of example images. Dreambooth [45] or LoRA [26] can be simply applied to customizing video foundation models to generate videos with certain subjects or styles, given a set of reference video clips or images. The recently proposed VideoCrafter [19] has explored this, which we categorize as appearance customization. In addition to appearances, videos are also characterized by the motion dynamics of subjects and camera movements across frames. After we proposed the task of motion customization, some concurrent works [28, 39, 43, 66, 67, 72, 76] emerged in this field. DreamVideo [67] trains several adapters in U-Net to customize both the subject and motion in video generation. Lamp [72] learns a motion pattern from a set of videos to utilize image diffusion models to generate videos." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 327, + 318, + 338 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 327, + 318, + 338 + ], + "spans": [ + { + "bbox": [ + 132, + 327, + 318, + 338 + ], + "type": "text", + "content": "2.3 Controllable Video Generation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 350, + 482, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 350, + 482, + 495 + ], + "spans": [ + { + "bbox": [ + 130, + 350, + 482, + 495 + ], + "type": "text", + "content": "Controllable generation aims to ensure the generation results align with the given explicit control signals, such as depth maps, human pose, optical flows, etc. [5, 16, 29, 37, 38, 42, 61, 73, 77, 83, 85]. For the controllable text-to-video generation methods, i.e. the VideoCrafter [17], VideoComposer [63], Control-A-Video [6], they train additional branches that take condition signals to align the generated videos with them. Unlike the human poses for specifically controlling the generation of human bodies, the general control singles, such as depth maps, are typically extracted from reference videos and are coupled with both appearance and motion. This results in the generation results being influenced by both the appearance and motion in reference videos. Applying these methods directly in motion customization is challenging when it comes to generalizing the desired motions to diverse appearances." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 516, + 233, + 530 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 516, + 233, + 530 + ], + "spans": [ + { + "bbox": [ + 132, + 516, + 233, + 530 + ], + "type": "text", + "content": "3 Methodology" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 544, + 228, + 555 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 544, + 228, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 544, + 228, + 555 + ], + "type": "text", + "content": "3.1 Preliminaries" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 567, + 481, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 567, + 481, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 567, + 481, + 640 + ], + "type": "text", + "content": "Video Diffusion Model. Video diffusion models train a 3D U-Net to denoise from a randomly sampled sequence of Gaussian noises to generate videos, guided by text prompts. The 3D U-net basically consists of down-sample, middle, and up-sample blocks. Each block has several convolution layers, spatial transformers, and temporal transformers as shown in Fig 3. The 3D U-Net " + }, + { + "bbox": [ + 130, + 567, + 481, + 640 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 130, + 567, + 481, + 640 + ], + "type": "text", + "content": " and a text encoder " + }, + { + "bbox": [ + 130, + 567, + 481, + 640 + ], + "type": "inline_equation", + "content": "\\tau_{\\theta}" + }, + { + "bbox": [ + 130, + 567, + 481, + 640 + ], + "type": "text", + "content": " are jointly optimized by the noise-prediction loss, as detailed in [9]:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 197, + 653, + 481, + 667 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 653, + 481, + 667 + ], + "spans": [ + { + "bbox": [ + 197, + 653, + 481, + 667 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathbb {E} _ {z _ {0}, y, \\epsilon \\sim \\mathcal {N} (0, I), t \\sim \\mathcal {U} (0, T)} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} \\left(z _ {t}, t, \\tau_ {\\theta} (y)\\right) \\| _ {2} ^ {2} \\right], \\tag {1}", + "image_path": "9702d78e30a412886c28cf6f41d61d88292a667170dc620d73c329fd17c5deaa.jpg" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 383, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 383, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 383, + 91, + 447, + 101 + ], + "type": "text", + "content": "MotionDirector" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 115, + 480, + 243 + ], + "blocks": [ + { + "bbox": [ + 134, + 115, + 480, + 243 + ], + "lines": [ + { + "bbox": [ + 134, + 115, + 480, + 243 + ], + "spans": [ + { + "bbox": [ + 134, + 115, + 480, + 243 + ], + "type": "image", + "image_path": "73f3d477644768419b0a8c81f48f3ad3fb9d92540b2eeecc7f57ad3c9ce486fe.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 251, + 482, + 319 + ], + "lines": [ + { + "bbox": [ + 130, + 251, + 482, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 251, + 482, + 319 + ], + "type": "text", + "content": "Fig. 3: The dual-path architecture of the proposed method. All pre-trained weights of the base diffusion model remain fixed. In the spatial path, the spatial transformers are injected with trainable spatial LoRAs as shown on the right side. In the temporal path, the spatial transformers are injected with spatial LoRAs sharing weights with those ones in the spatial path, and the temporal transformers are injected with trainable temporal LoRAs." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 342, + 482, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 342, + 482, + 380 + ], + "spans": [ + { + "bbox": [ + 130, + 342, + 482, + 380 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 342, + 482, + 380 + ], + "type": "inline_equation", + "content": "z_0" + }, + { + "bbox": [ + 130, + 342, + 482, + 380 + ], + "type": "text", + "content": " is the latent code of the training videos, " + }, + { + "bbox": [ + 130, + 342, + 482, + 380 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 130, + 342, + 482, + 380 + ], + "type": "text", + "content": " is the text prompt, " + }, + { + "bbox": [ + 130, + 342, + 482, + 380 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 130, + 342, + 482, + 380 + ], + "type": "text", + "content": " is the Gaussian noise added to the latent code, and " + }, + { + "bbox": [ + 130, + 342, + 482, + 380 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 342, + 482, + 380 + ], + "type": "text", + "content": " is the time step. As discussed in [9], the noised latent code " + }, + { + "bbox": [ + 130, + 342, + 482, + 380 + ], + "type": "inline_equation", + "content": "z_t" + }, + { + "bbox": [ + 130, + 342, + 482, + 380 + ], + "type": "text", + "content": " is determined as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 225, + 388, + 481, + 422 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 388, + 481, + 422 + ], + "spans": [ + { + "bbox": [ + 225, + 388, + 481, + 422 + ], + "type": "interline_equation", + "content": "z _ {t} = \\sqrt {\\bar {\\alpha} _ {t}} z _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon , \\bar {\\alpha} _ {t} = \\prod_ {i = 1} ^ {t} \\alpha_ {t}, \\tag {2}", + "image_path": "99b0527373fc4f481f974c15fc7a775d0cda66a0de4f368335d8650b88823039.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 429, + 401, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 429, + 401, + 441 + ], + "spans": [ + { + "bbox": [ + 130, + 429, + 401, + 441 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 429, + 401, + 441 + ], + "type": "inline_equation", + "content": "\\alpha_{t}" + }, + { + "bbox": [ + 130, + 429, + 401, + 441 + ], + "type": "text", + "content": " is a hyper-parameter controlling the noise strength." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 441, + 482, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 441, + 482, + 501 + ], + "spans": [ + { + "bbox": [ + 130, + 441, + 482, + 501 + ], + "type": "text", + "content": "Low-Rank Adaption. Low-rank adaption (LoRA) [26] was proposed to adapt the pre-trained large language models to downstream tasks. Recently it has been applied in text-to-image generation and text-to-video generation tasks to achieve appearance customization [19, 46]. LoRA employs a low-rank factorization technique to update the weight matrix " + }, + { + "bbox": [ + 130, + 441, + 482, + 501 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 130, + 441, + 482, + 501 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 241, + 511, + 481, + 525 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 241, + 511, + 481, + 525 + ], + "spans": [ + { + "bbox": [ + 241, + 511, + 481, + 525 + ], + "type": "interline_equation", + "content": "W = W _ {0} + \\Delta W = W _ {0} + B A, \\tag {3}", + "image_path": "49acc485b193e447f13634c184c40ad92b68b9fbb64afc4c29d035b9d8b84b7b.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 533, + 482, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 604 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 604 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 533, + 482, + 604 + ], + "type": "inline_equation", + "content": "W_0 \\in \\mathbb{R}^{d \\times k}" + }, + { + "bbox": [ + 130, + 533, + 482, + 604 + ], + "type": "text", + "content": " represents the original weights of the pre-trained model, " + }, + { + "bbox": [ + 130, + 533, + 482, + 604 + ], + "type": "inline_equation", + "content": "B \\in \\mathbb{R}^{d \\times r}" + }, + { + "bbox": [ + 130, + 533, + 482, + 604 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 533, + 482, + 604 + ], + "type": "inline_equation", + "content": "A \\in \\mathbb{R}^{r \\times k}" + }, + { + "bbox": [ + 130, + 533, + 482, + 604 + ], + "type": "text", + "content": " represent the low-rank factors, where " + }, + { + "bbox": [ + 130, + 533, + 482, + 604 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 130, + 533, + 482, + 604 + ], + "type": "text", + "content": " is much smaller than original dimensions " + }, + { + "bbox": [ + 130, + 533, + 482, + 604 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 130, + 533, + 482, + 604 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 533, + 482, + 604 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 533, + 482, + 604 + ], + "type": "text", + "content": ". LoRA requires smaller computing sources than fine-tuning the weights of the entire network like DreamBooth [45], and it is convenient to spread and deploy as a plug-and-play plugin for pre-trained models." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "In the following sections, we introduce the proposed MotionDirector, containing the dual-path Low-Rank Adaptions (in Sec. 3.2) for learning the appearance and motion in reference videos in a decoupled manner, and appearance-debiased temporal loss (in Sec. 3.3)) to further mitigate the influence of appearance on the temporal training objective." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "type": "text", + "content": "Zhao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 115, + 480, + 252 + ], + "blocks": [ + { + "bbox": [ + 134, + 115, + 480, + 252 + ], + "lines": [ + { + "bbox": [ + 134, + 115, + 480, + 252 + ], + "spans": [ + { + "bbox": [ + 134, + 115, + 480, + 252 + ], + "type": "image", + "image_path": "bf379e347ad1d2416a527089af9d55c14e67a0e0b6f113aaef00e343a5ff23d3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 262, + 482, + 297 + ], + "lines": [ + { + "bbox": [ + 130, + 262, + 482, + 297 + ], + "spans": [ + { + "bbox": [ + 130, + 262, + 482, + 297 + ], + "type": "text", + "content": "Fig. 4: Comparison of videos generated by the base model, the model fine-tuned in the coupled manner, the model fine-tuned without the proposed Appearance-Debiased Temporal Loss (AD-Loss), and our full-version MotionDirector method." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 319, + 322, + 332 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 319, + 322, + 332 + ], + "spans": [ + { + "bbox": [ + 132, + 319, + 322, + 332 + ], + "type": "text", + "content": "3.2 Dual-Path Low-Rank Adaptions" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 341, + 482, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 341, + 482, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 341, + 482, + 544 + ], + "type": "text", + "content": "At each time-step " + }, + { + "bbox": [ + 130, + 341, + 482, + 544 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 341, + 482, + 544 + ], + "type": "text", + "content": ", the 3D U-Net takes in the latent code " + }, + { + "bbox": [ + 130, + 341, + 482, + 544 + ], + "type": "inline_equation", + "content": "z_{t} \\in \\mathbb{R}^{b \\times f \\times w \\times h \\times c}" + }, + { + "bbox": [ + 130, + 341, + 482, + 544 + ], + "type": "text", + "content": " and the conditional input " + }, + { + "bbox": [ + 130, + 341, + 482, + 544 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 130, + 341, + 482, + 544 + ], + "type": "text", + "content": " (e.g., text), where " + }, + { + "bbox": [ + 130, + 341, + 482, + 544 + ], + "type": "inline_equation", + "content": "b, f, w, h, c" + }, + { + "bbox": [ + 130, + 341, + 482, + 544 + ], + "type": "text", + "content": " represents the size of the batch, frame, width, height, and channel dimensions, respectively. The spatial transformers apply spatial self-attention along the spatial dimensions " + }, + { + "bbox": [ + 130, + 341, + 482, + 544 + ], + "type": "inline_equation", + "content": "w, h" + }, + { + "bbox": [ + 130, + 341, + 482, + 544 + ], + "type": "text", + "content": " to improve the correlation between pixels, and then leverage the cross-attention between the latent code and the conditional input " + }, + { + "bbox": [ + 130, + 341, + 482, + 544 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 130, + 341, + 482, + 544 + ], + "type": "text", + "content": " to improve textual alignment. The temporal transformers apply temporal self-attention along the frame dimension " + }, + { + "bbox": [ + 130, + 341, + 482, + 544 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 130, + 341, + 482, + 544 + ], + "type": "text", + "content": " to improve the temporal consistency between frames. However, spatial and temporal information in the latent code gradually become coupled with each other during the step-by-step denoising stage. Attempting to directly learn and fit the motions in reference videos will inevitably lead to fitting their limited appearances. As shown in the third row \"Coupled Tuned\" of Fig. 4. From the results on the left side, we can observe that the generated \"panda\" is coupled with the appearance of \"human\" in the reference videos and the generated background is also coupled with that of reference videos. Besides, the results on the right side show that the coupled tuned model can not generate the \"Eiffel Tower\" because it is overfitted on the limited backgrounds in the reference videos." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 545, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 628 + ], + "type": "text", + "content": "To address this problem, we propose to tune the spatial and temporal transformers in a dual-path way to learn the earn the appearance and motion in reference videos, respectively, as shown in Fig. 3. Specifically, for the spatial path, we inject LoRAs into spatial transformers to learn the appearance of training data, and for the temporal path, we inject LoRAs into temporal transformers to learn the motion in videos. The training of spatial and temporal LoRAs are as follows." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "content": "Spatial Training. For the spatial path, we inject unique spatial LoRAs into the spatial transformers for each training video while keeping the weights of pre-trained 3D U-Net fixed. To maintain the learned strong and diverse textual" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 383, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 383, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 383, + 91, + 447, + 100 + ], + "type": "text", + "content": "MotionDirector" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": "alignment ability, we do not inject LoRAs into cross-attention layers of spatial transformers, since their weights influence the correlations between the pixels and text prompts. On the other hand, we inject LoRAs into spatial self-attention layers and feed-forward layers to update the correlations in spatial dimensions to enable the model to reconstruct the appearance of training data. For each training step, the spatial LoRAs are trained on a single frame randomly sampled from the training video to fit its appearance while ignoring its motion, based on spatial loss, which is reformulated as" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 195, + 224, + 481, + 239 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 195, + 224, + 481, + 239 + ], + "spans": [ + { + "bbox": [ + 195, + 224, + 481, + 239 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {s p a t i a l}} = \\mathbb {E} _ {z _ {0}, y, \\epsilon , t, i \\sim \\mathcal {U} (0, F)} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} \\left(z _ {t, i}, t, \\tau_ {\\theta} (y)\\right) \\| _ {2} ^ {2} \\right], \\tag {4}", + "image_path": "3967a70538e76490cc253063e12bdc446569873ef5222c41ea155bd6c8b144b3.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 243, + 481, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 243, + 481, + 266 + ], + "spans": [ + { + "bbox": [ + 130, + 243, + 481, + 266 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 243, + 481, + 266 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 130, + 243, + 481, + 266 + ], + "type": "text", + "content": " is the number of frames of the training data and the " + }, + { + "bbox": [ + 130, + 243, + 481, + 266 + ], + "type": "inline_equation", + "content": "z_{t,i}" + }, + { + "bbox": [ + 130, + 243, + 481, + 266 + ], + "type": "text", + "content": " is the sampled frame from the latent code " + }, + { + "bbox": [ + 130, + 243, + 481, + 266 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 130, + 243, + 481, + 266 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 268, + 482, + 364 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 268, + 482, + 364 + ], + "spans": [ + { + "bbox": [ + 130, + 268, + 482, + 364 + ], + "type": "text", + "content": "Temporal Training. For the temporal path, we inject the temporal LoRAs into self-attention and feed-forward layers of temporal transformers to update the correlations along the frame dimension. Besides, the spatial transformers are injected with LoRAs sharing the same weights learned from the spatial path, to force the trainable temporal LoRAs to ignore the appearance of the training data. The temporal LoRAs could be simply trained on all frames of training data based on the temporal loss " + }, + { + "bbox": [ + 130, + 268, + 482, + 364 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{org-temp}" + }, + { + "bbox": [ + 130, + 268, + 482, + 364 + ], + "type": "text", + "content": ", formulated in the same way as equation (1)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 382, + 347, + 395 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 382, + 347, + 395 + ], + "spans": [ + { + "bbox": [ + 131, + 382, + 347, + 395 + ], + "type": "text", + "content": "3.3 Appearance-Debiased Temporal Loss" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 403, + 481, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 403, + 481, + 450 + ], + "spans": [ + { + "bbox": [ + 130, + 403, + 481, + 450 + ], + "type": "text", + "content": "The videos generated by models trained with original temporal loss are shown in the fourth row of Fig. 4. We can observe that the motions are not learned well. To explore the underlying reasons, we visualized the denoising process of several videos." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 451, + 482, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 451, + 482, + 559 + ], + "spans": [ + { + "bbox": [ + 130, + 451, + 482, + 559 + ], + "type": "text", + "content": "As illustrated in Fig. 5, when considering the latent codes of each frame " + }, + { + "bbox": [ + 130, + 451, + 482, + 559 + ], + "type": "inline_equation", + "content": "z_{t,i_{i = 1}^{F}}" + }, + { + "bbox": [ + 130, + 451, + 482, + 559 + ], + "type": "text", + "content": " as a set of points in the latent space, we found that motion primarily impacts the underlying dependencies between these point sets, whereas the distances between different sets of points are more influenced by appearance. The original temporal training objective does not take this into account. The learning of motion dynamics, embedded within the internal connectivity structure of latent codes, remains biased by positional differences of latent codes. These biases are primarily influenced by the appearance of training samples and pose a challenge to learning motions efficiently." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 559, + 482, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 559, + 482, + 608 + ], + "spans": [ + { + "bbox": [ + 131, + 559, + 482, + 608 + ], + "type": "text", + "content": "To further decouple the motion from appearance, we proposed to eliminate the appearance bias among the noises and predicted noises, and calculate the appearance-debiased temporal loss on them. The debiasing of each noise " + }, + { + "bbox": [ + 131, + 559, + 482, + 608 + ], + "type": "inline_equation", + "content": "\\epsilon_{i} \\in \\{\\epsilon_{i}\\}_{i=1}^{F}" + }, + { + "bbox": [ + 131, + 559, + 482, + 608 + ], + "type": "text", + "content": " is as follows," + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 241, + 617, + 481, + 632 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 241, + 617, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 241, + 617, + 481, + 632 + ], + "type": "interline_equation", + "content": "\\phi (\\epsilon_ {i}) = \\sqrt {\\beta^ {2} + 1} \\epsilon_ {i} - \\beta \\epsilon_ {a n c h o r}, \\tag {5}", + "image_path": "3240b18fabe941aae4c07f4b12655d6b06a79baac74c265ea23dc377d6f1cf76.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "text", + "content": " is the strength factor controlling the debiasing strength and a higher value leads to stronger debiasing effects. The " + }, + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\epsilon_{anchor}" + }, + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "text", + "content": " is the anchor among the" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 211, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 211, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 211, + 101 + ], + "type": "text", + "content": "Zhao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 114, + 479, + 247 + ], + "blocks": [ + { + "bbox": [ + 136, + 114, + 479, + 247 + ], + "lines": [ + { + "bbox": [ + 136, + 114, + 479, + 247 + ], + "spans": [ + { + "bbox": [ + 136, + 114, + 479, + 247 + ], + "type": "image", + "image_path": "6dae295d058d052c7b45fb02d1b226b792d3ff94cbf9b58a37a7ac22d83c6106.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 258, + 482, + 380 + ], + "lines": [ + { + "bbox": [ + 130, + 258, + 482, + 380 + ], + "spans": [ + { + "bbox": [ + 130, + 258, + 482, + 380 + ], + "type": "text", + "content": "Fig. 5: (a) Four example videos (the same as the videos in the first and fourth rows of Fig. 2) and their relationships in terms of motion and appearance. (b) We inverse the four videos based on the video diffusion model and visualize the denoising process. Each point corresponds to a latent code " + }, + { + "bbox": [ + 130, + 258, + 482, + 380 + ], + "type": "inline_equation", + "content": "z_{t,i,j}" + }, + { + "bbox": [ + 130, + 258, + 482, + 380 + ], + "type": "text", + "content": " at time step " + }, + { + "bbox": [ + 130, + 258, + 482, + 380 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 258, + 482, + 380 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 130, + 258, + 482, + 380 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 258, + 482, + 380 + ], + "type": "text", + "content": "-th frame of " + }, + { + "bbox": [ + 130, + 258, + 482, + 380 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 130, + 258, + 482, + 380 + ], + "type": "text", + "content": "-th video, reduced to 2 dimensions through PCA. (c) Take latent codes at time step 0 for example, the ones of the same video are connected in order of frames. We find that the internal connectivity structure between latent codes is more influenced by motion, while the distance between sets of latent codes is primarily affected by the difference in appearance. (d) Through the proposed appearance-debiased operation, as formulated in Equation 5, the latent codes eliminate distance bias introduced by appearance, while preserving the internal connectivity structure that represents motion dynamics." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 403, + 482, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 403, + 482, + 475 + ], + "spans": [ + { + "bbox": [ + 130, + 403, + 482, + 475 + ], + "type": "text", + "content": "frames from the same training data. If " + }, + { + "bbox": [ + 130, + 403, + 482, + 475 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 130, + 403, + 482, + 475 + ], + "type": "text", + "content": " equals 1, it indicates that we attribute equal importance to both appearance and motion in the video. The detailed mathematical derivation of this equation is provided in the supplementary materials. In practice, we find that simply setting " + }, + { + "bbox": [ + 130, + 403, + 482, + 475 + ], + "type": "inline_equation", + "content": "\\beta = 1" + }, + { + "bbox": [ + 130, + 403, + 482, + 475 + ], + "type": "text", + "content": " and randomly sampling " + }, + { + "bbox": [ + 130, + 403, + 482, + 475 + ], + "type": "inline_equation", + "content": "\\epsilon_{i} \\in \\{\\epsilon_{i}\\}_{i=1}^{F}" + }, + { + "bbox": [ + 130, + 403, + 482, + 475 + ], + "type": "text", + "content": " as the anchor can achieve good performance. The appearance-debiased temporal loss (AD-Loss) is reformulated as" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 197, + 483, + 481, + 499 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 483, + 481, + 499 + ], + "spans": [ + { + "bbox": [ + 197, + 483, + 481, + 499 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {a d - t e m p} = \\mathbb {E} _ {z _ {0}, y, \\epsilon , t} \\left[ \\| \\phi (\\epsilon) - \\phi \\left(\\epsilon_ {\\theta} \\left(z _ {t}, t, \\tau_ {\\theta} (y)\\right)\\right) \\| _ {2} ^ {2} \\right]. \\tag {6}", + "image_path": "8379b943b14071a1222f6a2bf2124075daa0fc9768a4ae03863e3960cf512575.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 507, + 482, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 507, + 482, + 532 + ], + "spans": [ + { + "bbox": [ + 130, + 507, + 482, + 532 + ], + "type": "text", + "content": "For temporal LoRAs, the loss function is the combination of temporal loss and appearance-debiased temporal loss as follows," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 232, + 542, + 481, + 556 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 232, + 542, + 481, + 556 + ], + "spans": [ + { + "bbox": [ + 232, + 542, + 481, + 556 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t e m p o r a l}} = \\mathcal {L} _ {\\text {o r g - t e m p}} + \\mathcal {L} _ {\\text {a d - t e m p}}. \\tag {7}", + "image_path": "cb3a130183e92327074b111e3be4a876b4f6a8f698f4c9fee3d60c4def64e3b6.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 572, + 303, + 583 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 572, + 303, + 583 + ], + "spans": [ + { + "bbox": [ + 131, + 572, + 303, + 583 + ], + "type": "text", + "content": "3.4 Inference of MotionDirector" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "In the inference stage, we inject the trained temporal LoRAs into the pre-trained video diffusion model to enable it to generate diverse videos with the learned motion from the training data. If the training data is a single video, the learned motion will be a specific motion, such as an object first moving forward and then turning to the left. If the training data is a set of videos, the learned motion will be the motion concept provided by them, like lifting weights or playing golf." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 383, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 383, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 383, + 91, + 447, + 100 + ], + "type": "text", + "content": "MotionDirector" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 114, + 481, + 278 + ], + "blocks": [ + { + "bbox": [ + 136, + 114, + 481, + 278 + ], + "lines": [ + { + "bbox": [ + 136, + 114, + 481, + 278 + ], + "spans": [ + { + "bbox": [ + 136, + 114, + 481, + 278 + ], + "type": "image", + "image_path": "4fc96f5fcd42560b73670b90cbcd8cab094aacc4b6cdd5589afa8609e2d9f8de.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 137, + 287, + 475, + 300 + ], + "lines": [ + { + "bbox": [ + 137, + 287, + 475, + 300 + ], + "spans": [ + { + "bbox": [ + 137, + 287, + 475, + 300 + ], + "type": "text", + "content": "Fig. 6: Qualitative comparison results of motion customization on multiple videos." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 134, + 373, + 480, + 456 + ], + "blocks": [ + { + "bbox": [ + 132, + 341, + 481, + 363 + ], + "lines": [ + { + "bbox": [ + 132, + 341, + 481, + 363 + ], + "spans": [ + { + "bbox": [ + 132, + 341, + 481, + 363 + ], + "type": "text", + "content": "Tab. 1: Automatic and human evaluations results of motion customization on multiple videos." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 373, + 480, + 456 + ], + "lines": [ + { + "bbox": [ + 134, + 373, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 134, + 373, + 480, + 456 + ], + "type": "table", + "html": "
Automatic EvaluationsHuman Evaluations
Appearance Diversity (↑)Temporal Consistency (↑)Pick Score (↑)Appearance DiversityTemporal ConsistencyMotion Fidelity
Tune-A-Video28.2292.4520.20v.s. Base Model (ModelScope)25.00 v.s. 75.0025.00 v.s. 75.0040.00 v.s. 60.00
v.s. Base Model (ZeroScope)44.00 v.s. 56.0016.67 v.s. 83.3353.33 v.s. 46.67
ModelScopeBase Model28.5592.5420.33v.s. Base Model (ModelScope)23.08 v.s. 76.9240.00 v.s. 60.0052.00 v.s. 48.00
Coupled Tuned25.66 (-2.89)90.6619.85
w/o AD-Loss28.32 (-0.23)91.1720.34v.s. Base Model (ModelScope)53.12 v.s. 46.8849.84 v.s. 50.1662.45 v.s. 37.55
ours28.66 (+0.11)92.3620.59v.s. Base Model (ModelScope)54.84 v.s. 45.1656.00 v.s. 44.0075.00 v.s. 25.00
ZeroScopeBase Model28.4092.9420.76v.s. Base Model (ZeroScope)37.81 v.s. 62.1941.67 v.s. 58.3354.55 v.s. 45.45
Coupled Tuned25.52 (-2.88)90.6719.99
w/o AD-Loss28.61 (+0.21)91.3720.56v.s. Base Model (ZeroScope)50.10 v.s. 49.9048.00 v.s. 52.0058.33 v.s. 41.67
ours28.94 (+0.54)92.6720.80v.s. Base Model (ZeroScope)52.94 v.s. 47.0655.00 v.s. 45.0076.47 v.s. 23.53
", + "image_path": "42bc7282cc3173da84a5af5c6410bcfa8c727724e993d605b037fdee4b5d6822.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": "The motion concepts can be ones preferred by users or ones that lie in the long-tailed distribution that can not be synthesized well by pre-trained models. Since appearance and motion are decoupled by our method, the spatial LoRAs can also be used to influence the appearance of generated videos, as shown in Fig. 2. Users can flexibly adjust the influence strength of learned appearance and motion on the generation according to their preferences by simply setting the strength of LoRAs as " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "W = W_{0} + \\gamma \\Delta W" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": " is called the LoRA scale, and " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\Delta W" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": " is the learned weights. For instance, a small scale of temporal LoRAs leads to the generated results weakly expressing the learned motion concepts, while a significantly large scale leads to discordant generation. More experimental results showing the effect of LoRA scales are provided in the supplementary materials. To ensure comparative fairness, we set the scale of temporal LoRAs to 1 in all experiments in Sec. 4." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "text", + "content": "Zhao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 230, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 230, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 230, + 129 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 145, + 367, + 157 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 145, + 367, + 157 + ], + "spans": [ + { + "bbox": [ + 132, + 145, + 367, + 157 + ], + "type": "text", + "content": "4.1 Motion customization on multiple videos" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 170, + 482, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 170, + 482, + 242 + ], + "spans": [ + { + "bbox": [ + 130, + 170, + 482, + 242 + ], + "type": "text", + "content": "Dataset. We conduct experiments on the adapted UCF Sports Action data set [53], which includes 95 videos of 12 different human motions, like playing golf, lifting weights, etc. For each type of motion, we label one original text prompt describing the motion, such as \"a person is playing golf, side view\". For these motions, we set 72 different text prompts in total as input to generate videos using comparison methods, such as \"a monkey is playing golf, side view\"." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 243, + 482, + 362 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 243, + 482, + 362 + ], + "spans": [ + { + "bbox": [ + 130, + 243, + 482, + 362 + ], + "type": "text", + "content": "Comparison Methods. We compare the proposed method with three baselines and the video generation method Tune-A-Video [70] that can be adapted to this task. Tune-A-Video was initially proposed for training temporal modules on a single video to learn its motion information, while here we adapt it to train on multiple videos. The baseline methods are compared with the proposed method on two different foundational text-to-video diffusion models, i.e. the ModelScope [60] and the ZeroScope [55]. We employ three baseline methods: the first is directly applying the vanilla foundation models, the second is tuning the foundation models with LoRAs in a coupled manner, and the third is the proposed dual-path method excluding the appearance-debiased temporal loss." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 363, + 482, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 363, + 482, + 532 + ], + "spans": [ + { + "bbox": [ + 130, + 363, + 482, + 532 + ], + "type": "text", + "content": "Qualitative Results As shown in Fig. 6, taking a set of videos with motions of playing golf as training data, the Tune-A-Video fails to generate diverse appearances with the learned motions, like a monkey playing golf. To compare the baseline methods and proposed method fairly, we feed the same initial Gaussian noise to these methods to generate videos. The pre-trained foundation model, ZeroScope, correctly generates the appearance but lacks the realistic motion that swings a golf club, as those desired motions in the reference videos. The coupled tuned model could generate the desired motion but the learned motion is coupled with too much appearance information causing the generated subject in the video to be more like a human rather than a monkey. The last two rows show that the proposed dual-path LoRAs can avoid hurting the appearance generation and the proposed appearance-debiased temporal loss enhances the learning of desired motion better. We could draw a similar conclusion from the second example showing the motion of riding a panda." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 532, + 481, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 532, + 481, + 555 + ], + "spans": [ + { + "bbox": [ + 131, + 532, + 481, + 555 + ], + "type": "text", + "content": "Quantitative Results. We evaluate the methods with automatic evaluations and human evaluations, and the results are shown in Table. 1." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 556, + 481, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 556, + 481, + 616 + ], + "spans": [ + { + "bbox": [ + 131, + 556, + 481, + 616 + ], + "type": "text", + "content": "Automatic Metrics. Following the LOVEU-TGVE competition [69], the appearance diversity is computing the average CLIP score [20] between the diverse text prompts and all frames of the generated videos, the temporal consistency is the average CLIP score between frames, and the Pick Score is the average PickScore [31] between all frames of output videos." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 617, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 617, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 617, + 481, + 666 + ], + "type": "text", + "content": "In Table 1, the automatic evaluation of temporal consistency is calculated as the CLIP score between frames. This may result in videos that are nearly static and without desired motion receiving high scores. Thus, we further apply human evaluation to evaluate generated videos more comprehensively and accurately." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 383, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 383, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 383, + 91, + 447, + 101 + ], + "type": "text", + "content": "MotionDirector" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "content": "Human Preference. On the Amazon MTurk " + }, + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "content": ", each generated video is evaluated by 5 human raters in terms of appearance diversity, temporal consistency, and motion fidelity, which evaluate whether the generated motion is similar to the references. To simplify the comparison for raters, they are asked to compare the results pairwise and select their preferred one, where the videos are shuffled and their source methods are anonymous. In Table. 1, the pairwise numbers \"p1 v.s. p2\" means " + }, + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "inline_equation", + "content": "p_1\\%" + }, + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "content": " results of the first method are preferred while " + }, + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "inline_equation", + "content": "p_2\\%" + }, + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "content": " results of the second method are preferred." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 213, + 482, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 213, + 482, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 213, + 482, + 248 + ], + "type": "text", + "content": "The evaluation results show that coupled tuning will destroy the appearance diversity of pre-trained models, while our method will preserve it and achieve the highest motion fidelity." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 267, + 358, + 280 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 267, + 358, + 280 + ], + "spans": [ + { + "bbox": [ + 131, + 267, + 358, + 280 + ], + "type": "text", + "content": "4.2 Motion customization on a single video" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 289, + 487, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 289, + 487, + 337 + ], + "spans": [ + { + "bbox": [ + 130, + 289, + 487, + 337 + ], + "type": "text", + "content": "Dataset. We conduct the comparison experiments on the open-sourced benchmark released by the LOVEU-TGVE competition at CVPR 2023 [69]. The dataset comprises 76 videos, each originally associated with 4 editing text prompts. Additionally, we introduced 3 more prompts with significant changes." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 338, + 482, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 338, + 482, + 397 + ], + "spans": [ + { + "bbox": [ + 130, + 338, + 482, + 397 + ], + "type": "text", + "content": "Comparison Methods. We compare the proposed method with SOTA controllable generation methods, the VideoCrafter [17], VideoComposer [63], and Control-A-Video [6], and the tuning-based method Tune-A-Video [70]. To ensure a fair comparison, we use the depth control mode of controllable generation methods, which is available in all of them." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 135, + 418, + 480, + 608 + ], + "blocks": [ + { + "bbox": [ + 135, + 418, + 480, + 608 + ], + "lines": [ + { + "bbox": [ + 135, + 418, + 480, + 608 + ], + "spans": [ + { + "bbox": [ + 135, + 418, + 480, + 608 + ], + "type": "image", + "image_path": "58d5620fac8497f224f5113feb5b0cf49363aa06b9f5242c48e71b4084ce4a60.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 140, + 610, + 473, + 622 + ], + "lines": [ + { + "bbox": [ + 140, + 610, + 473, + 622 + ], + "spans": [ + { + "bbox": [ + 140, + 610, + 473, + 622 + ], + "type": "text", + "content": "Fig. 7: Qualitative comparison results of motion customization on a single video." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "type": "text", + "content": "Zhao et al." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 134, + 652, + 265, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 652, + 265, + 666 + ], + "spans": [ + { + "bbox": [ + 134, + 652, + 265, + 666 + ], + "type": "text", + "content": "1 https://requester.mturk.com/" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 134, + 140, + 481, + 191 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 480, + 136 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 480, + 136 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 480, + 136 + ], + "type": "text", + "content": "Tab. 2: Automatic and human evaluations results of motion customization on a single video." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 140, + 481, + 191 + ], + "lines": [ + { + "bbox": [ + 134, + 140, + 481, + 191 + ], + "spans": [ + { + "bbox": [ + 134, + 140, + 481, + 191 + ], + "type": "table", + "html": "
Automatic EvaluationsHuman Evaluations
Text Alignment (↑)Appearance Diversity (↑)Temporal Consistency (↑)Pick Score (↑)Text AlignmentAppearance DiversityTemporal ConsistencyMotion Fidelity
VideoComposer27.6627.0392.2220.26ours v.s. VideoComposer54.55 v.s. 45.4572.83 v.s. 27.1761.57 v.s. 38.4361.24 v.s. 38.76
Control-a-Video26.5425.3592.6319.75ours v.s. Control-A-Video68.00 v.s. 32.0078.43 v.s. 21.5771.28 v.s. 29.7256.47 v.s. 43.53
VideoCrafter28.0327.6992.2620.12ours v.s. VideoCrafter52.72 v.s. 47.2871.11 v.s. 28.8960.22 v.s. 39.7860.00 v.s. 40.00
Tune-a-Video25.6425.9592.4220.09ours v.s. Tune-A-Video67.86 v.s. 32.1469.14 v.s. 30.8671.67 v.s. 28.3356.52 v.s. 43.48
ours27.8228.4893.0020.74
", + "image_path": "ddcc2d2cef987346087f3c37c5b04bff2b611475f6ab8253c1eb89be73aa5c00.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 133, + 202, + 480, + 325 + ], + "blocks": [ + { + "bbox": [ + 133, + 202, + 480, + 325 + ], + "lines": [ + { + "bbox": [ + 133, + 202, + 480, + 325 + ], + "spans": [ + { + "bbox": [ + 133, + 202, + 480, + 325 + ], + "type": "image", + "image_path": "f07b22b517f3924135187f8be2ea53bada783e21e4b60984dad8df35282af69c.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 333, + 480, + 367 + ], + "lines": [ + { + "bbox": [ + 130, + 333, + 480, + 367 + ], + "spans": [ + { + "bbox": [ + 130, + 333, + 480, + 367 + ], + "type": "text", + "content": "Fig. 8: Videos generated by MotionDirector. Their appearances are customized according to the reference images shown on the top-left side. Their motions are customized based on the reference videos shown on the top-right side." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 392, + 482, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 392, + 482, + 513 + ], + "spans": [ + { + "bbox": [ + 130, + 392, + 482, + 513 + ], + "type": "text", + "content": "Qualitative and Quantitative Results. As shown in Fig. 7, comparison methods fail to generalize the desired motions to diverse appearances, like the ears of bears and the Arc de Triomphe. Similar to the depth, using the edge and optical flow as control signals also introduces interference from the appearance into the generated videos. In Table. 2, we refer to the alignment between the generated videos and the original 4 editing text prompts as text alignment, and the alignment with the 3 new text prompts with significant changes as appearance diversity. The results show that our method outperforms other methods by a large margin when generalizing the motions to diverse appearances, and achieves competitive motion fidelity." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 533, + 378, + 547 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 533, + 378, + 547 + ], + "spans": [ + { + "bbox": [ + 132, + 533, + 378, + 547 + ], + "type": "text", + "content": "4.3 Customizing Both Appearance and Motion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": "Since MotionDirector learns the appearance and motion in a decoupled manner, it can also reorganize and combine the appearances and motions from different sources. For example, as shown in Fig. 2, MotionDirector can combine the appearance from one video with the motion from another video to generate a new video that exhibits similar appearance and motion characteristics to each of the input videos, respectively. Here we provide more results in Fig. 8. The base model is injected with the spatial LoRAs trained on the given images of \"Terracotta Warrior\" and injected with different temporal LoRAs trained on different sets of videos of different motion concepts, such as \"riding horse\"." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 383, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 383, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 383, + 91, + 447, + 100 + ], + "type": "text", + "content": "MotionDirector" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 116, + 276, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 276, + 128 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 276, + 128 + ], + "type": "text", + "content": "4.4 Efficiency Performance" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 134, + 136, + 481, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 136, + 481, + 243 + ], + "spans": [ + { + "bbox": [ + 134, + 136, + 481, + 243 + ], + "type": "text", + "content": "The lightweight LoRAs enable our method to tune the foundation models efficiently. Taking the foundation model ZeroScope for example, it has over 1.8 billion pre-trained parameters. Each set of trainable spatial and temporal LoRAs only adds 9 million and 12 million parameters, respectively. Requiring 14 GB VRAM, MotionDirector takes 20 minutes to converge on multiple reference videos, and 8 minutes for a single reference video, competitive to the 10 minutes required by Tune-A-Video [70]. After training, MotionDirector can generate diverse videos with learned motions. In the inference stage, MotionDirector does not introduce significant increases in computational costs or inference time." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 134, + 262, + 333, + 274 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 262, + 333, + 274 + ], + "spans": [ + { + "bbox": [ + 134, + 262, + 333, + 274 + ], + "type": "text", + "content": "5 Limitations and Future Works" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 287, + 481, + 359 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 287, + 481, + 359 + ], + "spans": [ + { + "bbox": [ + 134, + 287, + 481, + 359 + ], + "type": "text", + "content": "Despite the MotionDirector can learn the motions of one or two subjects in the reference videos, it is still hard to learn complex motions of multiple subjects, such as a group of boys playing soccer. Previous appearance customization methods suffer similar problems when generating multiple customized subjects [14]. A possible solution is to further decouple the motions of different subjects in the latent space and learn them separately." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 134, + 377, + 218, + 390 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 377, + 218, + 390 + ], + "spans": [ + { + "bbox": [ + 134, + 377, + 218, + 390 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 134, + 402, + 481, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 402, + 481, + 534 + ], + "spans": [ + { + "bbox": [ + 134, + 402, + 481, + 534 + ], + "type": "text", + "content": "We introduce and formulate the task of Motion Customization, which is adapting the pre-trained foundation text-to-video diffusion models to generate videos with desired motions. The challenge of this task is generalizing the customized motions to various appearances. To overcome this challenge, we propose the MotionDirector with a dual-path architecture and a novel appearance-debiased temporal training objective to decouple the learning of appearance and motion. Experimental results show that MotionDirector can learn motion concepts from a limited number of reference videos, and generalize them to diverse appearances. The automatic and human evaluations on two benchmarks demonstrate the MontionDirector outperforms other methods in terms of appearance diversity and motion fidelity." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 134, + 552, + 246, + 567 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 552, + 246, + 567 + ], + "spans": [ + { + "bbox": [ + 134, + 552, + 246, + 567 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 134, + 578, + 481, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 578, + 481, + 638 + ], + "spans": [ + { + "bbox": [ + 134, + 578, + 481, + 638 + ], + "type": "text", + "content": "This research is supported by National Research Foundation, Singapore and A\\*STAR, under its RIE2020 Industry Alignment Fund - Industry Collaboration Projects (IAF-ICP) grant call (Grant No. I2001E0059) - SIA-NUS Digital Aviation Corp Lab. Mike Shou is supported by the Ministry of Education, Singapore, under the Academic Research Fund Tier 1 (FY2023)." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 211, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 211, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 211, + 100 + ], + "type": "text", + "content": "Zhao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 198, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 198, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 198, + 127 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 141, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 138, + 141, + 481, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 141, + 481, + 175 + ], + "spans": [ + { + "bbox": [ + 138, + 141, + 481, + 175 + ], + "type": "text", + "content": "1. Bain, M., Nagrani, A., Varol, G., Zisserman, A.: Frozen in time: A joint video and image encoder for end-to-end retrieval. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1728-1738 (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 176, + 481, + 209 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 176, + 481, + 209 + ], + "spans": [ + { + "bbox": [ + 138, + 176, + 481, + 209 + ], + "type": "text", + "content": "2. Balaji, Y., Min, M.R., Bai, B., Chellappa, R., Graf, H.P.: Conditional gan with discriminative filter generation for text-to-video synthesis. In: IJCAI. vol. 1, p. 2 (2019)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 209, + 481, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 209, + 481, + 241 + ], + "spans": [ + { + "bbox": [ + 138, + 209, + 481, + 241 + ], + "type": "text", + "content": "3. Blattmann, A., Rombach, R., Ling, H., Dockhorn, T., Kim, S.W., Fidler, S., Kreis, K.: Align your latents: High-resolution video synthesis with latent diffusion models. In: CVPR (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 243, + 481, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 243, + 481, + 286 + ], + "spans": [ + { + "bbox": [ + 138, + 243, + 481, + 286 + ], + "type": "text", + "content": "4. Blattmann, A., Rombach, R., Ling, H., Dockhorn, T., Kim, S.W., Fidler, S., Kreis, K.: Align your latents: High-resolution video synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22563-22575 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 287, + 481, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 287, + 481, + 308 + ], + "spans": [ + { + "bbox": [ + 138, + 287, + 481, + 308 + ], + "type": "text", + "content": "5. Chen, T.S., Lin, C.H., Tseng, H.Y., Lin, T.Y., Yang, M.H.: Motion-conditioned diffusion model for controllable video synthesis. arXiv:2304.14404 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 310, + 481, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 310, + 481, + 342 + ], + "spans": [ + { + "bbox": [ + 138, + 310, + 481, + 342 + ], + "type": "text", + "content": "6. Chen, W., Wu, J., Xie, P., Wu, H., Li, J., Xia, X., Xiao, X., Lin, L.: Control-a-video: Controllable text-to-video generation with diffusion models. arXiv preprint arXiv:2305.13840 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 343, + 481, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 343, + 481, + 365 + ], + "spans": [ + { + "bbox": [ + 138, + 343, + 481, + 365 + ], + "type": "text", + "content": "7. Chen, X., Huang, L., Liu, Y., Shen, Y., Zhao, D., Zhao, H.: Anydoor: Zero-shot object-level image customization. arXiv preprint arXiv:2307.09481 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 365, + 481, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 365, + 481, + 398 + ], + "spans": [ + { + "bbox": [ + 138, + 365, + 481, + 398 + ], + "type": "text", + "content": "8. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 248-255. IEEE (2009)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 399, + 481, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 399, + 481, + 421 + ], + "spans": [ + { + "bbox": [ + 138, + 399, + 481, + 421 + ], + "type": "text", + "content": "9. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems 34, 8780-8794 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 422, + 481, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 422, + 481, + 453 + ], + "spans": [ + { + "bbox": [ + 138, + 422, + 481, + 453 + ], + "type": "text", + "content": "0. Duan, Z., You, L., Wang, C., Chen, C., Wu, Z., Qian, W., Huang, J., Chao, F., Ji, R.: Diffsynth: Latent in-iteration deflickering for realistic video synthesis. arXiv:2308.03463 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 454, + 481, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 454, + 481, + 487 + ], + "spans": [ + { + "bbox": [ + 138, + 454, + 481, + 487 + ], + "type": "text", + "content": "1. Esser, P., Chiu, J., Atighehchian, P., Granskog, J., Germanidis, A.: Structure and content-guided video synthesis with diffusion models. arXiv preprint arXiv:2302.03011 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 487, + 481, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 487, + 481, + 520 + ], + "spans": [ + { + "bbox": [ + 138, + 487, + 481, + 520 + ], + "type": "text", + "content": "2. Ge, S., Hayes, T., Yang, H., Yin, X., Pang, G., Jacobs, D., Huang, J.B., Parikh, D.: Long video generation with time-agnostic vqgan and time-sensitive transformer. arXiv preprint arXiv:2204.03638 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 521, + 481, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 521, + 481, + 553 + ], + "spans": [ + { + "bbox": [ + 138, + 521, + 481, + 553 + ], + "type": "text", + "content": "3. Ge, S., Nah, S., Liu, G., Poon, T., Tao, A., Catanzaro, B., Jacobs, D., Huang, J.B., Liu, M.Y., Balaji, Y.: Preserve your own correlation: A noise prior for video diffusion models. arXiv:2305.10474 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 555, + 481, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 555, + 481, + 597 + ], + "spans": [ + { + "bbox": [ + 138, + 555, + 481, + 597 + ], + "type": "text", + "content": "4. Gu, Y., Wang, X., Wu, J.Z., Shi, Y., Chen, Y., Fan, Z., Xiao, W., Zhao, R., Chang, S., Wu, W., et al.: Mix-of-show: Decentralized low-rank adaptation for multi-concept customization of diffusion models. arXiv preprint arXiv:2305.18292 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 599, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 599, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 599, + 481, + 632 + ], + "type": "text", + "content": "5. Guo, Y., Yang, C., Rao, A., Wang, Y., Qiao, Y., Lin, D., Dai, B.: Animatediff: Animate your personalized text-to-image diffusion models without specific tuning. arXiv preprint arXiv:2307.04725 (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 632, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 632, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 632, + 481, + 665 + ], + "type": "text", + "content": "6. He, Y., Xia, M., Chen, H., Cun, X., Gong, Y., Xing, J., Zhang, Y., Wang, X., Weng, C., Shan, Y., et al.: animate-a-story: Storytelling with retrieval-augmented video generation. arXiv:2307.06940 (2023)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 383, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 383, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 383, + 91, + 447, + 100 + ], + "type": "text", + "content": "MotionDirector" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "type": "text", + "content": "17. He, Y., Yang, T., Zhang, Y., Shan, Y., Chen, Q.: Latent video diffusion models for high-fidelity long video generation (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 139, + 480, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 139, + 480, + 160 + ], + "spans": [ + { + "bbox": [ + 133, + 139, + 480, + 160 + ], + "type": "text", + "content": "18. He, Y., Yang, T., Zhang, Y., Shan, Y., Chen, Q.: Latent video diffusion models for high-fidelity video generation with arbitrary lengths. arXiv:2211.13221 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 161, + 480, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 161, + 480, + 194 + ], + "spans": [ + { + "bbox": [ + 133, + 161, + 480, + 194 + ], + "type": "text", + "content": "19. He, Y., Yang, T., Zhang, Y., Shan, Y., Chen, Q.: Videocrafter: A toolkit for text-to-video generation and editing. https://github.com/AILab-CVC/VideoCrafter (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 194, + 480, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 480, + 216 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 480, + 216 + ], + "type": "text", + "content": "20. Hessel, J., Holtzman, A., Forbes, M., Bras, R.L., Choi, Y.: Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 216, + 480, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 480, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 480, + 248 + ], + "type": "text", + "content": "21. Ho, J., Chan, W., Sahara, C., Whang, J., Gao, R., Gritsenko, A., Kingma, D.P., Poole, B., Norouzi, M., Fleet, D.J., et al.: Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 249, + 480, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 480, + 281 + ], + "type": "text", + "content": "22. Ho, J., Chan, W., Sahara, C., Whang, J., Gao, R., Gritsenko, A., Kingma, D.P., Poole, B., Norouzi, M., Fleet, D.J., et al.: Imagen video: High definition video generation with diffusion models. arXiv:2210.02303 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 281, + 480, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 281, + 480, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 281, + 480, + 303 + ], + "type": "text", + "content": "23. Ho, J., Salimans, T., Gritsenko, A., Chan, W., Norouzi, M., Fleet, D.J.: Video diffusion models. In: NeurIPS (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 304, + 480, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 480, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 480, + 325 + ], + "type": "text", + "content": "24. Hong, S., Seo, J., Hong, S., Shin, H., Kim, S.: Large language models are frame-level directors for zero-shot text-to-video generation. arXiv:2305.14330 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 326, + 480, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 480, + 357 + ], + "type": "text", + "content": "25. Hong, W., Ding, M., Zheng, W., Liu, X., Tang, J.: Cogvideo: Large-scale pretraining for text-to-video generation via transformers. arXiv preprint arXiv:2205.15868 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 357, + 480, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 357, + 480, + 390 + ], + "spans": [ + { + "bbox": [ + 132, + 357, + 480, + 390 + ], + "type": "text", + "content": "26. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 391, + 480, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 391, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 391, + 480, + 413 + ], + "type": "text", + "content": "27. Huang, H., Feng, Y., Shi, C., Xu, L., Yu, J., Yang, S.: Free-bloom: Zero-shot text-to-video generator with lvm director and ldm animator. In: NeurIPS (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 414, + 480, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 414, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 414, + 480, + 445 + ], + "type": "text", + "content": "28. Jeong, H., Park, G.Y., Ye, J.C.: Vmc: Video motion customization using temporal attention adaption for text-to-video diffusion models. arXiv preprint arXiv:2312.00845 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 446, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 446, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 446, + 480, + 468 + ], + "type": "text", + "content": "29. Karras, J., Holynski, A., Wang, T.C., Kemelmacher-Shlizerman, I.: Dreampose: Fashion image-to-video synthesis via stable diffusion. arXiv:2304.06025 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 468, + 480, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 480, + 500 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 480, + 500 + ], + "type": "text", + "content": "30. Khachatryan, L., Movsisyan, A., Tadevosyan, V., Henschel, R., Wang, Z., Navasardyan, S., Shi, H.: Text2video-zero: Text-to-image diffusion models are zero-shot video generators. In: ICCV (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 501, + 480, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 480, + 533 + ], + "type": "text", + "content": "31. Kirstain, Y., Polyak, A., Singer, U., Matiana, S., Penna, J., Levy, O.: Pick-a-pic: An open dataset of user preferences for text-to-image generation. arXiv preprint arXiv:2305.01569 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 533, + 480, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 533, + 480, + 566 + ], + "spans": [ + { + "bbox": [ + 132, + 533, + 480, + 566 + ], + "type": "text", + "content": "32. Kumari, N., Zhang, B., Zhang, R., Shechtman, E., Zhu, J.Y.: Multi-concept customization of text-to-image diffusion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1931-1941 (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 567, + 480, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 480, + 588 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 480, + 588 + ], + "type": "text", + "content": "33. Le Moing, G., Ponce, J., Schmid, C.: Ccvs: context-aware controllable video synthesis. Advances in Neural Information Processing Systems 34, 14042-14055 (2021)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 589, + 480, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 589, + 480, + 621 + ], + "spans": [ + { + "bbox": [ + 132, + 589, + 480, + 621 + ], + "type": "text", + "content": "34. Li, X., Chu, W., Wu, Y., Yuan, W., Liu, F., Zhang, Q., Li, F., Feng, H., Ding, E., Wang, J.: Videogen: A reference-guided latent diffusion approach for high definition text-to-video generation. arXiv preprint arXiv:2309.00398 (2023)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 621, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 621, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 621, + 480, + 643 + ], + "type": "text", + "content": "35. Lian, L., Shi, B., Yala, A., Darrell, T., Li, B.: Llm-grounded video diffusion models. arXiv:2309.17444 (2023)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "text", + "content": "36. Luo, Z., Chen, D., Zhang, Y., Huang, Y., Wang, L., Shen, Y., Zhao, D., Zhou, J., Tan, T.: Videofusion: Decomposed diffusion models for high-quality video gen" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 211, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 211, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 211, + 100 + ], + "type": "text", + "content": "Zhao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 147, + 116, + 481, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 116, + 481, + 138 + ], + "spans": [ + { + "bbox": [ + 147, + 116, + 481, + 138 + ], + "type": "text", + "content": "eration. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 139, + 481, + 170 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 139, + 481, + 170 + ], + "spans": [ + { + "bbox": [ + 133, + 139, + 481, + 170 + ], + "type": "text", + "content": "37. Ma, Y., He, Y., Cun, X., Wang, X., Shan, Y., Li, X., Chen, Q.: Follow your pose: Pose-guided text-to-video generation using pose-free videos. arXiv preprint arXiv:2304.01186 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 171, + 481, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 171, + 481, + 203 + ], + "spans": [ + { + "bbox": [ + 132, + 171, + 481, + 203 + ], + "type": "text", + "content": "38. Ma, Y., He, Y., Cun, X., Wang, X., Shan, Y., Li, X., Chen, Q.: Follow your pose: Pose-guided text-to-video generation using pose-free videos. arXiv:2304.01186 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 203, + 481, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 203, + 481, + 236 + ], + "spans": [ + { + "bbox": [ + 132, + 203, + 481, + 236 + ], + "type": "text", + "content": "39. Materzynska, J., Sivic, J., Shechtman, E., Torralba, A., Zhang, R., Russell, B.: Customizing motion in text-to-video diffusion models. arXiv preprint arXiv:2312.04966 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 236, + 481, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 236, + 481, + 257 + ], + "spans": [ + { + "bbox": [ + 132, + 236, + 481, + 257 + ], + "type": "text", + "content": "40. Mei, K., Patel, V.: Vidm: Video implicit diffusion models. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 9117-9125 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 257, + 481, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 257, + 481, + 289 + ], + "spans": [ + { + "bbox": [ + 132, + 257, + 481, + 289 + ], + "type": "text", + "content": "41. Ni, H., Shi, C., Li, K., Huang, S.X., Min, M.R.: Conditional image-to-video generation with latent flow diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18444-18455 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 289, + 481, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 289, + 481, + 320 + ], + "spans": [ + { + "bbox": [ + 132, + 289, + 481, + 320 + ], + "type": "text", + "content": "42. Qin, B., Ye, W., Yu, Q., Tang, S., Zhuang, Y.: Dancing avatar: Pose and text-guided human motion videos synthesis with image diffusion model. arXiv:2308.07749 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 320, + 481, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 320, + 481, + 353 + ], + "spans": [ + { + "bbox": [ + 132, + 320, + 481, + 353 + ], + "type": "text", + "content": "43. Ren, Y., Zhou, Y., Yang, J., Shi, J., Liu, D., Liu, F., Kwon, M., Shrivastava, A.: Customize-a-video: One-shot motion customization of text-to-video diffusion models. arXiv preprint arXiv:2402.14780 (2024)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 353, + 481, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 353, + 481, + 374 + ], + "spans": [ + { + "bbox": [ + 132, + 353, + 481, + 374 + ], + "type": "text", + "content": "44. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: MICCAI (2015)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 374, + 481, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 374, + 481, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 374, + 481, + 418 + ], + "type": "text", + "content": "45. Ruiz, N., Li, Y., Jampani, V., Pritch, Y., Rubinstein, M., Aberman, K.: Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22500-22510 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 418, + 481, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 418, + 481, + 439 + ], + "spans": [ + { + "bbox": [ + 132, + 418, + 481, + 439 + ], + "type": "text", + "content": "46. Ryu, S.: Low-rank adaptation for fast text-to-image diffusion fine-tuning. https://github.com/cloneofsimo/lora (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 439, + 481, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 439, + 481, + 471 + ], + "spans": [ + { + "bbox": [ + 132, + 439, + 481, + 471 + ], + "type": "text", + "content": "47. Saito, M., Matsumoto, E., Saito, S.: Temporal generative adversarial nets with singular value clipping. In: Proceedings of the IEEE international conference on computer vision. pp. 2830-2839 (2017)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 471, + 481, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 471, + 481, + 514 + ], + "spans": [ + { + "bbox": [ + 132, + 471, + 481, + 514 + ], + "type": "text", + "content": "48. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems 35, 25278-25294 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 514, + 481, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 514, + 481, + 547 + ], + "spans": [ + { + "bbox": [ + 132, + 514, + 481, + 547 + ], + "type": "text", + "content": "49. Shen, X., Li, X., Elhoseiny, M.: Mostgan-v: Video generation with temporal motion styles. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5652-5661 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 547, + 481, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 547, + 481, + 579 + ], + "spans": [ + { + "bbox": [ + 132, + 547, + 481, + 579 + ], + "type": "text", + "content": "50. Singer, U., Polyak, A., Hayes, T., Yin, X., An, J., Zhang, S., Hu, Q., Yang, H., Ashual, O., Gafni, O., et al.: Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792 (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 579, + 481, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 579, + 481, + 611 + ], + "spans": [ + { + "bbox": [ + 132, + 579, + 481, + 611 + ], + "type": "text", + "content": "51. Skorokhodov, I., Tulyakov, S., Elhoseiny, M.: Stylegan-v: A continuous video generator with the price, image quality and perks of stylegan2. arXiv preprint arXiv:2112.14683 (2021)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 611, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 481, + 643 + ], + "type": "text", + "content": "52. Smith, J.S., Hsu, Y.C., Zhang, L., Hua, T., Kira, Z., Shen, Y., Jin, H.: Continual diffusion: Continual customization of text-to-image diffusion with c-lora. arXiv preprint arXiv:2304.06027 (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 643, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 481, + 666 + ], + "type": "text", + "content": "53. Soomro, K., Zamir, A.R.: Action recognition in realistic sports videos. In: Computer vision in sports, pp. 181-208. Springer (2015)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 383, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 383, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 383, + 91, + 447, + 100 + ], + "type": "text", + "content": "MotionDirector" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 666 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "text", + "content": "54. Srivastava, N., Mansimov, E., Salakhudinov, R.: Unsupervised learning of video representations using lstms. In: International conference on machine learning. pp. 843-852. PMLR (2015)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 480, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 480, + 172 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 480, + 172 + ], + "type": "text", + "content": "55. Sterling, S.: Zeroscope. https://huggingface.co/cerspense/zeroscope_v2_576w (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 172, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 172, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 133, + 172, + 480, + 205 + ], + "type": "text", + "content": "56. Tian, Y., Ren, J., Chai, M., Olszewski, K., Peng, X., Metaxas, D.N., Tulyakov, S.: A good image generator is what you need for high-resolution video synthesis. In: International Conference on Learning Representations (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 205, + 480, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 480, + 237 + ], + "type": "text", + "content": "57. Tulyakov, S., Liu, M.Y., Yang, X., Kautz, J.: Mocogan: Decomposing motion and content for video generation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1526-1535 (2018)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 237, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 237, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 237, + 480, + 270 + ], + "type": "text", + "content": "58. Voleti, V., Jolicoeur-Martineau, A., Pal, C.: Mcvd-masked conditional video diffusion for prediction, generation, and interpolation. Advances in Neural Information Processing Systems 35, 23371-23385 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 270, + 480, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 270, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 270, + 480, + 293 + ], + "type": "text", + "content": "59. Vondrick, C., Pirsiavash, H., Torralba, A.: Generating videos with scene dynamics. Advances in neural information processing systems 29 (2016)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 293, + 480, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 480, + 315 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 480, + 315 + ], + "type": "text", + "content": "60. Wang, J., Yuan, H., Chen, D., Zhang, Y., Wang, X., Zhang, S.: Modelscope text-to-video technical report. arXiv preprint arXiv:2308.06571 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 315, + 480, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 480, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 480, + 346 + ], + "type": "text", + "content": "61. Wang, T., Li, L., Lin, K., Lin, C.C., Yang, Z., Zhang, H., Liu, Z., Wang, L.: Disco: Disentangled control for referring human dance generation in real world. arXiv:2307.00040 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 347, + 480, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 480, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 480, + 380 + ], + "type": "text", + "content": "62. Wang, W., Yang, H., Tuo, Z., He, H., Zhu, J., Fu, J., Liu, J.: Videofactory: Swap attention in spatiotemporal diffusions for text-to-video generation. arXiv:2305.10874 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 380, + 480, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 480, + 413 + ], + "type": "text", + "content": "63. Wang, X., Yuan, H., Zhang, S., Chen, D., Wang, J., Zhang, Y., Shen, Y., Zhao, D., Zhou, J.: Videocomposer: Compositional video synthesis with motion controllability. arXiv preprint arXiv:2306.02018 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 413, + 480, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 413, + 480, + 466 + ], + "spans": [ + { + "bbox": [ + 132, + 413, + 480, + 466 + ], + "type": "text", + "content": "64. Wang, Y., Chen, X., Ma, X., Zhou, S., Huang, Z., Wang, Y., Yang, C., He, Y., Yu, J., der Yang, P., Guo, Y., Wu, T., Si, C., Jiang, Y., Chen, C., Loy, C.C., Dai, B., Lin, D., Qiao, Y., Liu, Z.: Lavie: High-quality video generation with cascaded latent diffusion models (2023), https://api-semanticscholar.org/CorpusID: 262823915" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 468, + 480, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 480, + 500 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 480, + 500 + ], + "type": "text", + "content": "65. Wang, Y., Chen, X., Ma, X., Zhou, S., Huang, Z., Wang, Y., Yang, C., He, Y., Yu, J., Yang, P., et al.: Lavie: High-quality video generation with cascaded latent diffusion models. arXiv:2309.15103 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 500, + 480, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 500, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 500, + 480, + 533 + ], + "type": "text", + "content": "66. Wang, Z., Yuan, Z., Wang, X., Chen, T., Xia, M., Luo, P., Shan, Y.: Motionctrl: A unified and flexible motion controller for video generation. arXiv preprint arXiv:2312.03641 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 533, + 480, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 533, + 480, + 566 + ], + "spans": [ + { + "bbox": [ + 132, + 533, + 480, + 566 + ], + "type": "text", + "content": "67. Wei, Y., Zhang, S., Qing, Z., Yuan, H., Liu, Z., Liu, Y., Zhang, Y., Zhou, J., Shan, H.: Dreamvideo: Composing your dream videos with customized subject and motion. arXiv preprint arXiv:2312.04433 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 566, + 480, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 566, + 480, + 600 + ], + "spans": [ + { + "bbox": [ + 132, + 566, + 480, + 600 + ], + "type": "text", + "content": "68. Wei, Y., Zhang, Y., Ji, Z., Bai, J., Zhang, L., Zuo, W.: Elite: Encoding visual concepts into textual embeddings for customized text-to-image generation. arXiv preprint arXiv:2302.13848 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "type": "text", + "content": "69. Wu, J.Z., Gao, D., Bai, J., Shou, M., Li, X., Dong, Z., Singh, A., Keutzer, K., Iandola, F.: The text-guided video editing benchmark at loveu 2023. https:// sites.google.com/view/loveucvpr23/track4 (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 632, + 480, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 666 + ], + "type": "text", + "content": "70. Wu, J.Z., Ge, Y., Wang, X., Lei, W., Gu, Y., Hsu, W., Shan, Y., Qie, X., Shou, M.Z.: Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. arXiv preprint arXiv:2212.11565 (2022)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 211, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 211, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 211, + 100 + ], + "type": "text", + "content": "Zhao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 589 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "71. Wu, J.Z., Ge, Y., Wang, X., Lei, W., Gu, Y., Hsu, W., Shan, Y., Qie, X., Shou, M.Z.: Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In: ICCV (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 481, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 481, + 172 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 481, + 172 + ], + "type": "text", + "content": "72. Wu, R., Chen, L., Yang, T., Guo, C., Li, C., Zhang, X.: Lamp: Learn a motion pattern for few-shot-based video generation. arXiv preprint arXiv:2310.10769 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 172, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 172, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 130, + 172, + 481, + 205 + ], + "type": "text", + "content": "73. Xing, J., Xia, M., Liu, Y., Zhang, Y., Zhang, Y., He, Y., Liu, H., Chen, H., Cun, X., Wang, X., et al.: Make-your-video: Customized video generation using textual and structural guidance. arXiv:2306.00943 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 205, + 481, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 205, + 481, + 226 + ], + "spans": [ + { + "bbox": [ + 130, + 205, + 481, + 226 + ], + "type": "text", + "content": "74. Xing, Z., Feng, Q., Chen, H., Dai, Q., Hu, H., Xu, H., Wu, Z., Jiang, Y.G.: A survey on video diffusion models. arXiv preprint arXiv:2310.10647 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 226, + 481, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 226, + 481, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 226, + 481, + 248 + ], + "type": "text", + "content": "75. Yan, W., Zhang, Y., Abbeel, P., Srinivas, A.: Videogpt: Video generation using vq-vae and transformers. arXiv preprint arXiv:2104.10157 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 248, + 481, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 248, + 481, + 282 + ], + "spans": [ + { + "bbox": [ + 130, + 248, + 481, + 282 + ], + "type": "text", + "content": "76. Yang, S., Hou, L., Huang, H., Ma, C., Wan, P., Zhang, D., Chen, X., Liao, J.: Direct-a-video: Customized video generation with user-directed camera movement and object motion. arXiv preprint arXiv:2402.03162 (2024)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 281, + 481, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 281, + 481, + 314 + ], + "spans": [ + { + "bbox": [ + 130, + 281, + 481, + 314 + ], + "type": "text", + "content": "77. Yin, S., Wu, C., Liang, J., Shi, J., Li, H., Ming, G., Duan, N.: Dragnuwa: Fine-grained control in video generation by integrating text, image, and trajectory. arXiv:2308.08089 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 314, + 481, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 314, + 481, + 347 + ], + "spans": [ + { + "bbox": [ + 130, + 314, + 481, + 347 + ], + "type": "text", + "content": "78. Yin, S., Wu, C., Yang, H., Wang, J., Wang, X., Ni, M., Yang, Z., Li, L., Liu, S., Yang, F., et al.: Nuwa-xl: Diffusion over diffusion for extremely long video generation. arXiv:2303.12346 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 347, + 481, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 347, + 481, + 380 + ], + "spans": [ + { + "bbox": [ + 130, + 347, + 481, + 380 + ], + "type": "text", + "content": "79. Yu, S., Sohn, K., Kim, S., Shin, J.: Video probabilistic diffusion models in projected latent space. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18456-18466 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 380, + 481, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 380, + 481, + 413 + ], + "spans": [ + { + "bbox": [ + 130, + 380, + 481, + 413 + ], + "type": "text", + "content": "80. Yu, S., Tack, J., Mo, S., Kim, H., Kim, J., Ha, J.W., Shin, J.: Generating videos with dynamics-aware implicit generative adversarial networks. In: International Conference on Learning Representations (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 413, + 481, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 413, + 481, + 445 + ], + "spans": [ + { + "bbox": [ + 130, + 413, + 481, + 445 + ], + "type": "text", + "content": "81. Zhang, D.J., Wu, J.Z., Liu, J.W., Zhao, R., Ran, L., Gu, Y., Gao, D., Shou, M.Z.: Show-1: Marrying pixel and latent diffusion models for text-to-video generation. arXiv:2309.15818 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 445, + 481, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 445, + 481, + 479 + ], + "spans": [ + { + "bbox": [ + 130, + 445, + 481, + 479 + ], + "type": "text", + "content": "82. Zhang, D.J., Wu, J.Z., Liu, J.W., Zhao, R., Ran, L., Gu, Y., Gao, D., Shou, M.Z.: Show-1: Marrying pixel and latent diffusion models for text-to-video generation (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 479, + 481, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 479, + 481, + 501 + ], + "spans": [ + { + "bbox": [ + 130, + 479, + 481, + 501 + ], + "type": "text", + "content": "83. Zhang, L., Agrawala, M.: Adding conditional control to text-to-image diffusion models. arXiv preprint arXiv:2302.05543 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 501, + 481, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 501, + 481, + 534 + ], + "spans": [ + { + "bbox": [ + 130, + 501, + 481, + 534 + ], + "type": "text", + "content": "84. Zhao, R., Gu, Y., Wu, J.Z., Zhang, D.J., Liu, J., Wu, W., Keppo, J., Shou, M.Z.: Motiondirector: Motion customization of text-to-video diffusion models. arXiv preprint arXiv:2310.08465 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 533, + 481, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 481, + 566 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 481, + 566 + ], + "type": "text", + "content": "85. Zhao, S., Chen, D., Chen, Y.C., Bao, J., Hao, S., Yuan, L., Wong, K.Y.K.: Uncontrolnet: All-in-one control to text-to-image diffusion models. arXiv preprint arXiv:2305.16322 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 566, + 481, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 566, + 481, + 589 + ], + "spans": [ + { + "bbox": [ + 130, + 566, + 481, + 589 + ], + "type": "text", + "content": "86. Zhou, D., Wang, W., Yan, H., Lv, W., Zhu, Y., Feng, J.: Magicvideo: Efficient video generation with latent diffusion models. arXiv:2211.11018 (2022)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 383, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 383, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 383, + 91, + 447, + 100 + ], + "type": "text", + "content": "MotionDirector" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/MotionLCM_ Real-time Controllable Motion Generation via Latent Consistency Model/69320bc9-c9fe-4fc1-8e78-3d61fd4c6087_content_list.json b/2024/MotionLCM_ Real-time Controllable Motion Generation via Latent Consistency Model/69320bc9-c9fe-4fc1-8e78-3d61fd4c6087_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..3142c0841cd889b69d11abb1180564fdf94c2371 --- /dev/null +++ b/2024/MotionLCM_ Real-time Controllable Motion Generation via Latent Consistency Model/69320bc9-c9fe-4fc1-8e78-3d61fd4c6087_content_list.json @@ -0,0 +1,1617 @@ +[ + { + "type": "text", + "text": "MotionLCM: Real-time Controllable Motion Generation via Latent Consistency Model", + "text_level": 1, + "bbox": [ + 246, + 140, + 754, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Wenxun Dai $^{1,2}$ , Ling-Hao Chen $^{2\\star}$ , Jingbo Wang $^{3\\boxtimes}$ , Jinpeng Liu $^{1,2}$ , Bo Dai $^{3\\boxtimes}$ , Yansong Tang $^{1,2}$", + "bbox": [ + 238, + 210, + 764, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1Shenzhen Key Laboratory of Ubiquitous Data Enabling, Tsinghua Shenzhen International Graduate School 2Tsinghua University 3Shanghai AI Laboratory {wx dai 2001, thu. lhchen, wangjingbo1219, liu. jinpeng.55}@gmail.com {doubledaibo, tangyansong15}@gmail.com Project page: https://dai-wenxun.github.io/MotionLCM-page", + "bbox": [ + 235, + 253, + 766, + 325 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/47c7ee0ac1d6278ba282f678ec7f939d12be6b7c6e0c3474f4d210e856f07a82.jpg", + "image_caption": [ + "Fig. 1: We propose MotionLCM, a real-time controllable motion latent consistency model. Our model uses the last few frames of the previous motion as temporal control signals to autoregressively generate the next motion in real-time under different text prompts. Green blocks denote the junctions. The numbers in red are the inference time." + ], + "image_footnote": [], + "bbox": [ + 240, + 359, + 763, + 614 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. This work introduces MotionLCM, extending controllable motion generation to a real-time level. Existing methods for spatial-temporal control in text-conditioned motion generation suffer from significant runtime inefficiency. To address this issue, we first propose the motion latent consistency model (MotionLCM) for motion generation,", + "bbox": [ + 259, + 746, + 740, + 816 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Project lead. ☑ Corresponding author.", + "bbox": [ + 218, + 824, + 509, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "building upon the latent diffusion model [9]. By adopting one-step (or few-step) inference, we further improve the runtime efficiency of the motion latent diffusion model for motion generation. To ensure effective controllability, we incorporate a motion ControlNet within the latent space of MotionLCM and enable explicit control signals (e.g., initial poses) in the vanilla motion space to control the generation process directly, similar to controlling other latent-free diffusion models [29, 73] for motion generation. By employing these techniques, our approach can generate human motions with text and control signals in real-time. Experimental results demonstrate the remarkable generation and controlling capabilities of MotionLCM while maintaining real-time runtime efficiency.", + "bbox": [ + 259, + 147, + 743, + 301 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Keywords: Text-to-Motion $\\cdot$ Real-time Control $\\cdot$ Consistency Model", + "bbox": [ + 261, + 313, + 727, + 327 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 354, + 375, + 369 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Text-to-motion generation (T2M) has attracted increasing attention [1, 15, 43, 49, 65] due to its important roles in many applications [70, 72]. Previous attempts mainly focus on GANs [1, 38], VAEs [3, 19, 48, 49] and diffusion models [9, 11, 39, 56, 65, 78, 83] via pairwise text-motion data [17, 26, 45, 52, 53, 57, 63, 74] and achieve impressive generation results. Existing approaches [9, 65, 83] mainly take diffusion models [20, 46, 55, 60] as a base generative model, owing to their powerful ability to model motion distribution. However, these diffusion fashions inevitably require considerable sampling steps for motion synthesis during inference, even with some sampling acceleration methods [61]. Specifically, MDM [65] and MLD [9] require $\\sim$ 12s and $\\sim$ 0.2s to generate", + "bbox": [ + 212, + 386, + 478, + 718 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/a4e5b6df9d8d0ad07f9ba9a2d1f4e85d197fee9e6e870e803aa016fd407bf094.jpg", + "image_caption": [ + "Fig.2: Comparison of the inference time costs on HumanML3D [17]. We compare the AITS and FID metrics with five SOTA methods. The closer the model is to the origin the better. Diffusion-based models are indicated by the blue dashed box. Our MotionLCM achieves real-time inference speed while ensuring high-quality motion generation." + ], + "image_footnote": [], + "bbox": [ + 488, + 415, + 787, + 556 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "a high-quality motion sequence. Such low efficiency blocks the applications of generating high-quality motions in various real-time scenarios.", + "bbox": [ + 212, + 719, + 787, + 748 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In addition to the language description itself serving as a coarse control signal, another line of research focuses on controlling the motion generation with spatial-temporal constraints [29, 56, 73]. Although these attempts enjoy impressive controlling ability in the T2M task, there still exists a significant gap towards real-time applications. For example, OmniControl [73] exhibits a relatively long inference time, $\\sim 81$ s per sequence. Therefore, trading-off between generation", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "W. Dai et al.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "quality and efficiency is a challenging problem. As a result, in this paper, we target the real-time controllable motion generation research problem.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recently, the concept of consistency models [44, 62] has been introduced in image generation, resulting in significant progress by enabling efficient and high-fidelity image synthesis with a minimal number of sampling steps (e.g., 4 steps vs. 50 steps). These properties perfectly align with our goal of accelerating motion generation without compromising generation quality. Therefore, we propose MotionLCM (Motion Latent Consistency Model) distilled from the motion latent diffusion model, MLD [9], to tackle the low-efficiency problem in diffusion sampling. To the best of our knowledge, we introduce consistency distillation into the motion generation area for the first time and accelerate motion generation to a real-time level via latent consistency distillation [44].", + "bbox": [ + 212, + 176, + 785, + 327 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Here, in MotionLCM, we are facing another challenge on how to control motions with spatial-temporal signals (e.g., initial poses) in the latent space. Previous methods [56, 73] model human motions in the vanilla motion space and can manipulate the motion directly in the denoising process. However, for our latent-diffusion-based MotionLCM, it is non-trivial to feed the control signals into the latent space. This is because the latent has no explicit motion semantics, which cannot be manipulated directly by the control signals. Inspired by the notable success of [82] in controllable image generation [55], we introduce a motion ControlNet to control motion generation in the latent space. However, the naïve motion ControlNet is not sufficient to provide supervision for the control signals. The main reason is the lack of explicit supervision in the motion space. Therefore, during the training phase, we decode the predicted latent through the frozen VAE [30] decoder into the vanilla motion space to provide explicit control supervision on the generated motion. Thanks to the powerful one-step inference capability of MotionLCM, the latent generated by MotionLCM can significantly facilitate control supervision both in the latent space and motion space for training the motion ControlNet compared to MLD [9].", + "bbox": [ + 212, + 328, + 787, + 584 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "With our key designs, our proposed MotionLCM successfully enjoys the balance between the generation quality and efficiency in controllable motion generation. Before delivering into detail, we sum up our core contributions as follows.", + "bbox": [ + 212, + 584, + 785, + 630 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose the Motion Latent Consistency Model (MotionLCM) via consistency distillation on the motion latent diffusion model extending controllable motion generation to a real-time level.", + "- Building upon our achievement of real-time motion generation, we introduce a motion ControlNet, enabling high-quality controllable motion generation.", + "- Extensive experimental results show that MotionLCM enjoys a good balance of generation quality, controlling capability, and real-time efficiency." + ], + "bbox": [ + 225, + 636, + 782, + 739 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 760, + 387, + 776 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Human Motion Generation", + "text_level": 1, + "bbox": [ + 215, + 787, + 486, + 801 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Generating human motions can be divided into three main fashions according to inputs: motion synthesis (1) without any condition [54,65,77,84], (2) with some", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "MotionLCM", + "bbox": [ + 645, + 114, + 730, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "given multi-modal conditions, such as action labels [6, 12, 19, 31, 48, 75], textual description [1-5, 7, 7, 9-11, 13-15, 17, 18, 24, 27, 29, 37-39, 43, 49-51, 56, 64, 65, 68-70, 72, 78, 81, 83, 85, 86], audio or music [32-35, 59, 66], (3) with user-defined trajectories [22, 23, 29, 36, 56, 58, 68, 71, 73]. To generate diverse, natural, and high-quality human motions, many generative models have been explored by [2, 38, 49, 79, 80]. Recently, diffusion-based models significantly improved the motion generation performance and diversity [8, 9, 11, 16, 65, 76, 83] with stable training. Specifically, MotionDiffuse [83] represents the first text-based motion diffusion model that provides fine-grained instructions on body parts and achieves arbitrary-length motion synthesis with time-varied text prompts. MDM [65] introduces a motion diffusion model that operates on raw motion data, enabling both high-quality generation and generic conditioning that together comprise a good baseline for new motion generation tasks. Based on MDM [65], OmniControl [73] integrates flexible spatial-temporal control signals across different joints by combining analytic spatial guidance and realism guidance into the diffusion model, ensuring that the generated motion closely conforms to the input control signals. The work most relevant to ours is MLD [9], which introduces a motion latent-based diffusion model to enhance generation quality and reduce computational resource requirements. The key idea is training a VAE [30] for motion embedding, followed by implementing latent diffusion [55] within the learned latent space. However, these diffusion fashions inevitably require considerable sampling steps for motion synthesis during inference, even with some sampling acceleration methods [61]. Thus, we propose MotionLCM, which not only guarantees high-quality controllable motion generation but also achieves real-time runtime efficiency.", + "bbox": [ + 212, + 146, + 787, + 508 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 536, + 330, + 553 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we first briefly introduce preliminaries about latent consistency models in Sec. 3.1. Then, we describe how to conduct latent consistency distillation for motion generation in Sec. 3.2, followed by our implementation of motion control in latent space in Sec. 3.3. The overall pipeline is illustrated in Fig. 4.", + "bbox": [ + 212, + 571, + 787, + 633 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Preliminaries", + "text_level": 1, + "bbox": [ + 215, + 659, + 372, + 674 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The Consistency Model (CM) [62] introduces a kind of efficient generative model designed for efficient one-step or few-step generation. Given a Probability Flow ODE (a.k.a. PF-ODE) that smoothly converts data to noise, the CM is to learn the function $\\pmb{f}(\\cdot, \\cdot)$ that maps any points on the ODE trajectory to its origin distribution (i.e., the solution of the PF-ODE). The consistency function is formally defined as $\\pmb{f}:(\\mathbf{x}_t,t)\\longmapsto \\mathbf{x}_{\\epsilon}$ , where $t\\in [\\epsilon ,T]$ , $T > 0$ is a fixed constant and $\\epsilon$ is a small positive number to avoid numerical instability. According to [62], the consistency function should satisfy the self-consistency property:", + "bbox": [ + 212, + 688, + 787, + 811 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {f} \\left(\\mathbf {x} _ {t}, t\\right) = \\boldsymbol {f} \\left(\\mathbf {x} _ {t ^ {\\prime}}, t ^ {\\prime}\\right), \\forall t, t ^ {\\prime} \\in [ \\epsilon , T ]. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 383, + 825, + 785, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "W. Dai et al.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/5e7d75e984cbe2d5e45177e83441b38e9f8635bbd8f86ef4e4da593e1f9802ab.jpg", + "image_caption": [ + "Distillation target of $f_{\\Theta}$ : Project any $\\mathbf{z}_t$ to $\\mathbf{z}_0$ consistently. $t \\in [0, T]$" + ], + "image_footnote": [], + "bbox": [ + 248, + 146, + 560, + 195 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/b475d73358e88bd5d634c16733218e43460cfb9ce878977325639f72136dd0db.jpg", + "image_caption": [ + "Fig. 3: The training objective of consistency distillation is to learn a consistency function $f_{\\Theta}$ , initialized with the parameters of a pre-trained diffusion model (e.g., MLD [9]). This function $f_{\\Theta}$ should projects any points (i.e., $\\mathbf{z}_t$ ) on the ODE trajectory to its solution (i.e., $\\mathbf{z}_0$ ). Once the pre-trained model [9] is distilled, unlike the traditional denoising model [65, 83] that requires considerable sampling steps, our MotionLCM can generate high-quality motion sequences with one-step sampling and further improve the generation quality through multi-step inference." + ], + "image_footnote": [], + "bbox": [ + 246, + 196, + 578, + 327 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/846916882c9ecd29231b43b048d6d08e6ffc143c2beee9eb0b33183dd4d1a56a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 588, + 198, + 751, + 327 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in Eq. (1), the self-consistency property indicates that for arbitrary pairs of $(\\mathbf{x}_t,t)$ on the same PF-ODE trajectory, the outputs of the model should be consistent. The goal of a parameterized consistency model $f_{\\Theta}$ is to learn a consistency function from data by enforcing the self-consistency property in Eq. (1). To ensure that $f_{\\Theta}(\\mathbf{x},\\epsilon) = \\mathbf{x}$ , the consistency model $f_{\\Theta}$ is parameterized as,", + "bbox": [ + 212, + 464, + 787, + 542 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {f} _ {\\Theta} (\\mathbf {x}, t) = c _ {\\text {s k i p}} (t) \\mathbf {x} + c _ {\\text {o u t}} (t) \\boldsymbol {F} _ {\\Theta} (\\mathbf {x}, t), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 359, + 551, + 784, + 566 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $c_{\\mathrm{skip}}(t)$ and $c_{\\mathrm{out}}(t)$ are differentiable functions with $c_{\\mathrm{skip}}(\\epsilon) = 1$ and $c_{\\mathrm{out}}(\\epsilon) = 0$ , and $\\pmb{F}_{\\Theta}(\\cdot, \\cdot)$ is a deep neural network to learn the self-consistency. The CM trained from distilling the knowledge of pre-trained diffusion models is called Consistency Distillation. The consistency loss is defined as follows,", + "bbox": [ + 212, + 577, + 787, + 638 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} (\\boldsymbol {\\Theta}, \\boldsymbol {\\Theta} ^ {-}; \\boldsymbol {\\Phi}) = \\mathbb {E} \\left[ d \\left(\\boldsymbol {f} _ {\\boldsymbol {\\Theta}} \\left(\\mathbf {x} _ {t _ {n + 1}}, t _ {n + 1}\\right), \\boldsymbol {f} _ {\\boldsymbol {\\Theta} ^ {-}} \\left(\\hat {\\mathbf {x}} _ {t _ {n}} ^ {\\boldsymbol {\\Phi}}, t _ {n}\\right)\\right) \\right], \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 305, + 646, + 784, + 666 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $d(\\cdot, \\cdot)$ is a chosen metric function for measuring the distance between two samples. $f_{\\Theta}(\\cdot, \\cdot)$ and $f_{\\Theta^{-}}(\\cdot, \\cdot)$ are referred to as \"online network\" and \"target network\" according to [62]. Besides, $\\Theta^{-}$ is updated with the exponential moving average (EMA) of the parameters of $\\Theta^{1}$ . In Eq. (3), $\\hat{\\mathbf{x}}_{t_n}^{\\Phi}$ is a one-step estimation of $\\mathbf{x}_{t_n}$ from $\\mathbf{x}_{t_{n+1}}$ , which is formulated as,", + "bbox": [ + 212, + 674, + 787, + 750 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {x}} _ {t _ {n}} ^ {\\Phi} \\leftarrow \\mathbf {x} _ {t _ {n + 1}} + (t _ {n} - t _ {n + 1}) \\Phi (\\mathbf {x} _ {t _ {n + 1}}, t _ {n + 1}, \\emptyset), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 341, + 761, + 784, + 779 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\Phi$ is a one-step ODE solver applied to PF-ODE.", + "bbox": [ + 212, + 787, + 612, + 803 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "MotionLCM", + "bbox": [ + 645, + 114, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "$^{1}$ EMA operation: $\\Theta^{-} \\gets \\mathbf{sg}(\\mu \\Theta^{-} + (1 - \\mu)\\Theta)$ , where $\\mathbf{sg}(\\cdot)$ denotes the stopgrad operation and $\\mu$ satisfies $0 \\leq \\mu < 1$ .", + "bbox": [ + 217, + 809, + 785, + 840 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/eab86df2a1587c50a0a037091a3594a96ccb0429d2d45e4eae44dd1bdba9eaeb.jpg", + "image_caption": [ + "(a) Motion Latent Consistency Distillation" + ], + "image_footnote": [], + "bbox": [ + 218, + 147, + 571, + 281 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7c2e2dde8b41f797545ffbbe271b2039f5bdac06f2bbd77b7c859b86f03f9a58.jpg", + "image_caption": [ + "(b) Motion Control in Latent Space", + "Fig. 4: The overview of MotionLCM. (a) Motion Latent Consistency Distillation (Sec. 3.2). Given a raw motion sequence $\\mathbf{x}_0^{1:N}$ , a pre-trained VAE [30] encoder first compresses it into the latent space, then a forward diffusion operation is performed to add $n + k$ steps of noise. Then, the noisy $\\mathbf{z}_{n + k}$ is fed into the online network and teacher network to predict the clean latent. The target network takes the $k$ -step estimation results of the teacher output to predict the clean latent. To learn self-consistency, a loss is applied to enforce the output of the online network and target network to be consistent. (b) Motion Control in Latent Space (Sec. 3.3). With the powerful MotionLCM trained in the first stage, we incorporate a motion ControlNet into the MotionLCM to achieve controllable motion generation. Furthermore, we leverage the decoded motion to explicitly supervise the spatial-temporal control signals (i.e., initial poses $\\mathbf{g}^{1:\\tau}$ )." + ], + "image_footnote": [], + "bbox": [ + 575, + 143, + 785, + 282 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The Latent Consistency Model (LCM) [44] learns the self-consistency property in the latent space $D_{\\mathbf{z}} = \\{(\\mathbf{z},\\mathbf{c})|\\mathbf{z} = \\mathcal{E}(\\mathbf{x}),(\\mathbf{x},\\mathbf{c})\\in D\\}$ , where $D$ denotes the dataset, $\\mathbf{c}$ is the given condition, and $\\mathcal{E}$ is the pre-trained encoder. Compared to CMs [62] using the numerical continuous PF-ODE solver [28], LCMs [44] adopt the discrete-time schedule [41,42,61] to adapt to Stable Diffusion [55]. Instead of ensuring consistency between adjacent time steps $t_{n + 1}\\rightarrow t_n$ , LCMs [44] are designed to ensure consistency between the current time step and $k$ -step away, i.e., $t_{n + k}\\to t_n$ , thereby significantly reducing convergence time costs. As classifier-free guidance (CFG) [21] plays a crucial role in synthesizing high-quality text-aligned images, LCMs integrate CFG into the distillation as follows,", + "bbox": [ + 212, + 492, + 787, + 645 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {z}} _ {t _ {n}} ^ {\\boldsymbol {\\Phi}, w} \\gets \\mathbf {z} _ {t _ {n + k}} + (1 + w) \\boldsymbol {\\Phi} (\\mathbf {z} _ {t _ {n + k}}, t _ {n + k}, t _ {n}, \\mathbf {c}) - w \\boldsymbol {\\Phi} (\\mathbf {z} _ {t _ {n + k}}, t _ {n + k}, t _ {n}, \\emptyset), \\qquad (5)\n$$\n", + "text_format": "latex", + "bbox": [ + 243, + 669, + 785, + 689 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $w$ denotes the CFG scale which is uniformly sampled from $[w_{\\min}, w_{\\max}]$ and $k$ is the skipping interval. To efficiently perform the above $k$ -step guided distillation, LCMs augment the consistency function to $f: (\\mathbf{z}_t, t, w, \\mathbf{c}) \\longmapsto \\mathbf{z}_0$ , which is also the form adopted by our MotionLCM.", + "bbox": [ + 212, + 699, + 787, + 761 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 MotionLCM: Motion Latent Consistency Model", + "text_level": 1, + "bbox": [ + 214, + 782, + 658, + 799 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Motion compression into the latent space. Motivated by [44, 62], we propose MotionLCM (Motion Latent Consistency Model) to tackle the low-efficiency", + "bbox": [ + 212, + 809, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "W. Dai et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "problem in motion diffusion models [65, 83], unleashing the potential of LCM in the motion generation task. Similar to MLD [9], our MotionLCM adopts a consistency model in the motion latent space. We choose MLD [9] as the underlying diffusion model to distill from. We aim to achieve few-step $(2\\sim 4)$ and even one-step inference without compromising motion quality. In MLD, an autoencoder $(\\mathcal{E},\\mathcal{D})$ is first trained to compress a high dimensional motion into a low dimensional latent vector $\\mathbf{z} = \\mathcal{E}(\\mathbf{x})$ , which is then decoded to reconstruct the motion as $\\hat{\\mathbf{x}} = \\mathcal{D}(\\mathbf{z})$ . Training diffusion models in the motion latent space greatly reduces the computational resources compared to the vanilla diffusion models trained on raw motion sequences (i.e., motion space) and speeds up the inference process. Thus, we effectively leverage the motion latent space for consistency distillation. Motion latent consistency distillation. An overview of our motion latent consistency distillation is described in Fig. 4 (a). A raw motion sequence $\\mathbf{x}_0^{1:N} = \\{\\mathbf{x}^i\\}_{i=1}^N$ is a sequence of human poses, where $N$ is the number of frames. We follow [17] to use the redundant motion representation for our experiments, which is widely used in previous work [9, 65, 83]. As shown in Fig. 4 (a), given a raw motion sequence $\\mathbf{x}_0^{1:N}$ , a pre-trained VAE [30] encoder first compresses it into the latent space, $\\mathbf{z}_0 = \\mathcal{E}(\\mathbf{x}_0)$ . Then, a forward diffusion operation with $n + k$ steps is conducted to add noise on $\\mathbf{z}_0$ , where $k$ is the skipping interval illustrated in Sec. 3.1. The noisy $\\mathbf{z}_{n+k}$ is fed to the frozen teacher network and trainable online network to predict the clean $\\hat{\\mathbf{z}}_0^*$ , and $\\hat{\\mathbf{z}}_0$ . The target network uses the cleaner $\\hat{\\mathbf{z}}_n$ obtained by a $k$ -step ODE solver $\\Phi$ , such as DDIM [61] to predict the $\\hat{\\mathbf{z}}_0^-$ . Since the classifier-free guidance (CFG) [21] is essential for condition alignment in diffusion models [9, 55, 65], we integrate CFG into the distillation,", + "bbox": [ + 212, + 146, + 787, + 508 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {z}} _ {n} \\leftarrow \\mathbf {z} _ {n + k} + (1 + w) \\boldsymbol {\\Phi} \\left(\\mathbf {z} _ {n + k}, t _ {n + k}, t _ {n}, \\mathbf {c}\\right) - w \\boldsymbol {\\Phi} \\left(\\mathbf {z} _ {n + k}, t _ {n + k}, t _ {n}, \\emptyset\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 269, + 520, + 785, + 536 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\mathbf{c}$ is the text condition and $w$ denotes the guidance scale. To ensure the self-consistency property defined in Eq. (1), the latent consistency distillation loss $\\mathcal{L}_{\\mathrm{LCD}}$ is designed as follows,", + "bbox": [ + 212, + 546, + 787, + 592 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {L C D}} (\\boldsymbol {\\Theta}, \\boldsymbol {\\Theta} ^ {-}) = \\mathbb {E} \\left[ d \\left(\\boldsymbol {f} _ {\\boldsymbol {\\Theta}} (\\mathbf {z} _ {n + k}, t _ {n + k}, w, \\mathbf {c}), \\boldsymbol {f} _ {\\boldsymbol {\\Theta} ^ {-}} (\\hat {\\mathbf {z}} _ {n}, t _ {n}, w, \\mathbf {c})\\right) \\right], \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 279, + 602, + 785, + 619 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $d(\\cdot, \\cdot)$ is a distance measuring function, such as L2 loss or Huber loss [25]. As discussed in Sec. 3.1, the target network $\\Theta^{-}$ is updated with the exponential moving average (EMA) of the trainable parameters of the online network $\\Theta$ . Here we define the teacher network $\\Theta^{*}$ as the pre-trained motion latent diffusion model, i.e., MLD [9]. According to [44], the online network and target network are initialized with the parameters of the teacher network. During the inference phase, as shown in Fig. 5, our MotionLCM can generate high-quality motions with one-step sampling and achieve the fastest runtime ( $\\sim$ 30ms per motion sequence) compared to other motion diffusion models [9,65].", + "bbox": [ + 212, + 628, + 787, + 765 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Controllable Motion Generation in Latent Space", + "text_level": 1, + "bbox": [ + 214, + 785, + 663, + 801 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "After addressing the low-efficiency issue in the motion latent diffusion model [9], we delve into another exploration of real-time motion control. Inspired by the", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "MotionLCM", + "bbox": [ + 645, + 114, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "great success of ControlNet [82] in controllable image generation [55], we introduce a motion ControlNet $\\Theta^a$ in the latent space of MotionLCM and initialize the motion ControlNet with a trainable copy of MotionLCM. Specifically, each layer in the motion ControlNet is appended with a zero-initialized linear layer for eliminating random noise in the initial training steps. To achieve an autoregressive motion generation paradigm, as depicted in Fig. 1, we define the motion control task as generating motions given the initial $\\tau$ poses and textual description. As shown in Fig. 4 (b), the initial $\\tau$ poses are defined by the trajectories of $K$ control joints, $\\mathbf{g}^{1:\\tau} = \\{\\mathbf{g}^i\\}_{i=1}^\\tau$ , where $\\mathbf{g}^i \\in \\mathbb{R}^{K \\times 3}$ denotes the global absolute locations of each control joint. In our motion control pipeline, we design a Trajectory Encoder $\\Theta^b$ consisting of stacked transformer [67] layers to encode the trajectory signals. We append a global token (i.e., [CLS]) before the start of the trajectory sequence as the output feature of the encoder, which is added to the noisy $\\mathbf{z}_n$ and fed into the trainable motion ControlNet $\\Theta^a$ . Under the guidance of motion ControlNet, MotionLCM predicts the denoised $\\hat{\\mathbf{z}}_0$ through the consistency function $f_{\\Theta^s}$ , where $\\Theta^s$ is the combination of $\\Theta^a$ , $\\Theta^b$ and $\\Theta$ . The following reconstruction loss $\\mathcal{L}_{\\mathrm{recon}}$ optimizes the motion ControlNet $\\Theta^a$ and Trajectory Encoder $\\Theta^b$ ,", + "bbox": [ + 212, + 146, + 787, + 419 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {r e c o n}} \\left(\\boldsymbol {\\Theta} ^ {a}, \\boldsymbol {\\Theta} ^ {b}\\right) = \\mathbb {E} \\left[ d \\left(\\boldsymbol {f} _ {\\boldsymbol {\\Theta} ^ {s}} \\left(\\mathbf {z} _ {n}, t _ {n}, w, \\mathbf {c} ^ {*}\\right), \\mathbf {z} _ {0}\\right) \\right], \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 334, + 430, + 785, + 448 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\mathbf{c}^*$ includes the text condition and control guidance from the Trajectory Encoder and the motion ControlNet. However, during training, the sole reconstruction supervision in the latent space is insufficient. We argue this is because the controllable motion generation requires more detailed constraints, which cannot be effectively provided solely by the reconstruction loss in the latent space. Unlike previous methods like OmniControl [73], which directly diffuse in the motion space, allowing explicit supervision of control signals, effectively supervising control signals in the latent space is non-trivial. Therefore, we utilize the frozen VAE [30] decoder $\\mathcal{D}$ to decode the latent $\\hat{\\mathbf{z}}_0$ into the motion space, obtaining the predicted motion $\\hat{\\mathbf{x}}_0$ , thereby introducing the control loss $\\mathcal{L}_{\\mathrm{control}}$ as follows,", + "bbox": [ + 212, + 458, + 787, + 609 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {c o n t r o l}} (\\boldsymbol {\\Theta} ^ {a}, \\boldsymbol {\\Theta} ^ {b}) = \\mathbb {E} \\left[ \\frac {\\sum_ {i} \\sum_ {j} m _ {i j} | | R (\\hat {\\mathbf {x}} _ {0}) _ {i j} - R (\\mathbf {x} _ {0}) _ {i j} | | _ {2} ^ {2}}{\\sum_ {i} \\sum_ {j} m _ {i j}} \\right], \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 295, + 621, + 785, + 662 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $R(\\cdot)$ converts the joint's local positions to global absolute locations and $m_{ij} \\in \\{0,1\\}$ is the binary joint mask at frame $i$ for the joint $j$ . Then we optimize the motion ControlNet $\\Theta^a$ and Trajectory Encoder $\\Theta^b$ with the overall objective,", + "bbox": [ + 214, + 671, + 787, + 719 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\boldsymbol {\\Theta} ^ {a}, \\boldsymbol {\\Theta} ^ {b} = \\underset {\\boldsymbol {\\Theta} ^ {a}, \\boldsymbol {\\Theta} ^ {b}} {\\operatorname {a r g m i n}} (\\mathcal {L} _ {\\text {r e c o n}} + \\lambda \\mathcal {L} _ {\\text {c o n t r o l}}), \\end{array} \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 370, + 732, + 785, + 757 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\lambda$ is the weight to balance the two losses. This design enables explicit control signals to directly influence the generation process, similar to controlling other latent-free diffusion models for motion generation [29,73]. Extensive experiments demonstrate the introduced supervision is very helpful in improving the motion control performance, which will be introduced in the following section.", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "W. Dai et al.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 143, + 375, + 162 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this section, we first present the experimental setup details in Sec. 4.1. Subsequently, we provide quantitative and qualitative comparisons to evaluate the effectiveness of our proposed MotionLCM framework in Sec. 4.2 and Sec. 4.3. Finally, we conduct comprehensive ablation studies on MotionLCM in Sec. 4.4.", + "bbox": [ + 212, + 175, + 787, + 237 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Experimental setup", + "text_level": 1, + "bbox": [ + 215, + 257, + 426, + 273 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Datasets. We experiment on the popular HumanML3D [17] dataset, featuring 14,616 unique human motion sequences with 44,970 textual descriptions. For a fair comparison with previous methods [9,17,49,65,83], we take the redundant motion representation, including root velocity, root height, local joint positions, velocities, rotations in root space, and the foot contact binary labels.", + "bbox": [ + 212, + 280, + 787, + 356 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Evaluation metrics. We extend the evaluation metrics of previous works [9, 17, 73]. (1) Time cost: We follow [9] to report the Average Inference Time per Sentence (AITS) to evaluate the inference efficiency of models. (2) Motion quality: Frechet Inception Distance (FID) is adopted as a principal metric to evaluate the feature distributions between the generated and real motions. The feature extractor employed is from [17]. (3) Motion diversity: MultiModality (MModality) measures the generation diversity conditioned on the same text and Diversity calculates variance through features [17]. (4) Condition matching: Following [17], we calculate the motion-retrieval precision (R-Precision) to report the text-motion Top-1/2/3 matching accuracy and Multimodal Distance (MM Dist) calculates the mean distance between motions and texts. (5) Control error: Trajectory error (Traj. err.) quantifies the ratio of unsuccessful trajectories, characterized by any control joint location error surpassing a predetermined threshold. Location error (Loc. err.) represents the unsuccessful joints. Average error (Avg. err.) denotes the mean location error of the control joints.", + "bbox": [ + 212, + 357, + 787, + 583 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Implementation details. Our baseline motion diffusion model is based on MLD [9]. We reproduce MLD with higher performance. Unless otherwise specified, all our experiments are conducted on this model. For MotionLCM, we employ the AdamW [40] optimizer for 96K iterations using a cosine decay learning rate scheduler and 1K iterations of linear warm-up. A batch size of 256 and a learning rate of 2e-4 are used. We set the training guidance scale range as $[w_{\\mathrm{min}}, w_{\\mathrm{max}}] = [5, 15]$ , with the testing guidance scale set to 7.5, and adopt the EMA rate $\\mu = 0.95$ by default. We use the DDIM [61] solver with skipping interval $k = 20$ and choose the Huber [25] loss as the distance measuring function $d$ . For motion ControlNet, we use the AdamW [40] optimizer for 192K iterations with 1K iterations of linear warm-up. The batch size and learning rate are set to 128 and 1e-4. The learning rate scheduler is the same as the first stage. For the training objective, we employ $d$ as the L2 loss and set the control loss weight $\\lambda$ to 1.0 by default. We set the control ratio $\\tau$ as 0.25 and the number of control joints as $K = 6$ (i.e., Pelvis, Left foot, Right foot, Head, Left wrist, and Right wrist) in both training and testing. We implement our model using PyTorch [47] with training on an NVIDIA RTX 4090 GPU and testing on a Tesla V100 GPU.", + "bbox": [ + 212, + 583, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "MotionLCM", + "bbox": [ + 645, + 114, + 730, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/7234ee009fbf2e4a1f06398b5ee0af8c68305ba2dc32e2903ed87c696769f9a0.jpg", + "table_caption": [ + "Table 1: Comparison of text-conditional motion synthesis on HumanML3D [17] dataset. We compute suggested metrics following [17]. We repeat the evaluation 20 times for each metric and report the average with a $95\\%$ confidence interval. “ $\\rightarrow$ ” indicates that the closer to the real data, the better. **Bold and underline indicate the best and the second best result.** denotes the reproduced version of MLD [9]. The MotionLCM in one-step inference (30ms) surpasses all state-of-the-art models." + ], + "table_footnote": [], + "table_body": "
MethodsAITS ↓R-Precision ↑FID ↓MM Dist ↓Diversity →MModality ↑
Top 1Top 2Top 3
Real-0.511±.0030.703±.0030.797±.0020.002±.0002.794±.0089.503±.065-
Seq2Seq [37]-0.180±.0020.300±.0020.396±.00211.75±.0355.529±.0076.223±.061-
JL2P [2]-0.246±.0010.387±.0020.486±.00211.02±.0465.296±.0087.676±.058-
T2G [5]-0.165±.0010.267±.0020.345±.0027.664±.0306.030±.0086.409±.071-
Hier [14]-0.301±.0020.425±.0020.552±.0046.532±.0245.012±.0188.332±.042-
TEMOS [49]0.0170.424±.0020.612±.0020.722±.0023.734±.0283.703±.0088.973±.0710.368±.018
T2M [17]0.0380.457±.0020.639±.0030.740±.0031.067±.0023.340±.0089.188±.0022.090±.083
MDM [65]24.740.320±.0050.498±.0040.611±.0070.544±.0445.566±.0279.559±.0862.799±.072
MotionDiffuse [83]14.740.491±.0010.681±.0010.782±.0010.630±.0013.113±.0019.410±.0491.553±.042
MLD [9]0.2170.481±.0030.673±.0030.772±.0020.473±.0133.196±.0109.724±.0822.413±.079
MLD* [9]0.2250.504±.0020.698±.0030.796±.0020.450±.0113.052±.0099.634±.0642.267±.082
MotionLCM (1-step)0.0300.502±.0030.701±.0020.803±.0020.467±.0123.022±.0099.631±.0662.172±.082
MotionLCM (2-step)0.0350.505±.0030.705±.0020.805±.0020.368±.0112.986±.0089.640±.0522.187±.094
MotionLCM (4-step)0.0430.502±.0030.698±.0020.798±.0020.304±.0123.012±.0079.607±.0662.259±.092
", + "bbox": [ + 222, + 239, + 784, + 406 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 Comparisons on Text-to-motion", + "text_level": 1, + "bbox": [ + 215, + 435, + 524, + 450 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "In the following part, we first evaluate our MotionLCM on the text-to-motion (T2M) task. We compare our method with some T2M baselines on HumanML3D [17] with suggested metrics [17] under the $95\\%$ confidence interval from 20 times running. As MotionLCM is based on MLD, we mainly focus on the performance compared with MLD. For evaluating time efficiency, we compare the Average Inference Time per Sentence (AITS) with TEMOS [49], T2M [17], MDM [65], MotionDiffuse [83] and MLD [9]. The results are borrowed from MLD [9]. The deterministic methods [2,5,14,37], are unable to produce diverse results from a single input text and thus we leave their MModality metrics empty. For the quantitative results, as shown in Tab. 1, our MotionLCM boasts an impressive real-time runtime efficiency, averaging around 30ms per motion sequence during inference. This performance exceeds that of previous diffusion-based methods [9,65,83] and even surpasses MLD [9] by an order of magnitude. Furthermore, despite employing only one-step inference, our MotionLCM can approximate or even surpass the performance of MLD [9] (DDIM [61] 50 steps). With two-step inference, we achieve the best R-Precision and MM Dist metrics, while increasing the sampling steps to four yields the best FID. The above results demonstrate the effectiveness of latent consistency distillation. For the qualitative results, as shown in Fig. 5, MotionLCM not only accelerates motion generation to real-time speed but also delivers high-quality outputs, closely aligning with the textual descriptions.", + "bbox": [ + 212, + 459, + 787, + 763 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.3 Comparisons on Controllable Motion Generation", + "text_level": 1, + "bbox": [ + 215, + 784, + 665, + 800 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "As shown in Tab. 2, we compare our MotionLCM with OmniControl [73] and MLD [9]. We observe that OmniControl struggles with multi-joint control and", + "bbox": [ + 212, + 808, + 785, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "W. Dai et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/819822b0bcfce68a51cca71300f44024fff863c75f425e09031d019cd43091cb.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
RealMotionLCM (Ours)MLDMDM
“a person walks clockwise in a large curve while swinging their arms.”
0.033s0.201s22.58s
“a person walking like a bird and then sniffing the air.”
0.031s0.192s21.16s
“a person slightly bent over with right hand pressing against the air walks forward slowly”
0.031s0.197s23.44s
", + "bbox": [ + 218, + 143, + 784, + 422 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Fig. 5: Qualitative comparison of the state-of-the-art methods in the text-to-motion task. With only one-step inference, MotionLCM achieves the fastest motion generation while producing high-quality movements that closely match the textual descriptions.", + "bbox": [ + 215, + 446, + 784, + 488 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "falls short in both generation quality and control performance compared to MotionLCM. To verify the effectiveness of the latent generated by our MotionLCM for training motion ControlNet, we conducted the following two sets of experiments: \"LC\" and \"MC\", which indicate introducing control supervision in the latent space and motion space. It can be observed that under the same experimental settings, MotionLCM maintains higher fidelity and significantly outperforms MLD [9] in motion control performance. This demonstrates that the latent generated by MotionLCM is more effective for training motion ControlNet compared to MLD [9]. In terms of inference speed, MotionLCM (1-step) is $1929 \\times$ faster compared to OmniControl [73] and $13 \\times$ faster than MLD [9]. For qualitative results, as shown in Fig. 6, OmniControl fails to control the initial poses in the second example and does not generate motion that aligns with the text in the third case. However, our MotionLCM not only adheres to the control of the initial poses but also generates motions that match the textual descriptions.", + "bbox": [ + 212, + 518, + 787, + 731 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.4 Ablation Studies", + "text_level": 1, + "bbox": [ + 215, + 753, + 401, + 767 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Impact of the hyperparameters of training MotionLCM. We conduct a comprehensive analysis of the training hyperparameters of MotionLCM, including the training guidance scale range $[w_{\\mathrm{min}}, w_{\\mathrm{max}}]$ , EMA rate $\\mu$ , skipping interval $k$ , and the type of loss. We summarize the evaluation results based on one-step", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "MotionLCM", + "bbox": [ + 645, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/c613453f758194c1580f836fd9fa424344ec7cec9e015d32657cf5c138bbe779.jpg", + "table_caption": [ + "Table 2: Comparison of motion control on HumanML3D [17] dataset. Bold indicates the best result. Our MotionLCM outperforms OmniControl [73] and MLD [9] regarding generation quality, control performance, and inference speed. \"LC\" and \"MC\" refer to the control supervision introduced in the latent space and motion space." + ], + "table_footnote": [], + "table_body": "
MethodsAITS ↓FID ↓R-Precision ↑ Top 3Diversity →Traj. err. ↓ (50cm)Loc. err. ↓ (50cm)Avg. err. ↓
Real-0.0020.7979.5030.00000.00000.0000
OmniControl [73]81.002.3280.5578.8670.33620.03220.0977
MLD [9] (LC)0.5520.4690.7239.4760.42300.06530.1690
MotionLCM (1-step, LC)0.0420.3190.7529.4240.29860.03440.1410
MotionLCM (2-step, LC)0.0470.3150.7709.4270.28400.03280.1365
MotionLCM (4-step, LC)0.0630.3280.7459.4410.29730.03390.1398
MLD [9] (LC&MC)0.5520.5550.7549.3730.27220.02150.1265
MotionLCM (1-step, LC&MC)0.0420.4190.7569.3900.19880.01470.1127
MotionLCM (2-step, LC&MC)0.0470.3970.7599.4690.19600.01430.1092
MotionLCM (4-step, LC&MC)0.0630.4440.7539.3550.20890.01720.1140
", + "bbox": [ + 223, + 212, + 784, + 345 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/70c5612c722d3f3c932a9fc676fc0d237bac9c22496e1004a5c4f956e13d7bd4.jpg", + "table_caption": [ + "Table 3: Ablation study on different training guidance scale ranges $[w_{\\mathrm{min}}, w_{\\mathrm{max}}]$ , EMA rates $\\mu$ , skipping intervals $k$ and types of loss. We use metrics in Tab. 1 and adopt a one-step inference setting with the testing CFG scale of 7.5 for fair comparison." + ], + "table_footnote": [], + "table_body": "
MethodsR-Precision ↑ Top 1FID ↓MM Dist ↓Diversity →MModality ↑
Real0.511±.0030.002±.0002.794±.0089.503±.065-
MotionLCM (w ∈ [5,15])0.502±.0030.467±.0123.022±.0099.631±.0662.172±.082
MotionLCM (w ∈ [2,18])0.497±.0030.481±.0093.030±.0109.644±.0732.226±.091
MotionLCM (w = 7.5)0.486±.0020.479±.0093.094±.0099.610±.0722.320±.097
MotionLCM (μ = 0.95)0.502±.0030.467±.0123.022±.0099.631±.0662.172±.082
MotionLCM (μ = 0.50)0.498±.0030.478±.0093.022±.0109.655±.0712.188±.087
MotionLCM (μ = 0)0.499±.0030.505±.0083.018±.0099.706±.0702.123±.085
MotionLCM (k = 50)0.488±.0030.547±.0113.096±.0109.511±.0742.324±.091
MotionLCM (k = 20)0.502±.0030.467±.0123.022±.0099.631±.0662.172±.082
MotionLCM (k = 10)0.497±.0030.449±.0093.017±.0109.693±.0752.133±.086
MotionLCM (k = 5)0.488±.0030.438±.0093.044±.0099.647±.0742.147±.083
MotionLCM (k = 1)0.442±.0020.635±.0113.255±.0089.384±.0802.146±.075
MotionLCM (w/ Huber)0.502±.0030.467±.0123.022±.0099.631±.0662.172±.082
MotionLCM (w/ L2)0.486±.0020.622±.0103.114±.0099.573±.0692.218±.086
", + "bbox": [ + 253, + 422, + 754, + 619 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "inference in Tab. 3. We find out that using a dynamic training guidance scale (e.g., $w \\in [5,15]$ ) during training leads to an improvement in model performance compared to using a static training guidance scale (e.g., $w = 7.5$ ). Additionally, an excessively large range for the training guidance scale can also negatively impact the performance of the model (e.g., $w \\in [2,18]$ ). Regarding the EMA rate $\\mu$ , we observe that the larger the value of $\\mu$ , the better the performance of the model. This indicates that maintaining a slower update rate for the target network $\\Theta^{-}$ helps improve the performance of latent consistency distillation. When the skipping interval $k$ continues to increase, the performance of the distillation model improves progressively, but larger values of $k$ (e.g., $k = 50$ ) may result in inferior results. As for the type of loss, the Huber loss [25] significantly outperforms the L2 loss, demonstrating its superior robustness.", + "bbox": [ + 212, + 657, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "W. Dai et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/4e31046014c5e58e0ee7f7dd7a4be938ed534f29a314ffc93a0ad01e23e8e3db.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MotionLCM (w/o ControlNet)0.029s0.033s0.030s0.033s0.029s
OmniControl87.81s90.00s85.11s80.15s86.77s
MotionLCM (w/ ControlNet)0.043s0.045s0.041s0.045s0.047s
Reala man lifts his left arm up.the person is widewalking.a person walks three steps to his right, then five steps to his left, and finally three steps to his right.this person is walking in circles then moves across room to step around an object.the person is holding their head while walking.
", + "bbox": [ + 218, + 143, + 787, + 450 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Fig. 6: Qualitative comparison of the state-of-the-art methods in the motion control task. We provide the visualized motion results and real references from five prompts. Compared to OmniControl [73], MotionLCM with ControlNet not only generates the initial poses that accurately follow the given multi-joint trajectories (i.e., the green poses in real references) but also produces motions that closely align with the texts.", + "bbox": [ + 212, + 472, + 787, + 542 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Impact of control loss weights $\\lambda$ . To verify the impact of different control loss weights $\\lambda$ on the control performance of MotionLCM, we report the experimental results in Tab. 4. We also include experiments of MotionLCM without ControlNet (i.e., only text-to-motion) for comparison. We found a significant improvement in control-related metrics (e.g., Loc. err.) after introducing motion ControlNet (i.e., $\\lambda = 0$ ). Furthermore, control performance can be further improved by introducing control loss (i.e., $\\lambda > 0$ ). Increasing the weight $\\lambda$ enhances control performance but leads to a decline in the generation quality, which is reflected in higher FID scores. To balance these two aspects, we adopt $\\lambda = 1$ as our default setting for training motion ControlNet.", + "bbox": [ + 212, + 579, + 787, + 731 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Impact of different control ratios $\\tau$ and number of control joints $K$ . In Tab. 5, we present the results of all models with the testing control ratio as 0.25 and keep the number of control joints $K$ equal in both training and testing. We found that the model with the fixed training control ratio (i.e., $\\tau = 0.25$ ) performs better compared to a dynamic ratio (e.g., $\\tau \\in [0.1, 0.5]$ ), and we discover that our model maintains good performance when incorporating additional redundant control signals, such as whole-body joints with $K = 22$ .", + "bbox": [ + 212, + 734, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "MotionLCM", + "bbox": [ + 645, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/1118fbac6341c6c1766093e7dab5ae2f25d029a84a3ac14825edfd4185d85cd6.jpg", + "table_caption": [ + "Table 4: Ablation study on different control loss weights $\\lambda$ . We present the results of (1, 2, 4)-step inference. We add the MotionLCM without ControlNet for comparison." + ], + "table_footnote": [], + "table_body": "
MethodsFID ↓ Top 3R-Precision ↑ Diversity →Traj. err. ↓ (50cm)Loc. err. ↓ (50cm)Avg. err. ↓
Real0.0020.7979.5030.00000.00000.0000
MotionLCM (1-step, w/o control)0.4670.8039.6310.76050.23020.3493
MotionLCM (2-step, w/o control)0.3680.8059.6400.76460.22140.3386
MotionLCM (4-step, w/o control)0.3040.7989.6070.77390.22070.3359
MotionLCM (1-step, λ = 0)0.3190.7529.4240.29860.03440.1410
MotionLCM (2-step, λ = 0)0.3150.7709.4270.28400.03280.1365
MotionLCM (4-step, λ = 0)0.3280.7459.4410.29730.03390.1398
MotionLCM (1-step, λ = 0.1)0.3440.7539.3860.27110.02750.1310
MotionLCM (2-step, λ = 0.1)0.3240.7599.4280.26310.02560.1268
MotionLCM (4-step, λ = 0.1)0.3570.7439.4150.27130.02680.1309
MotionLCM (1-step, λ = 1.0)0.4190.7569.3900.19880.01470.1127
MotionLCM (2-step, λ = 1.0)0.3970.7599.4690.19600.01430.1092
MotionLCM (4-step, λ = 1.0)0.4440.7539.3550.20890.01720.1140
MotionLCM (1-step, λ = 10.0)0.6360.7449.4790.14650.00970.0967
MotionLCM (2-step, λ = 10.0)0.5510.7579.5690.15900.01070.0987
MotionLCM (4-step, λ = 10.0)0.5680.7429.4860.17230.01320.1045
", + "bbox": [ + 245, + 184, + 761, + 373 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/26ea8229db741eeea9335aa71179766413b138d14beb55975be5ebf95740c450.jpg", + "table_caption": [ + "Table 5: Ablation study on different control ratios $\\tau$ and number of control joints $K$ . We report the results of (1,2,4)-step inference. \" $*$ \" denotes the default training setting." + ], + "table_footnote": [], + "table_body": "
MethodsFID ↓R-Precision ↑ Top 3Diversity →Traj. err. ↓ (50cm)Loc. err. ↓ (50cm)Avg. err. ↓
Real0.0020.7979.5030.00000.00000.0000
MotionLCM* (1-step, τ = 0.25, K = 6)0.4190.7569.3900.19880.01470.1127
MotionLCM* (2-step, τ = 0.25, K = 6)0.3970.7599.4690.19600.01430.1092
MotionLCM* (4-step, τ = 0.25, K = 6)0.4440.7539.3550.20890.01720.1140
MotionLCM (1-step, τ ∈ [0.1, 0.25])0.4560.7579.4770.28210.02340.1214
MotionLCM (2-step, τ ∈ [0.1, 0.25])0.4090.7699.5920.27070.02300.1179
MotionLCM (4-step, τ ∈ [0.1, 0.25])0.4570.7579.5400.29280.02560.1228
MotionLCM (1-step, τ ∈ [0.1, 0.5])0.4480.7639.5380.23900.01820.1182
MotionLCM (2-step, τ ∈ [0.1, 0.5])0.4130.7689.5170.23490.01800.1153
MotionLCM (4-step, τ ∈ [0.1, 0.5])0.4460.7539.4980.25170.01990.1196
MotionLCM (1-step, K = 12)0.4120.7539.4120.20720.01100.1029
MotionLCM (2-step, K = 12)0.4100.7589.5090.19790.01080.1000
MotionLCM (4-step, K = 12)0.4420.7559.3800.21690.01320.1048
MotionLCM (1-step, K = 22(whole-body))0.4360.7489.3790.21430.00830.0914
MotionLCM (2-step, K = 22(whole-body))0.4130.7589.4920.20610.00820.0881
MotionLCM (4-step, K = 22(whole-body))0.4610.7459.4590.21730.00970.0918
", + "bbox": [ + 245, + 436, + 761, + 611 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 217, + 647, + 359, + 662 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "This work proposes an efficient controllable motion generation framework, MotionLCM. By introducing latent consistency distillation, MotionLCM enjoys the trade-off between runtime efficiency and generation quality. Moreover, thanks to the motion ControlNet manipulation in the latent space, our method obtains excellent controlling ability with given conditions. Extensive experimental results show the effectiveness of the proposed method. As the VAE of MLD lacks explicit temporal modeling, the MotionLCM cannot achieve a good temporal explanation. Therefore, our future work will lie in developing a more explainable compression architecture for efficient motion control.", + "bbox": [ + 215, + 703, + 785, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "W. Dai et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "The research is supported by Shenzhen Ubiquitous Data Enabling Key Lab under grant ZDSYS20220527171406015 and CCF-Tencent Rhino-Bird Open Research Fund. This project is also supported by Shanghai Artificial Intelligence Laboratory. The author team would like to acknowledge Yiming Xie, Zhiyang Dou, and Shunlin Lu for their helpful technical discussions and suggestions.", + "bbox": [ + 212, + 176, + 787, + 251 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 273, + 321, + 289 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Ahn, H., Ha, T., Choi, Y., Yoo, H., Oh, S.: Text2action: Generative adversarial synthesis from language to action. In: ICRA. pp. 5915-5920 (2018)", + "2. Ahuja, C., Morency, L.P.: Language2pose: Natural language grounded pose forecasting. In: 3DV. pp. 719-728 (2019)", + "3. Athanasiou, N., Petrovich, M., Black, M.J., Varol, G.: Teach: Temporal action composition for 3d humans. In: 3DV. pp. 414-423 (2022)", + "4. Barquero, G., Escalera, S., Palmero, C.: Seamless human motion composition with blended positional encodings. In: CVPR. pp. 457-469 (2024)", + "5. Bhattacharya, U., Rewkowski, N., Banerjee, A., Guhan, P., Bera, A., Manocha, D.: Text2gestures: A transformer-based network for generating emotive body gestures for virtual agents. In: VR. pp. 1-10 (2021)", + "6. Cervantes, P., Sekikawa, Y., Sato, I., Shinoda, K.: Implicit neural representations for variable length human motion generation. In: ECCV. pp. 356-372 (2022)", + "7. Chen, L.H., Lu, S., Zeng, A., Zhang, H., Wang, B., Zhang, R., Zhang, L.: Motionllm: Understanding human behaviors from human motions and videos. arXiv preprint arXiv:2405.20340 (2024)", + "8. Chen, L.H., Zhang, J., Li, Y., Pang, Y., Xia, X., Liu, T.: Humanmac: Masked motion completion for human motion prediction. In: ICCV. pp. 9544-9555 (2023)", + "9. Chen, X., Jiang, B., Liu, W., Huang, Z., Fu, B., Chen, T., Yu, G.: Executing your commands via motion diffusion in latent space. In: CVPR. pp. 18000-18010 (2023)", + "0. Cong, P., Dou, Z.W., Ren, Y., Yin, W., Cheng, K., Sun, Y., Long, X., Zhu, X., Ma, Y.: Laserhuman: Language-guided scene-aware human motion generation in free environment. arXiv preprint arXiv:2403.13307 (2024)", + "1. Dabral, R., Mughal, M.H., Golyanik, V., Theobalt, C.: Mofusion: A framework for denoising-diffusion-based motion synthesis. In: CVPR. pp. 9760-9770 (2023)", + "2. Dou, Z., Chen, X., Fan, Q., Komura, T., Wang, W.: C-ase: Learning conditional adversarial skill embeddings for physics-based characters. In: SIGGRAPH Asia. pp. 1-11 (2023)", + "3. Fan, K., Tang, J., Cao, W., Yi, R., Li, M., Gong, J., Zhang, J., Wang, Y., Wang, C., Ma, L.: Freemotion: A unified framework for number-free text-to-motion synthesis. arXiv preprint arXiv:2405.15763 (2024)", + "4. Ghosh, A., Cheema, N., Oguz, C., Theobalt, C., Slusallek, P.: Synthesis of compositional animations from textual descriptions. In: ICCV. pp. 1396-1406 (2021)", + "5. Guo, C., Mu, Y., Javed, M.G., Wang, S., Cheng, L.: Momask: Generative masked modeling of 3d human motions. In: CVPR. pp. 1900-1910 (2024)", + "6. Guo, C., Mu, Y., Zuo, X., Dai, P., Yan, Y., Lu, J., Cheng, L.: Generative human motion stylization in latent space. arXiv preprint arXiv:2401.13505 (2024)", + "7. Guo, C., Zou, S., Zuo, X., Wang, S., Ji, W., Li, X., Cheng, L.: Generating diverse and natural 3d human motions from text. In: CVPR. pp. 5152-5161 (2022)" + ], + "bbox": [ + 225, + 303, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "MotionLCM", + "bbox": [ + 645, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "18. Guo, C., Zuo, X., Wang, S., Cheng, L.: Tm2t: Stochastic and tokenized modeling for the reciprocal generation of 3d human motions and texts. In: ECCV. pp. 580-597 (2022)", + "19. Guo, C., Zuo, X., Wang, S., Zou, S., Sun, Q., Deng, A., Gong, M., Cheng, L.: Action2motion: Conditioned generation of 3d human motions. In: ACMMM. pp. 2021-2029 (2020)", + "20. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. NeurIPS pp. 6840-6851 (2020)", + "21. Ho, J., Salimans, T.: Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598 (2022)", + "22. Holden, D., Komura, T., Saito, J.: Phase-functional neural networks for character control. TOG 36(4), 1-13 (2017)", + "23. Holden, D., Saito, J., Komura, T.: A deep learning framework for character motion synthesis and editing. TOG 35(4), 1-11 (2016)", + "24. Huang, Y., Yang, H., Luo, C., Wang, Y., Xu, S., Zhang, Z., Zhang, M., Peng, J.: Stablemofusion: Towards robust and efficient diffusion-based motion generation framework. arXiv preprint arXiv:2405.05691 (2024)", + "25. Huber, P.J.: Robust estimation of a location parameter. The Annals of Mathematical Statistics 35(1), 73-101 (1964)", + "26. Ji, Y., Xu, F., Yang, Y., Shen, F., Shen, H.T., Zheng, W.S.: A large-scale rgb-d database for arbitrary-view human action recognition. In: ACMM. pp. 1510-1518 (2018)", + "27. Jiang, B., Chen, X., Liu, W., Yu, J., Yu, G., Chen, T.: Motiongpt: Human motion as a foreign language. NeurIPS (2024)", + "28. Karras, T., Aittala, M., Aila, T., Laine, S.: Elucidating the design space of diffusion-based generative models. NeurIPS pp. 26565-26577 (2022)", + "29. Karunratanakul, K., Preechakul, K., Suwajanakorn, S., Tang, S.: Guided motion diffusion for controllable human motion synthesis. In: CVPR. pp. 2151-2162 (2023)", + "30. Kingma, D.P., Welling, M.: Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114 (2013)", + "31. Lee, T., Moon, G., Lee, K.M.: Multiact: Long-term 3d human motion generation from multiple action labels. In: AAAI. pp. 1231-1239 (2023)", + "32. Li, B., Zhao, Y., Zhelun, S., Sheng, L.: Danceformer: Music conditioned 3d dance generation with parametric motion transformer. In: AAAI. pp. 1272-1279 (2022)", + "33. Li, R., Zhang, Y., Zhang, Y., Zhang, H., Guo, J., Zhang, Y., Liu, Y., Li, X.: Lodge: A coarse to fine diffusion network for long dance generation guided by the characteristic dance primitives. In: CVPR. pp. 1524-1534 (2024)", + "34. Li, R., Zhao, J., Zhang, Y., Su, M., Ren, Z., Zhang, H., Tang, Y., Li, X.: Finedance: A fine-grained choreography dataset for 3d full body dance generation. In: ICCV. pp. 10234-10243 (2023)", + "35. Li, R., Yang, S., Ross, D.A., Kanazawa, A.: Ai choreographer: Music conditioned 3d dance generation with aist++. In: ICCV. pp. 13401-13412 (2021)", + "36. Li, T., Qiao, C., Ren, G., Yin, K., Ha, S.: Aamdm: Accelerated auto-regressive motion diffusion model. In: CVPR. pp. 1813-1823 (2024)", + "37. Lin, A.S., Wu, L., Corona, R., Tai, K., Huang, Q., Mooney, R.J.: Generating animated videos of human activities from natural language descriptions. Learning 1(2018), 1 (2018)", + "38. Lin, X., Amer, M.R.: Human motion modeling using dvgens. arXiv preprint arXiv:1804.10652 (2018)" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "W. Dai et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "39. Liu, J., Dai, W., Wang, C., Cheng, Y., Tang, Y., Tong, X.: Plan, posture and go: Towards open-world text-to-motion generation. arXiv preprint arXiv:2312.14828 (2023)", + "40. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101 (2017)", + "41. Lu, C., Zhou, Y., Bao, F., Chen, J., Li, C., Zhu, J.: Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. NeurIPS pp. 5775-5787 (2022)", + "42. Lu, C., Zhou, Y., Bao, F., Chen, J., Li, C., Zhu, J.: Dpm-solver++: Fast solver for guided sampling of diffusion probabilistic models. arXiv preprint arXiv:2211.01095 (2022)", + "43. Lu, S., Chen, L.H., Zeng, A., Lin, J., Zhang, R., Zhang, L., Shum, H.Y.: Humantomato: Text-aligned whole-body motion generation. arXiv preprint arXiv:2310.12978 (2023)", + "44. Luo, S., Tan, Y., Huang, L., Li, J., Zhao, H.: Latent consistency models: Synthesizing high-resolution images with few-step inference. arXiv preprint arXiv:2310.04378 (2023)", + "45. Mahmood, N., Ghorbani, N., Troje, N.F., Pons-Moll, G., Black, M.J.: Amass: Archive of motion capture as surface shapes. In: ICCV. pp. 5442-5451 (2019)", + "46. Nichol, A.Q., Dhariwal, P.: Improved denoising diffusion probabilistic models. In: ICML. pp. 8162-8171 (2021)", + "47. Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., et al.: Pytorch: An imperative style, high-performance deep learning library. NeurIPS (2019)", + "48. Petrovich, M., Black, M.J., Varol, G.: Action-conditioned 3d human motion synthesis with transformer vae. In: ICCV. pp. 10985-10995 (2021)", + "49. Petrovich, M., Black, M.J., Varol, G.: Temos: Generating diverse human motions from textual descriptions. In: ECCV. pp. 480-497 (2022)", + "50. Petrovich, M., Black, M.J., Varol, G.: Tmr: Text-to-motion retrieval using contrastive 3d human motion synthesis. In: ICCV. pp. 9488-9497 (2023)", + "51. Petrovich, M., Litany, O., Iqbal, U., Black, M.J., Varol, G., Bin Peng, X., Rempe, D.: Multi-track timeline control for text-driven 3d human motion generation. In: CVPRW. pp. 1911-1921 (2024)", + "52. Plappert, M., Mandery, C., Asfour, T.: The kit motion-language dataset. Big data 4(4), 236-252 (2016)", + "53. Punnakkal, A.R., Chandrasekaran, A., Athanasiou, N., Quiros-Ramirez, A., Black, M.J.: Babel: Bodies, action and behavior with english labels. In: CVPR. pp. 722-731 (2021)", + "54. Raab, S., Leibovitch, I., Li, P., Aberman, K., Sorkine-Hornung, O., Cohen-Or, D.: Modi: Unconditional motion synthesis from diverse data. In: CVPR. pp. 13873-13883 (2023)", + "55. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR. pp. 10684-10695 (2022)", + "56. Shafir, Y., Tevet, G., Kapon, R., Bermano, A.H.: Human motion diffusion as a generative prior. In: ICLR (2024)", + "57. Shahroudy, A., Liu, J., Ng, T.T., Wang, G.: Ntu $\\mathrm{rgb + }$ d: A large scale dataset for 3d human activity analysis. In: CVPR. pp. 1010-1019 (2016)", + "58. Shi, Y., Wang, J., Jiang, X., Dai, B.: Controllable motion diffusion model. arXiv preprint arXiv:2306.00416 (2023)" + ], + "bbox": [ + 212, + 146, + 787, + 840 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "MotionLCM", + "bbox": [ + 645, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "59. Siyao, L., Yu, W., Gu, T., Lin, C., Wang, Q., Qian, C., Loy, C.C., Liu, Z.: Bailando: 3d dance generation by actor-critic gpt with choreographic memory. In: CVPR. pp. 11050-11059 (2022)", + "60. Sohl-Dickstein, J., Weiss, E., Maheswaranathan, N., Ganguli, S.: Deep unsupervised learning using nonequilibrium thermodynamics. In: ICML. pp. 2256-2265 (2015)", + "61. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. In: ICLR (2021)", + "62. Song, Y., Dhariwal, P., Chen, M., Sutskever, I.: Consistency models. In: ICML (2023)", + "63. Tang, Y., Liu, J., Liu, A., Yang, B., Dai, W., Rao, Y., Lu, J., Zhou, J., Li, X.: Flag3d: A 3d fitness activity dataset with language instruction. In: CVPR. pp. 22106-22117 (2023)", + "64. Tevet, G., Gordon, B., Hertz, A., Bermano, A.H., Cohen-Or, D.: Motionclip: Exposing human motion generation to clip space. In: ECCV. pp. 358-374 (2022)", + "65. Tevet, G., Raab, S., Gordon, B., Shafir, Y., Cohen-Or, D., Bermano, A.H.: Human motion diffusion model. In: ICLR (2022)", + "66. Tseng, J., Castellon, R., Liu, K.: Edge: Editable dance generation from music. In: CVPR. pp. 448-458 (2023)", + "67. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. NeurIPS (2017)", + "68. Wan, W., Dou, Z., Komura, T., Wang, W., Jayaraman, D., Liu, L.: Tlcontrol: Trajectory and language control for human motion synthesis. arXiv preprint arXiv:2311.17135 (2023)", + "69. Wang, Z., Chen, Y., Jia, B., Li, P., Zhang, J., Zhang, J., Liu, T., Zhu, Y., Liang, W., Huang, S.: Move as you say interact as you can: Language-guided human motion generation with scene affordance. In: CVPR. pp. 433-444 (2024)", + "70. Wang, Z., Chen, Y., Liu, T., Zhu, Y., Liang, W., Huang, S.: Humanise: Language-conditioned human motion generation in 3d scenes. NeurIPS pp. 14959-14971 (2022)", + "71. Wang, Z., Wang, J., Lin, D., Dai, B.: Intercontrol: Generate human motion interactions by controlling every joint. arXiv preprint arXiv:2311.15864 (2023)", + "72. Xiao, Z., Wang, T., Wang, J., Cao, J., Zhang, W., Dai, B., Lin, D., Pang, J.: Unified human-scene interaction via prompted chain-of-contacts. In: ICLR (2024)", + "73. Xie, Y., Jampani, V., Zhong, L., Sun, D., Jiang, H.: Omnicontrol: Control any joint at any time for human motion generation. In: ICLR (2024)", + "74. Xu, L., Lv, X., Yan, Y., Jin, X., Wu, S., Xu, C., Liu, Y., Zhou, Y., Rao, F., Sheng, X., et al.: Inter-x: Towards versatile human-human interaction analysis. In: CVPR. pp. 22260-22271 (2024)", + "75. Xu, L., Song, Z., Wang, D., Su, J., Fang, Z., Ding, C., Gan, W., Yan, Y., Jin, X., Yang, X., et al.: Actformer: A gan-based transformer towards general action-conditioned 3d human motion generation. In: ICCV. pp. 2228-2238 (2023)", + "76. Xu, L., Zhou, Y., Yan, Y., Jin, X., Zhu, W., Rao, F., Yang, X., Zeng, W.: Regennet: Towards human action-reaction synthesis. In: CVPR. pp. 1759-1769 (2024)", + "77. Yan, S., Li, Z., Xiong, Y., Yan, H., Lin, D.: Convolutional sequence generation for skeleton-based action synthesis. In: ICCV. pp. 4394-4402 (2019)", + "78. Yuan, Y., Song, J., Iqbal, U., Vahdat, A., Kautz, J.: Physdiff: Physics-guided human motion diffusion model. In: ICCV. pp. 16010-16021 (2023)", + "79. Zhang, B., Cheng, Y., Wang, C., Zhang, T., Yang, J., Tang, Y., Zhao, F., Chen, D., Guo, B.: Rodinhd: High-fidelity 3d avatar generation with diffusion models. arXiv preprint arXiv:2407.06938 (2024)" + ], + "bbox": [ + 215, + 146, + 784, + 840 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "W. Dai et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "80. Zhang, B., Cheng, Y., Yang, J., Wang, C., Zhao, F., Tang, Y., Chen, D., Guo, B.: Gaussiancube: Structuring gaussian splatting using optimal transport for 3d generative modeling. arXiv preprint arXiv:2403.19655 (2024)", + "81. Zhang, J., Zhang, Y., Cun, X., Zhang, Y., Zhao, H., Lu, H., Shen, X., Shan, Y.: Generating human motion from textual descriptions with discrete representations. In: CVPR. pp. 14730-14740 (2023)", + "82. Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models. In: ICCV. pp. 3836-3847 (2023)", + "83. Zhang, M., Cai, Z., Pan, L., Hong, F., Guo, X., Yang, L., Liu, Z.: Motiondiffuse: Text-driven human motion generation with diffusion model. arXiv preprint arXiv:2208.15001 (2022)", + "84. Zhao, R., Su, H., Ji, Q.: Bayesian adversarial human motion synthesis. In: CVPR. pp. 6225-6234 (2020)", + "85. Zhong, L., Xie, Y., Jampani, V., Sun, D., Jiang, H.: Smoodi: Stylized motion diffusion model. arXiv preprint arXiv:2407.12783 (2024)", + "86. Zhou, W., Dou, Z., Cao, Z., Liao, Z., Wang, J., Wang, W., Liu, Y., Komura, T., Wang, W., Liu, L.: Emdm: Efficient motion diffusion model for fast, high-quality motion generation. arXiv preprint arXiv:2312.02256 (2023)" + ], + "bbox": [ + 212, + 146, + 787, + 397 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "MotionLCM", + "bbox": [ + 645, + 114, + 730, + 126 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 18 + } +] \ No newline at end of file diff --git a/2024/MotionLCM_ Real-time Controllable Motion Generation via Latent Consistency Model/69320bc9-c9fe-4fc1-8e78-3d61fd4c6087_model.json b/2024/MotionLCM_ Real-time Controllable Motion Generation via Latent Consistency Model/69320bc9-c9fe-4fc1-8e78-3d61fd4c6087_model.json new file mode 100644 index 0000000000000000000000000000000000000000..9609baad3f7aa08f87838d0cb87e6e3a59263c05 --- /dev/null +++ b/2024/MotionLCM_ Real-time Controllable Motion Generation via Latent Consistency Model/69320bc9-c9fe-4fc1-8e78-3d61fd4c6087_model.json @@ -0,0 +1,2581 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.248, + 0.141, + 0.756, + 0.187 + ], + "angle": 0, + "content": "MotionLCM: Real-time Controllable Motion Generation via Latent Consistency Model" + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.212, + 0.765, + 0.245 + ], + "angle": 0, + "content": "Wenxun Dai\\(^{1,2}\\), Ling-Hao Chen\\(^{2\\star}\\), Jingbo Wang\\(^{3\\boxtimes}\\), Jinpeng Liu\\(^{1,2}\\), Bo Dai\\(^{3\\boxtimes}\\), Yansong Tang\\(^{1,2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.236, + 0.255, + 0.767, + 0.327 + ], + "angle": 0, + "content": "1Shenzhen Key Laboratory of Ubiquitous Data Enabling, Tsinghua Shenzhen International Graduate School 2Tsinghua University 3Shanghai AI Laboratory {wx dai 2001, thu. lhchen, wangjingbo1219, liu. jinpeng.55}@gmail.com {doubledaibo, tangyansong15}@gmail.com Project page: https://dai-wenxun.github.io/MotionLCM-page" + }, + { + "type": "image", + "bbox": [ + 0.241, + 0.361, + 0.764, + 0.616 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.639, + 0.788, + 0.696 + ], + "angle": 0, + "content": "Fig. 1: We propose MotionLCM, a real-time controllable motion latent consistency model. Our model uses the last few frames of the previous motion as temporal control signals to autoregressively generate the next motion in real-time under different text prompts. Green blocks denote the junctions. The numbers in red are the inference time." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.747, + 0.741, + 0.817 + ], + "angle": 0, + "content": "Abstract. This work introduces MotionLCM, extending controllable motion generation to a real-time level. Existing methods for spatial-temporal control in text-conditioned motion generation suffer from significant runtime inefficiency. To address this issue, we first propose the motion latent consistency model (MotionLCM) for motion generation," + }, + { + "type": "page_footnote", + "bbox": [ + 0.22, + 0.825, + 0.51, + 0.84 + ], + "angle": 0, + "content": "* Project lead. ☑ Corresponding author." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.128 + ], + "angle": 0, + "content": "W. Dai et al." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.148, + 0.744, + 0.302 + ], + "angle": 0, + "content": "building upon the latent diffusion model [9]. By adopting one-step (or few-step) inference, we further improve the runtime efficiency of the motion latent diffusion model for motion generation. To ensure effective controllability, we incorporate a motion ControlNet within the latent space of MotionLCM and enable explicit control signals (e.g., initial poses) in the vanilla motion space to control the generation process directly, similar to controlling other latent-free diffusion models [29, 73] for motion generation. By employing these techniques, our approach can generate human motions with text and control signals in real-time. Experimental results demonstrate the remarkable generation and controlling capabilities of MotionLCM while maintaining real-time runtime efficiency." + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.314, + 0.728, + 0.328 + ], + "angle": 0, + "content": "Keywords: Text-to-Motion \\(\\cdot\\) Real-time Control \\(\\cdot\\) Consistency Model" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.355, + 0.376, + 0.371 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.387, + 0.479, + 0.719 + ], + "angle": 0, + "content": "Text-to-motion generation (T2M) has attracted increasing attention [1, 15, 43, 49, 65] due to its important roles in many applications [70, 72]. Previous attempts mainly focus on GANs [1, 38], VAEs [3, 19, 48, 49] and diffusion models [9, 11, 39, 56, 65, 78, 83] via pairwise text-motion data [17, 26, 45, 52, 53, 57, 63, 74] and achieve impressive generation results. Existing approaches [9, 65, 83] mainly take diffusion models [20, 46, 55, 60] as a base generative model, owing to their powerful ability to model motion distribution. However, these diffusion fashions inevitably require considerable sampling steps for motion synthesis during inference, even with some sampling acceleration methods [61]. Specifically, MDM [65] and MLD [9] require \\(\\sim\\)12s and \\(\\sim\\)0.2s to generate" + }, + { + "type": "image", + "bbox": [ + 0.489, + 0.416, + 0.789, + 0.557 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.486, + 0.567, + 0.789, + 0.68 + ], + "angle": 0, + "content": "Fig.2: Comparison of the inference time costs on HumanML3D [17]. We compare the AITS and FID metrics with five SOTA methods. The closer the model is to the origin the better. Diffusion-based models are indicated by the blue dashed box. Our MotionLCM achieves real-time inference speed while ensuring high-quality motion generation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.749 + ], + "angle": 0, + "content": "a high-quality motion sequence. Such low efficiency blocks the applications of generating high-quality motions in various real-time scenarios." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In addition to the language description itself serving as a coarse control signal, another line of research focuses on controlling the motion generation with spatial-temporal constraints [29, 56, 73]. Although these attempts enjoy impressive controlling ability in the T2M task, there still exists a significant gap towards real-time applications. For example, OmniControl [73] exhibits a relatively long inference time, \\(\\sim 81\\)s per sequence. Therefore, trading-off between generation" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.647, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "MotionLCM" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "quality and efficiency is a challenging problem. As a result, in this paper, we target the real-time controllable motion generation research problem." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.787, + 0.328 + ], + "angle": 0, + "content": "Recently, the concept of consistency models [44, 62] has been introduced in image generation, resulting in significant progress by enabling efficient and high-fidelity image synthesis with a minimal number of sampling steps (e.g., 4 steps vs. 50 steps). These properties perfectly align with our goal of accelerating motion generation without compromising generation quality. Therefore, we propose MotionLCM (Motion Latent Consistency Model) distilled from the motion latent diffusion model, MLD [9], to tackle the low-efficiency problem in diffusion sampling. To the best of our knowledge, we introduce consistency distillation into the motion generation area for the first time and accelerate motion generation to a real-time level via latent consistency distillation [44]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.329, + 0.788, + 0.585 + ], + "angle": 0, + "content": "Here, in MotionLCM, we are facing another challenge on how to control motions with spatial-temporal signals (e.g., initial poses) in the latent space. Previous methods [56, 73] model human motions in the vanilla motion space and can manipulate the motion directly in the denoising process. However, for our latent-diffusion-based MotionLCM, it is non-trivial to feed the control signals into the latent space. This is because the latent has no explicit motion semantics, which cannot be manipulated directly by the control signals. Inspired by the notable success of [82] in controllable image generation [55], we introduce a motion ControlNet to control motion generation in the latent space. However, the naïve motion ControlNet is not sufficient to provide supervision for the control signals. The main reason is the lack of explicit supervision in the motion space. Therefore, during the training phase, we decode the predicted latent through the frozen VAE [30] decoder into the vanilla motion space to provide explicit control supervision on the generated motion. Thanks to the powerful one-step inference capability of MotionLCM, the latent generated by MotionLCM can significantly facilitate control supervision both in the latent space and motion space for training the motion ControlNet compared to MLD [9]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.585, + 0.787, + 0.631 + ], + "angle": 0, + "content": "With our key designs, our proposed MotionLCM successfully enjoys the balance between the generation quality and efficiency in controllable motion generation. Before delivering into detail, we sum up our core contributions as follows." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.637, + 0.784, + 0.68 + ], + "angle": 0, + "content": "- We propose the Motion Latent Consistency Model (MotionLCM) via consistency distillation on the motion latent diffusion model extending controllable motion generation to a real-time level." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.682, + 0.784, + 0.71 + ], + "angle": 0, + "content": "- Building upon our achievement of real-time motion generation, we introduce a motion ControlNet, enabling high-quality controllable motion generation." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.711, + 0.784, + 0.74 + ], + "angle": 0, + "content": "- Extensive experimental results show that MotionLCM enjoys a good balance of generation quality, controlling capability, and real-time efficiency." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.637, + 0.784, + 0.74 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.761, + 0.388, + 0.777 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.789, + 0.488, + 0.803 + ], + "angle": 0, + "content": "2.1 Human Motion Generation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Generating human motions can be divided into three main fashions according to inputs: motion synthesis (1) without any condition [54,65,77,84], (2) with some" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.128 + ], + "angle": 0, + "content": "W. Dai et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.51 + ], + "angle": 0, + "content": "given multi-modal conditions, such as action labels [6, 12, 19, 31, 48, 75], textual description [1-5, 7, 7, 9-11, 13-15, 17, 18, 24, 27, 29, 37-39, 43, 49-51, 56, 64, 65, 68-70, 72, 78, 81, 83, 85, 86], audio or music [32-35, 59, 66], (3) with user-defined trajectories [22, 23, 29, 36, 56, 58, 68, 71, 73]. To generate diverse, natural, and high-quality human motions, many generative models have been explored by [2, 38, 49, 79, 80]. Recently, diffusion-based models significantly improved the motion generation performance and diversity [8, 9, 11, 16, 65, 76, 83] with stable training. Specifically, MotionDiffuse [83] represents the first text-based motion diffusion model that provides fine-grained instructions on body parts and achieves arbitrary-length motion synthesis with time-varied text prompts. MDM [65] introduces a motion diffusion model that operates on raw motion data, enabling both high-quality generation and generic conditioning that together comprise a good baseline for new motion generation tasks. Based on MDM [65], OmniControl [73] integrates flexible spatial-temporal control signals across different joints by combining analytic spatial guidance and realism guidance into the diffusion model, ensuring that the generated motion closely conforms to the input control signals. The work most relevant to ours is MLD [9], which introduces a motion latent-based diffusion model to enhance generation quality and reduce computational resource requirements. The key idea is training a VAE [30] for motion embedding, followed by implementing latent diffusion [55] within the learned latent space. However, these diffusion fashions inevitably require considerable sampling steps for motion synthesis during inference, even with some sampling acceleration methods [61]. Thus, we propose MotionLCM, which not only guarantees high-quality controllable motion generation but also achieves real-time runtime efficiency." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.537, + 0.331, + 0.554 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.573, + 0.788, + 0.635 + ], + "angle": 0, + "content": "In this section, we first briefly introduce preliminaries about latent consistency models in Sec. 3.1. Then, we describe how to conduct latent consistency distillation for motion generation in Sec. 3.2, followed by our implementation of motion control in latent space in Sec. 3.3. The overall pipeline is illustrated in Fig. 4." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.66, + 0.374, + 0.675 + ], + "angle": 0, + "content": "3.1 Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.812 + ], + "angle": 0, + "content": "The Consistency Model (CM) [62] introduces a kind of efficient generative model designed for efficient one-step or few-step generation. Given a Probability Flow ODE (a.k.a. PF-ODE) that smoothly converts data to noise, the CM is to learn the function \\( \\pmb{f}(\\cdot, \\cdot) \\) that maps any points on the ODE trajectory to its origin distribution (i.e., the solution of the PF-ODE). The consistency function is formally defined as \\( \\pmb{f}:(\\mathbf{x}_t,t)\\longmapsto \\mathbf{x}_{\\epsilon} \\), where \\( t\\in [\\epsilon ,T] \\), \\( T > 0 \\) is a fixed constant and \\( \\epsilon \\) is a small positive number to avoid numerical instability. According to [62], the consistency function should satisfy the self-consistency property:" + }, + { + "type": "equation", + "bbox": [ + 0.384, + 0.826, + 0.786, + 0.842 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {f} \\left(\\mathbf {x} _ {t}, t\\right) = \\boldsymbol {f} \\left(\\mathbf {x} _ {t ^ {\\prime}}, t ^ {\\prime}\\right), \\forall t, t ^ {\\prime} \\in [ \\epsilon , T ]. \\tag {1}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.647, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "MotionLCM" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.249, + 0.147, + 0.562, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.565, + 0.153, + 0.752, + 0.193 + ], + "angle": 0, + "content": "Distillation target of \\( f_{\\Theta} \\): Project any \\( \\mathbf{z}_t \\) to \\( \\mathbf{z}_0 \\) consistently. \\( t \\in [0, T] \\)" + }, + { + "type": "image", + "bbox": [ + 0.247, + 0.197, + 0.579, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.589, + 0.199, + 0.753, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.339, + 0.788, + 0.437 + ], + "angle": 0, + "content": "Fig. 3: The training objective of consistency distillation is to learn a consistency function \\( f_{\\Theta} \\), initialized with the parameters of a pre-trained diffusion model (e.g., MLD [9]). This function \\( f_{\\Theta} \\) should projects any points (i.e., \\( \\mathbf{z}_t \\)) on the ODE trajectory to its solution (i.e., \\( \\mathbf{z}_0 \\)). Once the pre-trained model [9] is distilled, unlike the traditional denoising model [65, 83] that requires considerable sampling steps, our MotionLCM can generate high-quality motion sequences with one-step sampling and further improve the generation quality through multi-step inference." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.465, + 0.788, + 0.543 + ], + "angle": 0, + "content": "As shown in Eq. (1), the self-consistency property indicates that for arbitrary pairs of \\((\\mathbf{x}_t,t)\\) on the same PF-ODE trajectory, the outputs of the model should be consistent. The goal of a parameterized consistency model \\(f_{\\Theta}\\) is to learn a consistency function from data by enforcing the self-consistency property in Eq. (1). To ensure that \\(f_{\\Theta}(\\mathbf{x},\\epsilon) = \\mathbf{x}\\), the consistency model \\(f_{\\Theta}\\) is parameterized as," + }, + { + "type": "equation", + "bbox": [ + 0.36, + 0.552, + 0.785, + 0.568 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {f} _ {\\Theta} (\\mathbf {x}, t) = c _ {\\text {s k i p}} (t) \\mathbf {x} + c _ {\\text {o u t}} (t) \\boldsymbol {F} _ {\\Theta} (\\mathbf {x}, t), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.578, + 0.788, + 0.639 + ], + "angle": 0, + "content": "where \\( c_{\\mathrm{skip}}(t) \\) and \\( c_{\\mathrm{out}}(t) \\) are differentiable functions with \\( c_{\\mathrm{skip}}(\\epsilon) = 1 \\) and \\( c_{\\mathrm{out}}(\\epsilon) = 0 \\), and \\( \\pmb{F}_{\\Theta}(\\cdot, \\cdot) \\) is a deep neural network to learn the self-consistency. The CM trained from distilling the knowledge of pre-trained diffusion models is called Consistency Distillation. The consistency loss is defined as follows," + }, + { + "type": "equation", + "bbox": [ + 0.306, + 0.647, + 0.785, + 0.667 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} (\\boldsymbol {\\Theta}, \\boldsymbol {\\Theta} ^ {-}; \\boldsymbol {\\Phi}) = \\mathbb {E} \\left[ d \\left(\\boldsymbol {f} _ {\\boldsymbol {\\Theta}} \\left(\\mathbf {x} _ {t _ {n + 1}}, t _ {n + 1}\\right), \\boldsymbol {f} _ {\\boldsymbol {\\Theta} ^ {-}} \\left(\\hat {\\mathbf {x}} _ {t _ {n}} ^ {\\boldsymbol {\\Phi}}, t _ {n}\\right)\\right) \\right], \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.675, + 0.788, + 0.751 + ], + "angle": 0, + "content": "where \\(d(\\cdot, \\cdot)\\) is a chosen metric function for measuring the distance between two samples. \\(f_{\\Theta}(\\cdot, \\cdot)\\) and \\(f_{\\Theta^{-}}(\\cdot, \\cdot)\\) are referred to as \"online network\" and \"target network\" according to [62]. Besides, \\(\\Theta^{-}\\) is updated with the exponential moving average (EMA) of the parameters of \\(\\Theta^{1}\\). In Eq. (3), \\(\\hat{\\mathbf{x}}_{t_n}^{\\Phi}\\) is a one-step estimation of \\(\\mathbf{x}_{t_n}\\) from \\(\\mathbf{x}_{t_{n+1}}\\), which is formulated as," + }, + { + "type": "equation", + "bbox": [ + 0.343, + 0.762, + 0.785, + 0.78 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {x}} _ {t _ {n}} ^ {\\Phi} \\leftarrow \\mathbf {x} _ {t _ {n + 1}} + (t _ {n} - t _ {n + 1}) \\Phi (\\mathbf {x} _ {t _ {n + 1}}, t _ {n + 1}, \\emptyset), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.789, + 0.614, + 0.804 + ], + "angle": 0, + "content": "where \\(\\Phi\\) is a one-step ODE solver applied to PF-ODE." + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.81, + 0.787, + 0.841 + ], + "angle": 0, + "content": "\\(^{1}\\) EMA operation: \\(\\Theta^{-} \\gets \\mathbf{sg}(\\mu \\Theta^{-} + (1 - \\mu)\\Theta)\\), where \\(\\mathbf{sg}(\\cdot)\\) denotes the stopgrad operation and \\(\\mu\\) satisfies \\(0 \\leq \\mu < 1\\)." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "W. Dai et al." + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.148, + 0.573, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.287, + 0.287, + 0.506, + 0.299 + ], + "angle": 0, + "content": "(a) Motion Latent Consistency Distillation" + }, + { + "type": "image", + "bbox": [ + 0.576, + 0.145, + 0.787, + 0.283 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.592, + 0.287, + 0.774, + 0.3 + ], + "angle": 0, + "content": "(b) Motion Control in Latent Space" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.31, + 0.789, + 0.465 + ], + "angle": 0, + "content": "Fig. 4: The overview of MotionLCM. (a) Motion Latent Consistency Distillation (Sec. 3.2). Given a raw motion sequence \\(\\mathbf{x}_0^{1:N}\\), a pre-trained VAE [30] encoder first compresses it into the latent space, then a forward diffusion operation is performed to add \\(n + k\\) steps of noise. Then, the noisy \\(\\mathbf{z}_{n + k}\\) is fed into the online network and teacher network to predict the clean latent. The target network takes the \\(k\\)-step estimation results of the teacher output to predict the clean latent. To learn self-consistency, a loss is applied to enforce the output of the online network and target network to be consistent. (b) Motion Control in Latent Space (Sec. 3.3). With the powerful MotionLCM trained in the first stage, we incorporate a motion ControlNet into the MotionLCM to achieve controllable motion generation. Furthermore, we leverage the decoded motion to explicitly supervise the spatial-temporal control signals (i.e., initial poses \\(\\mathbf{g}^{1:\\tau}\\))." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.493, + 0.789, + 0.646 + ], + "angle": 0, + "content": "The Latent Consistency Model (LCM) [44] learns the self-consistency property in the latent space \\( D_{\\mathbf{z}} = \\{(\\mathbf{z},\\mathbf{c})|\\mathbf{z} = \\mathcal{E}(\\mathbf{x}),(\\mathbf{x},\\mathbf{c})\\in D\\} \\), where \\( D \\) denotes the dataset, \\( \\mathbf{c} \\) is the given condition, and \\( \\mathcal{E} \\) is the pre-trained encoder. Compared to CMs [62] using the numerical continuous PF-ODE solver [28], LCMs [44] adopt the discrete-time schedule [41,42,61] to adapt to Stable Diffusion [55]. Instead of ensuring consistency between adjacent time steps \\( t_{n + 1}\\rightarrow t_n \\), LCMs [44] are designed to ensure consistency between the current time step and \\( k \\)-step away, i.e., \\( t_{n + k}\\to t_n \\), thereby significantly reducing convergence time costs. As classifier-free guidance (CFG) [21] plays a crucial role in synthesizing high-quality text-aligned images, LCMs integrate CFG into the distillation as follows," + }, + { + "type": "equation", + "bbox": [ + 0.245, + 0.67, + 0.787, + 0.69 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {z}} _ {t _ {n}} ^ {\\boldsymbol {\\Phi}, w} \\gets \\mathbf {z} _ {t _ {n + k}} + (1 + w) \\boldsymbol {\\Phi} (\\mathbf {z} _ {t _ {n + k}}, t _ {n + k}, t _ {n}, \\mathbf {c}) - w \\boldsymbol {\\Phi} (\\mathbf {z} _ {t _ {n + k}}, t _ {n + k}, t _ {n}, \\emptyset), \\qquad (5)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.7, + 0.788, + 0.762 + ], + "angle": 0, + "content": "where \\( w \\) denotes the CFG scale which is uniformly sampled from \\( [w_{\\min}, w_{\\max}] \\) and \\( k \\) is the skipping interval. To efficiently perform the above \\( k \\)-step guided distillation, LCMs augment the consistency function to \\( f: (\\mathbf{z}_t, t, w, \\mathbf{c}) \\longmapsto \\mathbf{z}_0 \\), which is also the form adopted by our MotionLCM." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.784, + 0.66, + 0.8 + ], + "angle": 0, + "content": "3.2 MotionLCM: Motion Latent Consistency Model" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Motion compression into the latent space. Motivated by [44, 62], we propose MotionLCM (Motion Latent Consistency Model) to tackle the low-efficiency" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.647, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "MotionLCM" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.51 + ], + "angle": 0, + "content": "problem in motion diffusion models [65, 83], unleashing the potential of LCM in the motion generation task. Similar to MLD [9], our MotionLCM adopts a consistency model in the motion latent space. We choose MLD [9] as the underlying diffusion model to distill from. We aim to achieve few-step \\((2\\sim 4)\\) and even one-step inference without compromising motion quality. In MLD, an autoencoder \\((\\mathcal{E},\\mathcal{D})\\) is first trained to compress a high dimensional motion into a low dimensional latent vector \\(\\mathbf{z} = \\mathcal{E}(\\mathbf{x})\\), which is then decoded to reconstruct the motion as \\(\\hat{\\mathbf{x}} = \\mathcal{D}(\\mathbf{z})\\). Training diffusion models in the motion latent space greatly reduces the computational resources compared to the vanilla diffusion models trained on raw motion sequences (i.e., motion space) and speeds up the inference process. Thus, we effectively leverage the motion latent space for consistency distillation. Motion latent consistency distillation. An overview of our motion latent consistency distillation is described in Fig. 4 (a). A raw motion sequence \\(\\mathbf{x}_0^{1:N} = \\{\\mathbf{x}^i\\}_{i=1}^N\\) is a sequence of human poses, where \\(N\\) is the number of frames. We follow [17] to use the redundant motion representation for our experiments, which is widely used in previous work [9, 65, 83]. As shown in Fig. 4 (a), given a raw motion sequence \\(\\mathbf{x}_0^{1:N}\\), a pre-trained VAE [30] encoder first compresses it into the latent space, \\(\\mathbf{z}_0 = \\mathcal{E}(\\mathbf{x}_0)\\). Then, a forward diffusion operation with \\(n + k\\) steps is conducted to add noise on \\(\\mathbf{z}_0\\), where \\(k\\) is the skipping interval illustrated in Sec. 3.1. The noisy \\(\\mathbf{z}_{n+k}\\) is fed to the frozen teacher network and trainable online network to predict the clean \\(\\hat{\\mathbf{z}}_0^*\\), and \\(\\hat{\\mathbf{z}}_0\\). The target network uses the cleaner \\(\\hat{\\mathbf{z}}_n\\) obtained by a \\(k\\)-step ODE solver \\(\\Phi\\), such as DDIM [61] to predict the \\(\\hat{\\mathbf{z}}_0^-\\). Since the classifier-free guidance (CFG) [21] is essential for condition alignment in diffusion models [9, 55, 65], we integrate CFG into the distillation," + }, + { + "type": "equation", + "bbox": [ + 0.271, + 0.521, + 0.786, + 0.537 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {z}} _ {n} \\leftarrow \\mathbf {z} _ {n + k} + (1 + w) \\boldsymbol {\\Phi} \\left(\\mathbf {z} _ {n + k}, t _ {n + k}, t _ {n}, \\mathbf {c}\\right) - w \\boldsymbol {\\Phi} \\left(\\mathbf {z} _ {n + k}, t _ {n + k}, t _ {n}, \\emptyset\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.547, + 0.788, + 0.593 + ], + "angle": 0, + "content": "where \\(\\mathbf{c}\\) is the text condition and \\(w\\) denotes the guidance scale. To ensure the self-consistency property defined in Eq. (1), the latent consistency distillation loss \\(\\mathcal{L}_{\\mathrm{LCD}}\\) is designed as follows," + }, + { + "type": "equation", + "bbox": [ + 0.28, + 0.603, + 0.786, + 0.62 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {L C D}} (\\boldsymbol {\\Theta}, \\boldsymbol {\\Theta} ^ {-}) = \\mathbb {E} \\left[ d \\left(\\boldsymbol {f} _ {\\boldsymbol {\\Theta}} (\\mathbf {z} _ {n + k}, t _ {n + k}, w, \\mathbf {c}), \\boldsymbol {f} _ {\\boldsymbol {\\Theta} ^ {-}} (\\hat {\\mathbf {z}} _ {n}, t _ {n}, w, \\mathbf {c})\\right) \\right], \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.788, + 0.766 + ], + "angle": 0, + "content": "where \\(d(\\cdot, \\cdot)\\) is a distance measuring function, such as L2 loss or Huber loss [25]. As discussed in Sec. 3.1, the target network \\(\\Theta^{-}\\) is updated with the exponential moving average (EMA) of the trainable parameters of the online network \\(\\Theta\\). Here we define the teacher network \\(\\Theta^{*}\\) as the pre-trained motion latent diffusion model, i.e., MLD [9]. According to [44], the online network and target network are initialized with the parameters of the teacher network. During the inference phase, as shown in Fig. 5, our MotionLCM can generate high-quality motions with one-step sampling and achieve the fastest runtime (\\(\\sim\\)30ms per motion sequence) compared to other motion diffusion models [9,65]." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.786, + 0.664, + 0.803 + ], + "angle": 0, + "content": "3.3 Controllable Motion Generation in Latent Space" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "After addressing the low-efficiency issue in the motion latent diffusion model [9], we delve into another exploration of real-time motion control. Inspired by the" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.128 + ], + "angle": 0, + "content": "W. Dai et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.42 + ], + "angle": 0, + "content": "great success of ControlNet [82] in controllable image generation [55], we introduce a motion ControlNet \\(\\Theta^a\\) in the latent space of MotionLCM and initialize the motion ControlNet with a trainable copy of MotionLCM. Specifically, each layer in the motion ControlNet is appended with a zero-initialized linear layer for eliminating random noise in the initial training steps. To achieve an autoregressive motion generation paradigm, as depicted in Fig. 1, we define the motion control task as generating motions given the initial \\(\\tau\\) poses and textual description. As shown in Fig. 4 (b), the initial \\(\\tau\\) poses are defined by the trajectories of \\(K\\) control joints, \\(\\mathbf{g}^{1:\\tau} = \\{\\mathbf{g}^i\\}_{i=1}^\\tau\\), where \\(\\mathbf{g}^i \\in \\mathbb{R}^{K \\times 3}\\) denotes the global absolute locations of each control joint. In our motion control pipeline, we design a Trajectory Encoder \\(\\Theta^b\\) consisting of stacked transformer [67] layers to encode the trajectory signals. We append a global token (i.e., [CLS]) before the start of the trajectory sequence as the output feature of the encoder, which is added to the noisy \\(\\mathbf{z}_n\\) and fed into the trainable motion ControlNet \\(\\Theta^a\\). Under the guidance of motion ControlNet, MotionLCM predicts the denoised \\(\\hat{\\mathbf{z}}_0\\) through the consistency function \\(f_{\\Theta^s}\\), where \\(\\Theta^s\\) is the combination of \\(\\Theta^a\\), \\(\\Theta^b\\) and \\(\\Theta\\). The following reconstruction loss \\(\\mathcal{L}_{\\mathrm{recon}}\\) optimizes the motion ControlNet \\(\\Theta^a\\) and Trajectory Encoder \\(\\Theta^b\\)," + }, + { + "type": "equation", + "bbox": [ + 0.335, + 0.431, + 0.786, + 0.449 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {r e c o n}} \\left(\\boldsymbol {\\Theta} ^ {a}, \\boldsymbol {\\Theta} ^ {b}\\right) = \\mathbb {E} \\left[ d \\left(\\boldsymbol {f} _ {\\boldsymbol {\\Theta} ^ {s}} \\left(\\mathbf {z} _ {n}, t _ {n}, w, \\mathbf {c} ^ {*}\\right), \\mathbf {z} _ {0}\\right) \\right], \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.459, + 0.788, + 0.611 + ], + "angle": 0, + "content": "where \\(\\mathbf{c}^*\\) includes the text condition and control guidance from the Trajectory Encoder and the motion ControlNet. However, during training, the sole reconstruction supervision in the latent space is insufficient. We argue this is because the controllable motion generation requires more detailed constraints, which cannot be effectively provided solely by the reconstruction loss in the latent space. Unlike previous methods like OmniControl [73], which directly diffuse in the motion space, allowing explicit supervision of control signals, effectively supervising control signals in the latent space is non-trivial. Therefore, we utilize the frozen VAE [30] decoder \\(\\mathcal{D}\\) to decode the latent \\(\\hat{\\mathbf{z}}_0\\) into the motion space, obtaining the predicted motion \\(\\hat{\\mathbf{x}}_0\\), thereby introducing the control loss \\(\\mathcal{L}_{\\mathrm{control}}\\) as follows," + }, + { + "type": "equation", + "bbox": [ + 0.297, + 0.622, + 0.787, + 0.663 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {c o n t r o l}} (\\boldsymbol {\\Theta} ^ {a}, \\boldsymbol {\\Theta} ^ {b}) = \\mathbb {E} \\left[ \\frac {\\sum_ {i} \\sum_ {j} m _ {i j} | | R (\\hat {\\mathbf {x}} _ {0}) _ {i j} - R (\\mathbf {x} _ {0}) _ {i j} | | _ {2} ^ {2}}{\\sum_ {i} \\sum_ {j} m _ {i j}} \\right], \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.672, + 0.788, + 0.72 + ], + "angle": 0, + "content": "where \\( R(\\cdot) \\) converts the joint's local positions to global absolute locations and \\( m_{ij} \\in \\{0,1\\} \\) is the binary joint mask at frame \\( i \\) for the joint \\( j \\). Then we optimize the motion ControlNet \\( \\Theta^a \\) and Trajectory Encoder \\( \\Theta^b \\) with the overall objective," + }, + { + "type": "equation", + "bbox": [ + 0.371, + 0.733, + 0.787, + 0.758 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\boldsymbol {\\Theta} ^ {a}, \\boldsymbol {\\Theta} ^ {b} = \\underset {\\boldsymbol {\\Theta} ^ {a}, \\boldsymbol {\\Theta} ^ {b}} {\\operatorname {a r g m i n}} (\\mathcal {L} _ {\\text {r e c o n}} + \\lambda \\mathcal {L} _ {\\text {c o n t r o l}}), \\end{array} \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.841 + ], + "angle": 0, + "content": "where \\(\\lambda\\) is the weight to balance the two losses. This design enables explicit control signals to directly influence the generation process, similar to controlling other latent-free diffusion models for motion generation [29,73]. Extensive experiments demonstrate the introduced supervision is very helpful in improving the motion control performance, which will be introduced in the following section." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.647, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "MotionLCM" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.377, + 0.163 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.176, + 0.788, + 0.238 + ], + "angle": 0, + "content": "In this section, we first present the experimental setup details in Sec. 4.1. Subsequently, we provide quantitative and qualitative comparisons to evaluate the effectiveness of our proposed MotionLCM framework in Sec. 4.2 and Sec. 4.3. Finally, we conduct comprehensive ablation studies on MotionLCM in Sec. 4.4." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.258, + 0.427, + 0.274 + ], + "angle": 0, + "content": "4.1 Experimental setup" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.281, + 0.788, + 0.357 + ], + "angle": 0, + "content": "Datasets. We experiment on the popular HumanML3D [17] dataset, featuring 14,616 unique human motion sequences with 44,970 textual descriptions. For a fair comparison with previous methods [9,17,49,65,83], we take the redundant motion representation, including root velocity, root height, local joint positions, velocities, rotations in root space, and the foot contact binary labels." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.358, + 0.789, + 0.584 + ], + "angle": 0, + "content": "Evaluation metrics. We extend the evaluation metrics of previous works [9, 17, 73]. (1) Time cost: We follow [9] to report the Average Inference Time per Sentence (AITS) to evaluate the inference efficiency of models. (2) Motion quality: Frechet Inception Distance (FID) is adopted as a principal metric to evaluate the feature distributions between the generated and real motions. The feature extractor employed is from [17]. (3) Motion diversity: MultiModality (MModality) measures the generation diversity conditioned on the same text and Diversity calculates variance through features [17]. (4) Condition matching: Following [17], we calculate the motion-retrieval precision (R-Precision) to report the text-motion Top-1/2/3 matching accuracy and Multimodal Distance (MM Dist) calculates the mean distance between motions and texts. (5) Control error: Trajectory error (Traj. err.) quantifies the ratio of unsuccessful trajectories, characterized by any control joint location error surpassing a predetermined threshold. Location error (Loc. err.) represents the unsuccessful joints. Average error (Avg. err.) denotes the mean location error of the control joints." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.584, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Implementation details. Our baseline motion diffusion model is based on MLD [9]. We reproduce MLD with higher performance. Unless otherwise specified, all our experiments are conducted on this model. For MotionLCM, we employ the AdamW [40] optimizer for 96K iterations using a cosine decay learning rate scheduler and 1K iterations of linear warm-up. A batch size of 256 and a learning rate of 2e-4 are used. We set the training guidance scale range as \\([w_{\\mathrm{min}}, w_{\\mathrm{max}}] = [5, 15]\\), with the testing guidance scale set to 7.5, and adopt the EMA rate \\(\\mu = 0.95\\) by default. We use the DDIM [61] solver with skipping interval \\(k = 20\\) and choose the Huber [25] loss as the distance measuring function \\(d\\). For motion ControlNet, we use the AdamW [40] optimizer for 192K iterations with 1K iterations of linear warm-up. The batch size and learning rate are set to 128 and 1e-4. The learning rate scheduler is the same as the first stage. For the training objective, we employ \\(d\\) as the L2 loss and set the control loss weight \\(\\lambda\\) to 1.0 by default. We set the control ratio \\(\\tau\\) as 0.25 and the number of control joints as \\(K = 6\\) (i.e., Pelvis, Left foot, Right foot, Head, Left wrist, and Right wrist) in both training and testing. We implement our model using PyTorch [47] with training on an NVIDIA RTX 4090 GPU and testing on a Tesla V100 GPU." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "W. Dai et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.229 + ], + "angle": 0, + "content": "Table 1: Comparison of text-conditional motion synthesis on HumanML3D [17] dataset. We compute suggested metrics following [17]. We repeat the evaluation 20 times for each metric and report the average with a \\(95\\%\\) confidence interval. “\\(\\rightarrow\\)” indicates that the closer to the real data, the better. **Bold and underline indicate the best and the second best result.** denotes the reproduced version of MLD [9]. The MotionLCM in one-step inference (30ms) surpasses all state-of-the-art models." + }, + { + "type": "table", + "bbox": [ + 0.223, + 0.24, + 0.785, + 0.407 + ], + "angle": 0, + "content": "
MethodsAITS ↓R-Precision ↑FID ↓MM Dist ↓Diversity →MModality ↑
Top 1Top 2Top 3
Real-0.511±.0030.703±.0030.797±.0020.002±.0002.794±.0089.503±.065-
Seq2Seq [37]-0.180±.0020.300±.0020.396±.00211.75±.0355.529±.0076.223±.061-
JL2P [2]-0.246±.0010.387±.0020.486±.00211.02±.0465.296±.0087.676±.058-
T2G [5]-0.165±.0010.267±.0020.345±.0027.664±.0306.030±.0086.409±.071-
Hier [14]-0.301±.0020.425±.0020.552±.0046.532±.0245.012±.0188.332±.042-
TEMOS [49]0.0170.424±.0020.612±.0020.722±.0023.734±.0283.703±.0088.973±.0710.368±.018
T2M [17]0.0380.457±.0020.639±.0030.740±.0031.067±.0023.340±.0089.188±.0022.090±.083
MDM [65]24.740.320±.0050.498±.0040.611±.0070.544±.0445.566±.0279.559±.0862.799±.072
MotionDiffuse [83]14.740.491±.0010.681±.0010.782±.0010.630±.0013.113±.0019.410±.0491.553±.042
MLD [9]0.2170.481±.0030.673±.0030.772±.0020.473±.0133.196±.0109.724±.0822.413±.079
MLD* [9]0.2250.504±.0020.698±.0030.796±.0020.450±.0113.052±.0099.634±.0642.267±.082
MotionLCM (1-step)0.0300.502±.0030.701±.0020.803±.0020.467±.0123.022±.0099.631±.0662.172±.082
MotionLCM (2-step)0.0350.505±.0030.705±.0020.805±.0020.368±.0112.986±.0089.640±.0522.187±.094
MotionLCM (4-step)0.0430.502±.0030.698±.0020.798±.0020.304±.0123.012±.0079.607±.0662.259±.092
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.436, + 0.525, + 0.451 + ], + "angle": 0, + "content": "4.2 Comparisons on Text-to-motion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.46, + 0.789, + 0.765 + ], + "angle": 0, + "content": "In the following part, we first evaluate our MotionLCM on the text-to-motion (T2M) task. We compare our method with some T2M baselines on HumanML3D [17] with suggested metrics [17] under the \\(95\\%\\) confidence interval from 20 times running. As MotionLCM is based on MLD, we mainly focus on the performance compared with MLD. For evaluating time efficiency, we compare the Average Inference Time per Sentence (AITS) with TEMOS [49], T2M [17], MDM [65], MotionDiffuse [83] and MLD [9]. The results are borrowed from MLD [9]. The deterministic methods [2,5,14,37], are unable to produce diverse results from a single input text and thus we leave their MModality metrics empty. For the quantitative results, as shown in Tab. 1, our MotionLCM boasts an impressive real-time runtime efficiency, averaging around 30ms per motion sequence during inference. This performance exceeds that of previous diffusion-based methods [9,65,83] and even surpasses MLD [9] by an order of magnitude. Furthermore, despite employing only one-step inference, our MotionLCM can approximate or even surpass the performance of MLD [9] (DDIM [61] 50 steps). With two-step inference, we achieve the best R-Precision and MM Dist metrics, while increasing the sampling steps to four yields the best FID. The above results demonstrate the effectiveness of latent consistency distillation. For the qualitative results, as shown in Fig. 5, MotionLCM not only accelerates motion generation to real-time speed but also delivers high-quality outputs, closely aligning with the textual descriptions." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.785, + 0.666, + 0.801 + ], + "angle": 0, + "content": "4.3 Comparisons on Controllable Motion Generation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.809, + 0.787, + 0.842 + ], + "angle": 0, + "content": "As shown in Tab. 2, we compare our MotionLCM with OmniControl [73] and MLD [9]. We observe that OmniControl struggles with multi-joint control and" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.647, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "MotionLCM" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table", + "bbox": [ + 0.219, + 0.145, + 0.785, + 0.423 + ], + "angle": 0, + "content": "
RealMotionLCM (Ours)MLDMDM
“a person walks clockwise in a large curve while swinging their arms.”
0.033s0.201s22.58s
“a person walking like a bird and then sniffing the air.”
0.031s0.192s21.16s
“a person slightly bent over with right hand pressing against the air walks forward slowly”
0.031s0.197s23.44s
" + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.447, + 0.785, + 0.489 + ], + "angle": 0, + "content": "Fig. 5: Qualitative comparison of the state-of-the-art methods in the text-to-motion task. With only one-step inference, MotionLCM achieves the fastest motion generation while producing high-quality movements that closely match the textual descriptions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.519, + 0.788, + 0.732 + ], + "angle": 0, + "content": "falls short in both generation quality and control performance compared to MotionLCM. To verify the effectiveness of the latent generated by our MotionLCM for training motion ControlNet, we conducted the following two sets of experiments: \"LC\" and \"MC\", which indicate introducing control supervision in the latent space and motion space. It can be observed that under the same experimental settings, MotionLCM maintains higher fidelity and significantly outperforms MLD [9] in motion control performance. This demonstrates that the latent generated by MotionLCM is more effective for training motion ControlNet compared to MLD [9]. In terms of inference speed, MotionLCM (1-step) is \\( 1929 \\times \\) faster compared to OmniControl [73] and \\( 13 \\times \\) faster than MLD [9]. For qualitative results, as shown in Fig. 6, OmniControl fails to control the initial poses in the second example and does not generate motion that aligns with the text in the third case. However, our MotionLCM not only adheres to the control of the initial poses but also generates motions that match the textual descriptions." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.754, + 0.402, + 0.768 + ], + "angle": 0, + "content": "4.4 Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Impact of the hyperparameters of training MotionLCM. We conduct a comprehensive analysis of the training hyperparameters of MotionLCM, including the training guidance scale range \\([w_{\\mathrm{min}}, w_{\\mathrm{max}}]\\), EMA rate \\(\\mu\\), skipping interval \\(k\\), and the type of loss. We summarize the evaluation results based on one-step" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "W. Dai et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.203 + ], + "angle": 0, + "content": "Table 2: Comparison of motion control on HumanML3D [17] dataset. Bold indicates the best result. Our MotionLCM outperforms OmniControl [73] and MLD [9] regarding generation quality, control performance, and inference speed. \"LC\" and \"MC\" refer to the control supervision introduced in the latent space and motion space." + }, + { + "type": "table", + "bbox": [ + 0.224, + 0.213, + 0.785, + 0.347 + ], + "angle": 0, + "content": "
MethodsAITS ↓FID ↓R-Precision ↑ Top 3Diversity →Traj. err. ↓ (50cm)Loc. err. ↓ (50cm)Avg. err. ↓
Real-0.0020.7979.5030.00000.00000.0000
OmniControl [73]81.002.3280.5578.8670.33620.03220.0977
MLD [9] (LC)0.5520.4690.7239.4760.42300.06530.1690
MotionLCM (1-step, LC)0.0420.3190.7529.4240.29860.03440.1410
MotionLCM (2-step, LC)0.0470.3150.7709.4270.28400.03280.1365
MotionLCM (4-step, LC)0.0630.3280.7459.4410.29730.03390.1398
MLD [9] (LC&MC)0.5520.5550.7549.3730.27220.02150.1265
MotionLCM (1-step, LC&MC)0.0420.4190.7569.3900.19880.01470.1127
MotionLCM (2-step, LC&MC)0.0470.3970.7599.4690.19600.01430.1092
MotionLCM (4-step, LC&MC)0.0630.4440.7539.3550.20890.01720.1140
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.369, + 0.788, + 0.412 + ], + "angle": 0, + "content": "Table 3: Ablation study on different training guidance scale ranges \\([w_{\\mathrm{min}}, w_{\\mathrm{max}}]\\), EMA rates \\(\\mu\\), skipping intervals \\(k\\) and types of loss. We use metrics in Tab. 1 and adopt a one-step inference setting with the testing CFG scale of 7.5 for fair comparison." + }, + { + "type": "table", + "bbox": [ + 0.254, + 0.424, + 0.756, + 0.621 + ], + "angle": 0, + "content": "
MethodsR-Precision ↑ Top 1FID ↓MM Dist ↓Diversity →MModality ↑
Real0.511±.0030.002±.0002.794±.0089.503±.065-
MotionLCM (w ∈ [5,15])0.502±.0030.467±.0123.022±.0099.631±.0662.172±.082
MotionLCM (w ∈ [2,18])0.497±.0030.481±.0093.030±.0109.644±.0732.226±.091
MotionLCM (w = 7.5)0.486±.0020.479±.0093.094±.0099.610±.0722.320±.097
MotionLCM (μ = 0.95)0.502±.0030.467±.0123.022±.0099.631±.0662.172±.082
MotionLCM (μ = 0.50)0.498±.0030.478±.0093.022±.0109.655±.0712.188±.087
MotionLCM (μ = 0)0.499±.0030.505±.0083.018±.0099.706±.0702.123±.085
MotionLCM (k = 50)0.488±.0030.547±.0113.096±.0109.511±.0742.324±.091
MotionLCM (k = 20)0.502±.0030.467±.0123.022±.0099.631±.0662.172±.082
MotionLCM (k = 10)0.497±.0030.449±.0093.017±.0109.693±.0752.133±.086
MotionLCM (k = 5)0.488±.0030.438±.0093.044±.0099.647±.0742.147±.083
MotionLCM (k = 1)0.442±.0020.635±.0113.255±.0089.384±.0802.146±.075
MotionLCM (w/ Huber)0.502±.0030.467±.0123.022±.0099.631±.0662.172±.082
MotionLCM (w/ L2)0.486±.0020.622±.0103.114±.0099.573±.0692.218±.086
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.841 + ], + "angle": 0, + "content": "inference in Tab. 3. We find out that using a dynamic training guidance scale (e.g., \\( w \\in [5,15] \\)) during training leads to an improvement in model performance compared to using a static training guidance scale (e.g., \\( w = 7.5 \\)). Additionally, an excessively large range for the training guidance scale can also negatively impact the performance of the model (e.g., \\( w \\in [2,18] \\)). Regarding the EMA rate \\( \\mu \\), we observe that the larger the value of \\( \\mu \\), the better the performance of the model. This indicates that maintaining a slower update rate for the target network \\( \\Theta^{-} \\) helps improve the performance of latent consistency distillation. When the skipping interval \\( k \\) continues to increase, the performance of the distillation model improves progressively, but larger values of \\( k \\) (e.g., \\( k = 50 \\)) may result in inferior results. As for the type of loss, the Huber loss [25] significantly outperforms the L2 loss, demonstrating its superior robustness." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.647, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "MotionLCM" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table", + "bbox": [ + 0.219, + 0.144, + 0.788, + 0.451 + ], + "angle": 0, + "content": "
MotionLCM (w/o ControlNet)0.029s0.033s0.030s0.033s0.029s
OmniControl87.81s90.00s85.11s80.15s86.77s
MotionLCM (w/ ControlNet)0.043s0.045s0.041s0.045s0.047s
Reala man lifts his left arm up.the person is widewalking.a person walks three steps to his right, then five steps to his left, and finally three steps to his right.this person is walking in circles then moves across room to step around an object.the person is holding their head while walking.
" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.473, + 0.788, + 0.543 + ], + "angle": 0, + "content": "Fig. 6: Qualitative comparison of the state-of-the-art methods in the motion control task. We provide the visualized motion results and real references from five prompts. Compared to OmniControl [73], MotionLCM with ControlNet not only generates the initial poses that accurately follow the given multi-joint trajectories (i.e., the green poses in real references) but also produces motions that closely align with the texts." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.58, + 0.788, + 0.732 + ], + "angle": 0, + "content": "Impact of control loss weights \\(\\lambda\\). To verify the impact of different control loss weights \\(\\lambda\\) on the control performance of MotionLCM, we report the experimental results in Tab. 4. We also include experiments of MotionLCM without ControlNet (i.e., only text-to-motion) for comparison. We found a significant improvement in control-related metrics (e.g., Loc. err.) after introducing motion ControlNet (i.e., \\(\\lambda = 0\\)). Furthermore, control performance can be further improved by introducing control loss (i.e., \\(\\lambda > 0\\)). Increasing the weight \\(\\lambda\\) enhances control performance but leads to a decline in the generation quality, which is reflected in higher FID scores. To balance these two aspects, we adopt \\(\\lambda = 1\\) as our default setting for training motion ControlNet." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Impact of different control ratios \\(\\tau\\) and number of control joints \\(K\\). In Tab. 5, we present the results of all models with the testing control ratio as 0.25 and keep the number of control joints \\(K\\) equal in both training and testing. We found that the model with the fixed training control ratio (i.e., \\(\\tau = 0.25\\)) performs better compared to a dynamic ratio (e.g., \\(\\tau \\in [0.1, 0.5]\\)), and we discover that our model maintains good performance when incorporating additional redundant control signals, such as whole-body joints with \\(K = 22\\)." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "W. Dai et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.145, + 0.788, + 0.174 + ], + "angle": 0, + "content": "Table 4: Ablation study on different control loss weights \\( \\lambda \\). We present the results of (1, 2, 4)-step inference. We add the MotionLCM without ControlNet for comparison." + }, + { + "type": "table", + "bbox": [ + 0.246, + 0.185, + 0.763, + 0.374 + ], + "angle": 0, + "content": "
MethodsFID ↓ Top 3R-Precision ↑ Diversity →Traj. err. ↓ (50cm)Loc. err. ↓ (50cm)Avg. err. ↓
Real0.0020.7979.5030.00000.00000.0000
MotionLCM (1-step, w/o control)0.4670.8039.6310.76050.23020.3493
MotionLCM (2-step, w/o control)0.3680.8059.6400.76460.22140.3386
MotionLCM (4-step, w/o control)0.3040.7989.6070.77390.22070.3359
MotionLCM (1-step, λ = 0)0.3190.7529.4240.29860.03440.1410
MotionLCM (2-step, λ = 0)0.3150.7709.4270.28400.03280.1365
MotionLCM (4-step, λ = 0)0.3280.7459.4410.29730.03390.1398
MotionLCM (1-step, λ = 0.1)0.3440.7539.3860.27110.02750.1310
MotionLCM (2-step, λ = 0.1)0.3240.7599.4280.26310.02560.1268
MotionLCM (4-step, λ = 0.1)0.3570.7439.4150.27130.02680.1309
MotionLCM (1-step, λ = 1.0)0.4190.7569.3900.19880.01470.1127
MotionLCM (2-step, λ = 1.0)0.3970.7599.4690.19600.01430.1092
MotionLCM (4-step, λ = 1.0)0.4440.7539.3550.20890.01720.1140
MotionLCM (1-step, λ = 10.0)0.6360.7449.4790.14650.00970.0967
MotionLCM (2-step, λ = 10.0)0.5510.7579.5690.15900.01070.0987
MotionLCM (4-step, λ = 10.0)0.5680.7429.4860.17230.01320.1045
" + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.396, + 0.785, + 0.425 + ], + "angle": 0, + "content": "Table 5: Ablation study on different control ratios \\( \\tau \\) and number of control joints \\( K \\) . We report the results of (1,2,4)-step inference. \" \\( * \\) \" denotes the default training setting." + }, + { + "type": "table", + "bbox": [ + 0.246, + 0.437, + 0.763, + 0.612 + ], + "angle": 0, + "content": "
MethodsFID ↓R-Precision ↑ Top 3Diversity →Traj. err. ↓ (50cm)Loc. err. ↓ (50cm)Avg. err. ↓
Real0.0020.7979.5030.00000.00000.0000
MotionLCM* (1-step, τ = 0.25, K = 6)0.4190.7569.3900.19880.01470.1127
MotionLCM* (2-step, τ = 0.25, K = 6)0.3970.7599.4690.19600.01430.1092
MotionLCM* (4-step, τ = 0.25, K = 6)0.4440.7539.3550.20890.01720.1140
MotionLCM (1-step, τ ∈ [0.1, 0.25])0.4560.7579.4770.28210.02340.1214
MotionLCM (2-step, τ ∈ [0.1, 0.25])0.4090.7699.5920.27070.02300.1179
MotionLCM (4-step, τ ∈ [0.1, 0.25])0.4570.7579.5400.29280.02560.1228
MotionLCM (1-step, τ ∈ [0.1, 0.5])0.4480.7639.5380.23900.01820.1182
MotionLCM (2-step, τ ∈ [0.1, 0.5])0.4130.7689.5170.23490.01800.1153
MotionLCM (4-step, τ ∈ [0.1, 0.5])0.4460.7539.4980.25170.01990.1196
MotionLCM (1-step, K = 12)0.4120.7539.4120.20720.01100.1029
MotionLCM (2-step, K = 12)0.4100.7589.5090.19790.01080.1000
MotionLCM (4-step, K = 12)0.4420.7559.3800.21690.01320.1048
MotionLCM (1-step, K = 22(whole-body))0.4360.7489.3790.21430.00830.0914
MotionLCM (2-step, K = 22(whole-body))0.4130.7589.4920.20610.00820.0881
MotionLCM (4-step, K = 22(whole-body))0.4610.7459.4590.21730.00970.0918
" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.648, + 0.36, + 0.664 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.704, + 0.787, + 0.84 + ], + "angle": 0, + "content": "This work proposes an efficient controllable motion generation framework, MotionLCM. By introducing latent consistency distillation, MotionLCM enjoys the trade-off between runtime efficiency and generation quality. Moreover, thanks to the motion ControlNet manipulation in the latent space, our method obtains excellent controlling ability with given conditions. Extensive experimental results show the effectiveness of the proposed method. As the VAE of MLD lacks explicit temporal modeling, the MotionLCM cannot achieve a good temporal explanation. Therefore, our future work will lie in developing a more explainable compression architecture for efficient motion control." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.647, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "MotionLCM" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.177, + 0.788, + 0.252 + ], + "angle": 0, + "content": "The research is supported by Shenzhen Ubiquitous Data Enabling Key Lab under grant ZDSYS20220527171406015 and CCF-Tencent Rhino-Bird Open Research Fund. This project is also supported by Shanghai Artificial Intelligence Laboratory. The author team would like to acknowledge Yiming Xie, Zhiyang Dou, and Shunlin Lu for their helpful technical discussions and suggestions." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.274, + 0.323, + 0.29 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.304, + 0.785, + 0.331 + ], + "angle": 0, + "content": "1. Ahn, H., Ha, T., Choi, Y., Yoo, H., Oh, S.: Text2action: Generative adversarial synthesis from language to action. In: ICRA. pp. 5915-5920 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.333, + 0.785, + 0.359 + ], + "angle": 0, + "content": "2. Ahuja, C., Morency, L.P.: Language2pose: Natural language grounded pose forecasting. In: 3DV. pp. 719-728 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.361, + 0.785, + 0.387 + ], + "angle": 0, + "content": "3. Athanasiou, N., Petrovich, M., Black, M.J., Varol, G.: Teach: Temporal action composition for 3d humans. In: 3DV. pp. 414-423 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.389, + 0.785, + 0.414 + ], + "angle": 0, + "content": "4. Barquero, G., Escalera, S., Palmero, C.: Seamless human motion composition with blended positional encodings. In: CVPR. pp. 457-469 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.416, + 0.785, + 0.455 + ], + "angle": 0, + "content": "5. Bhattacharya, U., Rewkowski, N., Banerjee, A., Guhan, P., Bera, A., Manocha, D.: Text2gestures: A transformer-based network for generating emotive body gestures for virtual agents. In: VR. pp. 1-10 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.457, + 0.785, + 0.482 + ], + "angle": 0, + "content": "6. Cervantes, P., Sekikawa, Y., Sato, I., Shinoda, K.: Implicit neural representations for variable length human motion generation. In: ECCV. pp. 356-372 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.484, + 0.785, + 0.524 + ], + "angle": 0, + "content": "7. Chen, L.H., Lu, S., Zeng, A., Zhang, H., Wang, B., Zhang, R., Zhang, L.: Motionllm: Understanding human behaviors from human motions and videos. arXiv preprint arXiv:2405.20340 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.526, + 0.785, + 0.551 + ], + "angle": 0, + "content": "8. Chen, L.H., Zhang, J., Li, Y., Pang, Y., Xia, X., Liu, T.: Humanmac: Masked motion completion for human motion prediction. In: ICCV. pp. 9544-9555 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.553, + 0.785, + 0.579 + ], + "angle": 0, + "content": "9. Chen, X., Jiang, B., Liu, W., Huang, Z., Fu, B., Chen, T., Yu, G.: Executing your commands via motion diffusion in latent space. In: CVPR. pp. 18000-18010 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.581, + 0.785, + 0.62 + ], + "angle": 0, + "content": "0. Cong, P., Dou, Z.W., Ren, Y., Yin, W., Cheng, K., Sun, Y., Long, X., Zhu, X., Ma, Y.: Laserhuman: Language-guided scene-aware human motion generation in free environment. arXiv preprint arXiv:2403.13307 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.622, + 0.785, + 0.647 + ], + "angle": 0, + "content": "1. Dabral, R., Mughal, M.H., Golyanik, V., Theobalt, C.: Mofusion: A framework for denoising-diffusion-based motion synthesis. In: CVPR. pp. 9760-9770 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.649, + 0.785, + 0.688 + ], + "angle": 0, + "content": "2. Dou, Z., Chen, X., Fan, Q., Komura, T., Wang, W.: C-ase: Learning conditional adversarial skill embeddings for physics-based characters. In: SIGGRAPH Asia. pp. 1-11 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.69, + 0.785, + 0.731 + ], + "angle": 0, + "content": "3. Fan, K., Tang, J., Cao, W., Yi, R., Li, M., Gong, J., Zhang, J., Wang, Y., Wang, C., Ma, L.: Freemotion: A unified framework for number-free text-to-motion synthesis. arXiv preprint arXiv:2405.15763 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.732, + 0.785, + 0.757 + ], + "angle": 0, + "content": "4. Ghosh, A., Cheema, N., Oguz, C., Theobalt, C., Slusallek, P.: Synthesis of compositional animations from textual descriptions. In: ICCV. pp. 1396-1406 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.759, + 0.785, + 0.784 + ], + "angle": 0, + "content": "5. Guo, C., Mu, Y., Javed, M.G., Wang, S., Cheng, L.: Momask: Generative masked modeling of 3d human motions. In: CVPR. pp. 1900-1910 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.787, + 0.785, + 0.812 + ], + "angle": 0, + "content": "6. Guo, C., Mu, Y., Zuo, X., Dai, P., Yan, Y., Lu, J., Cheng, L.: Generative human motion stylization in latent space. arXiv preprint arXiv:2401.13505 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.814, + 0.785, + 0.84 + ], + "angle": 0, + "content": "7. Guo, C., Zou, S., Zuo, X., Wang, S., Ji, W., Li, X., Cheng, L.: Generating diverse and natural 3d human motions from text. In: CVPR. pp. 5152-5161 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.304, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "W. Dai et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "18. Guo, C., Zuo, X., Wang, S., Cheng, L.: Tm2t: Stochastic and tokenized modeling for the reciprocal generation of 3d human motions and texts. In: ECCV. pp. 580-597 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.786, + 0.232 + ], + "angle": 0, + "content": "19. Guo, C., Zuo, X., Wang, S., Zou, S., Sun, Q., Deng, A., Gong, M., Cheng, L.: Action2motion: Conditioned generation of 3d human motions. In: ACMMM. pp. 2021-2029 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.233, + 0.785, + 0.26 + ], + "angle": 0, + "content": "20. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. NeurIPS pp. 6840-6851 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.261, + 0.785, + 0.288 + ], + "angle": 0, + "content": "21. Ho, J., Salimans, T.: Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.289, + 0.785, + 0.317 + ], + "angle": 0, + "content": "22. Holden, D., Komura, T., Saito, J.: Phase-functional neural networks for character control. TOG 36(4), 1-13 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.318, + 0.785, + 0.346 + ], + "angle": 0, + "content": "23. Holden, D., Saito, J., Komura, T.: A deep learning framework for character motion synthesis and editing. TOG 35(4), 1-11 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.346, + 0.785, + 0.387 + ], + "angle": 0, + "content": "24. Huang, Y., Yang, H., Luo, C., Wang, Y., Xu, S., Zhang, Z., Zhang, M., Peng, J.: Stablemofusion: Towards robust and efficient diffusion-based motion generation framework. arXiv preprint arXiv:2405.05691 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.388, + 0.785, + 0.416 + ], + "angle": 0, + "content": "25. Huber, P.J.: Robust estimation of a location parameter. The Annals of Mathematical Statistics 35(1), 73-101 (1964)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.417, + 0.785, + 0.457 + ], + "angle": 0, + "content": "26. Ji, Y., Xu, F., Yang, Y., Shen, F., Shen, H.T., Zheng, W.S.: A large-scale rgb-d database for arbitrary-view human action recognition. In: ACMM. pp. 1510-1518 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.459, + 0.785, + 0.486 + ], + "angle": 0, + "content": "27. Jiang, B., Chen, X., Liu, W., Yu, J., Yu, G., Chen, T.: Motiongpt: Human motion as a foreign language. NeurIPS (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.487, + 0.785, + 0.515 + ], + "angle": 0, + "content": "28. Karras, T., Aittala, M., Aila, T., Laine, S.: Elucidating the design space of diffusion-based generative models. NeurIPS pp. 26565-26577 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.516, + 0.785, + 0.543 + ], + "angle": 0, + "content": "29. Karunratanakul, K., Preechakul, K., Suwajanakorn, S., Tang, S.: Guided motion diffusion for controllable human motion synthesis. In: CVPR. pp. 2151-2162 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.544, + 0.785, + 0.572 + ], + "angle": 0, + "content": "30. Kingma, D.P., Welling, M.: Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114 (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.573, + 0.785, + 0.6 + ], + "angle": 0, + "content": "31. Lee, T., Moon, G., Lee, K.M.: Multiact: Long-term 3d human motion generation from multiple action labels. In: AAAI. pp. 1231-1239 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.601, + 0.785, + 0.629 + ], + "angle": 0, + "content": "32. Li, B., Zhao, Y., Zhelun, S., Sheng, L.: Danceformer: Music conditioned 3d dance generation with parametric motion transformer. In: AAAI. pp. 1272-1279 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.63, + 0.785, + 0.67 + ], + "angle": 0, + "content": "33. Li, R., Zhang, Y., Zhang, Y., Zhang, H., Guo, J., Zhang, Y., Liu, Y., Li, X.: Lodge: A coarse to fine diffusion network for long dance generation guided by the characteristic dance primitives. In: CVPR. pp. 1524-1534 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.671, + 0.785, + 0.712 + ], + "angle": 0, + "content": "34. Li, R., Zhao, J., Zhang, Y., Su, M., Ren, Z., Zhang, H., Tang, Y., Li, X.: Finedance: A fine-grained choreography dataset for 3d full body dance generation. In: ICCV. pp. 10234-10243 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.713, + 0.785, + 0.742 + ], + "angle": 0, + "content": "35. Li, R., Yang, S., Ross, D.A., Kanazawa, A.: Ai choreographer: Music conditioned 3d dance generation with aist++. In: ICCV. pp. 13401-13412 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.743, + 0.785, + 0.77 + ], + "angle": 0, + "content": "36. Li, T., Qiao, C., Ren, G., Yin, K., Ha, S.: Aamdm: Accelerated auto-regressive motion diffusion model. In: CVPR. pp. 1813-1823 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.771, + 0.785, + 0.811 + ], + "angle": 0, + "content": "37. Lin, A.S., Wu, L., Corona, R., Tai, K., Huang, Q., Mooney, R.J.: Generating animated videos of human activities from natural language descriptions. Learning 1(2018), 1 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.812, + 0.785, + 0.84 + ], + "angle": 0, + "content": "38. Lin, X., Amer, M.R.: Human motion modeling using dvgens. arXiv preprint arXiv:1804.10652 (2018)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.647, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "MotionLCM" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "39. Liu, J., Dai, W., Wang, C., Cheng, Y., Tang, Y., Tong, X.: Plan, posture and go: Towards open-world text-to-motion generation. arXiv preprint arXiv:2312.14828 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.787, + 0.218 + ], + "angle": 0, + "content": "40. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.218, + 0.787, + 0.261 + ], + "angle": 0, + "content": "41. Lu, C., Zhou, Y., Bao, F., Chen, J., Li, C., Zhu, J.: Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. NeurIPS pp. 5775-5787 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.261, + 0.787, + 0.302 + ], + "angle": 0, + "content": "42. Lu, C., Zhou, Y., Bao, F., Chen, J., Li, C., Zhu, J.: Dpm-solver++: Fast solver for guided sampling of diffusion probabilistic models. arXiv preprint arXiv:2211.01095 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.302, + 0.787, + 0.346 + ], + "angle": 0, + "content": "43. Lu, S., Chen, L.H., Zeng, A., Lin, J., Zhang, R., Zhang, L., Shum, H.Y.: Humantomato: Text-aligned whole-body motion generation. arXiv preprint arXiv:2310.12978 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.346, + 0.787, + 0.387 + ], + "angle": 0, + "content": "44. Luo, S., Tan, Y., Huang, L., Li, J., Zhao, H.: Latent consistency models: Synthesizing high-resolution images with few-step inference. arXiv preprint arXiv:2310.04378 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.387, + 0.787, + 0.416 + ], + "angle": 0, + "content": "45. Mahmood, N., Ghorbani, N., Troje, N.F., Pons-Moll, G., Black, M.J.: Amass: Archive of motion capture as surface shapes. In: ICCV. pp. 5442-5451 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.416, + 0.787, + 0.444 + ], + "angle": 0, + "content": "46. Nichol, A.Q., Dhariwal, P.: Improved denoising diffusion probabilistic models. In: ICML. pp. 8162-8171 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.444, + 0.787, + 0.487 + ], + "angle": 0, + "content": "47. Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., et al.: Pytorch: An imperative style, high-performance deep learning library. NeurIPS (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.487, + 0.787, + 0.515 + ], + "angle": 0, + "content": "48. Petrovich, M., Black, M.J., Varol, G.: Action-conditioned 3d human motion synthesis with transformer vae. In: ICCV. pp. 10985-10995 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.515, + 0.787, + 0.543 + ], + "angle": 0, + "content": "49. Petrovich, M., Black, M.J., Varol, G.: Temos: Generating diverse human motions from textual descriptions. In: ECCV. pp. 480-497 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.543, + 0.787, + 0.572 + ], + "angle": 0, + "content": "50. Petrovich, M., Black, M.J., Varol, G.: Tmr: Text-to-motion retrieval using contrastive 3d human motion synthesis. In: ICCV. pp. 9488-9497 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.572, + 0.787, + 0.614 + ], + "angle": 0, + "content": "51. Petrovich, M., Litany, O., Iqbal, U., Black, M.J., Varol, G., Bin Peng, X., Rempe, D.: Multi-track timeline control for text-driven 3d human motion generation. In: CVPRW. pp. 1911-1921 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.614, + 0.787, + 0.643 + ], + "angle": 0, + "content": "52. Plappert, M., Mandery, C., Asfour, T.: The kit motion-language dataset. Big data 4(4), 236-252 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.643, + 0.787, + 0.685 + ], + "angle": 0, + "content": "53. Punnakkal, A.R., Chandrasekaran, A., Athanasiou, N., Quiros-Ramirez, A., Black, M.J.: Babel: Bodies, action and behavior with english labels. In: CVPR. pp. 722-731 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.685, + 0.787, + 0.727 + ], + "angle": 0, + "content": "54. Raab, S., Leibovitch, I., Li, P., Aberman, K., Sorkine-Hornung, O., Cohen-Or, D.: Modi: Unconditional motion synthesis from diverse data. In: CVPR. pp. 13873-13883 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.727, + 0.787, + 0.755 + ], + "angle": 0, + "content": "55. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR. pp. 10684-10695 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.755, + 0.787, + 0.784 + ], + "angle": 0, + "content": "56. Shafir, Y., Tevet, G., Kapon, R., Bermano, A.H.: Human motion diffusion as a generative prior. In: ICLR (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.784, + 0.787, + 0.812 + ], + "angle": 0, + "content": "57. Shahroudy, A., Liu, J., Ng, T.T., Wang, G.: Ntu \\(\\mathrm{rgb + }\\) d: A large scale dataset for 3d human activity analysis. In: CVPR. pp. 1010-1019 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.812, + 0.787, + 0.841 + ], + "angle": 0, + "content": "58. Shi, Y., Wang, J., Jiang, X., Dai, B.: Controllable motion diffusion model. arXiv preprint arXiv:2306.00416 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "W. Dai et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "59. Siyao, L., Yu, W., Gu, T., Lin, C., Wang, Q., Qian, C., Loy, C.C., Liu, Z.: Bailando: 3d dance generation by actor-critic gpt with choreographic memory. In: CVPR. pp. 11050-11059 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.785, + 0.232 + ], + "angle": 0, + "content": "60. Sohl-Dickstein, J., Weiss, E., Maheswaranathan, N., Ganguli, S.: Deep unsupervised learning using nonequilibrium thermodynamics. In: ICML. pp. 2256-2265 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.785, + 0.245 + ], + "angle": 0, + "content": "61. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. In: ICLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.245, + 0.785, + 0.272 + ], + "angle": 0, + "content": "62. Song, Y., Dhariwal, P., Chen, M., Sutskever, I.: Consistency models. In: ICML (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.785, + 0.313 + ], + "angle": 0, + "content": "63. Tang, Y., Liu, J., Liu, A., Yang, B., Dai, W., Rao, Y., Lu, J., Zhou, J., Li, X.: Flag3d: A 3d fitness activity dataset with language instruction. In: CVPR. pp. 22106-22117 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.314, + 0.785, + 0.342 + ], + "angle": 0, + "content": "64. Tevet, G., Gordon, B., Hertz, A., Bermano, A.H., Cohen-Or, D.: Motionclip: Exposing human motion generation to clip space. In: ECCV. pp. 358-374 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.343, + 0.785, + 0.37 + ], + "angle": 0, + "content": "65. Tevet, G., Raab, S., Gordon, B., Shafir, Y., Cohen-Or, D., Bermano, A.H.: Human motion diffusion model. In: ICLR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.371, + 0.785, + 0.397 + ], + "angle": 0, + "content": "66. Tseng, J., Castellon, R., Liu, K.: Edge: Editable dance generation from music. In: CVPR. pp. 448-458 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.785, + 0.424 + ], + "angle": 0, + "content": "67. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. NeurIPS (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.425, + 0.785, + 0.466 + ], + "angle": 0, + "content": "68. Wan, W., Dou, Z., Komura, T., Wang, W., Jayaraman, D., Liu, L.: Tlcontrol: Trajectory and language control for human motion synthesis. arXiv preprint arXiv:2311.17135 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.467, + 0.785, + 0.508 + ], + "angle": 0, + "content": "69. Wang, Z., Chen, Y., Jia, B., Li, P., Zhang, J., Zhang, J., Liu, T., Zhu, Y., Liang, W., Huang, S.: Move as you say interact as you can: Language-guided human motion generation with scene affordance. In: CVPR. pp. 433-444 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.509, + 0.785, + 0.55 + ], + "angle": 0, + "content": "70. Wang, Z., Chen, Y., Liu, T., Zhu, Y., Liang, W., Huang, S.: Humanise: Language-conditioned human motion generation in 3d scenes. NeurIPS pp. 14959-14971 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.55, + 0.785, + 0.577 + ], + "angle": 0, + "content": "71. Wang, Z., Wang, J., Lin, D., Dai, B.: Intercontrol: Generate human motion interactions by controlling every joint. arXiv preprint arXiv:2311.15864 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.578, + 0.785, + 0.605 + ], + "angle": 0, + "content": "72. Xiao, Z., Wang, T., Wang, J., Cao, J., Zhang, W., Dai, B., Lin, D., Pang, J.: Unified human-scene interaction via prompted chain-of-contacts. In: ICLR (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.606, + 0.785, + 0.633 + ], + "angle": 0, + "content": "73. Xie, Y., Jampani, V., Zhong, L., Sun, D., Jiang, H.: Omnicontrol: Control any joint at any time for human motion generation. In: ICLR (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.633, + 0.785, + 0.674 + ], + "angle": 0, + "content": "74. Xu, L., Lv, X., Yan, Y., Jin, X., Wu, S., Xu, C., Liu, Y., Zhou, Y., Rao, F., Sheng, X., et al.: Inter-x: Towards versatile human-human interaction analysis. In: CVPR. pp. 22260-22271 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.675, + 0.785, + 0.715 + ], + "angle": 0, + "content": "75. Xu, L., Song, Z., Wang, D., Su, J., Fang, Z., Ding, C., Gan, W., Yan, Y., Jin, X., Yang, X., et al.: Actformer: A gan-based transformer towards general action-conditioned 3d human motion generation. In: ICCV. pp. 2228-2238 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.716, + 0.785, + 0.743 + ], + "angle": 0, + "content": "76. Xu, L., Zhou, Y., Yan, Y., Jin, X., Zhu, W., Rao, F., Yang, X., Zeng, W.: Regennet: Towards human action-reaction synthesis. In: CVPR. pp. 1759-1769 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.744, + 0.785, + 0.771 + ], + "angle": 0, + "content": "77. Yan, S., Li, Z., Xiong, Y., Yan, H., Lin, D.: Convolutional sequence generation for skeleton-based action synthesis. In: ICCV. pp. 4394-4402 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.772, + 0.785, + 0.799 + ], + "angle": 0, + "content": "78. Yuan, Y., Song, J., Iqbal, U., Vahdat, A., Kautz, J.: Physdiff: Physics-guided human motion diffusion model. In: ICCV. pp. 16010-16021 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.841 + ], + "angle": 0, + "content": "79. Zhang, B., Cheng, Y., Wang, C., Zhang, T., Yang, J., Tang, Y., Zhao, F., Chen, D., Guo, B.: Rodinhd: High-fidelity 3d avatar generation with diffusion models. arXiv preprint arXiv:2407.06938 (2024)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.647, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "MotionLCM" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "80. Zhang, B., Cheng, Y., Yang, J., Wang, C., Zhao, F., Tang, Y., Chen, D., Guo, B.: Gaussiancube: Structuring gaussian splatting using optimal transport for 3d generative modeling. arXiv preprint arXiv:2403.19655 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.788, + 0.232 + ], + "angle": 0, + "content": "81. Zhang, J., Zhang, Y., Cun, X., Zhang, Y., Zhao, H., Lu, H., Shen, X., Shan, Y.: Generating human motion from textual descriptions with discrete representations. In: CVPR. pp. 14730-14740 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.232, + 0.788, + 0.259 + ], + "angle": 0, + "content": "82. Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models. In: ICCV. pp. 3836-3847 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.259, + 0.788, + 0.3 + ], + "angle": 0, + "content": "83. Zhang, M., Cai, Z., Pan, L., Hong, F., Guo, X., Yang, L., Liu, Z.: Motiondiffuse: Text-driven human motion generation with diffusion model. arXiv preprint arXiv:2208.15001 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.3, + 0.788, + 0.328 + ], + "angle": 0, + "content": "84. Zhao, R., Su, H., Ji, Q.: Bayesian adversarial human motion synthesis. In: CVPR. pp. 6225-6234 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.328, + 0.788, + 0.356 + ], + "angle": 0, + "content": "85. Zhong, L., Xie, Y., Jampani, V., Sun, D., Jiang, H.: Smoodi: Stylized motion diffusion model. arXiv preprint arXiv:2407.12783 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.356, + 0.788, + 0.398 + ], + "angle": 0, + "content": "86. Zhou, W., Dou, Z., Cao, Z., Liao, Z., Wang, J., Wang, W., Liu, Y., Komura, T., Wang, W., Liu, L.: Emdm: Efficient motion diffusion model for fast, high-quality motion generation. arXiv preprint arXiv:2312.02256 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.398 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/MotionLCM_ Real-time Controllable Motion Generation via Latent Consistency Model/69320bc9-c9fe-4fc1-8e78-3d61fd4c6087_origin.pdf b/2024/MotionLCM_ Real-time Controllable Motion Generation via Latent Consistency Model/69320bc9-c9fe-4fc1-8e78-3d61fd4c6087_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..c312e1736a614b6cb841e7edf25112955a24eba9 --- /dev/null +++ b/2024/MotionLCM_ Real-time Controllable Motion Generation via Latent Consistency Model/69320bc9-c9fe-4fc1-8e78-3d61fd4c6087_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:215277d0ec0f4da0b88cba68c3e8569ac8b135ae274440605d96935a5a57ed1f +size 1729489 diff --git a/2024/MotionLCM_ Real-time Controllable Motion Generation via Latent Consistency Model/full.md b/2024/MotionLCM_ Real-time Controllable Motion Generation via Latent Consistency Model/full.md new file mode 100644 index 0000000000000000000000000000000000000000..db0c4352df2bb3d07ed9c3c30b9c6151f4567af7 --- /dev/null +++ b/2024/MotionLCM_ Real-time Controllable Motion Generation via Latent Consistency Model/full.md @@ -0,0 +1,303 @@ +# MotionLCM: Real-time Controllable Motion Generation via Latent Consistency Model + +Wenxun Dai $^{1,2}$ , Ling-Hao Chen $^{2\star}$ , Jingbo Wang $^{3\boxtimes}$ , Jinpeng Liu $^{1,2}$ , Bo Dai $^{3\boxtimes}$ , Yansong Tang $^{1,2}$ + +1Shenzhen Key Laboratory of Ubiquitous Data Enabling, Tsinghua Shenzhen International Graduate School 2Tsinghua University 3Shanghai AI Laboratory {wx dai 2001, thu. lhchen, wangjingbo1219, liu. jinpeng.55}@gmail.com {doubledaibo, tangyansong15}@gmail.com Project page: https://dai-wenxun.github.io/MotionLCM-page + +![](images/47c7ee0ac1d6278ba282f678ec7f939d12be6b7c6e0c3474f4d210e856f07a82.jpg) +Fig. 1: We propose MotionLCM, a real-time controllable motion latent consistency model. Our model uses the last few frames of the previous motion as temporal control signals to autoregressively generate the next motion in real-time under different text prompts. Green blocks denote the junctions. The numbers in red are the inference time. + +Abstract. This work introduces MotionLCM, extending controllable motion generation to a real-time level. Existing methods for spatial-temporal control in text-conditioned motion generation suffer from significant runtime inefficiency. To address this issue, we first propose the motion latent consistency model (MotionLCM) for motion generation, + +building upon the latent diffusion model [9]. By adopting one-step (or few-step) inference, we further improve the runtime efficiency of the motion latent diffusion model for motion generation. To ensure effective controllability, we incorporate a motion ControlNet within the latent space of MotionLCM and enable explicit control signals (e.g., initial poses) in the vanilla motion space to control the generation process directly, similar to controlling other latent-free diffusion models [29, 73] for motion generation. By employing these techniques, our approach can generate human motions with text and control signals in real-time. Experimental results demonstrate the remarkable generation and controlling capabilities of MotionLCM while maintaining real-time runtime efficiency. + +Keywords: Text-to-Motion $\cdot$ Real-time Control $\cdot$ Consistency Model + +# 1 Introduction + +Text-to-motion generation (T2M) has attracted increasing attention [1, 15, 43, 49, 65] due to its important roles in many applications [70, 72]. Previous attempts mainly focus on GANs [1, 38], VAEs [3, 19, 48, 49] and diffusion models [9, 11, 39, 56, 65, 78, 83] via pairwise text-motion data [17, 26, 45, 52, 53, 57, 63, 74] and achieve impressive generation results. Existing approaches [9, 65, 83] mainly take diffusion models [20, 46, 55, 60] as a base generative model, owing to their powerful ability to model motion distribution. However, these diffusion fashions inevitably require considerable sampling steps for motion synthesis during inference, even with some sampling acceleration methods [61]. Specifically, MDM [65] and MLD [9] require $\sim$ 12s and $\sim$ 0.2s to generate + +![](images/a4e5b6df9d8d0ad07f9ba9a2d1f4e85d197fee9e6e870e803aa016fd407bf094.jpg) +Fig.2: Comparison of the inference time costs on HumanML3D [17]. We compare the AITS and FID metrics with five SOTA methods. The closer the model is to the origin the better. Diffusion-based models are indicated by the blue dashed box. Our MotionLCM achieves real-time inference speed while ensuring high-quality motion generation. + +a high-quality motion sequence. Such low efficiency blocks the applications of generating high-quality motions in various real-time scenarios. + +In addition to the language description itself serving as a coarse control signal, another line of research focuses on controlling the motion generation with spatial-temporal constraints [29, 56, 73]. Although these attempts enjoy impressive controlling ability in the T2M task, there still exists a significant gap towards real-time applications. For example, OmniControl [73] exhibits a relatively long inference time, $\sim 81$ s per sequence. Therefore, trading-off between generation + +quality and efficiency is a challenging problem. As a result, in this paper, we target the real-time controllable motion generation research problem. + +Recently, the concept of consistency models [44, 62] has been introduced in image generation, resulting in significant progress by enabling efficient and high-fidelity image synthesis with a minimal number of sampling steps (e.g., 4 steps vs. 50 steps). These properties perfectly align with our goal of accelerating motion generation without compromising generation quality. Therefore, we propose MotionLCM (Motion Latent Consistency Model) distilled from the motion latent diffusion model, MLD [9], to tackle the low-efficiency problem in diffusion sampling. To the best of our knowledge, we introduce consistency distillation into the motion generation area for the first time and accelerate motion generation to a real-time level via latent consistency distillation [44]. + +Here, in MotionLCM, we are facing another challenge on how to control motions with spatial-temporal signals (e.g., initial poses) in the latent space. Previous methods [56, 73] model human motions in the vanilla motion space and can manipulate the motion directly in the denoising process. However, for our latent-diffusion-based MotionLCM, it is non-trivial to feed the control signals into the latent space. This is because the latent has no explicit motion semantics, which cannot be manipulated directly by the control signals. Inspired by the notable success of [82] in controllable image generation [55], we introduce a motion ControlNet to control motion generation in the latent space. However, the naïve motion ControlNet is not sufficient to provide supervision for the control signals. The main reason is the lack of explicit supervision in the motion space. Therefore, during the training phase, we decode the predicted latent through the frozen VAE [30] decoder into the vanilla motion space to provide explicit control supervision on the generated motion. Thanks to the powerful one-step inference capability of MotionLCM, the latent generated by MotionLCM can significantly facilitate control supervision both in the latent space and motion space for training the motion ControlNet compared to MLD [9]. + +With our key designs, our proposed MotionLCM successfully enjoys the balance between the generation quality and efficiency in controllable motion generation. Before delivering into detail, we sum up our core contributions as follows. + +- We propose the Motion Latent Consistency Model (MotionLCM) via consistency distillation on the motion latent diffusion model extending controllable motion generation to a real-time level. +- Building upon our achievement of real-time motion generation, we introduce a motion ControlNet, enabling high-quality controllable motion generation. +- Extensive experimental results show that MotionLCM enjoys a good balance of generation quality, controlling capability, and real-time efficiency. + +# 2 Related Work + +# 2.1 Human Motion Generation + +Generating human motions can be divided into three main fashions according to inputs: motion synthesis (1) without any condition [54,65,77,84], (2) with some + +given multi-modal conditions, such as action labels [6, 12, 19, 31, 48, 75], textual description [1-5, 7, 7, 9-11, 13-15, 17, 18, 24, 27, 29, 37-39, 43, 49-51, 56, 64, 65, 68-70, 72, 78, 81, 83, 85, 86], audio or music [32-35, 59, 66], (3) with user-defined trajectories [22, 23, 29, 36, 56, 58, 68, 71, 73]. To generate diverse, natural, and high-quality human motions, many generative models have been explored by [2, 38, 49, 79, 80]. Recently, diffusion-based models significantly improved the motion generation performance and diversity [8, 9, 11, 16, 65, 76, 83] with stable training. Specifically, MotionDiffuse [83] represents the first text-based motion diffusion model that provides fine-grained instructions on body parts and achieves arbitrary-length motion synthesis with time-varied text prompts. MDM [65] introduces a motion diffusion model that operates on raw motion data, enabling both high-quality generation and generic conditioning that together comprise a good baseline for new motion generation tasks. Based on MDM [65], OmniControl [73] integrates flexible spatial-temporal control signals across different joints by combining analytic spatial guidance and realism guidance into the diffusion model, ensuring that the generated motion closely conforms to the input control signals. The work most relevant to ours is MLD [9], which introduces a motion latent-based diffusion model to enhance generation quality and reduce computational resource requirements. The key idea is training a VAE [30] for motion embedding, followed by implementing latent diffusion [55] within the learned latent space. However, these diffusion fashions inevitably require considerable sampling steps for motion synthesis during inference, even with some sampling acceleration methods [61]. Thus, we propose MotionLCM, which not only guarantees high-quality controllable motion generation but also achieves real-time runtime efficiency. + +# 3 Method + +In this section, we first briefly introduce preliminaries about latent consistency models in Sec. 3.1. Then, we describe how to conduct latent consistency distillation for motion generation in Sec. 3.2, followed by our implementation of motion control in latent space in Sec. 3.3. The overall pipeline is illustrated in Fig. 4. + +# 3.1 Preliminaries + +The Consistency Model (CM) [62] introduces a kind of efficient generative model designed for efficient one-step or few-step generation. Given a Probability Flow ODE (a.k.a. PF-ODE) that smoothly converts data to noise, the CM is to learn the function $\pmb{f}(\cdot, \cdot)$ that maps any points on the ODE trajectory to its origin distribution (i.e., the solution of the PF-ODE). The consistency function is formally defined as $\pmb{f}:(\mathbf{x}_t,t)\longmapsto \mathbf{x}_{\epsilon}$ , where $t\in [\epsilon ,T]$ , $T > 0$ is a fixed constant and $\epsilon$ is a small positive number to avoid numerical instability. According to [62], the consistency function should satisfy the self-consistency property: + +$$ +\boldsymbol {f} \left(\mathbf {x} _ {t}, t\right) = \boldsymbol {f} \left(\mathbf {x} _ {t ^ {\prime}}, t ^ {\prime}\right), \forall t, t ^ {\prime} \in [ \epsilon , T ]. \tag {1} +$$ + +![](images/5e7d75e984cbe2d5e45177e83441b38e9f8635bbd8f86ef4e4da593e1f9802ab.jpg) +Distillation target of $f_{\Theta}$ : Project any $\mathbf{z}_t$ to $\mathbf{z}_0$ consistently. $t \in [0, T]$ + +![](images/b475d73358e88bd5d634c16733218e43460cfb9ce878977325639f72136dd0db.jpg) +Fig. 3: The training objective of consistency distillation is to learn a consistency function $f_{\Theta}$ , initialized with the parameters of a pre-trained diffusion model (e.g., MLD [9]). This function $f_{\Theta}$ should projects any points (i.e., $\mathbf{z}_t$ ) on the ODE trajectory to its solution (i.e., $\mathbf{z}_0$ ). Once the pre-trained model [9] is distilled, unlike the traditional denoising model [65, 83] that requires considerable sampling steps, our MotionLCM can generate high-quality motion sequences with one-step sampling and further improve the generation quality through multi-step inference. + +![](images/846916882c9ecd29231b43b048d6d08e6ffc143c2beee9eb0b33183dd4d1a56a.jpg) + +As shown in Eq. (1), the self-consistency property indicates that for arbitrary pairs of $(\mathbf{x}_t,t)$ on the same PF-ODE trajectory, the outputs of the model should be consistent. The goal of a parameterized consistency model $f_{\Theta}$ is to learn a consistency function from data by enforcing the self-consistency property in Eq. (1). To ensure that $f_{\Theta}(\mathbf{x},\epsilon) = \mathbf{x}$ , the consistency model $f_{\Theta}$ is parameterized as, + +$$ +\boldsymbol {f} _ {\Theta} (\mathbf {x}, t) = c _ {\text {s k i p}} (t) \mathbf {x} + c _ {\text {o u t}} (t) \boldsymbol {F} _ {\Theta} (\mathbf {x}, t), \tag {2} +$$ + +where $c_{\mathrm{skip}}(t)$ and $c_{\mathrm{out}}(t)$ are differentiable functions with $c_{\mathrm{skip}}(\epsilon) = 1$ and $c_{\mathrm{out}}(\epsilon) = 0$ , and $\pmb{F}_{\Theta}(\cdot, \cdot)$ is a deep neural network to learn the self-consistency. The CM trained from distilling the knowledge of pre-trained diffusion models is called Consistency Distillation. The consistency loss is defined as follows, + +$$ +\mathcal {L} (\boldsymbol {\Theta}, \boldsymbol {\Theta} ^ {-}; \boldsymbol {\Phi}) = \mathbb {E} \left[ d \left(\boldsymbol {f} _ {\boldsymbol {\Theta}} \left(\mathbf {x} _ {t _ {n + 1}}, t _ {n + 1}\right), \boldsymbol {f} _ {\boldsymbol {\Theta} ^ {-}} \left(\hat {\mathbf {x}} _ {t _ {n}} ^ {\boldsymbol {\Phi}}, t _ {n}\right)\right) \right], \tag {3} +$$ + +where $d(\cdot, \cdot)$ is a chosen metric function for measuring the distance between two samples. $f_{\Theta}(\cdot, \cdot)$ and $f_{\Theta^{-}}(\cdot, \cdot)$ are referred to as "online network" and "target network" according to [62]. Besides, $\Theta^{-}$ is updated with the exponential moving average (EMA) of the parameters of $\Theta^{1}$ . In Eq. (3), $\hat{\mathbf{x}}_{t_n}^{\Phi}$ is a one-step estimation of $\mathbf{x}_{t_n}$ from $\mathbf{x}_{t_{n+1}}$ , which is formulated as, + +$$ +\hat {\mathbf {x}} _ {t _ {n}} ^ {\Phi} \leftarrow \mathbf {x} _ {t _ {n + 1}} + (t _ {n} - t _ {n + 1}) \Phi (\mathbf {x} _ {t _ {n + 1}}, t _ {n + 1}, \emptyset), \tag {4} +$$ + +where $\Phi$ is a one-step ODE solver applied to PF-ODE. + +![](images/eab86df2a1587c50a0a037091a3594a96ccb0429d2d45e4eae44dd1bdba9eaeb.jpg) +(a) Motion Latent Consistency Distillation + +![](images/7c2e2dde8b41f797545ffbbe271b2039f5bdac06f2bbd77b7c859b86f03f9a58.jpg) +(b) Motion Control in Latent Space +Fig. 4: The overview of MotionLCM. (a) Motion Latent Consistency Distillation (Sec. 3.2). Given a raw motion sequence $\mathbf{x}_0^{1:N}$ , a pre-trained VAE [30] encoder first compresses it into the latent space, then a forward diffusion operation is performed to add $n + k$ steps of noise. Then, the noisy $\mathbf{z}_{n + k}$ is fed into the online network and teacher network to predict the clean latent. The target network takes the $k$ -step estimation results of the teacher output to predict the clean latent. To learn self-consistency, a loss is applied to enforce the output of the online network and target network to be consistent. (b) Motion Control in Latent Space (Sec. 3.3). With the powerful MotionLCM trained in the first stage, we incorporate a motion ControlNet into the MotionLCM to achieve controllable motion generation. Furthermore, we leverage the decoded motion to explicitly supervise the spatial-temporal control signals (i.e., initial poses $\mathbf{g}^{1:\tau}$ ). + +The Latent Consistency Model (LCM) [44] learns the self-consistency property in the latent space $D_{\mathbf{z}} = \{(\mathbf{z},\mathbf{c})|\mathbf{z} = \mathcal{E}(\mathbf{x}),(\mathbf{x},\mathbf{c})\in D\}$ , where $D$ denotes the dataset, $\mathbf{c}$ is the given condition, and $\mathcal{E}$ is the pre-trained encoder. Compared to CMs [62] using the numerical continuous PF-ODE solver [28], LCMs [44] adopt the discrete-time schedule [41,42,61] to adapt to Stable Diffusion [55]. Instead of ensuring consistency between adjacent time steps $t_{n + 1}\rightarrow t_n$ , LCMs [44] are designed to ensure consistency between the current time step and $k$ -step away, i.e., $t_{n + k}\to t_n$ , thereby significantly reducing convergence time costs. As classifier-free guidance (CFG) [21] plays a crucial role in synthesizing high-quality text-aligned images, LCMs integrate CFG into the distillation as follows, + +$$ +\hat {\mathbf {z}} _ {t _ {n}} ^ {\boldsymbol {\Phi}, w} \gets \mathbf {z} _ {t _ {n + k}} + (1 + w) \boldsymbol {\Phi} (\mathbf {z} _ {t _ {n + k}}, t _ {n + k}, t _ {n}, \mathbf {c}) - w \boldsymbol {\Phi} (\mathbf {z} _ {t _ {n + k}}, t _ {n + k}, t _ {n}, \emptyset), \qquad (5) +$$ + +where $w$ denotes the CFG scale which is uniformly sampled from $[w_{\min}, w_{\max}]$ and $k$ is the skipping interval. To efficiently perform the above $k$ -step guided distillation, LCMs augment the consistency function to $f: (\mathbf{z}_t, t, w, \mathbf{c}) \longmapsto \mathbf{z}_0$ , which is also the form adopted by our MotionLCM. + +# 3.2 MotionLCM: Motion Latent Consistency Model + +Motion compression into the latent space. Motivated by [44, 62], we propose MotionLCM (Motion Latent Consistency Model) to tackle the low-efficiency + +problem in motion diffusion models [65, 83], unleashing the potential of LCM in the motion generation task. Similar to MLD [9], our MotionLCM adopts a consistency model in the motion latent space. We choose MLD [9] as the underlying diffusion model to distill from. We aim to achieve few-step $(2\sim 4)$ and even one-step inference without compromising motion quality. In MLD, an autoencoder $(\mathcal{E},\mathcal{D})$ is first trained to compress a high dimensional motion into a low dimensional latent vector $\mathbf{z} = \mathcal{E}(\mathbf{x})$ , which is then decoded to reconstruct the motion as $\hat{\mathbf{x}} = \mathcal{D}(\mathbf{z})$ . Training diffusion models in the motion latent space greatly reduces the computational resources compared to the vanilla diffusion models trained on raw motion sequences (i.e., motion space) and speeds up the inference process. Thus, we effectively leverage the motion latent space for consistency distillation. Motion latent consistency distillation. An overview of our motion latent consistency distillation is described in Fig. 4 (a). A raw motion sequence $\mathbf{x}_0^{1:N} = \{\mathbf{x}^i\}_{i=1}^N$ is a sequence of human poses, where $N$ is the number of frames. We follow [17] to use the redundant motion representation for our experiments, which is widely used in previous work [9, 65, 83]. As shown in Fig. 4 (a), given a raw motion sequence $\mathbf{x}_0^{1:N}$ , a pre-trained VAE [30] encoder first compresses it into the latent space, $\mathbf{z}_0 = \mathcal{E}(\mathbf{x}_0)$ . Then, a forward diffusion operation with $n + k$ steps is conducted to add noise on $\mathbf{z}_0$ , where $k$ is the skipping interval illustrated in Sec. 3.1. The noisy $\mathbf{z}_{n+k}$ is fed to the frozen teacher network and trainable online network to predict the clean $\hat{\mathbf{z}}_0^*$ , and $\hat{\mathbf{z}}_0$ . The target network uses the cleaner $\hat{\mathbf{z}}_n$ obtained by a $k$ -step ODE solver $\Phi$ , such as DDIM [61] to predict the $\hat{\mathbf{z}}_0^-$ . Since the classifier-free guidance (CFG) [21] is essential for condition alignment in diffusion models [9, 55, 65], we integrate CFG into the distillation, + +$$ +\hat {\mathbf {z}} _ {n} \leftarrow \mathbf {z} _ {n + k} + (1 + w) \boldsymbol {\Phi} \left(\mathbf {z} _ {n + k}, t _ {n + k}, t _ {n}, \mathbf {c}\right) - w \boldsymbol {\Phi} \left(\mathbf {z} _ {n + k}, t _ {n + k}, t _ {n}, \emptyset\right), \tag {6} +$$ + +where $\mathbf{c}$ is the text condition and $w$ denotes the guidance scale. To ensure the self-consistency property defined in Eq. (1), the latent consistency distillation loss $\mathcal{L}_{\mathrm{LCD}}$ is designed as follows, + +$$ +\mathcal {L} _ {\mathrm {L C D}} (\boldsymbol {\Theta}, \boldsymbol {\Theta} ^ {-}) = \mathbb {E} \left[ d \left(\boldsymbol {f} _ {\boldsymbol {\Theta}} (\mathbf {z} _ {n + k}, t _ {n + k}, w, \mathbf {c}), \boldsymbol {f} _ {\boldsymbol {\Theta} ^ {-}} (\hat {\mathbf {z}} _ {n}, t _ {n}, w, \mathbf {c})\right) \right], \tag {7} +$$ + +where $d(\cdot, \cdot)$ is a distance measuring function, such as L2 loss or Huber loss [25]. As discussed in Sec. 3.1, the target network $\Theta^{-}$ is updated with the exponential moving average (EMA) of the trainable parameters of the online network $\Theta$ . Here we define the teacher network $\Theta^{*}$ as the pre-trained motion latent diffusion model, i.e., MLD [9]. According to [44], the online network and target network are initialized with the parameters of the teacher network. During the inference phase, as shown in Fig. 5, our MotionLCM can generate high-quality motions with one-step sampling and achieve the fastest runtime ( $\sim$ 30ms per motion sequence) compared to other motion diffusion models [9,65]. + +# 3.3 Controllable Motion Generation in Latent Space + +After addressing the low-efficiency issue in the motion latent diffusion model [9], we delve into another exploration of real-time motion control. Inspired by the + +great success of ControlNet [82] in controllable image generation [55], we introduce a motion ControlNet $\Theta^a$ in the latent space of MotionLCM and initialize the motion ControlNet with a trainable copy of MotionLCM. Specifically, each layer in the motion ControlNet is appended with a zero-initialized linear layer for eliminating random noise in the initial training steps. To achieve an autoregressive motion generation paradigm, as depicted in Fig. 1, we define the motion control task as generating motions given the initial $\tau$ poses and textual description. As shown in Fig. 4 (b), the initial $\tau$ poses are defined by the trajectories of $K$ control joints, $\mathbf{g}^{1:\tau} = \{\mathbf{g}^i\}_{i=1}^\tau$ , where $\mathbf{g}^i \in \mathbb{R}^{K \times 3}$ denotes the global absolute locations of each control joint. In our motion control pipeline, we design a Trajectory Encoder $\Theta^b$ consisting of stacked transformer [67] layers to encode the trajectory signals. We append a global token (i.e., [CLS]) before the start of the trajectory sequence as the output feature of the encoder, which is added to the noisy $\mathbf{z}_n$ and fed into the trainable motion ControlNet $\Theta^a$ . Under the guidance of motion ControlNet, MotionLCM predicts the denoised $\hat{\mathbf{z}}_0$ through the consistency function $f_{\Theta^s}$ , where $\Theta^s$ is the combination of $\Theta^a$ , $\Theta^b$ and $\Theta$ . The following reconstruction loss $\mathcal{L}_{\mathrm{recon}}$ optimizes the motion ControlNet $\Theta^a$ and Trajectory Encoder $\Theta^b$ , + +$$ +\mathcal {L} _ {\mathrm {r e c o n}} \left(\boldsymbol {\Theta} ^ {a}, \boldsymbol {\Theta} ^ {b}\right) = \mathbb {E} \left[ d \left(\boldsymbol {f} _ {\boldsymbol {\Theta} ^ {s}} \left(\mathbf {z} _ {n}, t _ {n}, w, \mathbf {c} ^ {*}\right), \mathbf {z} _ {0}\right) \right], \tag {8} +$$ + +where $\mathbf{c}^*$ includes the text condition and control guidance from the Trajectory Encoder and the motion ControlNet. However, during training, the sole reconstruction supervision in the latent space is insufficient. We argue this is because the controllable motion generation requires more detailed constraints, which cannot be effectively provided solely by the reconstruction loss in the latent space. Unlike previous methods like OmniControl [73], which directly diffuse in the motion space, allowing explicit supervision of control signals, effectively supervising control signals in the latent space is non-trivial. Therefore, we utilize the frozen VAE [30] decoder $\mathcal{D}$ to decode the latent $\hat{\mathbf{z}}_0$ into the motion space, obtaining the predicted motion $\hat{\mathbf{x}}_0$ , thereby introducing the control loss $\mathcal{L}_{\mathrm{control}}$ as follows, + +$$ +\mathcal {L} _ {\mathrm {c o n t r o l}} (\boldsymbol {\Theta} ^ {a}, \boldsymbol {\Theta} ^ {b}) = \mathbb {E} \left[ \frac {\sum_ {i} \sum_ {j} m _ {i j} | | R (\hat {\mathbf {x}} _ {0}) _ {i j} - R (\mathbf {x} _ {0}) _ {i j} | | _ {2} ^ {2}}{\sum_ {i} \sum_ {j} m _ {i j}} \right], \tag {9} +$$ + +where $R(\cdot)$ converts the joint's local positions to global absolute locations and $m_{ij} \in \{0,1\}$ is the binary joint mask at frame $i$ for the joint $j$ . Then we optimize the motion ControlNet $\Theta^a$ and Trajectory Encoder $\Theta^b$ with the overall objective, + +$$ +\begin{array}{l} \boldsymbol {\Theta} ^ {a}, \boldsymbol {\Theta} ^ {b} = \underset {\boldsymbol {\Theta} ^ {a}, \boldsymbol {\Theta} ^ {b}} {\operatorname {a r g m i n}} (\mathcal {L} _ {\text {r e c o n}} + \lambda \mathcal {L} _ {\text {c o n t r o l}}), \end{array} \tag {10} +$$ + +where $\lambda$ is the weight to balance the two losses. This design enables explicit control signals to directly influence the generation process, similar to controlling other latent-free diffusion models for motion generation [29,73]. Extensive experiments demonstrate the introduced supervision is very helpful in improving the motion control performance, which will be introduced in the following section. + +# 4 Experiments + +In this section, we first present the experimental setup details in Sec. 4.1. Subsequently, we provide quantitative and qualitative comparisons to evaluate the effectiveness of our proposed MotionLCM framework in Sec. 4.2 and Sec. 4.3. Finally, we conduct comprehensive ablation studies on MotionLCM in Sec. 4.4. + +# 4.1 Experimental setup + +Datasets. We experiment on the popular HumanML3D [17] dataset, featuring 14,616 unique human motion sequences with 44,970 textual descriptions. For a fair comparison with previous methods [9,17,49,65,83], we take the redundant motion representation, including root velocity, root height, local joint positions, velocities, rotations in root space, and the foot contact binary labels. + +Evaluation metrics. We extend the evaluation metrics of previous works [9, 17, 73]. (1) Time cost: We follow [9] to report the Average Inference Time per Sentence (AITS) to evaluate the inference efficiency of models. (2) Motion quality: Frechet Inception Distance (FID) is adopted as a principal metric to evaluate the feature distributions between the generated and real motions. The feature extractor employed is from [17]. (3) Motion diversity: MultiModality (MModality) measures the generation diversity conditioned on the same text and Diversity calculates variance through features [17]. (4) Condition matching: Following [17], we calculate the motion-retrieval precision (R-Precision) to report the text-motion Top-1/2/3 matching accuracy and Multimodal Distance (MM Dist) calculates the mean distance between motions and texts. (5) Control error: Trajectory error (Traj. err.) quantifies the ratio of unsuccessful trajectories, characterized by any control joint location error surpassing a predetermined threshold. Location error (Loc. err.) represents the unsuccessful joints. Average error (Avg. err.) denotes the mean location error of the control joints. + +Implementation details. Our baseline motion diffusion model is based on MLD [9]. We reproduce MLD with higher performance. Unless otherwise specified, all our experiments are conducted on this model. For MotionLCM, we employ the AdamW [40] optimizer for 96K iterations using a cosine decay learning rate scheduler and 1K iterations of linear warm-up. A batch size of 256 and a learning rate of 2e-4 are used. We set the training guidance scale range as $[w_{\mathrm{min}}, w_{\mathrm{max}}] = [5, 15]$ , with the testing guidance scale set to 7.5, and adopt the EMA rate $\mu = 0.95$ by default. We use the DDIM [61] solver with skipping interval $k = 20$ and choose the Huber [25] loss as the distance measuring function $d$ . For motion ControlNet, we use the AdamW [40] optimizer for 192K iterations with 1K iterations of linear warm-up. The batch size and learning rate are set to 128 and 1e-4. The learning rate scheduler is the same as the first stage. For the training objective, we employ $d$ as the L2 loss and set the control loss weight $\lambda$ to 1.0 by default. We set the control ratio $\tau$ as 0.25 and the number of control joints as $K = 6$ (i.e., Pelvis, Left foot, Right foot, Head, Left wrist, and Right wrist) in both training and testing. We implement our model using PyTorch [47] with training on an NVIDIA RTX 4090 GPU and testing on a Tesla V100 GPU. + +Table 1: Comparison of text-conditional motion synthesis on HumanML3D [17] dataset. We compute suggested metrics following [17]. We repeat the evaluation 20 times for each metric and report the average with a $95\%$ confidence interval. “ $\rightarrow$ ” indicates that the closer to the real data, the better. **Bold and underline indicate the best and the second best result.** denotes the reproduced version of MLD [9]. The MotionLCM in one-step inference (30ms) surpasses all state-of-the-art models. + +
MethodsAITS ↓R-Precision ↑FID ↓MM Dist ↓Diversity →MModality ↑
Top 1Top 2Top 3
Real-0.511±.0030.703±.0030.797±.0020.002±.0002.794±.0089.503±.065-
Seq2Seq [37]-0.180±.0020.300±.0020.396±.00211.75±.0355.529±.0076.223±.061-
JL2P [2]-0.246±.0010.387±.0020.486±.00211.02±.0465.296±.0087.676±.058-
T2G [5]-0.165±.0010.267±.0020.345±.0027.664±.0306.030±.0086.409±.071-
Hier [14]-0.301±.0020.425±.0020.552±.0046.532±.0245.012±.0188.332±.042-
TEMOS [49]0.0170.424±.0020.612±.0020.722±.0023.734±.0283.703±.0088.973±.0710.368±.018
T2M [17]0.0380.457±.0020.639±.0030.740±.0031.067±.0023.340±.0089.188±.0022.090±.083
MDM [65]24.740.320±.0050.498±.0040.611±.0070.544±.0445.566±.0279.559±.0862.799±.072
MotionDiffuse [83]14.740.491±.0010.681±.0010.782±.0010.630±.0013.113±.0019.410±.0491.553±.042
MLD [9]0.2170.481±.0030.673±.0030.772±.0020.473±.0133.196±.0109.724±.0822.413±.079
MLD* [9]0.2250.504±.0020.698±.0030.796±.0020.450±.0113.052±.0099.634±.0642.267±.082
MotionLCM (1-step)0.0300.502±.0030.701±.0020.803±.0020.467±.0123.022±.0099.631±.0662.172±.082
MotionLCM (2-step)0.0350.505±.0030.705±.0020.805±.0020.368±.0112.986±.0089.640±.0522.187±.094
MotionLCM (4-step)0.0430.502±.0030.698±.0020.798±.0020.304±.0123.012±.0079.607±.0662.259±.092
+ +# 4.2 Comparisons on Text-to-motion + +In the following part, we first evaluate our MotionLCM on the text-to-motion (T2M) task. We compare our method with some T2M baselines on HumanML3D [17] with suggested metrics [17] under the $95\%$ confidence interval from 20 times running. As MotionLCM is based on MLD, we mainly focus on the performance compared with MLD. For evaluating time efficiency, we compare the Average Inference Time per Sentence (AITS) with TEMOS [49], T2M [17], MDM [65], MotionDiffuse [83] and MLD [9]. The results are borrowed from MLD [9]. The deterministic methods [2,5,14,37], are unable to produce diverse results from a single input text and thus we leave their MModality metrics empty. For the quantitative results, as shown in Tab. 1, our MotionLCM boasts an impressive real-time runtime efficiency, averaging around 30ms per motion sequence during inference. This performance exceeds that of previous diffusion-based methods [9,65,83] and even surpasses MLD [9] by an order of magnitude. Furthermore, despite employing only one-step inference, our MotionLCM can approximate or even surpass the performance of MLD [9] (DDIM [61] 50 steps). With two-step inference, we achieve the best R-Precision and MM Dist metrics, while increasing the sampling steps to four yields the best FID. The above results demonstrate the effectiveness of latent consistency distillation. For the qualitative results, as shown in Fig. 5, MotionLCM not only accelerates motion generation to real-time speed but also delivers high-quality outputs, closely aligning with the textual descriptions. + +# 4.3 Comparisons on Controllable Motion Generation + +As shown in Tab. 2, we compare our MotionLCM with OmniControl [73] and MLD [9]. We observe that OmniControl struggles with multi-joint control and + +
RealMotionLCM (Ours)MLDMDM
“a person walks clockwise in a large curve while swinging their arms.”
0.033s0.201s22.58s
“a person walking like a bird and then sniffing the air.”
0.031s0.192s21.16s
“a person slightly bent over with right hand pressing against the air walks forward slowly”
0.031s0.197s23.44s
+ +Fig. 5: Qualitative comparison of the state-of-the-art methods in the text-to-motion task. With only one-step inference, MotionLCM achieves the fastest motion generation while producing high-quality movements that closely match the textual descriptions. + +falls short in both generation quality and control performance compared to MotionLCM. To verify the effectiveness of the latent generated by our MotionLCM for training motion ControlNet, we conducted the following two sets of experiments: "LC" and "MC", which indicate introducing control supervision in the latent space and motion space. It can be observed that under the same experimental settings, MotionLCM maintains higher fidelity and significantly outperforms MLD [9] in motion control performance. This demonstrates that the latent generated by MotionLCM is more effective for training motion ControlNet compared to MLD [9]. In terms of inference speed, MotionLCM (1-step) is $1929 \times$ faster compared to OmniControl [73] and $13 \times$ faster than MLD [9]. For qualitative results, as shown in Fig. 6, OmniControl fails to control the initial poses in the second example and does not generate motion that aligns with the text in the third case. However, our MotionLCM not only adheres to the control of the initial poses but also generates motions that match the textual descriptions. + +# 4.4 Ablation Studies + +Impact of the hyperparameters of training MotionLCM. We conduct a comprehensive analysis of the training hyperparameters of MotionLCM, including the training guidance scale range $[w_{\mathrm{min}}, w_{\mathrm{max}}]$ , EMA rate $\mu$ , skipping interval $k$ , and the type of loss. We summarize the evaluation results based on one-step + +Table 2: Comparison of motion control on HumanML3D [17] dataset. Bold indicates the best result. Our MotionLCM outperforms OmniControl [73] and MLD [9] regarding generation quality, control performance, and inference speed. "LC" and "MC" refer to the control supervision introduced in the latent space and motion space. + +
MethodsAITS ↓FID ↓R-Precision ↑ Top 3Diversity →Traj. err. ↓ (50cm)Loc. err. ↓ (50cm)Avg. err. ↓
Real-0.0020.7979.5030.00000.00000.0000
OmniControl [73]81.002.3280.5578.8670.33620.03220.0977
MLD [9] (LC)0.5520.4690.7239.4760.42300.06530.1690
MotionLCM (1-step, LC)0.0420.3190.7529.4240.29860.03440.1410
MotionLCM (2-step, LC)0.0470.3150.7709.4270.28400.03280.1365
MotionLCM (4-step, LC)0.0630.3280.7459.4410.29730.03390.1398
MLD [9] (LC&MC)0.5520.5550.7549.3730.27220.02150.1265
MotionLCM (1-step, LC&MC)0.0420.4190.7569.3900.19880.01470.1127
MotionLCM (2-step, LC&MC)0.0470.3970.7599.4690.19600.01430.1092
MotionLCM (4-step, LC&MC)0.0630.4440.7539.3550.20890.01720.1140
+ +Table 3: Ablation study on different training guidance scale ranges $[w_{\mathrm{min}}, w_{\mathrm{max}}]$ , EMA rates $\mu$ , skipping intervals $k$ and types of loss. We use metrics in Tab. 1 and adopt a one-step inference setting with the testing CFG scale of 7.5 for fair comparison. + +
MethodsR-Precision ↑ Top 1FID ↓MM Dist ↓Diversity →MModality ↑
Real0.511±.0030.002±.0002.794±.0089.503±.065-
MotionLCM (w ∈ [5,15])0.502±.0030.467±.0123.022±.0099.631±.0662.172±.082
MotionLCM (w ∈ [2,18])0.497±.0030.481±.0093.030±.0109.644±.0732.226±.091
MotionLCM (w = 7.5)0.486±.0020.479±.0093.094±.0099.610±.0722.320±.097
MotionLCM (μ = 0.95)0.502±.0030.467±.0123.022±.0099.631±.0662.172±.082
MotionLCM (μ = 0.50)0.498±.0030.478±.0093.022±.0109.655±.0712.188±.087
MotionLCM (μ = 0)0.499±.0030.505±.0083.018±.0099.706±.0702.123±.085
MotionLCM (k = 50)0.488±.0030.547±.0113.096±.0109.511±.0742.324±.091
MotionLCM (k = 20)0.502±.0030.467±.0123.022±.0099.631±.0662.172±.082
MotionLCM (k = 10)0.497±.0030.449±.0093.017±.0109.693±.0752.133±.086
MotionLCM (k = 5)0.488±.0030.438±.0093.044±.0099.647±.0742.147±.083
MotionLCM (k = 1)0.442±.0020.635±.0113.255±.0089.384±.0802.146±.075
MotionLCM (w/ Huber)0.502±.0030.467±.0123.022±.0099.631±.0662.172±.082
MotionLCM (w/ L2)0.486±.0020.622±.0103.114±.0099.573±.0692.218±.086
+ +inference in Tab. 3. We find out that using a dynamic training guidance scale (e.g., $w \in [5,15]$ ) during training leads to an improvement in model performance compared to using a static training guidance scale (e.g., $w = 7.5$ ). Additionally, an excessively large range for the training guidance scale can also negatively impact the performance of the model (e.g., $w \in [2,18]$ ). Regarding the EMA rate $\mu$ , we observe that the larger the value of $\mu$ , the better the performance of the model. This indicates that maintaining a slower update rate for the target network $\Theta^{-}$ helps improve the performance of latent consistency distillation. When the skipping interval $k$ continues to increase, the performance of the distillation model improves progressively, but larger values of $k$ (e.g., $k = 50$ ) may result in inferior results. As for the type of loss, the Huber loss [25] significantly outperforms the L2 loss, demonstrating its superior robustness. + +
MotionLCM (w/o ControlNet)0.029s0.033s0.030s0.033s0.029s
OmniControl87.81s90.00s85.11s80.15s86.77s
MotionLCM (w/ ControlNet)0.043s0.045s0.041s0.045s0.047s
Reala man lifts his left arm up.the person is widewalking.a person walks three steps to his right, then five steps to his left, and finally three steps to his right.this person is walking in circles then moves across room to step around an object.the person is holding their head while walking.
+ +Fig. 6: Qualitative comparison of the state-of-the-art methods in the motion control task. We provide the visualized motion results and real references from five prompts. Compared to OmniControl [73], MotionLCM with ControlNet not only generates the initial poses that accurately follow the given multi-joint trajectories (i.e., the green poses in real references) but also produces motions that closely align with the texts. + +Impact of control loss weights $\lambda$ . To verify the impact of different control loss weights $\lambda$ on the control performance of MotionLCM, we report the experimental results in Tab. 4. We also include experiments of MotionLCM without ControlNet (i.e., only text-to-motion) for comparison. We found a significant improvement in control-related metrics (e.g., Loc. err.) after introducing motion ControlNet (i.e., $\lambda = 0$ ). Furthermore, control performance can be further improved by introducing control loss (i.e., $\lambda > 0$ ). Increasing the weight $\lambda$ enhances control performance but leads to a decline in the generation quality, which is reflected in higher FID scores. To balance these two aspects, we adopt $\lambda = 1$ as our default setting for training motion ControlNet. + +Impact of different control ratios $\tau$ and number of control joints $K$ . In Tab. 5, we present the results of all models with the testing control ratio as 0.25 and keep the number of control joints $K$ equal in both training and testing. We found that the model with the fixed training control ratio (i.e., $\tau = 0.25$ ) performs better compared to a dynamic ratio (e.g., $\tau \in [0.1, 0.5]$ ), and we discover that our model maintains good performance when incorporating additional redundant control signals, such as whole-body joints with $K = 22$ . + +Table 4: Ablation study on different control loss weights $\lambda$ . We present the results of (1, 2, 4)-step inference. We add the MotionLCM without ControlNet for comparison. + +
MethodsFID ↓ Top 3R-Precision ↑ Diversity →Traj. err. ↓ (50cm)Loc. err. ↓ (50cm)Avg. err. ↓
Real0.0020.7979.5030.00000.00000.0000
MotionLCM (1-step, w/o control)0.4670.8039.6310.76050.23020.3493
MotionLCM (2-step, w/o control)0.3680.8059.6400.76460.22140.3386
MotionLCM (4-step, w/o control)0.3040.7989.6070.77390.22070.3359
MotionLCM (1-step, λ = 0)0.3190.7529.4240.29860.03440.1410
MotionLCM (2-step, λ = 0)0.3150.7709.4270.28400.03280.1365
MotionLCM (4-step, λ = 0)0.3280.7459.4410.29730.03390.1398
MotionLCM (1-step, λ = 0.1)0.3440.7539.3860.27110.02750.1310
MotionLCM (2-step, λ = 0.1)0.3240.7599.4280.26310.02560.1268
MotionLCM (4-step, λ = 0.1)0.3570.7439.4150.27130.02680.1309
MotionLCM (1-step, λ = 1.0)0.4190.7569.3900.19880.01470.1127
MotionLCM (2-step, λ = 1.0)0.3970.7599.4690.19600.01430.1092
MotionLCM (4-step, λ = 1.0)0.4440.7539.3550.20890.01720.1140
MotionLCM (1-step, λ = 10.0)0.6360.7449.4790.14650.00970.0967
MotionLCM (2-step, λ = 10.0)0.5510.7579.5690.15900.01070.0987
MotionLCM (4-step, λ = 10.0)0.5680.7429.4860.17230.01320.1045
+ +Table 5: Ablation study on different control ratios $\tau$ and number of control joints $K$ . We report the results of (1,2,4)-step inference. " $*$ " denotes the default training setting. + +
MethodsFID ↓R-Precision ↑ Top 3Diversity →Traj. err. ↓ (50cm)Loc. err. ↓ (50cm)Avg. err. ↓
Real0.0020.7979.5030.00000.00000.0000
MotionLCM* (1-step, τ = 0.25, K = 6)0.4190.7569.3900.19880.01470.1127
MotionLCM* (2-step, τ = 0.25, K = 6)0.3970.7599.4690.19600.01430.1092
MotionLCM* (4-step, τ = 0.25, K = 6)0.4440.7539.3550.20890.01720.1140
MotionLCM (1-step, τ ∈ [0.1, 0.25])0.4560.7579.4770.28210.02340.1214
MotionLCM (2-step, τ ∈ [0.1, 0.25])0.4090.7699.5920.27070.02300.1179
MotionLCM (4-step, τ ∈ [0.1, 0.25])0.4570.7579.5400.29280.02560.1228
MotionLCM (1-step, τ ∈ [0.1, 0.5])0.4480.7639.5380.23900.01820.1182
MotionLCM (2-step, τ ∈ [0.1, 0.5])0.4130.7689.5170.23490.01800.1153
MotionLCM (4-step, τ ∈ [0.1, 0.5])0.4460.7539.4980.25170.01990.1196
MotionLCM (1-step, K = 12)0.4120.7539.4120.20720.01100.1029
MotionLCM (2-step, K = 12)0.4100.7589.5090.19790.01080.1000
MotionLCM (4-step, K = 12)0.4420.7559.3800.21690.01320.1048
MotionLCM (1-step, K = 22(whole-body))0.4360.7489.3790.21430.00830.0914
MotionLCM (2-step, K = 22(whole-body))0.4130.7589.4920.20610.00820.0881
MotionLCM (4-step, K = 22(whole-body))0.4610.7459.4590.21730.00970.0918
+ +# 5 Conclusion + +This work proposes an efficient controllable motion generation framework, MotionLCM. By introducing latent consistency distillation, MotionLCM enjoys the trade-off between runtime efficiency and generation quality. Moreover, thanks to the motion ControlNet manipulation in the latent space, our method obtains excellent controlling ability with given conditions. Extensive experimental results show the effectiveness of the proposed method. As the VAE of MLD lacks explicit temporal modeling, the MotionLCM cannot achieve a good temporal explanation. Therefore, our future work will lie in developing a more explainable compression architecture for efficient motion control. + +# Acknowledgements + +The research is supported by Shenzhen Ubiquitous Data Enabling Key Lab under grant ZDSYS20220527171406015 and CCF-Tencent Rhino-Bird Open Research Fund. This project is also supported by Shanghai Artificial Intelligence Laboratory. The author team would like to acknowledge Yiming Xie, Zhiyang Dou, and Shunlin Lu for their helpful technical discussions and suggestions. + +# References + +1. Ahn, H., Ha, T., Choi, Y., Yoo, H., Oh, S.: Text2action: Generative adversarial synthesis from language to action. In: ICRA. pp. 5915-5920 (2018) +2. Ahuja, C., Morency, L.P.: Language2pose: Natural language grounded pose forecasting. In: 3DV. pp. 719-728 (2019) +3. Athanasiou, N., Petrovich, M., Black, M.J., Varol, G.: Teach: Temporal action composition for 3d humans. In: 3DV. pp. 414-423 (2022) +4. Barquero, G., Escalera, S., Palmero, C.: Seamless human motion composition with blended positional encodings. In: CVPR. pp. 457-469 (2024) +5. Bhattacharya, U., Rewkowski, N., Banerjee, A., Guhan, P., Bera, A., Manocha, D.: Text2gestures: A transformer-based network for generating emotive body gestures for virtual agents. In: VR. pp. 1-10 (2021) +6. Cervantes, P., Sekikawa, Y., Sato, I., Shinoda, K.: Implicit neural representations for variable length human motion generation. In: ECCV. pp. 356-372 (2022) +7. Chen, L.H., Lu, S., Zeng, A., Zhang, H., Wang, B., Zhang, R., Zhang, L.: Motionllm: Understanding human behaviors from human motions and videos. arXiv preprint arXiv:2405.20340 (2024) +8. Chen, L.H., Zhang, J., Li, Y., Pang, Y., Xia, X., Liu, T.: Humanmac: Masked motion completion for human motion prediction. In: ICCV. pp. 9544-9555 (2023) +9. Chen, X., Jiang, B., Liu, W., Huang, Z., Fu, B., Chen, T., Yu, G.: Executing your commands via motion diffusion in latent space. In: CVPR. pp. 18000-18010 (2023) +0. Cong, P., Dou, Z.W., Ren, Y., Yin, W., Cheng, K., Sun, Y., Long, X., Zhu, X., Ma, Y.: Laserhuman: Language-guided scene-aware human motion generation in free environment. arXiv preprint arXiv:2403.13307 (2024) +1. Dabral, R., Mughal, M.H., Golyanik, V., Theobalt, C.: Mofusion: A framework for denoising-diffusion-based motion synthesis. In: CVPR. pp. 9760-9770 (2023) +2. Dou, Z., Chen, X., Fan, Q., Komura, T., Wang, W.: C-ase: Learning conditional adversarial skill embeddings for physics-based characters. In: SIGGRAPH Asia. pp. 1-11 (2023) +3. Fan, K., Tang, J., Cao, W., Yi, R., Li, M., Gong, J., Zhang, J., Wang, Y., Wang, C., Ma, L.: Freemotion: A unified framework for number-free text-to-motion synthesis. arXiv preprint arXiv:2405.15763 (2024) +4. Ghosh, A., Cheema, N., Oguz, C., Theobalt, C., Slusallek, P.: Synthesis of compositional animations from textual descriptions. In: ICCV. pp. 1396-1406 (2021) +5. Guo, C., Mu, Y., Javed, M.G., Wang, S., Cheng, L.: Momask: Generative masked modeling of 3d human motions. In: CVPR. pp. 1900-1910 (2024) +6. Guo, C., Mu, Y., Zuo, X., Dai, P., Yan, Y., Lu, J., Cheng, L.: Generative human motion stylization in latent space. arXiv preprint arXiv:2401.13505 (2024) +7. Guo, C., Zou, S., Zuo, X., Wang, S., Ji, W., Li, X., Cheng, L.: Generating diverse and natural 3d human motions from text. In: CVPR. pp. 5152-5161 (2022) + +18. Guo, C., Zuo, X., Wang, S., Cheng, L.: Tm2t: Stochastic and tokenized modeling for the reciprocal generation of 3d human motions and texts. In: ECCV. pp. 580-597 (2022) +19. Guo, C., Zuo, X., Wang, S., Zou, S., Sun, Q., Deng, A., Gong, M., Cheng, L.: Action2motion: Conditioned generation of 3d human motions. In: ACMMM. pp. 2021-2029 (2020) +20. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. NeurIPS pp. 6840-6851 (2020) +21. Ho, J., Salimans, T.: Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598 (2022) +22. Holden, D., Komura, T., Saito, J.: Phase-functional neural networks for character control. TOG 36(4), 1-13 (2017) +23. Holden, D., Saito, J., Komura, T.: A deep learning framework for character motion synthesis and editing. TOG 35(4), 1-11 (2016) +24. Huang, Y., Yang, H., Luo, C., Wang, Y., Xu, S., Zhang, Z., Zhang, M., Peng, J.: Stablemofusion: Towards robust and efficient diffusion-based motion generation framework. arXiv preprint arXiv:2405.05691 (2024) +25. Huber, P.J.: Robust estimation of a location parameter. The Annals of Mathematical Statistics 35(1), 73-101 (1964) +26. Ji, Y., Xu, F., Yang, Y., Shen, F., Shen, H.T., Zheng, W.S.: A large-scale rgb-d database for arbitrary-view human action recognition. In: ACMM. pp. 1510-1518 (2018) +27. Jiang, B., Chen, X., Liu, W., Yu, J., Yu, G., Chen, T.: Motiongpt: Human motion as a foreign language. NeurIPS (2024) +28. Karras, T., Aittala, M., Aila, T., Laine, S.: Elucidating the design space of diffusion-based generative models. NeurIPS pp. 26565-26577 (2022) +29. Karunratanakul, K., Preechakul, K., Suwajanakorn, S., Tang, S.: Guided motion diffusion for controllable human motion synthesis. In: CVPR. pp. 2151-2162 (2023) +30. Kingma, D.P., Welling, M.: Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114 (2013) +31. Lee, T., Moon, G., Lee, K.M.: Multiact: Long-term 3d human motion generation from multiple action labels. In: AAAI. pp. 1231-1239 (2023) +32. Li, B., Zhao, Y., Zhelun, S., Sheng, L.: Danceformer: Music conditioned 3d dance generation with parametric motion transformer. In: AAAI. pp. 1272-1279 (2022) +33. Li, R., Zhang, Y., Zhang, Y., Zhang, H., Guo, J., Zhang, Y., Liu, Y., Li, X.: Lodge: A coarse to fine diffusion network for long dance generation guided by the characteristic dance primitives. In: CVPR. pp. 1524-1534 (2024) +34. Li, R., Zhao, J., Zhang, Y., Su, M., Ren, Z., Zhang, H., Tang, Y., Li, X.: Finedance: A fine-grained choreography dataset for 3d full body dance generation. In: ICCV. pp. 10234-10243 (2023) +35. Li, R., Yang, S., Ross, D.A., Kanazawa, A.: Ai choreographer: Music conditioned 3d dance generation with aist++. In: ICCV. pp. 13401-13412 (2021) +36. Li, T., Qiao, C., Ren, G., Yin, K., Ha, S.: Aamdm: Accelerated auto-regressive motion diffusion model. In: CVPR. pp. 1813-1823 (2024) +37. Lin, A.S., Wu, L., Corona, R., Tai, K., Huang, Q., Mooney, R.J.: Generating animated videos of human activities from natural language descriptions. Learning 1(2018), 1 (2018) +38. Lin, X., Amer, M.R.: Human motion modeling using dvgens. arXiv preprint arXiv:1804.10652 (2018) + +39. Liu, J., Dai, W., Wang, C., Cheng, Y., Tang, Y., Tong, X.: Plan, posture and go: Towards open-world text-to-motion generation. arXiv preprint arXiv:2312.14828 (2023) +40. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101 (2017) +41. Lu, C., Zhou, Y., Bao, F., Chen, J., Li, C., Zhu, J.: Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. NeurIPS pp. 5775-5787 (2022) +42. Lu, C., Zhou, Y., Bao, F., Chen, J., Li, C., Zhu, J.: Dpm-solver++: Fast solver for guided sampling of diffusion probabilistic models. arXiv preprint arXiv:2211.01095 (2022) +43. Lu, S., Chen, L.H., Zeng, A., Lin, J., Zhang, R., Zhang, L., Shum, H.Y.: Humantomato: Text-aligned whole-body motion generation. arXiv preprint arXiv:2310.12978 (2023) +44. Luo, S., Tan, Y., Huang, L., Li, J., Zhao, H.: Latent consistency models: Synthesizing high-resolution images with few-step inference. arXiv preprint arXiv:2310.04378 (2023) +45. Mahmood, N., Ghorbani, N., Troje, N.F., Pons-Moll, G., Black, M.J.: Amass: Archive of motion capture as surface shapes. In: ICCV. pp. 5442-5451 (2019) +46. Nichol, A.Q., Dhariwal, P.: Improved denoising diffusion probabilistic models. In: ICML. pp. 8162-8171 (2021) +47. Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., et al.: Pytorch: An imperative style, high-performance deep learning library. NeurIPS (2019) +48. Petrovich, M., Black, M.J., Varol, G.: Action-conditioned 3d human motion synthesis with transformer vae. In: ICCV. pp. 10985-10995 (2021) +49. Petrovich, M., Black, M.J., Varol, G.: Temos: Generating diverse human motions from textual descriptions. In: ECCV. pp. 480-497 (2022) +50. Petrovich, M., Black, M.J., Varol, G.: Tmr: Text-to-motion retrieval using contrastive 3d human motion synthesis. In: ICCV. pp. 9488-9497 (2023) +51. Petrovich, M., Litany, O., Iqbal, U., Black, M.J., Varol, G., Bin Peng, X., Rempe, D.: Multi-track timeline control for text-driven 3d human motion generation. In: CVPRW. pp. 1911-1921 (2024) +52. Plappert, M., Mandery, C., Asfour, T.: The kit motion-language dataset. Big data 4(4), 236-252 (2016) +53. Punnakkal, A.R., Chandrasekaran, A., Athanasiou, N., Quiros-Ramirez, A., Black, M.J.: Babel: Bodies, action and behavior with english labels. In: CVPR. pp. 722-731 (2021) +54. Raab, S., Leibovitch, I., Li, P., Aberman, K., Sorkine-Hornung, O., Cohen-Or, D.: Modi: Unconditional motion synthesis from diverse data. In: CVPR. pp. 13873-13883 (2023) +55. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR. pp. 10684-10695 (2022) +56. Shafir, Y., Tevet, G., Kapon, R., Bermano, A.H.: Human motion diffusion as a generative prior. In: ICLR (2024) +57. Shahroudy, A., Liu, J., Ng, T.T., Wang, G.: Ntu $\mathrm{rgb + }$ d: A large scale dataset for 3d human activity analysis. In: CVPR. pp. 1010-1019 (2016) +58. Shi, Y., Wang, J., Jiang, X., Dai, B.: Controllable motion diffusion model. arXiv preprint arXiv:2306.00416 (2023) + +59. Siyao, L., Yu, W., Gu, T., Lin, C., Wang, Q., Qian, C., Loy, C.C., Liu, Z.: Bailando: 3d dance generation by actor-critic gpt with choreographic memory. In: CVPR. pp. 11050-11059 (2022) +60. Sohl-Dickstein, J., Weiss, E., Maheswaranathan, N., Ganguli, S.: Deep unsupervised learning using nonequilibrium thermodynamics. In: ICML. pp. 2256-2265 (2015) +61. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. In: ICLR (2021) +62. Song, Y., Dhariwal, P., Chen, M., Sutskever, I.: Consistency models. In: ICML (2023) +63. Tang, Y., Liu, J., Liu, A., Yang, B., Dai, W., Rao, Y., Lu, J., Zhou, J., Li, X.: Flag3d: A 3d fitness activity dataset with language instruction. In: CVPR. pp. 22106-22117 (2023) +64. Tevet, G., Gordon, B., Hertz, A., Bermano, A.H., Cohen-Or, D.: Motionclip: Exposing human motion generation to clip space. In: ECCV. pp. 358-374 (2022) +65. Tevet, G., Raab, S., Gordon, B., Shafir, Y., Cohen-Or, D., Bermano, A.H.: Human motion diffusion model. In: ICLR (2022) +66. Tseng, J., Castellon, R., Liu, K.: Edge: Editable dance generation from music. In: CVPR. pp. 448-458 (2023) +67. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. NeurIPS (2017) +68. Wan, W., Dou, Z., Komura, T., Wang, W., Jayaraman, D., Liu, L.: Tlcontrol: Trajectory and language control for human motion synthesis. arXiv preprint arXiv:2311.17135 (2023) +69. Wang, Z., Chen, Y., Jia, B., Li, P., Zhang, J., Zhang, J., Liu, T., Zhu, Y., Liang, W., Huang, S.: Move as you say interact as you can: Language-guided human motion generation with scene affordance. In: CVPR. pp. 433-444 (2024) +70. Wang, Z., Chen, Y., Liu, T., Zhu, Y., Liang, W., Huang, S.: Humanise: Language-conditioned human motion generation in 3d scenes. NeurIPS pp. 14959-14971 (2022) +71. Wang, Z., Wang, J., Lin, D., Dai, B.: Intercontrol: Generate human motion interactions by controlling every joint. arXiv preprint arXiv:2311.15864 (2023) +72. Xiao, Z., Wang, T., Wang, J., Cao, J., Zhang, W., Dai, B., Lin, D., Pang, J.: Unified human-scene interaction via prompted chain-of-contacts. In: ICLR (2024) +73. Xie, Y., Jampani, V., Zhong, L., Sun, D., Jiang, H.: Omnicontrol: Control any joint at any time for human motion generation. In: ICLR (2024) +74. Xu, L., Lv, X., Yan, Y., Jin, X., Wu, S., Xu, C., Liu, Y., Zhou, Y., Rao, F., Sheng, X., et al.: Inter-x: Towards versatile human-human interaction analysis. In: CVPR. pp. 22260-22271 (2024) +75. Xu, L., Song, Z., Wang, D., Su, J., Fang, Z., Ding, C., Gan, W., Yan, Y., Jin, X., Yang, X., et al.: Actformer: A gan-based transformer towards general action-conditioned 3d human motion generation. In: ICCV. pp. 2228-2238 (2023) +76. Xu, L., Zhou, Y., Yan, Y., Jin, X., Zhu, W., Rao, F., Yang, X., Zeng, W.: Regennet: Towards human action-reaction synthesis. In: CVPR. pp. 1759-1769 (2024) +77. Yan, S., Li, Z., Xiong, Y., Yan, H., Lin, D.: Convolutional sequence generation for skeleton-based action synthesis. In: ICCV. pp. 4394-4402 (2019) +78. Yuan, Y., Song, J., Iqbal, U., Vahdat, A., Kautz, J.: Physdiff: Physics-guided human motion diffusion model. In: ICCV. pp. 16010-16021 (2023) +79. Zhang, B., Cheng, Y., Wang, C., Zhang, T., Yang, J., Tang, Y., Zhao, F., Chen, D., Guo, B.: Rodinhd: High-fidelity 3d avatar generation with diffusion models. arXiv preprint arXiv:2407.06938 (2024) + +80. Zhang, B., Cheng, Y., Yang, J., Wang, C., Zhao, F., Tang, Y., Chen, D., Guo, B.: Gaussiancube: Structuring gaussian splatting using optimal transport for 3d generative modeling. arXiv preprint arXiv:2403.19655 (2024) +81. Zhang, J., Zhang, Y., Cun, X., Zhang, Y., Zhao, H., Lu, H., Shen, X., Shan, Y.: Generating human motion from textual descriptions with discrete representations. In: CVPR. pp. 14730-14740 (2023) +82. Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models. In: ICCV. pp. 3836-3847 (2023) +83. Zhang, M., Cai, Z., Pan, L., Hong, F., Guo, X., Yang, L., Liu, Z.: Motiondiffuse: Text-driven human motion generation with diffusion model. arXiv preprint arXiv:2208.15001 (2022) +84. Zhao, R., Su, H., Ji, Q.: Bayesian adversarial human motion synthesis. In: CVPR. pp. 6225-6234 (2020) +85. Zhong, L., Xie, Y., Jampani, V., Sun, D., Jiang, H.: Smoodi: Stylized motion diffusion model. arXiv preprint arXiv:2407.12783 (2024) +86. Zhou, W., Dou, Z., Cao, Z., Liao, Z., Wang, J., Wang, W., Liu, Y., Komura, T., Wang, W., Liu, L.: Emdm: Efficient motion diffusion model for fast, high-quality motion generation. arXiv preprint arXiv:2312.02256 (2023) \ No newline at end of file diff --git a/2024/MotionLCM_ Real-time Controllable Motion Generation via Latent Consistency Model/images.zip b/2024/MotionLCM_ Real-time Controllable Motion Generation via Latent Consistency Model/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..8ca705ca6b37e8cd9437d9a29f53ae7955ffff23 --- /dev/null +++ b/2024/MotionLCM_ Real-time Controllable Motion Generation via Latent Consistency Model/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a7a64320ade9482c9ef57eb73a3afe15a77690d44a7281b808de4fd603863683 +size 757315 diff --git a/2024/MotionLCM_ Real-time Controllable Motion Generation via Latent Consistency Model/layout.json b/2024/MotionLCM_ Real-time Controllable Motion Generation via Latent Consistency Model/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..067a3da74889dea620447f6ba200499e2a247a1f --- /dev/null +++ b/2024/MotionLCM_ Real-time Controllable Motion Generation via Latent Consistency Model/layout.json @@ -0,0 +1,11179 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 151, + 111, + 462, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 111, + 462, + 148 + ], + "spans": [ + { + "bbox": [ + 151, + 111, + 462, + 148 + ], + "type": "text", + "content": "MotionLCM: Real-time Controllable Motion Generation via Latent Consistency Model" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 167, + 468, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 167, + 468, + 194 + ], + "spans": [ + { + "bbox": [ + 146, + 167, + 468, + 194 + ], + "type": "text", + "content": "Wenxun Dai" + }, + { + "bbox": [ + 146, + 167, + 468, + 194 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 146, + 167, + 468, + 194 + ], + "type": "text", + "content": ", Ling-Hao Chen" + }, + { + "bbox": [ + 146, + 167, + 468, + 194 + ], + "type": "inline_equation", + "content": "^{2\\star}" + }, + { + "bbox": [ + 146, + 167, + 468, + 194 + ], + "type": "text", + "content": ", Jingbo Wang" + }, + { + "bbox": [ + 146, + 167, + 468, + 194 + ], + "type": "inline_equation", + "content": "^{3\\boxtimes}" + }, + { + "bbox": [ + 146, + 167, + 468, + 194 + ], + "type": "text", + "content": ", Jinpeng Liu" + }, + { + "bbox": [ + 146, + 167, + 468, + 194 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 146, + 167, + 468, + 194 + ], + "type": "text", + "content": ", Bo Dai" + }, + { + "bbox": [ + 146, + 167, + 468, + 194 + ], + "type": "inline_equation", + "content": "^{3\\boxtimes}" + }, + { + "bbox": [ + 146, + 167, + 468, + 194 + ], + "type": "text", + "content": ", Yansong Tang" + }, + { + "bbox": [ + 146, + 167, + 468, + 194 + ], + "type": "inline_equation", + "content": "^{1,2}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 144, + 201, + 469, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 144, + 201, + 469, + 258 + ], + "spans": [ + { + "bbox": [ + 144, + 201, + 469, + 258 + ], + "type": "text", + "content": "1Shenzhen Key Laboratory of Ubiquitous Data Enabling, Tsinghua Shenzhen International Graduate School 2Tsinghua University 3Shanghai AI Laboratory {wx dai 2001, thu. lhchen, wangjingbo1219, liu. jinpeng.55}@gmail.com {doubledaibo, tangyansong15}@gmail.com Project page: https://dai-wenxun.github.io/MotionLCM-page" + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 147, + 285, + 467, + 487 + ], + "blocks": [ + { + "bbox": [ + 147, + 285, + 467, + 487 + ], + "lines": [ + { + "bbox": [ + 147, + 285, + 467, + 487 + ], + "spans": [ + { + "bbox": [ + 147, + 285, + 467, + 487 + ], + "type": "image", + "image_path": "47c7ee0ac1d6278ba282f678ec7f939d12be6b7c6e0c3474f4d210e856f07a82.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 506, + 482, + 551 + ], + "lines": [ + { + "bbox": [ + 130, + 506, + 482, + 551 + ], + "spans": [ + { + "bbox": [ + 130, + 506, + 482, + 551 + ], + "type": "text", + "content": "Fig. 1: We propose MotionLCM, a real-time controllable motion latent consistency model. Our model uses the last few frames of the previous motion as temporal control signals to autoregressively generate the next motion in real-time under different text prompts. Green blocks denote the junctions. The numbers in red are the inference time." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 159, + 591, + 453, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 591, + 453, + 647 + ], + "spans": [ + { + "bbox": [ + 159, + 591, + 453, + 647 + ], + "type": "text", + "content": "Abstract. This work introduces MotionLCM, extending controllable motion generation to a real-time level. Existing methods for spatial-temporal control in text-conditioned motion generation suffer from significant runtime inefficiency. To address this issue, we first propose the motion latent consistency model (MotionLCM) for motion generation," + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 134, + 653, + 312, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 653, + 312, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 653, + 312, + 665 + ], + "type": "text", + "content": "* Project lead. ☑ Corresponding author." + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 159, + 117, + 455, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 117, + 455, + 239 + ], + "spans": [ + { + "bbox": [ + 159, + 117, + 455, + 239 + ], + "type": "text", + "content": "building upon the latent diffusion model [9]. By adopting one-step (or few-step) inference, we further improve the runtime efficiency of the motion latent diffusion model for motion generation. To ensure effective controllability, we incorporate a motion ControlNet within the latent space of MotionLCM and enable explicit control signals (e.g., initial poses) in the vanilla motion space to control the generation process directly, similar to controlling other latent-free diffusion models [29, 73] for motion generation. By employing these techniques, our approach can generate human motions with text and control signals in real-time. Experimental results demonstrate the remarkable generation and controlling capabilities of MotionLCM while maintaining real-time runtime efficiency." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 248, + 445, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 248, + 445, + 259 + ], + "spans": [ + { + "bbox": [ + 160, + 248, + 445, + 259 + ], + "type": "text", + "content": "Keywords: Text-to-Motion " + }, + { + "bbox": [ + 160, + 248, + 445, + 259 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 248, + 445, + 259 + ], + "type": "text", + "content": " Real-time Control " + }, + { + "bbox": [ + 160, + 248, + 445, + 259 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 248, + 445, + 259 + ], + "type": "text", + "content": " Consistency Model" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 281, + 230, + 293 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 281, + 230, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 281, + 230, + 293 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 306, + 293, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 306, + 293, + 569 + ], + "spans": [ + { + "bbox": [ + 130, + 306, + 293, + 569 + ], + "type": "text", + "content": "Text-to-motion generation (T2M) has attracted increasing attention [1, 15, 43, 49, 65] due to its important roles in many applications [70, 72]. Previous attempts mainly focus on GANs [1, 38], VAEs [3, 19, 48, 49] and diffusion models [9, 11, 39, 56, 65, 78, 83] via pairwise text-motion data [17, 26, 45, 52, 53, 57, 63, 74] and achieve impressive generation results. Existing approaches [9, 65, 83] mainly take diffusion models [20, 46, 55, 60] as a base generative model, owing to their powerful ability to model motion distribution. However, these diffusion fashions inevitably require considerable sampling steps for motion synthesis during inference, even with some sampling acceleration methods [61]. Specifically, MDM [65] and MLD [9] require " + }, + { + "bbox": [ + 130, + 306, + 293, + 569 + ], + "type": "inline_equation", + "content": "\\sim" + }, + { + "bbox": [ + 130, + 306, + 293, + 569 + ], + "type": "text", + "content": "12s and " + }, + { + "bbox": [ + 130, + 306, + 293, + 569 + ], + "type": "inline_equation", + "content": "\\sim" + }, + { + "bbox": [ + 130, + 306, + 293, + 569 + ], + "type": "text", + "content": "0.2s to generate" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 299, + 329, + 482, + 441 + ], + "blocks": [ + { + "bbox": [ + 299, + 329, + 482, + 441 + ], + "lines": [ + { + "bbox": [ + 299, + 329, + 482, + 441 + ], + "spans": [ + { + "bbox": [ + 299, + 329, + 482, + 441 + ], + "type": "image", + "image_path": "a4e5b6df9d8d0ad07f9ba9a2d1f4e85d197fee9e6e870e803aa016fd407bf094.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 297, + 449, + 482, + 538 + ], + "lines": [ + { + "bbox": [ + 297, + 449, + 482, + 538 + ], + "spans": [ + { + "bbox": [ + 297, + 449, + 482, + 538 + ], + "type": "text", + "content": "Fig.2: Comparison of the inference time costs on HumanML3D [17]. We compare the AITS and FID metrics with five SOTA methods. The closer the model is to the origin the better. Diffusion-based models are indicated by the blue dashed box. Our MotionLCM achieves real-time inference speed while ensuring high-quality motion generation." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 570, + 482, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 593 + ], + "type": "text", + "content": "a high-quality motion sequence. Such low efficiency blocks the applications of generating high-quality motions in various real-time scenarios." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "In addition to the language description itself serving as a coarse control signal, another line of research focuses on controlling the motion generation with spatial-temporal constraints [29, 56, 73]. Although these attempts enjoy impressive controlling ability in the T2M task, there still exists a significant gap towards real-time applications. For example, OmniControl [73] exhibits a relatively long inference time, " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\sim 81" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "s per sequence. Therefore, trading-off between generation" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "W. Dai et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "quality and efficiency is a challenging problem. As a result, in this paper, we target the real-time controllable motion generation research problem." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 481, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 481, + 259 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 481, + 259 + ], + "type": "text", + "content": "Recently, the concept of consistency models [44, 62] has been introduced in image generation, resulting in significant progress by enabling efficient and high-fidelity image synthesis with a minimal number of sampling steps (e.g., 4 steps vs. 50 steps). These properties perfectly align with our goal of accelerating motion generation without compromising generation quality. Therefore, we propose MotionLCM (Motion Latent Consistency Model) distilled from the motion latent diffusion model, MLD [9], to tackle the low-efficiency problem in diffusion sampling. To the best of our knowledge, we introduce consistency distillation into the motion generation area for the first time and accelerate motion generation to a real-time level via latent consistency distillation [44]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 260, + 482, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 260, + 482, + 463 + ], + "spans": [ + { + "bbox": [ + 130, + 260, + 482, + 463 + ], + "type": "text", + "content": "Here, in MotionLCM, we are facing another challenge on how to control motions with spatial-temporal signals (e.g., initial poses) in the latent space. Previous methods [56, 73] model human motions in the vanilla motion space and can manipulate the motion directly in the denoising process. However, for our latent-diffusion-based MotionLCM, it is non-trivial to feed the control signals into the latent space. This is because the latent has no explicit motion semantics, which cannot be manipulated directly by the control signals. Inspired by the notable success of [82] in controllable image generation [55], we introduce a motion ControlNet to control motion generation in the latent space. However, the naïve motion ControlNet is not sufficient to provide supervision for the control signals. The main reason is the lack of explicit supervision in the motion space. Therefore, during the training phase, we decode the predicted latent through the frozen VAE [30] decoder into the vanilla motion space to provide explicit control supervision on the generated motion. Thanks to the powerful one-step inference capability of MotionLCM, the latent generated by MotionLCM can significantly facilitate control supervision both in the latent space and motion space for training the motion ControlNet compared to MLD [9]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 463, + 481, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 463, + 481, + 499 + ], + "spans": [ + { + "bbox": [ + 130, + 463, + 481, + 499 + ], + "type": "text", + "content": "With our key designs, our proposed MotionLCM successfully enjoys the balance between the generation quality and efficiency in controllable motion generation. Before delivering into detail, we sum up our core contributions as follows." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 504, + 479, + 586 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 138, + 504, + 479, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 504, + 479, + 538 + ], + "spans": [ + { + "bbox": [ + 138, + 504, + 479, + 538 + ], + "type": "text", + "content": "- We propose the Motion Latent Consistency Model (MotionLCM) via consistency distillation on the motion latent diffusion model extending controllable motion generation to a real-time level." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 540, + 479, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 540, + 479, + 562 + ], + "spans": [ + { + "bbox": [ + 138, + 540, + 479, + 562 + ], + "type": "text", + "content": "- Building upon our achievement of real-time motion generation, we introduce a motion ControlNet, enabling high-quality controllable motion generation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 563, + 479, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 563, + 479, + 586 + ], + "spans": [ + { + "bbox": [ + 138, + 563, + 479, + 586 + ], + "type": "text", + "content": "- Extensive experimental results show that MotionLCM enjoys a good balance of generation quality, controlling capability, and real-time efficiency." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 602, + 237, + 615 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 602, + 237, + 615 + ], + "spans": [ + { + "bbox": [ + 132, + 602, + 237, + 615 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 624, + 298, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 624, + 298, + 635 + ], + "spans": [ + { + "bbox": [ + 132, + 624, + 298, + 635 + ], + "type": "text", + "content": "2.1 Human Motion Generation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "Generating human motions can be divided into three main fashions according to inputs: motion synthesis (1) without any condition [54,65,77,84], (2) with some" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 395, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 395, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 395, + 91, + 447, + 100 + ], + "type": "text", + "content": "MotionLCM" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 403 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 403 + ], + "type": "text", + "content": "given multi-modal conditions, such as action labels [6, 12, 19, 31, 48, 75], textual description [1-5, 7, 7, 9-11, 13-15, 17, 18, 24, 27, 29, 37-39, 43, 49-51, 56, 64, 65, 68-70, 72, 78, 81, 83, 85, 86], audio or music [32-35, 59, 66], (3) with user-defined trajectories [22, 23, 29, 36, 56, 58, 68, 71, 73]. To generate diverse, natural, and high-quality human motions, many generative models have been explored by [2, 38, 49, 79, 80]. Recently, diffusion-based models significantly improved the motion generation performance and diversity [8, 9, 11, 16, 65, 76, 83] with stable training. Specifically, MotionDiffuse [83] represents the first text-based motion diffusion model that provides fine-grained instructions on body parts and achieves arbitrary-length motion synthesis with time-varied text prompts. MDM [65] introduces a motion diffusion model that operates on raw motion data, enabling both high-quality generation and generic conditioning that together comprise a good baseline for new motion generation tasks. Based on MDM [65], OmniControl [73] integrates flexible spatial-temporal control signals across different joints by combining analytic spatial guidance and realism guidance into the diffusion model, ensuring that the generated motion closely conforms to the input control signals. The work most relevant to ours is MLD [9], which introduces a motion latent-based diffusion model to enhance generation quality and reduce computational resource requirements. The key idea is training a VAE [30] for motion embedding, followed by implementing latent diffusion [55] within the learned latent space. However, these diffusion fashions inevitably require considerable sampling steps for motion synthesis during inference, even with some sampling acceleration methods [61]. Thus, we propose MotionLCM, which not only guarantees high-quality controllable motion generation but also achieves real-time runtime efficiency." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 425, + 202, + 438 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 425, + 202, + 438 + ], + "spans": [ + { + "bbox": [ + 132, + 425, + 202, + 438 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 453, + 482, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 453, + 482, + 502 + ], + "spans": [ + { + "bbox": [ + 130, + 453, + 482, + 502 + ], + "type": "text", + "content": "In this section, we first briefly introduce preliminaries about latent consistency models in Sec. 3.1. Then, we describe how to conduct latent consistency distillation for motion generation in Sec. 3.2, followed by our implementation of motion control in latent space in Sec. 3.3. The overall pipeline is illustrated in Fig. 4." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 522, + 228, + 534 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 522, + 228, + 534 + ], + "spans": [ + { + "bbox": [ + 132, + 522, + 228, + 534 + ], + "type": "text", + "content": "3.1 Preliminaries" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 545, + 482, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 643 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 643 + ], + "type": "text", + "content": "The Consistency Model (CM) [62] introduces a kind of efficient generative model designed for efficient one-step or few-step generation. Given a Probability Flow ODE (a.k.a. PF-ODE) that smoothly converts data to noise, the CM is to learn the function " + }, + { + "bbox": [ + 130, + 545, + 482, + 643 + ], + "type": "inline_equation", + "content": "\\pmb{f}(\\cdot, \\cdot)" + }, + { + "bbox": [ + 130, + 545, + 482, + 643 + ], + "type": "text", + "content": " that maps any points on the ODE trajectory to its origin distribution (i.e., the solution of the PF-ODE). The consistency function is formally defined as " + }, + { + "bbox": [ + 130, + 545, + 482, + 643 + ], + "type": "inline_equation", + "content": "\\pmb{f}:(\\mathbf{x}_t,t)\\longmapsto \\mathbf{x}_{\\epsilon}" + }, + { + "bbox": [ + 130, + 545, + 482, + 643 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 545, + 482, + 643 + ], + "type": "inline_equation", + "content": "t\\in [\\epsilon ,T]" + }, + { + "bbox": [ + 130, + 545, + 482, + 643 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 545, + 482, + 643 + ], + "type": "inline_equation", + "content": "T > 0" + }, + { + "bbox": [ + 130, + 545, + 482, + 643 + ], + "type": "text", + "content": " is a fixed constant and " + }, + { + "bbox": [ + 130, + 545, + 482, + 643 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 130, + 545, + 482, + 643 + ], + "type": "text", + "content": " is a small positive number to avoid numerical instability. According to [62], the consistency function should satisfy the self-consistency property:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 235, + 654, + 481, + 666 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 654, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 235, + 654, + 481, + 666 + ], + "type": "interline_equation", + "content": "\\boldsymbol {f} \\left(\\mathbf {x} _ {t}, t\\right) = \\boldsymbol {f} \\left(\\mathbf {x} _ {t ^ {\\prime}}, t ^ {\\prime}\\right), \\forall t, t ^ {\\prime} \\in [ \\epsilon , T ]. \\tag {1}", + "image_path": "07faa52b6a84bb6c40fa792c387f4e645c47eb02c00d7721bb3fde3f4b5fdcef.jpg" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "W. Dai et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 152, + 116, + 343, + 155 + ], + "blocks": [ + { + "bbox": [ + 152, + 116, + 343, + 155 + ], + "lines": [ + { + "bbox": [ + 152, + 116, + 343, + 155 + ], + "spans": [ + { + "bbox": [ + 152, + 116, + 343, + 155 + ], + "type": "image", + "image_path": "5e7d75e984cbe2d5e45177e83441b38e9f8635bbd8f86ef4e4da593e1f9802ab.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 345, + 121, + 460, + 152 + ], + "lines": [ + { + "bbox": [ + 345, + 121, + 460, + 152 + ], + "spans": [ + { + "bbox": [ + 345, + 121, + 460, + 152 + ], + "type": "text", + "content": "Distillation target of " + }, + { + "bbox": [ + 345, + 121, + 460, + 152 + ], + "type": "inline_equation", + "content": "f_{\\Theta}" + }, + { + "bbox": [ + 345, + 121, + 460, + 152 + ], + "type": "text", + "content": ": Project any " + }, + { + "bbox": [ + 345, + 121, + 460, + 152 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_t" + }, + { + "bbox": [ + 345, + 121, + 460, + 152 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 345, + 121, + 460, + 152 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_0" + }, + { + "bbox": [ + 345, + 121, + 460, + 152 + ], + "type": "text", + "content": " consistently. " + }, + { + "bbox": [ + 345, + 121, + 460, + 152 + ], + "type": "inline_equation", + "content": "t \\in [0, T]" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 151, + 156, + 354, + 259 + ], + "blocks": [ + { + "bbox": [ + 151, + 156, + 354, + 259 + ], + "lines": [ + { + "bbox": [ + 151, + 156, + 354, + 259 + ], + "spans": [ + { + "bbox": [ + 151, + 156, + 354, + 259 + ], + "type": "image", + "image_path": "b475d73358e88bd5d634c16733218e43460cfb9ce878977325639f72136dd0db.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 268, + 482, + 346 + ], + "lines": [ + { + "bbox": [ + 130, + 268, + 482, + 346 + ], + "spans": [ + { + "bbox": [ + 130, + 268, + 482, + 346 + ], + "type": "text", + "content": "Fig. 3: The training objective of consistency distillation is to learn a consistency function " + }, + { + "bbox": [ + 130, + 268, + 482, + 346 + ], + "type": "inline_equation", + "content": "f_{\\Theta}" + }, + { + "bbox": [ + 130, + 268, + 482, + 346 + ], + "type": "text", + "content": ", initialized with the parameters of a pre-trained diffusion model (e.g., MLD [9]). This function " + }, + { + "bbox": [ + 130, + 268, + 482, + 346 + ], + "type": "inline_equation", + "content": "f_{\\Theta}" + }, + { + "bbox": [ + 130, + 268, + 482, + 346 + ], + "type": "text", + "content": " should projects any points (i.e., " + }, + { + "bbox": [ + 130, + 268, + 482, + 346 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_t" + }, + { + "bbox": [ + 130, + 268, + 482, + 346 + ], + "type": "text", + "content": ") on the ODE trajectory to its solution (i.e., " + }, + { + "bbox": [ + 130, + 268, + 482, + 346 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_0" + }, + { + "bbox": [ + 130, + 268, + 482, + 346 + ], + "type": "text", + "content": "). Once the pre-trained model [9] is distilled, unlike the traditional denoising model [65, 83] that requires considerable sampling steps, our MotionLCM can generate high-quality motion sequences with one-step sampling and further improve the generation quality through multi-step inference." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 360, + 157, + 460, + 259 + ], + "blocks": [ + { + "bbox": [ + 360, + 157, + 460, + 259 + ], + "lines": [ + { + "bbox": [ + 360, + 157, + 460, + 259 + ], + "spans": [ + { + "bbox": [ + 360, + 157, + 460, + 259 + ], + "type": "image", + "image_path": "846916882c9ecd29231b43b048d6d08e6ffc143c2beee9eb0b33183dd4d1a56a.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 368, + 482, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 368, + 482, + 430 + ], + "spans": [ + { + "bbox": [ + 130, + 368, + 482, + 430 + ], + "type": "text", + "content": "As shown in Eq. (1), the self-consistency property indicates that for arbitrary pairs of " + }, + { + "bbox": [ + 130, + 368, + 482, + 430 + ], + "type": "inline_equation", + "content": "(\\mathbf{x}_t,t)" + }, + { + "bbox": [ + 130, + 368, + 482, + 430 + ], + "type": "text", + "content": " on the same PF-ODE trajectory, the outputs of the model should be consistent. The goal of a parameterized consistency model " + }, + { + "bbox": [ + 130, + 368, + 482, + 430 + ], + "type": "inline_equation", + "content": "f_{\\Theta}" + }, + { + "bbox": [ + 130, + 368, + 482, + 430 + ], + "type": "text", + "content": " is to learn a consistency function from data by enforcing the self-consistency property in Eq. (1). To ensure that " + }, + { + "bbox": [ + 130, + 368, + 482, + 430 + ], + "type": "inline_equation", + "content": "f_{\\Theta}(\\mathbf{x},\\epsilon) = \\mathbf{x}" + }, + { + "bbox": [ + 130, + 368, + 482, + 430 + ], + "type": "text", + "content": ", the consistency model " + }, + { + "bbox": [ + 130, + 368, + 482, + 430 + ], + "type": "inline_equation", + "content": "f_{\\Theta}" + }, + { + "bbox": [ + 130, + 368, + 482, + 430 + ], + "type": "text", + "content": " is parameterized as," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 220, + 437, + 480, + 449 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 220, + 437, + 480, + 449 + ], + "spans": [ + { + "bbox": [ + 220, + 437, + 480, + 449 + ], + "type": "interline_equation", + "content": "\\boldsymbol {f} _ {\\Theta} (\\mathbf {x}, t) = c _ {\\text {s k i p}} (t) \\mathbf {x} + c _ {\\text {o u t}} (t) \\boldsymbol {F} _ {\\Theta} (\\mathbf {x}, t), \\tag {2}", + "image_path": "042ced61a9c106c8d4d2404d8cf0e530e409b82704a712ea38f4accc9d180361.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 457, + 482, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 457, + 482, + 506 + ], + "spans": [ + { + "bbox": [ + 130, + 457, + 482, + 506 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 457, + 482, + 506 + ], + "type": "inline_equation", + "content": "c_{\\mathrm{skip}}(t)" + }, + { + "bbox": [ + 130, + 457, + 482, + 506 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 457, + 482, + 506 + ], + "type": "inline_equation", + "content": "c_{\\mathrm{out}}(t)" + }, + { + "bbox": [ + 130, + 457, + 482, + 506 + ], + "type": "text", + "content": " are differentiable functions with " + }, + { + "bbox": [ + 130, + 457, + 482, + 506 + ], + "type": "inline_equation", + "content": "c_{\\mathrm{skip}}(\\epsilon) = 1" + }, + { + "bbox": [ + 130, + 457, + 482, + 506 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 457, + 482, + 506 + ], + "type": "inline_equation", + "content": "c_{\\mathrm{out}}(\\epsilon) = 0" + }, + { + "bbox": [ + 130, + 457, + 482, + 506 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 457, + 482, + 506 + ], + "type": "inline_equation", + "content": "\\pmb{F}_{\\Theta}(\\cdot, \\cdot)" + }, + { + "bbox": [ + 130, + 457, + 482, + 506 + ], + "type": "text", + "content": " is a deep neural network to learn the self-consistency. The CM trained from distilling the knowledge of pre-trained diffusion models is called Consistency Distillation. The consistency loss is defined as follows," + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 187, + 512, + 480, + 528 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 187, + 512, + 480, + 528 + ], + "spans": [ + { + "bbox": [ + 187, + 512, + 480, + 528 + ], + "type": "interline_equation", + "content": "\\mathcal {L} (\\boldsymbol {\\Theta}, \\boldsymbol {\\Theta} ^ {-}; \\boldsymbol {\\Phi}) = \\mathbb {E} \\left[ d \\left(\\boldsymbol {f} _ {\\boldsymbol {\\Theta}} \\left(\\mathbf {x} _ {t _ {n + 1}}, t _ {n + 1}\\right), \\boldsymbol {f} _ {\\boldsymbol {\\Theta} ^ {-}} \\left(\\hat {\\mathbf {x}} _ {t _ {n}} ^ {\\boldsymbol {\\Phi}}, t _ {n}\\right)\\right) \\right], \\tag {3}", + "image_path": "742e965fdf316dbf1829a4267c21022b7727b7107601b97ce07151702fa020f4.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 534, + 482, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 534, + 482, + 594 + ], + "spans": [ + { + "bbox": [ + 130, + 534, + 482, + 594 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 534, + 482, + 594 + ], + "type": "inline_equation", + "content": "d(\\cdot, \\cdot)" + }, + { + "bbox": [ + 130, + 534, + 482, + 594 + ], + "type": "text", + "content": " is a chosen metric function for measuring the distance between two samples. " + }, + { + "bbox": [ + 130, + 534, + 482, + 594 + ], + "type": "inline_equation", + "content": "f_{\\Theta}(\\cdot, \\cdot)" + }, + { + "bbox": [ + 130, + 534, + 482, + 594 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 534, + 482, + 594 + ], + "type": "inline_equation", + "content": "f_{\\Theta^{-}}(\\cdot, \\cdot)" + }, + { + "bbox": [ + 130, + 534, + 482, + 594 + ], + "type": "text", + "content": " are referred to as \"online network\" and \"target network\" according to [62]. Besides, " + }, + { + "bbox": [ + 130, + 534, + 482, + 594 + ], + "type": "inline_equation", + "content": "\\Theta^{-}" + }, + { + "bbox": [ + 130, + 534, + 482, + 594 + ], + "type": "text", + "content": " is updated with the exponential moving average (EMA) of the parameters of " + }, + { + "bbox": [ + 130, + 534, + 482, + 594 + ], + "type": "inline_equation", + "content": "\\Theta^{1}" + }, + { + "bbox": [ + 130, + 534, + 482, + 594 + ], + "type": "text", + "content": ". In Eq. (3), " + }, + { + "bbox": [ + 130, + 534, + 482, + 594 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_{t_n}^{\\Phi}" + }, + { + "bbox": [ + 130, + 534, + 482, + 594 + ], + "type": "text", + "content": " is a one-step estimation of " + }, + { + "bbox": [ + 130, + 534, + 482, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{t_n}" + }, + { + "bbox": [ + 130, + 534, + 482, + 594 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 130, + 534, + 482, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{t_{n+1}}" + }, + { + "bbox": [ + 130, + 534, + 482, + 594 + ], + "type": "text", + "content": ", which is formulated as," + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 209, + 603, + 480, + 617 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 209, + 603, + 480, + 617 + ], + "spans": [ + { + "bbox": [ + 209, + 603, + 480, + 617 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {x}} _ {t _ {n}} ^ {\\Phi} \\leftarrow \\mathbf {x} _ {t _ {n + 1}} + (t _ {n} - t _ {n + 1}) \\Phi (\\mathbf {x} _ {t _ {n + 1}}, t _ {n + 1}, \\emptyset), \\tag {4}", + "image_path": "508c80f51034b299cdced3376feba1a1037ce1f2b0eaa2606bfecd31c2b7e2c3.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 624, + 375, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 624, + 375, + 636 + ], + "spans": [ + { + "bbox": [ + 130, + 624, + 375, + 636 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 624, + 375, + 636 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 130, + 624, + 375, + 636 + ], + "type": "text", + "content": " is a one-step ODE solver applied to PF-ODE." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 395, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 395, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 395, + 91, + 447, + 100 + ], + "type": "text", + "content": "MotionLCM" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 641, + 481, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 133, + 641, + 481, + 666 + ], + "type": "text", + "content": " EMA operation: " + }, + { + "bbox": [ + 133, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\Theta^{-} \\gets \\mathbf{sg}(\\mu \\Theta^{-} + (1 - \\mu)\\Theta)" + }, + { + "bbox": [ + 133, + 641, + 481, + 666 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 133, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{sg}(\\cdot)" + }, + { + "bbox": [ + 133, + 641, + 481, + 666 + ], + "type": "text", + "content": " denotes the stopgrad operation and " + }, + { + "bbox": [ + 133, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 133, + 641, + 481, + 666 + ], + "type": "text", + "content": " satisfies " + }, + { + "bbox": [ + 133, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "0 \\leq \\mu < 1" + }, + { + "bbox": [ + 133, + 641, + 481, + 666 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 117, + 350, + 223 + ], + "blocks": [ + { + "bbox": [ + 134, + 117, + 350, + 223 + ], + "lines": [ + { + "bbox": [ + 134, + 117, + 350, + 223 + ], + "spans": [ + { + "bbox": [ + 134, + 117, + 350, + 223 + ], + "type": "image", + "image_path": "eab86df2a1587c50a0a037091a3594a96ccb0429d2d45e4eae44dd1bdba9eaeb.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 175, + 227, + 309, + 236 + ], + "lines": [ + { + "bbox": [ + 175, + 227, + 309, + 236 + ], + "spans": [ + { + "bbox": [ + 175, + 227, + 309, + 236 + ], + "type": "text", + "content": "(a) Motion Latent Consistency Distillation" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 352, + 114, + 481, + 224 + ], + "blocks": [ + { + "bbox": [ + 352, + 114, + 481, + 224 + ], + "lines": [ + { + "bbox": [ + 352, + 114, + 481, + 224 + ], + "spans": [ + { + "bbox": [ + 352, + 114, + 481, + 224 + ], + "type": "image", + "image_path": "7c2e2dde8b41f797545ffbbe271b2039f5bdac06f2bbd77b7c859b86f03f9a58.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 362, + 227, + 473, + 237 + ], + "lines": [ + { + "bbox": [ + 362, + 227, + 473, + 237 + ], + "spans": [ + { + "bbox": [ + 362, + 227, + 473, + 237 + ], + "type": "text", + "content": "(b) Motion Control in Latent Space" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 245, + 482, + 368 + ], + "lines": [ + { + "bbox": [ + 130, + 245, + 482, + 368 + ], + "spans": [ + { + "bbox": [ + 130, + 245, + 482, + 368 + ], + "type": "text", + "content": "Fig. 4: The overview of MotionLCM. (a) Motion Latent Consistency Distillation (Sec. 3.2). Given a raw motion sequence " + }, + { + "bbox": [ + 130, + 245, + 482, + 368 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0^{1:N}" + }, + { + "bbox": [ + 130, + 245, + 482, + 368 + ], + "type": "text", + "content": ", a pre-trained VAE [30] encoder first compresses it into the latent space, then a forward diffusion operation is performed to add " + }, + { + "bbox": [ + 130, + 245, + 482, + 368 + ], + "type": "inline_equation", + "content": "n + k" + }, + { + "bbox": [ + 130, + 245, + 482, + 368 + ], + "type": "text", + "content": " steps of noise. Then, the noisy " + }, + { + "bbox": [ + 130, + 245, + 482, + 368 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{n + k}" + }, + { + "bbox": [ + 130, + 245, + 482, + 368 + ], + "type": "text", + "content": " is fed into the online network and teacher network to predict the clean latent. The target network takes the " + }, + { + "bbox": [ + 130, + 245, + 482, + 368 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 245, + 482, + 368 + ], + "type": "text", + "content": "-step estimation results of the teacher output to predict the clean latent. To learn self-consistency, a loss is applied to enforce the output of the online network and target network to be consistent. (b) Motion Control in Latent Space (Sec. 3.3). With the powerful MotionLCM trained in the first stage, we incorporate a motion ControlNet into the MotionLCM to achieve controllable motion generation. Furthermore, we leverage the decoded motion to explicitly supervise the spatial-temporal control signals (i.e., initial poses " + }, + { + "bbox": [ + 130, + 245, + 482, + 368 + ], + "type": "inline_equation", + "content": "\\mathbf{g}^{1:\\tau}" + }, + { + "bbox": [ + 130, + 245, + 482, + 368 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 390, + 482, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 390, + 482, + 511 + ], + "spans": [ + { + "bbox": [ + 130, + 390, + 482, + 511 + ], + "type": "text", + "content": "The Latent Consistency Model (LCM) [44] learns the self-consistency property in the latent space " + }, + { + "bbox": [ + 130, + 390, + 482, + 511 + ], + "type": "inline_equation", + "content": "D_{\\mathbf{z}} = \\{(\\mathbf{z},\\mathbf{c})|\\mathbf{z} = \\mathcal{E}(\\mathbf{x}),(\\mathbf{x},\\mathbf{c})\\in D\\}" + }, + { + "bbox": [ + 130, + 390, + 482, + 511 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 390, + 482, + 511 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 130, + 390, + 482, + 511 + ], + "type": "text", + "content": " denotes the dataset, " + }, + { + "bbox": [ + 130, + 390, + 482, + 511 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 130, + 390, + 482, + 511 + ], + "type": "text", + "content": " is the given condition, and " + }, + { + "bbox": [ + 130, + 390, + 482, + 511 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 130, + 390, + 482, + 511 + ], + "type": "text", + "content": " is the pre-trained encoder. Compared to CMs [62] using the numerical continuous PF-ODE solver [28], LCMs [44] adopt the discrete-time schedule [41,42,61] to adapt to Stable Diffusion [55]. Instead of ensuring consistency between adjacent time steps " + }, + { + "bbox": [ + 130, + 390, + 482, + 511 + ], + "type": "inline_equation", + "content": "t_{n + 1}\\rightarrow t_n" + }, + { + "bbox": [ + 130, + 390, + 482, + 511 + ], + "type": "text", + "content": ", LCMs [44] are designed to ensure consistency between the current time step and " + }, + { + "bbox": [ + 130, + 390, + 482, + 511 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 390, + 482, + 511 + ], + "type": "text", + "content": "-step away, i.e., " + }, + { + "bbox": [ + 130, + 390, + 482, + 511 + ], + "type": "inline_equation", + "content": "t_{n + k}\\to t_n" + }, + { + "bbox": [ + 130, + 390, + 482, + 511 + ], + "type": "text", + "content": ", thereby significantly reducing convergence time costs. As classifier-free guidance (CFG) [21] plays a crucial role in synthesizing high-quality text-aligned images, LCMs integrate CFG into the distillation as follows," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 149, + 530, + 481, + 546 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 530, + 481, + 546 + ], + "spans": [ + { + "bbox": [ + 149, + 530, + 481, + 546 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {z}} _ {t _ {n}} ^ {\\boldsymbol {\\Phi}, w} \\gets \\mathbf {z} _ {t _ {n + k}} + (1 + w) \\boldsymbol {\\Phi} (\\mathbf {z} _ {t _ {n + k}}, t _ {n + k}, t _ {n}, \\mathbf {c}) - w \\boldsymbol {\\Phi} (\\mathbf {z} _ {t _ {n + k}}, t _ {n + k}, t _ {n}, \\emptyset), \\qquad (5)", + "image_path": "afa5ba9db57d9ca53141086426b4ffa2e81533ba79f274baef9785548978ef65.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 554, + 482, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 554, + 482, + 603 + ], + "spans": [ + { + "bbox": [ + 130, + 554, + 482, + 603 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 554, + 482, + 603 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 130, + 554, + 482, + 603 + ], + "type": "text", + "content": " denotes the CFG scale which is uniformly sampled from " + }, + { + "bbox": [ + 130, + 554, + 482, + 603 + ], + "type": "inline_equation", + "content": "[w_{\\min}, w_{\\max}]" + }, + { + "bbox": [ + 130, + 554, + 482, + 603 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 554, + 482, + 603 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 554, + 482, + 603 + ], + "type": "text", + "content": " is the skipping interval. To efficiently perform the above " + }, + { + "bbox": [ + 130, + 554, + 482, + 603 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 554, + 482, + 603 + ], + "type": "text", + "content": "-step guided distillation, LCMs augment the consistency function to " + }, + { + "bbox": [ + 130, + 554, + 482, + 603 + ], + "type": "inline_equation", + "content": "f: (\\mathbf{z}_t, t, w, \\mathbf{c}) \\longmapsto \\mathbf{z}_0" + }, + { + "bbox": [ + 130, + 554, + 482, + 603 + ], + "type": "text", + "content": ", which is also the form adopted by our MotionLCM." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 620, + 403, + 633 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 620, + 403, + 633 + ], + "spans": [ + { + "bbox": [ + 131, + 620, + 403, + 633 + ], + "type": "text", + "content": "3.2 MotionLCM: Motion Latent Consistency Model" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": "Motion compression into the latent space. Motivated by [44, 62], we propose MotionLCM (Motion Latent Consistency Model) to tackle the low-efficiency" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "W. Dai et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 403 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 403 + ], + "type": "text", + "content": "problem in motion diffusion models [65, 83], unleashing the potential of LCM in the motion generation task. Similar to MLD [9], our MotionLCM adopts a consistency model in the motion latent space. We choose MLD [9] as the underlying diffusion model to distill from. We aim to achieve few-step " + }, + { + "bbox": [ + 130, + 116, + 482, + 403 + ], + "type": "inline_equation", + "content": "(2\\sim 4)" + }, + { + "bbox": [ + 130, + 116, + 482, + 403 + ], + "type": "text", + "content": " and even one-step inference without compromising motion quality. In MLD, an autoencoder " + }, + { + "bbox": [ + 130, + 116, + 482, + 403 + ], + "type": "inline_equation", + "content": "(\\mathcal{E},\\mathcal{D})" + }, + { + "bbox": [ + 130, + 116, + 482, + 403 + ], + "type": "text", + "content": " is first trained to compress a high dimensional motion into a low dimensional latent vector " + }, + { + "bbox": [ + 130, + 116, + 482, + 403 + ], + "type": "inline_equation", + "content": "\\mathbf{z} = \\mathcal{E}(\\mathbf{x})" + }, + { + "bbox": [ + 130, + 116, + 482, + 403 + ], + "type": "text", + "content": ", which is then decoded to reconstruct the motion as " + }, + { + "bbox": [ + 130, + 116, + 482, + 403 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}} = \\mathcal{D}(\\mathbf{z})" + }, + { + "bbox": [ + 130, + 116, + 482, + 403 + ], + "type": "text", + "content": ". Training diffusion models in the motion latent space greatly reduces the computational resources compared to the vanilla diffusion models trained on raw motion sequences (i.e., motion space) and speeds up the inference process. Thus, we effectively leverage the motion latent space for consistency distillation. Motion latent consistency distillation. An overview of our motion latent consistency distillation is described in Fig. 4 (a). A raw motion sequence " + }, + { + "bbox": [ + 130, + 116, + 482, + 403 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0^{1:N} = \\{\\mathbf{x}^i\\}_{i=1}^N" + }, + { + "bbox": [ + 130, + 116, + 482, + 403 + ], + "type": "text", + "content": " is a sequence of human poses, where " + }, + { + "bbox": [ + 130, + 116, + 482, + 403 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 116, + 482, + 403 + ], + "type": "text", + "content": " is the number of frames. We follow [17] to use the redundant motion representation for our experiments, which is widely used in previous work [9, 65, 83]. As shown in Fig. 4 (a), given a raw motion sequence " + }, + { + "bbox": [ + 130, + 116, + 482, + 403 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0^{1:N}" + }, + { + "bbox": [ + 130, + 116, + 482, + 403 + ], + "type": "text", + "content": ", a pre-trained VAE [30] encoder first compresses it into the latent space, " + }, + { + "bbox": [ + 130, + 116, + 482, + 403 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_0 = \\mathcal{E}(\\mathbf{x}_0)" + }, + { + "bbox": [ + 130, + 116, + 482, + 403 + ], + "type": "text", + "content": ". Then, a forward diffusion operation with " + }, + { + "bbox": [ + 130, + 116, + 482, + 403 + ], + "type": "inline_equation", + "content": "n + k" + }, + { + "bbox": [ + 130, + 116, + 482, + 403 + ], + "type": "text", + "content": " steps is conducted to add noise on " + }, + { + "bbox": [ + 130, + 116, + 482, + 403 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_0" + }, + { + "bbox": [ + 130, + 116, + 482, + 403 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 116, + 482, + 403 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 116, + 482, + 403 + ], + "type": "text", + "content": " is the skipping interval illustrated in Sec. 3.1. The noisy " + }, + { + "bbox": [ + 130, + 116, + 482, + 403 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{n+k}" + }, + { + "bbox": [ + 130, + 116, + 482, + 403 + ], + "type": "text", + "content": " is fed to the frozen teacher network and trainable online network to predict the clean " + }, + { + "bbox": [ + 130, + 116, + 482, + 403 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{z}}_0^*" + }, + { + "bbox": [ + 130, + 116, + 482, + 403 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 116, + 482, + 403 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{z}}_0" + }, + { + "bbox": [ + 130, + 116, + 482, + 403 + ], + "type": "text", + "content": ". The target network uses the cleaner " + }, + { + "bbox": [ + 130, + 116, + 482, + 403 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{z}}_n" + }, + { + "bbox": [ + 130, + 116, + 482, + 403 + ], + "type": "text", + "content": " obtained by a " + }, + { + "bbox": [ + 130, + 116, + 482, + 403 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 116, + 482, + 403 + ], + "type": "text", + "content": "-step ODE solver " + }, + { + "bbox": [ + 130, + 116, + 482, + 403 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 130, + 116, + 482, + 403 + ], + "type": "text", + "content": ", such as DDIM [61] to predict the " + }, + { + "bbox": [ + 130, + 116, + 482, + 403 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{z}}_0^-" + }, + { + "bbox": [ + 130, + 116, + 482, + 403 + ], + "type": "text", + "content": ". Since the classifier-free guidance (CFG) [21] is essential for condition alignment in diffusion models [9, 55, 65], we integrate CFG into the distillation," + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 165, + 412, + 481, + 425 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 412, + 481, + 425 + ], + "spans": [ + { + "bbox": [ + 165, + 412, + 481, + 425 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {z}} _ {n} \\leftarrow \\mathbf {z} _ {n + k} + (1 + w) \\boldsymbol {\\Phi} \\left(\\mathbf {z} _ {n + k}, t _ {n + k}, t _ {n}, \\mathbf {c}\\right) - w \\boldsymbol {\\Phi} \\left(\\mathbf {z} _ {n + k}, t _ {n + k}, t _ {n}, \\emptyset\\right), \\tag {6}", + "image_path": "7fb5c7c614e5308ecec11fc4b1a5114625c11daec051018ecd82f705f379c6c5.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 433, + 482, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 433, + 482, + 469 + ], + "spans": [ + { + "bbox": [ + 130, + 433, + 482, + 469 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 433, + 482, + 469 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 130, + 433, + 482, + 469 + ], + "type": "text", + "content": " is the text condition and " + }, + { + "bbox": [ + 130, + 433, + 482, + 469 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 130, + 433, + 482, + 469 + ], + "type": "text", + "content": " denotes the guidance scale. To ensure the self-consistency property defined in Eq. (1), the latent consistency distillation loss " + }, + { + "bbox": [ + 130, + 433, + 482, + 469 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{LCD}}" + }, + { + "bbox": [ + 130, + 433, + 482, + 469 + ], + "type": "text", + "content": " is designed as follows," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 171, + 477, + 481, + 491 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 171, + 477, + 481, + 491 + ], + "spans": [ + { + "bbox": [ + 171, + 477, + 481, + 491 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {L C D}} (\\boldsymbol {\\Theta}, \\boldsymbol {\\Theta} ^ {-}) = \\mathbb {E} \\left[ d \\left(\\boldsymbol {f} _ {\\boldsymbol {\\Theta}} (\\mathbf {z} _ {n + k}, t _ {n + k}, w, \\mathbf {c}), \\boldsymbol {f} _ {\\boldsymbol {\\Theta} ^ {-}} (\\hat {\\mathbf {z}} _ {n}, t _ {n}, w, \\mathbf {c})\\right) \\right], \\tag {7}", + "image_path": "e27f744b94f34974810a631538638281fa063a82f17253de87c391395d540341.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 498, + 482, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 606 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 606 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 498, + 482, + 606 + ], + "type": "inline_equation", + "content": "d(\\cdot, \\cdot)" + }, + { + "bbox": [ + 130, + 498, + 482, + 606 + ], + "type": "text", + "content": " is a distance measuring function, such as L2 loss or Huber loss [25]. As discussed in Sec. 3.1, the target network " + }, + { + "bbox": [ + 130, + 498, + 482, + 606 + ], + "type": "inline_equation", + "content": "\\Theta^{-}" + }, + { + "bbox": [ + 130, + 498, + 482, + 606 + ], + "type": "text", + "content": " is updated with the exponential moving average (EMA) of the trainable parameters of the online network " + }, + { + "bbox": [ + 130, + 498, + 482, + 606 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 130, + 498, + 482, + 606 + ], + "type": "text", + "content": ". Here we define the teacher network " + }, + { + "bbox": [ + 130, + 498, + 482, + 606 + ], + "type": "inline_equation", + "content": "\\Theta^{*}" + }, + { + "bbox": [ + 130, + 498, + 482, + 606 + ], + "type": "text", + "content": " as the pre-trained motion latent diffusion model, i.e., MLD [9]. According to [44], the online network and target network are initialized with the parameters of the teacher network. During the inference phase, as shown in Fig. 5, our MotionLCM can generate high-quality motions with one-step sampling and achieve the fastest runtime (" + }, + { + "bbox": [ + 130, + 498, + 482, + 606 + ], + "type": "inline_equation", + "content": "\\sim" + }, + { + "bbox": [ + 130, + 498, + 482, + 606 + ], + "type": "text", + "content": "30ms per motion sequence) compared to other motion diffusion models [9,65]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 622, + 406, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 622, + 406, + 635 + ], + "spans": [ + { + "bbox": [ + 131, + 622, + 406, + 635 + ], + "type": "text", + "content": "3.3 Controllable Motion Generation in Latent Space" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "After addressing the low-efficiency issue in the motion latent diffusion model [9], we delve into another exploration of real-time motion control. Inspired by the" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 395, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 395, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 395, + 91, + 447, + 100 + ], + "type": "text", + "content": "MotionLCM" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 332 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 332 + ], + "type": "text", + "content": "great success of ControlNet [82] in controllable image generation [55], we introduce a motion ControlNet " + }, + { + "bbox": [ + 130, + 116, + 482, + 332 + ], + "type": "inline_equation", + "content": "\\Theta^a" + }, + { + "bbox": [ + 130, + 116, + 482, + 332 + ], + "type": "text", + "content": " in the latent space of MotionLCM and initialize the motion ControlNet with a trainable copy of MotionLCM. Specifically, each layer in the motion ControlNet is appended with a zero-initialized linear layer for eliminating random noise in the initial training steps. To achieve an autoregressive motion generation paradigm, as depicted in Fig. 1, we define the motion control task as generating motions given the initial " + }, + { + "bbox": [ + 130, + 116, + 482, + 332 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 130, + 116, + 482, + 332 + ], + "type": "text", + "content": " poses and textual description. As shown in Fig. 4 (b), the initial " + }, + { + "bbox": [ + 130, + 116, + 482, + 332 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 130, + 116, + 482, + 332 + ], + "type": "text", + "content": " poses are defined by the trajectories of " + }, + { + "bbox": [ + 130, + 116, + 482, + 332 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 116, + 482, + 332 + ], + "type": "text", + "content": " control joints, " + }, + { + "bbox": [ + 130, + 116, + 482, + 332 + ], + "type": "inline_equation", + "content": "\\mathbf{g}^{1:\\tau} = \\{\\mathbf{g}^i\\}_{i=1}^\\tau" + }, + { + "bbox": [ + 130, + 116, + 482, + 332 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 116, + 482, + 332 + ], + "type": "inline_equation", + "content": "\\mathbf{g}^i \\in \\mathbb{R}^{K \\times 3}" + }, + { + "bbox": [ + 130, + 116, + 482, + 332 + ], + "type": "text", + "content": " denotes the global absolute locations of each control joint. In our motion control pipeline, we design a Trajectory Encoder " + }, + { + "bbox": [ + 130, + 116, + 482, + 332 + ], + "type": "inline_equation", + "content": "\\Theta^b" + }, + { + "bbox": [ + 130, + 116, + 482, + 332 + ], + "type": "text", + "content": " consisting of stacked transformer [67] layers to encode the trajectory signals. We append a global token (i.e., [CLS]) before the start of the trajectory sequence as the output feature of the encoder, which is added to the noisy " + }, + { + "bbox": [ + 130, + 116, + 482, + 332 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_n" + }, + { + "bbox": [ + 130, + 116, + 482, + 332 + ], + "type": "text", + "content": " and fed into the trainable motion ControlNet " + }, + { + "bbox": [ + 130, + 116, + 482, + 332 + ], + "type": "inline_equation", + "content": "\\Theta^a" + }, + { + "bbox": [ + 130, + 116, + 482, + 332 + ], + "type": "text", + "content": ". Under the guidance of motion ControlNet, MotionLCM predicts the denoised " + }, + { + "bbox": [ + 130, + 116, + 482, + 332 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{z}}_0" + }, + { + "bbox": [ + 130, + 116, + 482, + 332 + ], + "type": "text", + "content": " through the consistency function " + }, + { + "bbox": [ + 130, + 116, + 482, + 332 + ], + "type": "inline_equation", + "content": "f_{\\Theta^s}" + }, + { + "bbox": [ + 130, + 116, + 482, + 332 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 116, + 482, + 332 + ], + "type": "inline_equation", + "content": "\\Theta^s" + }, + { + "bbox": [ + 130, + 116, + 482, + 332 + ], + "type": "text", + "content": " is the combination of " + }, + { + "bbox": [ + 130, + 116, + 482, + 332 + ], + "type": "inline_equation", + "content": "\\Theta^a" + }, + { + "bbox": [ + 130, + 116, + 482, + 332 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 116, + 482, + 332 + ], + "type": "inline_equation", + "content": "\\Theta^b" + }, + { + "bbox": [ + 130, + 116, + 482, + 332 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 482, + 332 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 130, + 116, + 482, + 332 + ], + "type": "text", + "content": ". The following reconstruction loss " + }, + { + "bbox": [ + 130, + 116, + 482, + 332 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{recon}}" + }, + { + "bbox": [ + 130, + 116, + 482, + 332 + ], + "type": "text", + "content": " optimizes the motion ControlNet " + }, + { + "bbox": [ + 130, + 116, + 482, + 332 + ], + "type": "inline_equation", + "content": "\\Theta^a" + }, + { + "bbox": [ + 130, + 116, + 482, + 332 + ], + "type": "text", + "content": " and Trajectory Encoder " + }, + { + "bbox": [ + 130, + 116, + 482, + 332 + ], + "type": "inline_equation", + "content": "\\Theta^b" + }, + { + "bbox": [ + 130, + 116, + 482, + 332 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 205, + 341, + 481, + 355 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 341, + 481, + 355 + ], + "spans": [ + { + "bbox": [ + 205, + 341, + 481, + 355 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {r e c o n}} \\left(\\boldsymbol {\\Theta} ^ {a}, \\boldsymbol {\\Theta} ^ {b}\\right) = \\mathbb {E} \\left[ d \\left(\\boldsymbol {f} _ {\\boldsymbol {\\Theta} ^ {s}} \\left(\\mathbf {z} _ {n}, t _ {n}, w, \\mathbf {c} ^ {*}\\right), \\mathbf {z} _ {0}\\right) \\right], \\tag {8}", + "image_path": "923360594998ab22947092a87c91b575c3552a5ca683d028f0beb49daa411e77.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 363, + 482, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 363, + 482, + 483 + ], + "spans": [ + { + "bbox": [ + 130, + 363, + 482, + 483 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 363, + 482, + 483 + ], + "type": "inline_equation", + "content": "\\mathbf{c}^*" + }, + { + "bbox": [ + 130, + 363, + 482, + 483 + ], + "type": "text", + "content": " includes the text condition and control guidance from the Trajectory Encoder and the motion ControlNet. However, during training, the sole reconstruction supervision in the latent space is insufficient. We argue this is because the controllable motion generation requires more detailed constraints, which cannot be effectively provided solely by the reconstruction loss in the latent space. Unlike previous methods like OmniControl [73], which directly diffuse in the motion space, allowing explicit supervision of control signals, effectively supervising control signals in the latent space is non-trivial. Therefore, we utilize the frozen VAE [30] decoder " + }, + { + "bbox": [ + 130, + 363, + 482, + 483 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 130, + 363, + 482, + 483 + ], + "type": "text", + "content": " to decode the latent " + }, + { + "bbox": [ + 130, + 363, + 482, + 483 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{z}}_0" + }, + { + "bbox": [ + 130, + 363, + 482, + 483 + ], + "type": "text", + "content": " into the motion space, obtaining the predicted motion " + }, + { + "bbox": [ + 130, + 363, + 482, + 483 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_0" + }, + { + "bbox": [ + 130, + 363, + 482, + 483 + ], + "type": "text", + "content": ", thereby introducing the control loss " + }, + { + "bbox": [ + 130, + 363, + 482, + 483 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{control}}" + }, + { + "bbox": [ + 130, + 363, + 482, + 483 + ], + "type": "text", + "content": " as follows," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 181, + 492, + 481, + 525 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 181, + 492, + 481, + 525 + ], + "spans": [ + { + "bbox": [ + 181, + 492, + 481, + 525 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {c o n t r o l}} (\\boldsymbol {\\Theta} ^ {a}, \\boldsymbol {\\Theta} ^ {b}) = \\mathbb {E} \\left[ \\frac {\\sum_ {i} \\sum_ {j} m _ {i j} | | R (\\hat {\\mathbf {x}} _ {0}) _ {i j} - R (\\mathbf {x} _ {0}) _ {i j} | | _ {2} ^ {2}}{\\sum_ {i} \\sum_ {j} m _ {i j}} \\right], \\tag {9}", + "image_path": "c8f357172577b0ac03f41d870526b4ba315b6922ff2be193e0ce5febde80446c.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 532, + 482, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 532, + 482, + 570 + ], + "spans": [ + { + "bbox": [ + 131, + 532, + 482, + 570 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 532, + 482, + 570 + ], + "type": "inline_equation", + "content": "R(\\cdot)" + }, + { + "bbox": [ + 131, + 532, + 482, + 570 + ], + "type": "text", + "content": " converts the joint's local positions to global absolute locations and " + }, + { + "bbox": [ + 131, + 532, + 482, + 570 + ], + "type": "inline_equation", + "content": "m_{ij} \\in \\{0,1\\}" + }, + { + "bbox": [ + 131, + 532, + 482, + 570 + ], + "type": "text", + "content": " is the binary joint mask at frame " + }, + { + "bbox": [ + 131, + 532, + 482, + 570 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 131, + 532, + 482, + 570 + ], + "type": "text", + "content": " for the joint " + }, + { + "bbox": [ + 131, + 532, + 482, + 570 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 131, + 532, + 482, + 570 + ], + "type": "text", + "content": ". Then we optimize the motion ControlNet " + }, + { + "bbox": [ + 131, + 532, + 482, + 570 + ], + "type": "inline_equation", + "content": "\\Theta^a" + }, + { + "bbox": [ + 131, + 532, + 482, + 570 + ], + "type": "text", + "content": " and Trajectory Encoder " + }, + { + "bbox": [ + 131, + 532, + 482, + 570 + ], + "type": "inline_equation", + "content": "\\Theta^b" + }, + { + "bbox": [ + 131, + 532, + 482, + 570 + ], + "type": "text", + "content": " with the overall objective," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 227, + 580, + 481, + 600 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 227, + 580, + 481, + 600 + ], + "spans": [ + { + "bbox": [ + 227, + 580, + 481, + 600 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\boldsymbol {\\Theta} ^ {a}, \\boldsymbol {\\Theta} ^ {b} = \\underset {\\boldsymbol {\\Theta} ^ {a}, \\boldsymbol {\\Theta} ^ {b}} {\\operatorname {a r g m i n}} (\\mathcal {L} _ {\\text {r e c o n}} + \\lambda \\mathcal {L} _ {\\text {c o n t r o l}}), \\end{array} \\tag {10}", + "image_path": "9ffc6313503d04cb920d452688e61abb974eed779c95b26a541ad0d33f8341e9.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": " is the weight to balance the two losses. This design enables explicit control signals to directly influence the generation process, similar to controlling other latent-free diffusion models for motion generation [29,73]. Extensive experiments demonstrate the introduced supervision is very helpful in improving the motion control performance, which will be introduced in the following section." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "W. Dai et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 230, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 230, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 230, + 129 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 139, + 482, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 139, + 482, + 188 + ], + "spans": [ + { + "bbox": [ + 130, + 139, + 482, + 188 + ], + "type": "text", + "content": "In this section, we first present the experimental setup details in Sec. 4.1. Subsequently, we provide quantitative and qualitative comparisons to evaluate the effectiveness of our proposed MotionLCM framework in Sec. 4.2 and Sec. 4.3. Finally, we conduct comprehensive ablation studies on MotionLCM in Sec. 4.4." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 204, + 261, + 217 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 204, + 261, + 217 + ], + "spans": [ + { + "bbox": [ + 132, + 204, + 261, + 217 + ], + "type": "text", + "content": "4.1 Experimental setup" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 222, + 482, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 222, + 482, + 282 + ], + "spans": [ + { + "bbox": [ + 130, + 222, + 482, + 282 + ], + "type": "text", + "content": "Datasets. We experiment on the popular HumanML3D [17] dataset, featuring 14,616 unique human motion sequences with 44,970 textual descriptions. For a fair comparison with previous methods [9,17,49,65,83], we take the redundant motion representation, including root velocity, root height, local joint positions, velocities, rotations in root space, and the foot contact binary labels." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 283, + 482, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 283, + 482, + 462 + ], + "spans": [ + { + "bbox": [ + 130, + 283, + 482, + 462 + ], + "type": "text", + "content": "Evaluation metrics. We extend the evaluation metrics of previous works [9, 17, 73]. (1) Time cost: We follow [9] to report the Average Inference Time per Sentence (AITS) to evaluate the inference efficiency of models. (2) Motion quality: Frechet Inception Distance (FID) is adopted as a principal metric to evaluate the feature distributions between the generated and real motions. The feature extractor employed is from [17]. (3) Motion diversity: MultiModality (MModality) measures the generation diversity conditioned on the same text and Diversity calculates variance through features [17]. (4) Condition matching: Following [17], we calculate the motion-retrieval precision (R-Precision) to report the text-motion Top-1/2/3 matching accuracy and Multimodal Distance (MM Dist) calculates the mean distance between motions and texts. (5) Control error: Trajectory error (Traj. err.) quantifies the ratio of unsuccessful trajectories, characterized by any control joint location error surpassing a predetermined threshold. Location error (Loc. err.) represents the unsuccessful joints. Average error (Avg. err.) denotes the mean location error of the control joints." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "type": "text", + "content": "Implementation details. Our baseline motion diffusion model is based on MLD [9]. We reproduce MLD with higher performance. Unless otherwise specified, all our experiments are conducted on this model. For MotionLCM, we employ the AdamW [40] optimizer for 96K iterations using a cosine decay learning rate scheduler and 1K iterations of linear warm-up. A batch size of 256 and a learning rate of 2e-4 are used. We set the training guidance scale range as " + }, + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "type": "inline_equation", + "content": "[w_{\\mathrm{min}}, w_{\\mathrm{max}}] = [5, 15]" + }, + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "type": "text", + "content": ", with the testing guidance scale set to 7.5, and adopt the EMA rate " + }, + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mu = 0.95" + }, + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "type": "text", + "content": " by default. We use the DDIM [61] solver with skipping interval " + }, + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "type": "inline_equation", + "content": "k = 20" + }, + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "type": "text", + "content": " and choose the Huber [25] loss as the distance measuring function " + }, + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "type": "text", + "content": ". For motion ControlNet, we use the AdamW [40] optimizer for 192K iterations with 1K iterations of linear warm-up. The batch size and learning rate are set to 128 and 1e-4. The learning rate scheduler is the same as the first stage. For the training objective, we employ " + }, + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "type": "text", + "content": " as the L2 loss and set the control loss weight " + }, + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "type": "text", + "content": " to 1.0 by default. We set the control ratio " + }, + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "type": "text", + "content": " as 0.25 and the number of control joints as " + }, + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "type": "inline_equation", + "content": "K = 6" + }, + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "type": "text", + "content": " (i.e., Pelvis, Left foot, Right foot, Head, Left wrist, and Right wrist) in both training and testing. We implement our model using PyTorch [47] with training on an NVIDIA RTX 4090 GPU and testing on a Tesla V100 GPU." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 395, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 395, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 395, + 91, + 447, + 100 + ], + "type": "text", + "content": "MotionLCM" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 136, + 190, + 480, + 322 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 181 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 181 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 181 + ], + "type": "text", + "content": "Table 1: Comparison of text-conditional motion synthesis on HumanML3D [17] dataset. We compute suggested metrics following [17]. We repeat the evaluation 20 times for each metric and report the average with a " + }, + { + "bbox": [ + 130, + 114, + 482, + 181 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 130, + 114, + 482, + 181 + ], + "type": "text", + "content": " confidence interval. “" + }, + { + "bbox": [ + 130, + 114, + 482, + 181 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 130, + 114, + 482, + 181 + ], + "type": "text", + "content": "” indicates that the closer to the real data, the better. **Bold and underline indicate the best and the second best result.** denotes the reproduced version of MLD [9]. The MotionLCM in one-step inference (30ms) surpasses all state-of-the-art models." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 136, + 190, + 480, + 322 + ], + "lines": [ + { + "bbox": [ + 136, + 190, + 480, + 322 + ], + "spans": [ + { + "bbox": [ + 136, + 190, + 480, + 322 + ], + "type": "table", + "html": "
MethodsAITS ↓R-Precision ↑FID ↓MM Dist ↓Diversity →MModality ↑
Top 1Top 2Top 3
Real-0.511±.0030.703±.0030.797±.0020.002±.0002.794±.0089.503±.065-
Seq2Seq [37]-0.180±.0020.300±.0020.396±.00211.75±.0355.529±.0076.223±.061-
JL2P [2]-0.246±.0010.387±.0020.486±.00211.02±.0465.296±.0087.676±.058-
T2G [5]-0.165±.0010.267±.0020.345±.0027.664±.0306.030±.0086.409±.071-
Hier [14]-0.301±.0020.425±.0020.552±.0046.532±.0245.012±.0188.332±.042-
TEMOS [49]0.0170.424±.0020.612±.0020.722±.0023.734±.0283.703±.0088.973±.0710.368±.018
T2M [17]0.0380.457±.0020.639±.0030.740±.0031.067±.0023.340±.0089.188±.0022.090±.083
MDM [65]24.740.320±.0050.498±.0040.611±.0070.544±.0445.566±.0279.559±.0862.799±.072
MotionDiffuse [83]14.740.491±.0010.681±.0010.782±.0010.630±.0013.113±.0019.410±.0491.553±.042
MLD [9]0.2170.481±.0030.673±.0030.772±.0020.473±.0133.196±.0109.724±.0822.413±.079
MLD* [9]0.2250.504±.0020.698±.0030.796±.0020.450±.0113.052±.0099.634±.0642.267±.082
MotionLCM (1-step)0.0300.502±.0030.701±.0020.803±.0020.467±.0123.022±.0099.631±.0662.172±.082
MotionLCM (2-step)0.0350.505±.0030.705±.0020.805±.0020.368±.0112.986±.0089.640±.0522.187±.094
MotionLCM (4-step)0.0430.502±.0030.698±.0020.798±.0020.304±.0123.012±.0079.607±.0662.259±.092
", + "image_path": "7234ee009fbf2e4a1f06398b5ee0af8c68305ba2dc32e2903ed87c696769f9a0.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 345, + 321, + 357 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 345, + 321, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 345, + 321, + 357 + ], + "type": "text", + "content": "4.2 Comparisons on Text-to-motion" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 364, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 364, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 364, + 482, + 605 + ], + "type": "text", + "content": "In the following part, we first evaluate our MotionLCM on the text-to-motion (T2M) task. We compare our method with some T2M baselines on HumanML3D [17] with suggested metrics [17] under the " + }, + { + "bbox": [ + 130, + 364, + 482, + 605 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 130, + 364, + 482, + 605 + ], + "type": "text", + "content": " confidence interval from 20 times running. As MotionLCM is based on MLD, we mainly focus on the performance compared with MLD. For evaluating time efficiency, we compare the Average Inference Time per Sentence (AITS) with TEMOS [49], T2M [17], MDM [65], MotionDiffuse [83] and MLD [9]. The results are borrowed from MLD [9]. The deterministic methods [2,5,14,37], are unable to produce diverse results from a single input text and thus we leave their MModality metrics empty. For the quantitative results, as shown in Tab. 1, our MotionLCM boasts an impressive real-time runtime efficiency, averaging around 30ms per motion sequence during inference. This performance exceeds that of previous diffusion-based methods [9,65,83] and even surpasses MLD [9] by an order of magnitude. Furthermore, despite employing only one-step inference, our MotionLCM can approximate or even surpass the performance of MLD [9] (DDIM [61] 50 steps). With two-step inference, we achieve the best R-Precision and MM Dist metrics, while increasing the sampling steps to four yields the best FID. The above results demonstrate the effectiveness of latent consistency distillation. For the qualitative results, as shown in Fig. 5, MotionLCM not only accelerates motion generation to real-time speed but also delivers high-quality outputs, closely aligning with the textual descriptions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 621, + 407, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 621, + 407, + 634 + ], + "spans": [ + { + "bbox": [ + 132, + 621, + 407, + 634 + ], + "type": "text", + "content": "4.3 Comparisons on Controllable Motion Generation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 640, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 640, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 640, + 481, + 666 + ], + "type": "text", + "content": "As shown in Tab. 2, we compare our MotionLCM with OmniControl [73] and MLD [9]. We observe that OmniControl struggles with multi-joint control and" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "W. Dai et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 134, + 114, + 480, + 335 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 480, + 335 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 480, + 335 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 480, + 335 + ], + "type": "table", + "html": "
RealMotionLCM (Ours)MLDMDM
“a person walks clockwise in a large curve while swinging their arms.”
0.033s0.201s22.58s
“a person walking like a bird and then sniffing the air.”
0.031s0.192s21.16s
“a person slightly bent over with right hand pressing against the air walks forward slowly”
0.031s0.197s23.44s
", + "image_path": "819822b0bcfce68a51cca71300f44024fff863c75f425e09031d019cd43091cb.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 354, + 480, + 387 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 354, + 480, + 387 + ], + "spans": [ + { + "bbox": [ + 132, + 354, + 480, + 387 + ], + "type": "text", + "content": "Fig. 5: Qualitative comparison of the state-of-the-art methods in the text-to-motion task. With only one-step inference, MotionLCM achieves the fastest motion generation while producing high-quality movements that closely match the textual descriptions." + } + ] + } + ], + "index": 3, + "type": "text" + }, + { + "bbox": [ + 130, + 411, + 482, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 411, + 482, + 579 + ], + "spans": [ + { + "bbox": [ + 130, + 411, + 482, + 579 + ], + "type": "text", + "content": "falls short in both generation quality and control performance compared to MotionLCM. To verify the effectiveness of the latent generated by our MotionLCM for training motion ControlNet, we conducted the following two sets of experiments: \"LC\" and \"MC\", which indicate introducing control supervision in the latent space and motion space. It can be observed that under the same experimental settings, MotionLCM maintains higher fidelity and significantly outperforms MLD [9] in motion control performance. This demonstrates that the latent generated by MotionLCM is more effective for training motion ControlNet compared to MLD [9]. In terms of inference speed, MotionLCM (1-step) is " + }, + { + "bbox": [ + 130, + 411, + 482, + 579 + ], + "type": "inline_equation", + "content": "1929 \\times" + }, + { + "bbox": [ + 130, + 411, + 482, + 579 + ], + "type": "text", + "content": " faster compared to OmniControl [73] and " + }, + { + "bbox": [ + 130, + 411, + 482, + 579 + ], + "type": "inline_equation", + "content": "13 \\times" + }, + { + "bbox": [ + 130, + 411, + 482, + 579 + ], + "type": "text", + "content": " faster than MLD [9]. For qualitative results, as shown in Fig. 6, OmniControl fails to control the initial poses in the second example and does not generate motion that aligns with the text in the third case. However, our MotionLCM not only adheres to the control of the initial poses but also generates motions that match the textual descriptions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 597, + 246, + 608 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 597, + 246, + 608 + ], + "spans": [ + { + "bbox": [ + 132, + 597, + 246, + 608 + ], + "type": "text", + "content": "4.4 Ablation Studies" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Impact of the hyperparameters of training MotionLCM. We conduct a comprehensive analysis of the training hyperparameters of MotionLCM, including the training guidance scale range " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "[w_{\\mathrm{min}}, w_{\\mathrm{max}}]" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": ", EMA rate " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": ", skipping interval " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": ", and the type of loss. We summarize the evaluation results based on one-step" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 395, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 395, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 395, + 91, + 447, + 100 + ], + "type": "text", + "content": "MotionLCM" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 137, + 168, + 480, + 274 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 160 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 160 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 160 + ], + "type": "text", + "content": "Table 2: Comparison of motion control on HumanML3D [17] dataset. Bold indicates the best result. Our MotionLCM outperforms OmniControl [73] and MLD [9] regarding generation quality, control performance, and inference speed. \"LC\" and \"MC\" refer to the control supervision introduced in the latent space and motion space." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 137, + 168, + 480, + 274 + ], + "lines": [ + { + "bbox": [ + 137, + 168, + 480, + 274 + ], + "spans": [ + { + "bbox": [ + 137, + 168, + 480, + 274 + ], + "type": "table", + "html": "
MethodsAITS ↓FID ↓R-Precision ↑ Top 3Diversity →Traj. err. ↓ (50cm)Loc. err. ↓ (50cm)Avg. err. ↓
Real-0.0020.7979.5030.00000.00000.0000
OmniControl [73]81.002.3280.5578.8670.33620.03220.0977
MLD [9] (LC)0.5520.4690.7239.4760.42300.06530.1690
MotionLCM (1-step, LC)0.0420.3190.7529.4240.29860.03440.1410
MotionLCM (2-step, LC)0.0470.3150.7709.4270.28400.03280.1365
MotionLCM (4-step, LC)0.0630.3280.7459.4410.29730.03390.1398
MLD [9] (LC&MC)0.5520.5550.7549.3730.27220.02150.1265
MotionLCM (1-step, LC&MC)0.0420.4190.7569.3900.19880.01470.1127
MotionLCM (2-step, LC&MC)0.0470.3970.7599.4690.19600.01430.1092
MotionLCM (4-step, LC&MC)0.0630.4440.7539.3550.20890.01720.1140
", + "image_path": "c613453f758194c1580f836fd9fa424344ec7cec9e015d32657cf5c138bbe779.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 155, + 335, + 462, + 491 + ], + "blocks": [ + { + "bbox": [ + 130, + 292, + 482, + 326 + ], + "lines": [ + { + "bbox": [ + 130, + 292, + 482, + 326 + ], + "spans": [ + { + "bbox": [ + 130, + 292, + 482, + 326 + ], + "type": "text", + "content": "Table 3: Ablation study on different training guidance scale ranges " + }, + { + "bbox": [ + 130, + 292, + 482, + 326 + ], + "type": "inline_equation", + "content": "[w_{\\mathrm{min}}, w_{\\mathrm{max}}]" + }, + { + "bbox": [ + 130, + 292, + 482, + 326 + ], + "type": "text", + "content": ", EMA rates " + }, + { + "bbox": [ + 130, + 292, + 482, + 326 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 130, + 292, + 482, + 326 + ], + "type": "text", + "content": ", skipping intervals " + }, + { + "bbox": [ + 130, + 292, + 482, + 326 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 292, + 482, + 326 + ], + "type": "text", + "content": " and types of loss. We use metrics in Tab. 1 and adopt a one-step inference setting with the testing CFG scale of 7.5 for fair comparison." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 155, + 335, + 462, + 491 + ], + "lines": [ + { + "bbox": [ + 155, + 335, + 462, + 491 + ], + "spans": [ + { + "bbox": [ + 155, + 335, + 462, + 491 + ], + "type": "table", + "html": "
MethodsR-Precision ↑ Top 1FID ↓MM Dist ↓Diversity →MModality ↑
Real0.511±.0030.002±.0002.794±.0089.503±.065-
MotionLCM (w ∈ [5,15])0.502±.0030.467±.0123.022±.0099.631±.0662.172±.082
MotionLCM (w ∈ [2,18])0.497±.0030.481±.0093.030±.0109.644±.0732.226±.091
MotionLCM (w = 7.5)0.486±.0020.479±.0093.094±.0099.610±.0722.320±.097
MotionLCM (μ = 0.95)0.502±.0030.467±.0123.022±.0099.631±.0662.172±.082
MotionLCM (μ = 0.50)0.498±.0030.478±.0093.022±.0109.655±.0712.188±.087
MotionLCM (μ = 0)0.499±.0030.505±.0083.018±.0099.706±.0702.123±.085
MotionLCM (k = 50)0.488±.0030.547±.0113.096±.0109.511±.0742.324±.091
MotionLCM (k = 20)0.502±.0030.467±.0123.022±.0099.631±.0662.172±.082
MotionLCM (k = 10)0.497±.0030.449±.0093.017±.0109.693±.0752.133±.086
MotionLCM (k = 5)0.488±.0030.438±.0093.044±.0099.647±.0742.147±.083
MotionLCM (k = 1)0.442±.0020.635±.0113.255±.0089.384±.0802.146±.075
MotionLCM (w/ Huber)0.502±.0030.467±.0123.022±.0099.631±.0662.172±.082
MotionLCM (w/ L2)0.486±.0020.622±.0103.114±.0099.573±.0692.218±.086
", + "image_path": "70c5612c722d3f3c932a9fc676fc0d237bac9c22496e1004a5c4f956e13d7bd4.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": "inference in Tab. 3. We find out that using a dynamic training guidance scale (e.g., " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "w \\in [5,15]" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": ") during training leads to an improvement in model performance compared to using a static training guidance scale (e.g., " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "w = 7.5" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": "). Additionally, an excessively large range for the training guidance scale can also negatively impact the performance of the model (e.g., " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "w \\in [2,18]" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": "). Regarding the EMA rate " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": ", we observe that the larger the value of " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": ", the better the performance of the model. This indicates that maintaining a slower update rate for the target network " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\Theta^{-}" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": " helps improve the performance of latent consistency distillation. When the skipping interval " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": " continues to increase, the performance of the distillation model improves progressively, but larger values of " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": " (e.g., " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "k = 50" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": ") may result in inferior results. As for the type of loss, the Huber loss [25] significantly outperforms the L2 loss, demonstrating its superior robustness." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "W. Dai et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 134, + 114, + 482, + 357 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 482, + 357 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 482, + 357 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 482, + 357 + ], + "type": "table", + "html": "
MotionLCM (w/o ControlNet)0.029s0.033s0.030s0.033s0.029s
OmniControl87.81s90.00s85.11s80.15s86.77s
MotionLCM (w/ ControlNet)0.043s0.045s0.041s0.045s0.047s
Reala man lifts his left arm up.the person is widewalking.a person walks three steps to his right, then five steps to his left, and finally three steps to his right.this person is walking in circles then moves across room to step around an object.the person is holding their head while walking.
", + "image_path": "4e31046014c5e58e0ee7f7dd7a4be938ed534f29a314ffc93a0ad01e23e8e3db.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 374, + 482, + 430 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 374, + 482, + 430 + ], + "spans": [ + { + "bbox": [ + 130, + 374, + 482, + 430 + ], + "type": "text", + "content": "Fig. 6: Qualitative comparison of the state-of-the-art methods in the motion control task. We provide the visualized motion results and real references from five prompts. Compared to OmniControl [73], MotionLCM with ControlNet not only generates the initial poses that accurately follow the given multi-joint trajectories (i.e., the green poses in real references) but also produces motions that closely align with the texts." + } + ] + } + ], + "index": 3, + "type": "text" + }, + { + "bbox": [ + 130, + 459, + 482, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 459, + 482, + 579 + ], + "spans": [ + { + "bbox": [ + 130, + 459, + 482, + 579 + ], + "type": "text", + "content": "Impact of control loss weights " + }, + { + "bbox": [ + 130, + 459, + 482, + 579 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 459, + 482, + 579 + ], + "type": "text", + "content": ". To verify the impact of different control loss weights " + }, + { + "bbox": [ + 130, + 459, + 482, + 579 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 459, + 482, + 579 + ], + "type": "text", + "content": " on the control performance of MotionLCM, we report the experimental results in Tab. 4. We also include experiments of MotionLCM without ControlNet (i.e., only text-to-motion) for comparison. We found a significant improvement in control-related metrics (e.g., Loc. err.) after introducing motion ControlNet (i.e., " + }, + { + "bbox": [ + 130, + 459, + 482, + 579 + ], + "type": "inline_equation", + "content": "\\lambda = 0" + }, + { + "bbox": [ + 130, + 459, + 482, + 579 + ], + "type": "text", + "content": "). Furthermore, control performance can be further improved by introducing control loss (i.e., " + }, + { + "bbox": [ + 130, + 459, + 482, + 579 + ], + "type": "inline_equation", + "content": "\\lambda > 0" + }, + { + "bbox": [ + 130, + 459, + 482, + 579 + ], + "type": "text", + "content": "). Increasing the weight " + }, + { + "bbox": [ + 130, + 459, + 482, + 579 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 459, + 482, + 579 + ], + "type": "text", + "content": " enhances control performance but leads to a decline in the generation quality, which is reflected in higher FID scores. To balance these two aspects, we adopt " + }, + { + "bbox": [ + 130, + 459, + 482, + 579 + ], + "type": "inline_equation", + "content": "\\lambda = 1" + }, + { + "bbox": [ + 130, + 459, + 482, + 579 + ], + "type": "text", + "content": " as our default setting for training motion ControlNet." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "Impact of different control ratios " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": " and number of control joints " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": ". In Tab. 5, we present the results of all models with the testing control ratio as 0.25 and keep the number of control joints " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": " equal in both training and testing. We found that the model with the fixed training control ratio (i.e., " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\tau = 0.25" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": ") performs better compared to a dynamic ratio (e.g., " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\tau \\in [0.1, 0.5]" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "), and we discover that our model maintains good performance when incorporating additional redundant control signals, such as whole-body joints with " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "K = 22" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 395, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 395, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 395, + 91, + 447, + 100 + ], + "type": "text", + "content": "MotionLCM" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 150, + 146, + 466, + 296 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "type": "text", + "content": "Table 4: Ablation study on different control loss weights " + }, + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "type": "text", + "content": ". We present the results of (1, 2, 4)-step inference. We add the MotionLCM without ControlNet for comparison." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 150, + 146, + 466, + 296 + ], + "lines": [ + { + "bbox": [ + 150, + 146, + 466, + 296 + ], + "spans": [ + { + "bbox": [ + 150, + 146, + 466, + 296 + ], + "type": "table", + "html": "
MethodsFID ↓ Top 3R-Precision ↑ Diversity →Traj. err. ↓ (50cm)Loc. err. ↓ (50cm)Avg. err. ↓
Real0.0020.7979.5030.00000.00000.0000
MotionLCM (1-step, w/o control)0.4670.8039.6310.76050.23020.3493
MotionLCM (2-step, w/o control)0.3680.8059.6400.76460.22140.3386
MotionLCM (4-step, w/o control)0.3040.7989.6070.77390.22070.3359
MotionLCM (1-step, λ = 0)0.3190.7529.4240.29860.03440.1410
MotionLCM (2-step, λ = 0)0.3150.7709.4270.28400.03280.1365
MotionLCM (4-step, λ = 0)0.3280.7459.4410.29730.03390.1398
MotionLCM (1-step, λ = 0.1)0.3440.7539.3860.27110.02750.1310
MotionLCM (2-step, λ = 0.1)0.3240.7599.4280.26310.02560.1268
MotionLCM (4-step, λ = 0.1)0.3570.7439.4150.27130.02680.1309
MotionLCM (1-step, λ = 1.0)0.4190.7569.3900.19880.01470.1127
MotionLCM (2-step, λ = 1.0)0.3970.7599.4690.19600.01430.1092
MotionLCM (4-step, λ = 1.0)0.4440.7539.3550.20890.01720.1140
MotionLCM (1-step, λ = 10.0)0.6360.7449.4790.14650.00970.0967
MotionLCM (2-step, λ = 10.0)0.5510.7579.5690.15900.01070.0987
MotionLCM (4-step, λ = 10.0)0.5680.7429.4860.17230.01320.1045
", + "image_path": "1118fbac6341c6c1766093e7dab5ae2f25d029a84a3ac14825edfd4185d85cd6.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 150, + 346, + 466, + 484 + ], + "blocks": [ + { + "bbox": [ + 132, + 313, + 480, + 336 + ], + "lines": [ + { + "bbox": [ + 132, + 313, + 480, + 336 + ], + "spans": [ + { + "bbox": [ + 132, + 313, + 480, + 336 + ], + "type": "text", + "content": "Table 5: Ablation study on different control ratios " + }, + { + "bbox": [ + 132, + 313, + 480, + 336 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 132, + 313, + 480, + 336 + ], + "type": "text", + "content": " and number of control joints " + }, + { + "bbox": [ + 132, + 313, + 480, + 336 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 132, + 313, + 480, + 336 + ], + "type": "text", + "content": " . We report the results of (1,2,4)-step inference. \" " + }, + { + "bbox": [ + 132, + 313, + 480, + 336 + ], + "type": "inline_equation", + "content": "*" + }, + { + "bbox": [ + 132, + 313, + 480, + 336 + ], + "type": "text", + "content": " \" denotes the default training setting." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 150, + 346, + 466, + 484 + ], + "lines": [ + { + "bbox": [ + 150, + 346, + 466, + 484 + ], + "spans": [ + { + "bbox": [ + 150, + 346, + 466, + 484 + ], + "type": "table", + "html": "
MethodsFID ↓R-Precision ↑ Top 3Diversity →Traj. err. ↓ (50cm)Loc. err. ↓ (50cm)Avg. err. ↓
Real0.0020.7979.5030.00000.00000.0000
MotionLCM* (1-step, τ = 0.25, K = 6)0.4190.7569.3900.19880.01470.1127
MotionLCM* (2-step, τ = 0.25, K = 6)0.3970.7599.4690.19600.01430.1092
MotionLCM* (4-step, τ = 0.25, K = 6)0.4440.7539.3550.20890.01720.1140
MotionLCM (1-step, τ ∈ [0.1, 0.25])0.4560.7579.4770.28210.02340.1214
MotionLCM (2-step, τ ∈ [0.1, 0.25])0.4090.7699.5920.27070.02300.1179
MotionLCM (4-step, τ ∈ [0.1, 0.25])0.4570.7579.5400.29280.02560.1228
MotionLCM (1-step, τ ∈ [0.1, 0.5])0.4480.7639.5380.23900.01820.1182
MotionLCM (2-step, τ ∈ [0.1, 0.5])0.4130.7689.5170.23490.01800.1153
MotionLCM (4-step, τ ∈ [0.1, 0.5])0.4460.7539.4980.25170.01990.1196
MotionLCM (1-step, K = 12)0.4120.7539.4120.20720.01100.1029
MotionLCM (2-step, K = 12)0.4100.7589.5090.19790.01080.1000
MotionLCM (4-step, K = 12)0.4420.7559.3800.21690.01320.1048
MotionLCM (1-step, K = 22(whole-body))0.4360.7489.3790.21430.00830.0914
MotionLCM (2-step, K = 22(whole-body))0.4130.7589.4920.20610.00820.0881
MotionLCM (4-step, K = 22(whole-body))0.4610.7459.4590.21730.00970.0918
", + "image_path": "26ea8229db741eeea9335aa71179766413b138d14beb55975be5ebf95740c450.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 513, + 220, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 513, + 220, + 525 + ], + "spans": [ + { + "bbox": [ + 133, + 513, + 220, + 525 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 557, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 557, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 557, + 481, + 665 + ], + "type": "text", + "content": "This work proposes an efficient controllable motion generation framework, MotionLCM. By introducing latent consistency distillation, MotionLCM enjoys the trade-off between runtime efficiency and generation quality. Moreover, thanks to the motion ControlNet manipulation in the latent space, our method obtains excellent controlling ability with given conditions. Extensive experimental results show the effectiveness of the proposed method. As the VAE of MLD lacks explicit temporal modeling, the MotionLCM cannot achieve a good temporal explanation. Therefore, our future work will lie in developing a more explainable compression architecture for efficient motion control." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "W. Dai et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 199 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 199 + ], + "type": "text", + "content": "The research is supported by Shenzhen Ubiquitous Data Enabling Key Lab under grant ZDSYS20220527171406015 and CCF-Tencent Rhino-Bird Open Research Fund. This project is also supported by Shanghai Artificial Intelligence Laboratory. The author team would like to acknowledge Yiming Xie, Zhiyang Dou, and Shunlin Lu for their helpful technical discussions and suggestions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 217, + 197, + 229 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 217, + 197, + 229 + ], + "spans": [ + { + "bbox": [ + 133, + 217, + 197, + 229 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 240, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 138, + 240, + 480, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 240, + 480, + 262 + ], + "spans": [ + { + "bbox": [ + 138, + 240, + 480, + 262 + ], + "type": "text", + "content": "1. Ahn, H., Ha, T., Choi, Y., Yoo, H., Oh, S.: Text2action: Generative adversarial synthesis from language to action. In: ICRA. pp. 5915-5920 (2018)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 263, + 480, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 263, + 480, + 284 + ], + "spans": [ + { + "bbox": [ + 138, + 263, + 480, + 284 + ], + "type": "text", + "content": "2. Ahuja, C., Morency, L.P.: Language2pose: Natural language grounded pose forecasting. In: 3DV. pp. 719-728 (2019)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 285, + 480, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 285, + 480, + 306 + ], + "spans": [ + { + "bbox": [ + 138, + 285, + 480, + 306 + ], + "type": "text", + "content": "3. Athanasiou, N., Petrovich, M., Black, M.J., Varol, G.: Teach: Temporal action composition for 3d humans. In: 3DV. pp. 414-423 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 308, + 480, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 308, + 480, + 327 + ], + "spans": [ + { + "bbox": [ + 138, + 308, + 480, + 327 + ], + "type": "text", + "content": "4. Barquero, G., Escalera, S., Palmero, C.: Seamless human motion composition with blended positional encodings. In: CVPR. pp. 457-469 (2024)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 329, + 480, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 329, + 480, + 360 + ], + "spans": [ + { + "bbox": [ + 138, + 329, + 480, + 360 + ], + "type": "text", + "content": "5. Bhattacharya, U., Rewkowski, N., Banerjee, A., Guhan, P., Bera, A., Manocha, D.: Text2gestures: A transformer-based network for generating emotive body gestures for virtual agents. In: VR. pp. 1-10 (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 361, + 480, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 361, + 480, + 381 + ], + "spans": [ + { + "bbox": [ + 138, + 361, + 480, + 381 + ], + "type": "text", + "content": "6. Cervantes, P., Sekikawa, Y., Sato, I., Shinoda, K.: Implicit neural representations for variable length human motion generation. In: ECCV. pp. 356-372 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 383, + 480, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 383, + 480, + 415 + ], + "spans": [ + { + "bbox": [ + 138, + 383, + 480, + 415 + ], + "type": "text", + "content": "7. Chen, L.H., Lu, S., Zeng, A., Zhang, H., Wang, B., Zhang, R., Zhang, L.: Motionllm: Understanding human behaviors from human motions and videos. arXiv preprint arXiv:2405.20340 (2024)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 416, + 480, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 416, + 480, + 436 + ], + "spans": [ + { + "bbox": [ + 138, + 416, + 480, + 436 + ], + "type": "text", + "content": "8. Chen, L.H., Zhang, J., Li, Y., Pang, Y., Xia, X., Liu, T.: Humanmac: Masked motion completion for human motion prediction. In: ICCV. pp. 9544-9555 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 437, + 480, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 437, + 480, + 458 + ], + "spans": [ + { + "bbox": [ + 138, + 437, + 480, + 458 + ], + "type": "text", + "content": "9. Chen, X., Jiang, B., Liu, W., Huang, Z., Fu, B., Chen, T., Yu, G.: Executing your commands via motion diffusion in latent space. In: CVPR. pp. 18000-18010 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 460, + 480, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 460, + 480, + 491 + ], + "spans": [ + { + "bbox": [ + 138, + 460, + 480, + 491 + ], + "type": "text", + "content": "0. Cong, P., Dou, Z.W., Ren, Y., Yin, W., Cheng, K., Sun, Y., Long, X., Zhu, X., Ma, Y.: Laserhuman: Language-guided scene-aware human motion generation in free environment. arXiv preprint arXiv:2403.13307 (2024)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 492, + 480, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 492, + 480, + 512 + ], + "spans": [ + { + "bbox": [ + 138, + 492, + 480, + 512 + ], + "type": "text", + "content": "1. Dabral, R., Mughal, M.H., Golyanik, V., Theobalt, C.: Mofusion: A framework for denoising-diffusion-based motion synthesis. In: CVPR. pp. 9760-9770 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 514, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 514, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 138, + 514, + 480, + 544 + ], + "type": "text", + "content": "2. Dou, Z., Chen, X., Fan, Q., Komura, T., Wang, W.: C-ase: Learning conditional adversarial skill embeddings for physics-based characters. In: SIGGRAPH Asia. pp. 1-11 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 546, + 480, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 546, + 480, + 578 + ], + "spans": [ + { + "bbox": [ + 138, + 546, + 480, + 578 + ], + "type": "text", + "content": "3. Fan, K., Tang, J., Cao, W., Yi, R., Li, M., Gong, J., Zhang, J., Wang, Y., Wang, C., Ma, L.: Freemotion: A unified framework for number-free text-to-motion synthesis. arXiv preprint arXiv:2405.15763 (2024)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 579, + 480, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 579, + 480, + 599 + ], + "spans": [ + { + "bbox": [ + 138, + 579, + 480, + 599 + ], + "type": "text", + "content": "4. Ghosh, A., Cheema, N., Oguz, C., Theobalt, C., Slusallek, P.: Synthesis of compositional animations from textual descriptions. In: ICCV. pp. 1396-1406 (2021)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 601, + 480, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 601, + 480, + 620 + ], + "spans": [ + { + "bbox": [ + 138, + 601, + 480, + 620 + ], + "type": "text", + "content": "5. Guo, C., Mu, Y., Javed, M.G., Wang, S., Cheng, L.: Momask: Generative masked modeling of 3d human motions. In: CVPR. pp. 1900-1910 (2024)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 623, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 623, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 138, + 623, + 480, + 643 + ], + "type": "text", + "content": "6. Guo, C., Mu, Y., Zuo, X., Dai, P., Yan, Y., Lu, J., Cheng, L.: Generative human motion stylization in latent space. arXiv preprint arXiv:2401.13505 (2024)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 138, + 644, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 644, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 644, + 480, + 665 + ], + "type": "text", + "content": "7. Guo, C., Zou, S., Zuo, X., Wang, S., Ji, W., Li, X., Cheng, L.: Generating diverse and natural 3d human motions from text. In: CVPR. pp. 5152-5161 (2022)" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 395, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 395, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 395, + 91, + 447, + 100 + ], + "type": "text", + "content": "MotionLCM" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "text", + "content": "18. Guo, C., Zuo, X., Wang, S., Cheng, L.: Tm2t: Stochastic and tokenized modeling for the reciprocal generation of 3d human motions and texts. In: ECCV. pp. 580-597 (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 481, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 481, + 183 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 481, + 183 + ], + "type": "text", + "content": "19. Guo, C., Zuo, X., Wang, S., Zou, S., Sun, Q., Deng, A., Gong, M., Cheng, L.: Action2motion: Conditioned generation of 3d human motions. In: ACMMM. pp. 2021-2029 (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 184, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 184, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 184, + 480, + 205 + ], + "type": "text", + "content": "20. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. NeurIPS pp. 6840-6851 (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 206, + 480, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 206, + 480, + 228 + ], + "spans": [ + { + "bbox": [ + 132, + 206, + 480, + 228 + ], + "type": "text", + "content": "21. Ho, J., Salimans, T.: Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 228, + 480, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 228, + 480, + 251 + ], + "spans": [ + { + "bbox": [ + 132, + 228, + 480, + 251 + ], + "type": "text", + "content": "22. Holden, D., Komura, T., Saito, J.: Phase-functional neural networks for character control. TOG 36(4), 1-13 (2017)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 251, + 480, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 251, + 480, + 274 + ], + "spans": [ + { + "bbox": [ + 132, + 251, + 480, + 274 + ], + "type": "text", + "content": "23. Holden, D., Saito, J., Komura, T.: A deep learning framework for character motion synthesis and editing. TOG 35(4), 1-11 (2016)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 274, + 480, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 274, + 480, + 306 + ], + "spans": [ + { + "bbox": [ + 132, + 274, + 480, + 306 + ], + "type": "text", + "content": "24. Huang, Y., Yang, H., Luo, C., Wang, Y., Xu, S., Zhang, Z., Zhang, M., Peng, J.: Stablemofusion: Towards robust and efficient diffusion-based motion generation framework. arXiv preprint arXiv:2405.05691 (2024)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 307, + 480, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 307, + 480, + 329 + ], + "spans": [ + { + "bbox": [ + 132, + 307, + 480, + 329 + ], + "type": "text", + "content": "25. Huber, P.J.: Robust estimation of a location parameter. The Annals of Mathematical Statistics 35(1), 73-101 (1964)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 330, + 480, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 330, + 480, + 361 + ], + "spans": [ + { + "bbox": [ + 132, + 330, + 480, + 361 + ], + "type": "text", + "content": "26. Ji, Y., Xu, F., Yang, Y., Shen, F., Shen, H.T., Zheng, W.S.: A large-scale rgb-d database for arbitrary-view human action recognition. In: ACMM. pp. 1510-1518 (2018)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 363, + 480, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 363, + 480, + 384 + ], + "spans": [ + { + "bbox": [ + 132, + 363, + 480, + 384 + ], + "type": "text", + "content": "27. Jiang, B., Chen, X., Liu, W., Yu, J., Yu, G., Chen, T.: Motiongpt: Human motion as a foreign language. NeurIPS (2024)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 385, + 480, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 385, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 385, + 480, + 407 + ], + "type": "text", + "content": "28. Karras, T., Aittala, M., Aila, T., Laine, S.: Elucidating the design space of diffusion-based generative models. NeurIPS pp. 26565-26577 (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 408, + 480, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 408, + 480, + 430 + ], + "spans": [ + { + "bbox": [ + 132, + 408, + 480, + 430 + ], + "type": "text", + "content": "29. Karunratanakul, K., Preechakul, K., Suwajanakorn, S., Tang, S.: Guided motion diffusion for controllable human motion synthesis. In: CVPR. pp. 2151-2162 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 430, + 480, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 430, + 480, + 453 + ], + "spans": [ + { + "bbox": [ + 132, + 430, + 480, + 453 + ], + "type": "text", + "content": "30. Kingma, D.P., Welling, M.: Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114 (2013)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 453, + 480, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 453, + 480, + 475 + ], + "spans": [ + { + "bbox": [ + 132, + 453, + 480, + 475 + ], + "type": "text", + "content": "31. Lee, T., Moon, G., Lee, K.M.: Multiact: Long-term 3d human motion generation from multiple action labels. In: AAAI. pp. 1231-1239 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 475, + 480, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 475, + 480, + 498 + ], + "spans": [ + { + "bbox": [ + 132, + 475, + 480, + 498 + ], + "type": "text", + "content": "32. Li, B., Zhao, Y., Zhelun, S., Sheng, L.: Danceformer: Music conditioned 3d dance generation with parametric motion transformer. In: AAAI. pp. 1272-1279 (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 498, + 480, + 530 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 498, + 480, + 530 + ], + "spans": [ + { + "bbox": [ + 132, + 498, + 480, + 530 + ], + "type": "text", + "content": "33. Li, R., Zhang, Y., Zhang, Y., Zhang, H., Guo, J., Zhang, Y., Liu, Y., Li, X.: Lodge: A coarse to fine diffusion network for long dance generation guided by the characteristic dance primitives. In: CVPR. pp. 1524-1534 (2024)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 531, + 480, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 531, + 480, + 563 + ], + "spans": [ + { + "bbox": [ + 132, + 531, + 480, + 563 + ], + "type": "text", + "content": "34. Li, R., Zhao, J., Zhang, Y., Su, M., Ren, Z., Zhang, H., Tang, Y., Li, X.: Finedance: A fine-grained choreography dataset for 3d full body dance generation. In: ICCV. pp. 10234-10243 (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 564, + 480, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 564, + 480, + 587 + ], + "spans": [ + { + "bbox": [ + 132, + 564, + 480, + 587 + ], + "type": "text", + "content": "35. Li, R., Yang, S., Ross, D.A., Kanazawa, A.: Ai choreographer: Music conditioned 3d dance generation with aist++. In: ICCV. pp. 13401-13412 (2021)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 588, + 480, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 588, + 480, + 609 + ], + "spans": [ + { + "bbox": [ + 132, + 588, + 480, + 609 + ], + "type": "text", + "content": "36. Li, T., Qiao, C., Ren, G., Yin, K., Ha, S.: Aamdm: Accelerated auto-regressive motion diffusion model. In: CVPR. pp. 1813-1823 (2024)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 610, + 480, + 642 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 610, + 480, + 642 + ], + "spans": [ + { + "bbox": [ + 132, + 610, + 480, + 642 + ], + "type": "text", + "content": "37. Lin, A.S., Wu, L., Corona, R., Tai, K., Huang, Q., Mooney, R.J.: Generating animated videos of human activities from natural language descriptions. Learning 1(2018), 1 (2018)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "text", + "content": "38. Lin, X., Amer, M.R.: Human motion modeling using dvgens. arXiv preprint arXiv:1804.10652 (2018)" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "W. Dai et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 666 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "39. Liu, J., Dai, W., Wang, C., Cheng, Y., Tang, Y., Tong, X.: Plan, posture and go: Towards open-world text-to-motion generation. arXiv preprint arXiv:2312.14828 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 481, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 481, + 172 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 481, + 172 + ], + "type": "text", + "content": "40. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101 (2017)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 481, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 481, + 206 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 481, + 206 + ], + "type": "text", + "content": "41. Lu, C., Zhou, Y., Bao, F., Chen, J., Li, C., Zhu, J.: Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. NeurIPS pp. 5775-5787 (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 206, + 481, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 206, + 481, + 239 + ], + "spans": [ + { + "bbox": [ + 132, + 206, + 481, + 239 + ], + "type": "text", + "content": "42. Lu, C., Zhou, Y., Bao, F., Chen, J., Li, C., Zhu, J.: Dpm-solver++: Fast solver for guided sampling of diffusion probabilistic models. arXiv preprint arXiv:2211.01095 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 239, + 481, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 239, + 481, + 274 + ], + "spans": [ + { + "bbox": [ + 132, + 239, + 481, + 274 + ], + "type": "text", + "content": "43. Lu, S., Chen, L.H., Zeng, A., Lin, J., Zhang, R., Zhang, L., Shum, H.Y.: Humantomato: Text-aligned whole-body motion generation. arXiv preprint arXiv:2310.12978 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 274, + 481, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 274, + 481, + 306 + ], + "spans": [ + { + "bbox": [ + 132, + 274, + 481, + 306 + ], + "type": "text", + "content": "44. Luo, S., Tan, Y., Huang, L., Li, J., Zhao, H.: Latent consistency models: Synthesizing high-resolution images with few-step inference. arXiv preprint arXiv:2310.04378 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 306, + 481, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 306, + 481, + 329 + ], + "spans": [ + { + "bbox": [ + 132, + 306, + 481, + 329 + ], + "type": "text", + "content": "45. Mahmood, N., Ghorbani, N., Troje, N.F., Pons-Moll, G., Black, M.J.: Amass: Archive of motion capture as surface shapes. In: ICCV. pp. 5442-5451 (2019)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 329, + 481, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 329, + 481, + 351 + ], + "spans": [ + { + "bbox": [ + 132, + 329, + 481, + 351 + ], + "type": "text", + "content": "46. Nichol, A.Q., Dhariwal, P.: Improved denoising diffusion probabilistic models. In: ICML. pp. 8162-8171 (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 351, + 481, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 351, + 481, + 385 + ], + "spans": [ + { + "bbox": [ + 132, + 351, + 481, + 385 + ], + "type": "text", + "content": "47. Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., et al.: Pytorch: An imperative style, high-performance deep learning library. NeurIPS (2019)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 385, + 481, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 385, + 481, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 385, + 481, + 407 + ], + "type": "text", + "content": "48. Petrovich, M., Black, M.J., Varol, G.: Action-conditioned 3d human motion synthesis with transformer vae. In: ICCV. pp. 10985-10995 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 407, + 481, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 407, + 481, + 430 + ], + "spans": [ + { + "bbox": [ + 132, + 407, + 481, + 430 + ], + "type": "text", + "content": "49. Petrovich, M., Black, M.J., Varol, G.: Temos: Generating diverse human motions from textual descriptions. In: ECCV. pp. 480-497 (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 430, + 481, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 430, + 481, + 453 + ], + "spans": [ + { + "bbox": [ + 132, + 430, + 481, + 453 + ], + "type": "text", + "content": "50. Petrovich, M., Black, M.J., Varol, G.: Tmr: Text-to-motion retrieval using contrastive 3d human motion synthesis. In: ICCV. pp. 9488-9497 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 453, + 481, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 453, + 481, + 486 + ], + "spans": [ + { + "bbox": [ + 132, + 453, + 481, + 486 + ], + "type": "text", + "content": "51. Petrovich, M., Litany, O., Iqbal, U., Black, M.J., Varol, G., Bin Peng, X., Rempe, D.: Multi-track timeline control for text-driven 3d human motion generation. In: CVPRW. pp. 1911-1921 (2024)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 486, + 481, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 486, + 481, + 509 + ], + "spans": [ + { + "bbox": [ + 132, + 486, + 481, + 509 + ], + "type": "text", + "content": "52. Plappert, M., Mandery, C., Asfour, T.: The kit motion-language dataset. Big data 4(4), 236-252 (2016)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 509, + 481, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 509, + 481, + 542 + ], + "spans": [ + { + "bbox": [ + 132, + 509, + 481, + 542 + ], + "type": "text", + "content": "53. Punnakkal, A.R., Chandrasekaran, A., Athanasiou, N., Quiros-Ramirez, A., Black, M.J.: Babel: Bodies, action and behavior with english labels. In: CVPR. pp. 722-731 (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 542, + 481, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 542, + 481, + 575 + ], + "spans": [ + { + "bbox": [ + 132, + 542, + 481, + 575 + ], + "type": "text", + "content": "54. Raab, S., Leibovitch, I., Li, P., Aberman, K., Sorkine-Hornung, O., Cohen-Or, D.: Modi: Unconditional motion synthesis from diverse data. In: CVPR. pp. 13873-13883 (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 575, + 481, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 575, + 481, + 597 + ], + "spans": [ + { + "bbox": [ + 132, + 575, + 481, + 597 + ], + "type": "text", + "content": "55. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR. pp. 10684-10695 (2022)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 597, + 481, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 597, + 481, + 620 + ], + "spans": [ + { + "bbox": [ + 132, + 597, + 481, + 620 + ], + "type": "text", + "content": "56. Shafir, Y., Tevet, G., Kapon, R., Bermano, A.H.: Human motion diffusion as a generative prior. In: ICLR (2024)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 620, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 620, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 620, + 481, + 643 + ], + "type": "text", + "content": "57. Shahroudy, A., Liu, J., Ng, T.T., Wang, G.: Ntu " + }, + { + "bbox": [ + 132, + 620, + 481, + 643 + ], + "type": "inline_equation", + "content": "\\mathrm{rgb + }" + }, + { + "bbox": [ + 132, + 620, + 481, + 643 + ], + "type": "text", + "content": " d: A large scale dataset for 3d human activity analysis. In: CVPR. pp. 1010-1019 (2016)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 643, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 481, + 666 + ], + "type": "text", + "content": "58. Shi, Y., Wang, J., Jiang, X., Dai, B.: Controllable motion diffusion model. arXiv preprint arXiv:2306.00416 (2023)" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 395, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 395, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 395, + 91, + 447, + 100 + ], + "type": "text", + "content": "MotionLCM" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 666 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "text", + "content": "59. Siyao, L., Yu, W., Gu, T., Lin, C., Wang, Q., Qian, C., Loy, C.C., Liu, Z.: Bailando: 3d dance generation by actor-critic gpt with choreographic memory. In: CVPR. pp. 11050-11059 (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 480, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 480, + 183 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 480, + 183 + ], + "type": "text", + "content": "60. Sohl-Dickstein, J., Weiss, E., Maheswaranathan, N., Ganguli, S.: Deep unsupervised learning using nonequilibrium thermodynamics. In: ICML. pp. 2256-2265 (2015)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 480, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 480, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 480, + 194 + ], + "type": "text", + "content": "61. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. In: ICLR (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 194, + 480, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 480, + 215 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 480, + 215 + ], + "type": "text", + "content": "62. Song, Y., Dhariwal, P., Chen, M., Sutskever, I.: Consistency models. In: ICML (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 216, + 480, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 480, + 247 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 480, + 247 + ], + "type": "text", + "content": "63. Tang, Y., Liu, J., Liu, A., Yang, B., Dai, W., Rao, Y., Lu, J., Zhou, J., Li, X.: Flag3d: A 3d fitness activity dataset with language instruction. In: CVPR. pp. 22106-22117 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 248, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 248, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 248, + 480, + 270 + ], + "type": "text", + "content": "64. Tevet, G., Gordon, B., Hertz, A., Bermano, A.H., Cohen-Or, D.: Motionclip: Exposing human motion generation to clip space. In: ECCV. pp. 358-374 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 271, + 480, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 480, + 293 + ], + "type": "text", + "content": "65. Tevet, G., Raab, S., Gordon, B., Shafir, Y., Cohen-Or, D., Bermano, A.H.: Human motion diffusion model. In: ICLR (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 293, + 480, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 480, + 314 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 480, + 314 + ], + "type": "text", + "content": "66. Tseng, J., Castellon, R., Liu, K.: Edge: Editable dance generation from music. In: CVPR. pp. 448-458 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 315, + 480, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 480, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 480, + 335 + ], + "type": "text", + "content": "67. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. NeurIPS (2017)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 336, + 480, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 480, + 369 + ], + "type": "text", + "content": "68. Wan, W., Dou, Z., Komura, T., Wang, W., Jayaraman, D., Liu, L.: Tlcontrol: Trajectory and language control for human motion synthesis. arXiv preprint arXiv:2311.17135 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 369, + 480, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 480, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 480, + 402 + ], + "type": "text", + "content": "69. Wang, Z., Chen, Y., Jia, B., Li, P., Zhang, J., Zhang, J., Liu, T., Zhu, Y., Liang, W., Huang, S.: Move as you say interact as you can: Language-guided human motion generation with scene affordance. In: CVPR. pp. 433-444 (2024)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 403, + 480, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 480, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 480, + 435 + ], + "type": "text", + "content": "70. Wang, Z., Chen, Y., Liu, T., Zhu, Y., Liang, W., Huang, S.: Humanise: Language-conditioned human motion generation in 3d scenes. NeurIPS pp. 14959-14971 (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 435, + 480, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 480, + 456 + ], + "type": "text", + "content": "71. Wang, Z., Wang, J., Lin, D., Dai, B.: Intercontrol: Generate human motion interactions by controlling every joint. arXiv preprint arXiv:2311.15864 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 457, + 480, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 480, + 479 + ], + "type": "text", + "content": "72. Xiao, Z., Wang, T., Wang, J., Cao, J., Zhang, W., Dai, B., Lin, D., Pang, J.: Unified human-scene interaction via prompted chain-of-contacts. In: ICLR (2024)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 479, + 480, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 480, + 501 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 480, + 501 + ], + "type": "text", + "content": "73. Xie, Y., Jampani, V., Zhong, L., Sun, D., Jiang, H.: Omnicontrol: Control any joint at any time for human motion generation. In: ICLR (2024)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 501, + 480, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 480, + 533 + ], + "type": "text", + "content": "74. Xu, L., Lv, X., Yan, Y., Jin, X., Wu, S., Xu, C., Liu, Y., Zhou, Y., Rao, F., Sheng, X., et al.: Inter-x: Towards versatile human-human interaction analysis. In: CVPR. pp. 22260-22271 (2024)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 534, + 480, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 480, + 566 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 480, + 566 + ], + "type": "text", + "content": "75. Xu, L., Song, Z., Wang, D., Su, J., Fang, Z., Ding, C., Gan, W., Yan, Y., Jin, X., Yang, X., et al.: Actformer: A gan-based transformer towards general action-conditioned 3d human motion generation. In: ICCV. pp. 2228-2238 (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 567, + 480, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 480, + 588 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 480, + 588 + ], + "type": "text", + "content": "76. Xu, L., Zhou, Y., Yan, Y., Jin, X., Zhu, W., Rao, F., Yang, X., Zeng, W.: Regennet: Towards human action-reaction synthesis. In: CVPR. pp. 1759-1769 (2024)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 589, + 480, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 589, + 480, + 610 + ], + "spans": [ + { + "bbox": [ + 132, + 589, + 480, + 610 + ], + "type": "text", + "content": "77. Yan, S., Li, Z., Xiong, Y., Yan, H., Lin, D.: Convolutional sequence generation for skeleton-based action synthesis. In: ICCV. pp. 4394-4402 (2019)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 611, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 480, + 632 + ], + "type": "text", + "content": "78. Yuan, Y., Song, J., Iqbal, U., Vahdat, A., Kautz, J.: Physdiff: Physics-guided human motion diffusion model. In: ICCV. pp. 16010-16021 (2023)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 632, + 480, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 666 + ], + "type": "text", + "content": "79. Zhang, B., Cheng, Y., Wang, C., Zhang, T., Yang, J., Tang, Y., Zhao, F., Chen, D., Guo, B.: Rodinhd: High-fidelity 3d avatar generation with diffusion models. arXiv preprint arXiv:2407.06938 (2024)" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "W. Dai et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 315 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "80. Zhang, B., Cheng, Y., Yang, J., Wang, C., Zhao, F., Tang, Y., Chen, D., Guo, B.: Gaussiancube: Structuring gaussian splatting using optimal transport for 3d generative modeling. arXiv preprint arXiv:2403.19655 (2024)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "type": "text", + "content": "81. Zhang, J., Zhang, Y., Cun, X., Zhang, Y., Zhao, H., Lu, H., Shen, X., Shan, Y.: Generating human motion from textual descriptions with discrete representations. In: CVPR. pp. 14730-14740 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 183, + 482, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 183, + 482, + 205 + ], + "spans": [ + { + "bbox": [ + 130, + 183, + 482, + 205 + ], + "type": "text", + "content": "82. Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models. In: ICCV. pp. 3836-3847 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 205, + 482, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 205, + 482, + 237 + ], + "spans": [ + { + "bbox": [ + 130, + 205, + 482, + 237 + ], + "type": "text", + "content": "83. Zhang, M., Cai, Z., Pan, L., Hong, F., Guo, X., Yang, L., Liu, Z.: Motiondiffuse: Text-driven human motion generation with diffusion model. arXiv preprint arXiv:2208.15001 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 237, + 482, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 237, + 482, + 259 + ], + "spans": [ + { + "bbox": [ + 130, + 237, + 482, + 259 + ], + "type": "text", + "content": "84. Zhao, R., Su, H., Ji, Q.: Bayesian adversarial human motion synthesis. In: CVPR. pp. 6225-6234 (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 259, + 482, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 259, + 482, + 281 + ], + "spans": [ + { + "bbox": [ + 130, + 259, + 482, + 281 + ], + "type": "text", + "content": "85. Zhong, L., Xie, Y., Jampani, V., Sun, D., Jiang, H.: Smoodi: Stylized motion diffusion model. arXiv preprint arXiv:2407.12783 (2024)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 281, + 482, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 281, + 482, + 315 + ], + "spans": [ + { + "bbox": [ + 130, + 281, + 482, + 315 + ], + "type": "text", + "content": "86. Zhou, W., Dou, Z., Cao, Z., Liao, Z., Wang, J., Wang, W., Liu, Y., Komura, T., Wang, W., Liu, L.: Emdm: Efficient motion diffusion model for fast, high-quality motion generation. arXiv preprint arXiv:2312.02256 (2023)" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 395, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 395, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 395, + 91, + 447, + 100 + ], + "type": "text", + "content": "MotionLCM" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Multi-Granularity Sparse Relationship Matrix Prediction Network for End-to-End Scene Graph Generation/6b85535a-3ad0-4a54-882a-f4d4244fbf81_content_list.json b/2024/Multi-Granularity Sparse Relationship Matrix Prediction Network for End-to-End Scene Graph Generation/6b85535a-3ad0-4a54-882a-f4d4244fbf81_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..1617a2a7b151f5a2cbb4fde09e1327d6367c2b3a --- /dev/null +++ b/2024/Multi-Granularity Sparse Relationship Matrix Prediction Network for End-to-End Scene Graph Generation/6b85535a-3ad0-4a54-882a-f4d4244fbf81_content_list.json @@ -0,0 +1,1921 @@ +[ + { + "type": "text", + "text": "Multi-Granularity Sparse Relationship Matrix Prediction Network for End-to-End Scene Graph Generation", + "text_level": 1, + "bbox": [ + 222, + 140, + 781, + 205 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Lei Wang, Zejian Yuan, and Badong Chen", + "bbox": [ + 328, + 234, + 671, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Institute of Artificial Intelligence and Robotics, Xi'an Jiaotong University, Xi'an 710049, China \nleiwangmail@stu.xjtu.edu.cn, {yuan.ze.jian, chenbd}@mail.xjtu.edu.cn", + "bbox": [ + 235, + 262, + 764, + 305 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Current end-to-end Scene Graph Generation (SGG) relies solely on visual representations to separately detect sparse relations and entities in an image. This leads to the issue where the predictions of entities do not contribute to the prediction of relations, necessitating post-processing to assign corresponding subjects and objects to the predicted relations. In this paper, we introduce a sparse relationship matrix that bridges entity detection and relation detection. Our approach not only eliminates the need for relation matching, but also leverages the semantics and positional information of predicted entities to enhance relation prediction. Specifically, a multi-granularity sparse relationship matrix prediction network is proposed, which utilizes three gated pooling modules focusing on filtering negative samples at different granularities, thereby obtaining a sparse relationship matrix containing entity pairs most likely to form relations. Finally, a set of sparse, most probable subject-object pairs can be constructed and used for relation decoding. Experimental results on multiple datasets demonstrate that our method achieves a new state-of-the-art overall performance. Our code is available at https://github.com/wanglei0618/Mg-RMPN.", + "bbox": [ + 259, + 340, + 743, + 592 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Scene Graph Generation $\\cdot$ End-to-End $\\cdot$ Sparse Relationship Matrix $\\cdot$ Multi-Granularity", + "bbox": [ + 259, + 604, + 740, + 632 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 657, + 375, + 672 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Scene Graph Generation (SGG) is a fundamental visual comprehension task that captures semantic information by detecting relation triplets $<$ subject entity, predicate, object entity $>$ in an image. This structured representation can facilitate many downstream tasks, such as image captioning [4, 41], visual question answering [9, 28], image retrieval [11, 25] and image generation [10, 19]. The end-to-end SGG methods [6, 16, 30] can directly generate sparse relations from an image based on a fixed number of relation queries, avoiding dense prediction that contains a lot of background relations in two-stage SGG. This leads to faster inference speed, and because fewer background relations are predicted, it exhibits superior predictive performance for rare samples of tail classes.", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ffba968c47daf7af70f546ad00059648ad59b6a4167a7122278183114fb04901.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 272, + 143, + 424, + 310 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/25b231acb2938b3714ad51294734b301f2375ed1f1dc452e95e18b3c647361ed.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 433, + 143, + 594, + 311 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/3ee8640a5287cfecdb9ad07fc0368e89580760db82751dbadde379c03aa80f58.jpg", + "image_caption": [ + "(c)", + "Fig. 1: (a) End-to-end SGG pipeline. (b) Our proposed end-to-end pipeline based on multi-granularity sparse relationship matrix prediction network. (c) The average number of entities and relations in the VG150, OIV6, and GQA200 datasets demonstrates the sparsity of relations in an image." + ], + "image_footnote": [], + "bbox": [ + 602, + 154, + 728, + 315 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Current end-to-end approaches [6,16,22] still lag behind two-stage methods in performance, which is attributed to their reliance solely on visual representations to separately detect entities and relations, as shown in Figure 1 (a). This leads to two drawbacks: (1) It requires post-processing to match the corresponding subjects and objects for the relations, and the model's performance is constrained by the capabilities of matching. (2) Due to the inability to determine the subject and object in advance, entity predictions cannot help with relation prediction, and relation prediction can only rely on visual features.", + "bbox": [ + 212, + 428, + 787, + 549 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we introduce a sparse relationship matrix that bridges entity detection and relation detection. Our approach not only eliminates the need for relation matching, but also leverages the semantics and positional information of predicted entities to enhance relation prediction. Figure 1 (b) shows our pipeline based on the \"pair then relation\" [30] framework, which first constructs a relationship matrix based on entity predictions to represent the possibility of forming a relation between two entities and then selects the sparsest, most likely subject-object pairs for relation prediction.", + "bbox": [ + 212, + 551, + 787, + 671 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Compared to all possible pairs formed by the entity proposals, the subject-object pairs with relations are extremely sparse. Figure 1 (c) illustrates the average entities and relations in the datasets. Taking VG150 [13] as an example, when generating $N_{e} = 100$ entities, there are a total of $(100(100 - 1)) = 9900$ possible combinations. But the average relations per image is only 5.8, which means that more than 99. $9\\%$ pairs are negative samples. Therefore, directly learning this extremely sparse relationship matrix is very challenging.", + "bbox": [ + 212, + 672, + 787, + 777 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we propose a Multi-granularity sparse Relationship Matrix Prediction Network (Mg-RMPN) to partition the negative samples into different granularities, achieving layer-by-layer filtering of dense negative samples. Specifically, Mg-RMPN employs three Gated Pooling Modules (GPM) with identical", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "L. Wang et al.", + "bbox": [ + 271, + 114, + 369, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "structures but independent parameters, each of which focuses on filtering negative samples at different granularities. The GPMs are used to generate relation matrices at different granularities and construct the corresponding ground truth matrices for supervised learning. The experimental results from various datasets reveal that our approach not only accurately identifies numerous head-class samples, but also excels in predicting sparse tail-class samples, achieving state-of-the-art overall performance.", + "bbox": [ + 212, + 146, + 787, + 252 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The contributions are summarized as follows: (1) An end-to-end SGG framework based on the relationship matrix is proposed that bridges entity detection and relation detection. (2) An Mg-RMPN is proposed to achieve sparse relationship matrix prediction based on multi-granularity negative sample learning. (3) Experimental results have confirmed that our method achieves state-of-the-art overall performance in both end-to-end and two-stage methods.", + "bbox": [ + 212, + 252, + 787, + 343 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Works", + "text_level": 1, + "bbox": [ + 214, + 361, + 397, + 378 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Scene Graph Generation", + "text_level": 1, + "bbox": [ + 214, + 388, + 467, + 402 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Early work [5, 27, 32, 36] in SGG is based on the two-stage pipeline, first employing a Faster-RCNN [24] object detector to generate entity proposals, and the class of each entity is predicted. Then, entities are paired to form relations, and all possible pairs are classified. Its drawback lies that dense relation pairs include numerous background relations, which dilute the sparse tail-class samples, thus aggravating the imbalance issue. Therefore, some recent efforts [8, 23] have focused on addressing the imbalanced predictions, developing a series of rebalancing strategies such as logit adjustment [2, 26], loss reweighting [18, 33, 35], and data resampling [17, 31, 37]. Unlike these methods, we introduce a relationship matrix to generate sparse subject-object pairs for relation prediction. The sparse relations generated in this way have a higher signal-to-noise ratio, which can alleviate the imbalance problem. The results of the experiments validate that our method maintains outstanding performance for head classes and also demonstrates high performance for tail classes.", + "bbox": [ + 212, + 409, + 787, + 621 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2 End-to-End Scene Graph Generation", + "text_level": 1, + "bbox": [ + 214, + 638, + 568, + 652 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "End-to-end SGG, inspired by DETR [1], generates sparse relations directly from an image based on queries. For instance, [6] introduced the Relation Transformer to directly generate a set of relations from visual features. However, the generated relations did not match the corresponding subjects and objects, hence [16] proposed a graph assembling module to address the relation matching. [30] further introduced a \"pair then relation\" framework that predetermines subject-object pairs, circumventing the relation matching. However, these approaches suffer from the issue of relying solely on visual representations to detect relations and entities separately. Therefore, this paper introduces Mg-RMPN to predict a sparse relationship matrix to link entity detection and relation detection, not only eliminating the need for relation matching but also utilizing the predicted entity semantics and positional information to assist in relation prediction.", + "bbox": [ + 212, + 657, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Mg-RMPN", + "bbox": [ + 653, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 143, + 330, + 160 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Problem Definition", + "text_level": 1, + "bbox": [ + 215, + 176, + 421, + 191 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "SGG aims to generate a scene graph $\\mathcal{G} = \\{\\mathcal{E},\\mathcal{R}\\}$ from an input image $\\mathcal{I}$ . The scene graph consists of a set of $n$ entities $\\mathcal{E} = \\{e_i\\}_{i=1}^n$ and a set of $m$ relations $\\mathcal{R} = \\{r_k\\}_{k=1}^m$ between entities. The set of entities $\\mathcal{E}$ can be further decomposed into a set of bounding boxes $\\mathcal{B} = \\{\\mathbf{b}_i\\}_{i=1}^n$ and a set of class labels $\\mathcal{C} = \\{c_i\\}_{i=1}^n$ . The generation of a scene graph $\\mathcal{G}$ can be formulated as the joint probability distribution:", + "bbox": [ + 212, + 202, + 787, + 291 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\Pr (\\mathcal {G} | \\mathcal {I}) = \\Pr (\\mathcal {B}, \\mathcal {C} | \\mathcal {I}) \\Pr (\\mathcal {R} | \\mathcal {I}, \\mathcal {B}, \\mathcal {C}), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 377, + 294, + 784, + 309 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\operatorname{Pr}(\\mathcal{B},\\mathcal{C}|\\mathcal{I})$ represents the entity representation obtained from the object detector, including the class labels and bounding boxes. $\\operatorname{Pr}(\\mathcal{R}|\\mathcal{I},\\mathcal{B},\\mathcal{C})$ represents the relation prediction based on pairs of entities by the relation decoder. The process from $\\operatorname{Pr}(\\mathcal{B},\\mathcal{C}|\\mathcal{I})$ to $\\operatorname{Pr}(\\mathcal{R}|\\mathcal{I},\\mathcal{B},\\mathcal{C})$ requires our Mg-RMPN to predict the pairs of entities most likely to form a relation from all possible pairs of entities.", + "bbox": [ + 212, + 316, + 787, + 393 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2 Overall Architecture", + "text_level": 1, + "bbox": [ + 215, + 415, + 433, + 429 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown in Figure 2 (a), our method comprises three modules: (1) entity detection, (2) relationship matrix prediction, and (3) relation prediction. The entity detection is responsible for generating a set of entities based on the object detector, including their visual representations, class labels, and bounding boxes. The relationship matrix prediction is responsible for producing a sparse relationship matrix that represents the relevance between two entities, and then obtaining the indices of the subject and object entities that are most likely to form a relation from all possible combinations. Finally, the relation of these sparse subject-object pairs is predicted through the relation decoder.", + "bbox": [ + 212, + 440, + 787, + 578 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3 Object Detector", + "text_level": 1, + "bbox": [ + 215, + 599, + 397, + 614 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This paper employs a deformable DETR [42] as the object detector, which contains a transformer encoder-decoder architecture with a set of entity queries. Here, the entity queries interact with encoder features through multi-scale deformable attention. Given an image $\\mathcal{I}$ , the object detector produces $N_{e}$ entity representations $\\mathbf{Q}_e = \\{\\mathbf{q}_i^e\\}_{i=1}^{N_e} \\in \\mathbb{R}^{N_e \\times d}$ from a set of learnable entity queries, where $d$ is the embedding dimensions. Based on $\\mathbf{Q}_e$ , the object detector then follows with a classification head and a regression head to predict the entity's class predictions $\\mathbf{C} = \\{\\hat{\\mathbf{c}}_i\\}_{i=1}^n \\in \\mathbb{R}^{N_e \\times C_e}$ and bounding boxes $\\mathbf{B} = \\{\\hat{\\mathbf{b}}_i\\}_{i=1}^n \\in \\mathbb{R}^{N_e \\times 4}$ , respectively, where $C_e$ is the number of entity classes.", + "bbox": [ + 212, + 625, + 787, + 762 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4 Multi-Granularity Relationship Matrix Prediction Network", + "text_level": 1, + "bbox": [ + 215, + 782, + 751, + 799 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given $N_{e}$ entities, Multi-granularity sparse Relationship Matrix Prediction Network (Mg-RMPN) generates a $N_{e} \\times N_{e}$ adjacency relationship matrix $\\mathbf{M}_r$ . Each", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "L. Wang et al.", + "bbox": [ + 271, + 114, + 369, + 128 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/bca1af44e3398e6d5ccd55d8f755be506a7c50a8e8f153e4d74f93c8d4a9011e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 230, + 143, + 769, + 253 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/6b05dda4c16792948784c67ff2c8ed53787ef133e697bc5080416e3c097918fc.jpg", + "image_caption": [ + "Fig. 2: (a) Overall pipeline of our end-to-end SGG method. (b) Multi-granularity sparse Relationship Matrix Prediction Network (Mg-RMPN). GPM is Gated Pooling Module." + ], + "image_footnote": [], + "bbox": [ + 236, + 255, + 733, + 424 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "node represents the probability of forming a relationship between two entities. Due to the sparse entities and relations in an image, $\\mathbf{M}_r$ contains a large number of pairs without relations, that is, negative samples.", + "bbox": [ + 212, + 491, + 782, + 536 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Therefore, Mg-RMPN defined three $\\mathbf{M}_r$ based on multi-granularity negative samples, where the three $\\mathbf{M}_r$ are learned through three structurally identical but parameter-independent Gated Pooling Modules (GPM), as shown in Figure 2 (b). With the outputs of the GPMs, Mg-RMPN is able to generate the final sparse relationship matrix, enabling the identification of most probable subject-object pairs to predict relations.", + "bbox": [ + 212, + 537, + 784, + 627 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Gated Pooling Module Unlike [30], which only uses visual features, our GPM determines the correlation of two entities based on their visual representation, linguistic prior and location information. Given a set of visual representations $\\mathbf{Q}_e$ of the entities in an image, GPM first uses two fully connected networks $\\mathcal{F}_{sub}$ and $\\mathcal{F}_{obj}$ to project it to the visual representations of the subject and object, $\\mathbf{Q}_{sub}$ and $\\mathbf{Q}_{obj}$ , respectively, as follows:", + "bbox": [ + 214, + 648, + 782, + 742 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Q} _ {s u b} = \\mathcal {F} _ {s u b} \\left(\\mathbf {Q} _ {e}\\right) = \\left\\{\\mathbf {q} _ {i} ^ {s u b} \\right\\} _ {i = 1} ^ {N _ {e}}, \\quad \\mathbf {Q} _ {o b j} = \\mathcal {F} _ {o b j} \\left(\\mathbf {Q} _ {e}\\right) = \\left\\{\\mathbf {q} _ {j} ^ {o b j} \\right\\} _ {j = 1} ^ {N _ {e}}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 272, + 750, + 784, + 771 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Subsequently, the visual similarity between the subject $i$ and the object $j$ can be calculated as $v_{ij}^{sim} = \\mathbf{q}_i^{sub} \\cdot (\\mathbf{q}_j^{obj})^{\\mathrm{T}}$ .", + "bbox": [ + 212, + 779, + 782, + 810 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given a set of class prediction distributions $\\mathbf{C}$ for entities, GPM uses the prior linguistic function $\\mathcal{F}_{prior}$ from [26] to predict the correlation between two", + "bbox": [ + 212, + 810, + 784, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Mg-RMPN", + "bbox": [ + 653, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "entities, $s_{ij}^{prior} = \\mathcal{F}_{prior}(\\hat{\\mathbf{c}}_i,\\hat{\\mathbf{c}}_j)$ . $\\mathcal{F}_{prior}$ aims to capture the linguistic inductive bias that exists between two entities to enhance the prediction of the relational matrices. Similarly to [26], $\\mathcal{F}_{prior}$ obtains a reasonable initialization based on the distribution of relations generated by entity pairs in the training dataset.", + "bbox": [ + 212, + 145, + 782, + 205 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Based on the positions $\\mathbf{B}$ of the entities, GPM uses the function $\\mathcal{F}_{\\text{overlap}}$ to obtain the relative position information between two entities $\\mathbf{b}_{ij}^{\\text{overlap}} = \\mathcal{F}_{\\text{overlap}}(\\hat{\\mathbf{b}}_i, \\hat{\\mathbf{b}}_j) \\in \\mathbb{R}^8$ , where 8 elements indicate difference in $x$ and $y$ coordinates, width ratio, height ratio, IOU, center distance, area ratio, and aspect ratio. Additionally, we can also obtain the union bounding box $\\hat{\\mathbf{b}}_{ij}^u \\in \\mathbb{R}^9$ for the entity pairs, where 9 elements indicate the union bounding box coordinates $(x_1, y_1, x_2, y_2)$ , center $(\\frac{x_1 + x_2}{2}, \\frac{y_1 + y_2}{2})$ , size $(x_2 - x_1, y_2 - y_1)$ and area $(x_2 - x_1)(y_2 - y_1)$ , respectively. Then, we can obtain the relevance representation of the subject $i$ and object $j$ as follows:", + "bbox": [ + 212, + 207, + 784, + 349 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {e} _ {i j} = \\left[ v _ {i j} ^ {\\text {s i m}}, s _ {i j} ^ {\\text {p r i o r}}, \\mathbf {b} _ {i j} ^ {\\text {o v e r l a p}}, \\hat {\\mathbf {b}} _ {i j} ^ {u} \\right], \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 388, + 359, + 784, + 380 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $[\\cdot, \\cdot]$ denotes the concatenation operation. Subsequently, the feature representation of the entity pair can be expressed as", + "bbox": [ + 214, + 388, + 782, + 420 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {f} _ {i j} = [ \\mathbf {q} _ {i} ^ {s u b}, \\mathbf {q} _ {j} ^ {o b j}, f _ {s} (s _ {i j} ^ {p r i o r}), p o s (\\mathbf {b} _ {i j} ^ {o v e r l a p}), p o s (\\hat {\\mathbf {b}} _ {i j} ^ {u}) ], \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 321, + 431, + 784, + 452 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Where $f_{s}$ is a fully-connected layer for language prior encoding, and $pos$ is a fully-connected layer for positional encoding. Then, GPM uses a gated pooling layer $f_{gate}$ to obtain the probability of relation between subject $i$ and object $j$ as follows:", + "bbox": [ + 214, + 460, + 782, + 518 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {p} _ {i j} = \\operatorname {S i g m o i d} \\left(\\sum f _ {\\text {g a t e}} \\left(\\mathbf {f} _ {i j}\\right) \\odot \\mathbf {e} _ {i j}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 377, + 518, + 784, + 540 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\odot$ represents element-wise multiplication. Finally, we can obtain a $N_{e} \\times N_{e}$ relationship matrix $\\mathbf{M}_{r} = \\mathrm{GPM}(\\mathbf{Q}_{e},\\mathbf{C},\\mathbf{B})$ . Note that we set the diagonal elements of $\\mathbf{M}_{r}$ to zero, removing the cases where an entity serves as both subject and object simultaneously.", + "bbox": [ + 214, + 546, + 782, + 604 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Multi-Granularity Sparse Relationship Matrix Prediction Based on the output of three GPMs, we can construct three relationship matrices based on the granularity of negative samples and build the corresponding ground truth matrices (see Section 3.6) for supervised learning, enabling each GPM to focus on identifying negative samples of specific granularity.", + "bbox": [ + 212, + 628, + 782, + 703 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Since the $N_{e}$ entity proposals are typically much greater than the number of entities contained in an image, resulting in a large number of background entities among all possible pairs of entities. Since background entities do not form relationships, we divide the negative samples in the relationship matrix into three levels of granularity: (1) high-confidence negative samples composed of background entities (background-background), (2) medium-confidence negative samples constructed from background and foreground entities (entity-background or background-entity), and (3) low-confidence negative samples composed of foreground entities (entity-entity).", + "bbox": [ + 212, + 704, + 784, + 839 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "L. Wang et al.", + "bbox": [ + 271, + 114, + 369, + 128 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Then we can define three relationship matrices based on the output of the GPMs as follows.", + "bbox": [ + 215, + 146, + 782, + 174 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {M} _ {r 1} = \\operatorname {G P M - 1} \\left(\\mathbf {Q} _ {e}, \\mathbf {C}, \\mathbf {B}\\right)\n$$\n", + "text_format": "latex", + "bbox": [ + 393, + 188, + 560, + 200 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {M} _ {r 2} = \\mathbf {M} _ {r 1} + \\operatorname {G P M - 2} \\left(\\mathbf {Q} _ {e}, \\mathbf {C}, \\mathbf {B}\\right) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 393, + 205, + 782, + 219 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {M} _ {r 3} = \\mathbf {M} _ {r 2} + \\operatorname {G P M - 3} (\\mathbf {Q} _ {e}, \\mathbf {C}, \\mathbf {B})\n$$\n", + "text_format": "latex", + "bbox": [ + 393, + 223, + 607, + 237 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Eq.6, each GPM can focus on learning the filtering of negative samples at different granularities, as follows: (1) GPM-1 focuses on learning from pairs between background and filters out a large number of high-confidence negative samples. (2) GPM-2 focuses on learning from pairs between background and entities and filters out medium-confidence negative samples. (3) GPM-3 focuses on learning from no relation pairs between entities and only needs to filter out entity pairs that do not form relationships from the remaining entity pairs.", + "bbox": [ + 212, + 253, + 782, + 358 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Based on the final sparse relationship matrix $\\mathbf{M}_{r3}$ , we can select the top- $N_{r}$ subject-object pairs most likely to form relations and obtain their indices in the entity visual representations $\\mathbf{Q}_e$ , the entity's class predictions $\\mathbf{C}$ , and the bounding boxes $\\mathbf{B}$ .", + "bbox": [ + 214, + 361, + 782, + 420 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.5 Relation Decoder", + "text_level": 1, + "bbox": [ + 215, + 450, + 405, + 464 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "With the $\\mathbf{Q}_e$ , $\\mathbf{C}$ and $\\mathbf{B}$ of subjects and objects, we can obtain the subject-object pair representation $\\mathbf{Q}_{pair}$ , which integrates visual, semantic, and positional encodings as follows:", + "bbox": [ + 212, + 483, + 782, + 527 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Q} _ {p a i r} = \\left[ \\mathbf {Q} _ {e} ^ {s u b}, e m b (\\mathbf {C} ^ {s u b}), p o s (\\mathbf {B} ^ {s u b}), \\mathbf {Q} _ {e} ^ {o b j}, e m b (\\mathbf {C} ^ {o b j}), p o s (\\mathbf {B} ^ {o b j}) \\right]. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 267, + 542, + 782, + 561 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $emb$ is a pre-trained Glove language model to acquire the word embedding.", + "bbox": [ + 214, + 575, + 782, + 590 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this paper, we adopt the relation decoder in [30], which consists of transformer decoders in the style of DETR, to predict relations. In the relation decoder, we initialize a relation query $\\mathbf{Q}_r \\in \\mathbb{R}^{N_r \\times d}$ as the query input, and the subject-object pair $\\mathbf{Q}_{pair}$ is projected as the key and value of cross-attention. Subsequently, the relation representation after the Relation Decoder (RD) can be expressed as $\\tilde{\\mathbf{Q}}_r = \\mathrm{RD}(\\mathbf{Q}_r, \\mathbf{Q}_{pair})$ . Then, the relation prediction for the entity pairs can be expressed as $\\hat{\\mathbf{r}} = W_{cls} \\tilde{\\mathbf{Q}}_r$ , where $W_{cls}$ is the liner relation classifier.", + "bbox": [ + 212, + 593, + 784, + 700 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.6 Training", + "text_level": 1, + "bbox": [ + 215, + 731, + 333, + 746 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our end-to-end SGG is divided into three subtasks: (1) an entity detection task based on the object detector, (2) a sparse relationship matrix prediction task based on the Mg-RMPN, and (3) a relation prediction task based on the relation decoder. During training, each subtask generates the corresponding supervisory information and losses, as follows.", + "bbox": [ + 212, + 763, + 782, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Mg-RMPN", + "bbox": [ + 653, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Entity Detection Loss We uses the end-to-end deformable DETR [42] as object detector, which uses the set prediction loss proposed in DETR [1] by assigning the ground truth entities to the predictions. A cost function is applied to compute the matching cost between a prediction and a ground-truth entity. With the cost matrix, the entity prediction-ground truth assignment is computed with the Hungarian Matching [14]. Given a set of $N_{e}$ entity proposals $\\{e_i\\}_{i=1}^{N_e}$ from object detector, the set prediction loss for can be presented as:", + "bbox": [ + 212, + 146, + 787, + 253 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {e} = \\Sigma_ {i = 1} ^ {N _ {e}} \\left[ \\mathcal {L} _ {c l s} ^ {e} + \\mathbf {1} _ {e _ {i} \\neq \\phi} \\mathcal {L} _ {b o x} ^ {e} \\right], \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 398, + 263, + 785, + 282 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{cls}^{e}$ denotes the cross-entropy loss for label classification and $e_i \\neq \\phi$ means that $<\\text{background}>$ is not assigned to the $i$ th entity prediction. $\\mathcal{L}_{box}^{e}$ consists of $L_1$ loss and generalized IoU loss for the box regression.", + "bbox": [ + 214, + 294, + 789, + 340 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Multi-granularity Relationship Matrices Loss To predict the final sparse relationship matrix, we need to define the supervisory information and loss functions during the training phase.", + "bbox": [ + 214, + 364, + 787, + 411 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Multi-granularity Supervision Matrices In the training phase, we assign labels to the entity proposals $\\{e_i\\}_{i=1}^{N_e}$ by Hungarian matching and obtain the ground truth for each node in $\\mathbf{M}_r$ regarding the subject, object, and whether they form a relation. In response to the sparsity of $\\mathbf{M}_r$ , we propose Mg-RMPN to achieve hierarchical identification of negative samples at different granularities.", + "bbox": [ + 214, + 428, + 787, + 502 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Consequently, based on the granularity of the negative samples that each GPM focuses on in Section 3.4, the corresponding ground truth can be obtained. (1) $\\mathbf{M}_{r1}^{gt}$ filters out all pairs composed of background entities as follows:", + "bbox": [ + 214, + 503, + 787, + 547 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {M} _ {r 1} ^ {g t} = \\left\\{ \\right.p _ {i j} \\mid p _ {i j} = \\left\\{\\begin{array}{l l}0&\\text {i f} (e _ {i} = \\phi \\text {a n d} e _ {j} = \\phi) \\text {o r} i = j\\\\1&\\text {o t h e r w i s e}\\end{array}, 1 \\leq i \\leq N _ {e}, 1 \\leq j \\leq N _ {e} \\right\\}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 240, + 559, + 785, + 594 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $e_i = \\phi$ ( $e_j = \\phi$ ) means that $<\\text{background}>$ is assigned to the $i$ th ( $j$ th) entity prediction, $i = j$ means removing self-connected entity pairs. (2) $\\mathbf{M}_{r2}^{gt}$ further filters out all pairs composed of backgrounds and entities as follows:", + "bbox": [ + 214, + 604, + 787, + 650 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {M} _ {r 2} ^ {g t} = \\left\\{ \\right.p _ {i j} \\mid p _ {i j} = \\left\\{\\begin{array}{l l}0&\\text {i f} (e _ {i} = \\phi \\text {o r} e _ {j} = \\phi) \\text {o r} i = j\\\\1&\\text {o t h e r w i s e}\\end{array}, 1 \\leq i \\leq N _ {e}, 1 \\leq j \\leq N _ {e} \\right\\}. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 241, + 661, + 785, + 696 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "(3) $\\mathbf{M}_{r3}^{gt}$ filters out all non-relationship pairs, and $\\mathbf{M}_{r3}^{gt}$ is also the final ground truth relationship matrix of the scene graph, expressed as follows", + "bbox": [ + 215, + 707, + 787, + 738 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {M} _ {r 3} ^ {g t} = \\left\\{p _ {i j} \\mid p _ {i j} = \\left\\{ \\begin{array}{l l} 0 & \\text {o t h e r w i s e} \\\\ 1 & \\text {i f} r _ {i j} = 1 \\end{array} , 1 \\leq i \\leq N _ {e}, 1 \\leq j \\leq N _ {e} \\right\\}, \\right. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 316, + 750, + 785, + 782 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $r_{ij} = 1$ implies that there is a relation between subject $i$ and object $j$ . Here, the relationships of the negative sample subsets in $\\mathbf{M}_{r1}^{gt}$ , $\\mathbf{M}_{r2}^{gt}$ and $\\mathbf{M}_{r3}^{gt}$ are $\\mathbf{M}_{r1|p_{ij} = 0}^{gt} \\subseteq \\mathbf{M}_{r2|p_{ij} = 0}^{gt} \\subseteq \\mathbf{M}_{r3|p_{ij} = 0}^{gt}$ .", + "bbox": [ + 214, + 794, + 787, + 844 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "L. Wang et al.", + "bbox": [ + 271, + 114, + 369, + 128 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Multi-granularity Relationship Matrices Loss Considering that the large number of negative samples in $\\mathbf{M}_r$ have different confidence, we modify the binary cross-entropy function using focal loss [20], which applies weighted discrimination on well-classified samples, forcing the model to focus on wrongly classified samples. Then, the multi-granularity learning loss is", + "bbox": [ + 214, + 146, + 782, + 217 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {M g} = \\mathcal {L} _ {R M} ^ {1} \\left(\\mathbf {M} _ {r 1}, \\mathbf {M} _ {r 1} ^ {g t}\\right) + \\mathcal {L} _ {R M} ^ {2} \\left(\\mathbf {M} _ {r 2}, \\mathbf {M} _ {r 2} ^ {g t}\\right) + \\mathcal {L} _ {R M} ^ {3} \\left(\\mathbf {M} _ {r 3}, \\mathbf {M} _ {r 3} ^ {g t}\\right), \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 292, + 226, + 782, + 242 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where the relationship matrices loss $\\mathcal{L}_{RM}^{k}(k = 1,2,3)$ is defined as", + "bbox": [ + 218, + 250, + 696, + 267 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {R M} ^ {k} \\left(\\mathbf {M} _ {r}, \\mathbf {M} _ {r} ^ {g t}\\right) = - \\sum_ {i} \\sum_ {j} \\left\\{\\alpha_ {k} \\left(1 - \\hat {p} _ {i j}\\right) ^ {\\gamma} p _ {i j} \\log (\\hat {p} _ {i j}) \\right. \\\\ + (1 - \\alpha_ {k}) \\hat {p} _ {i j} ^ {\\gamma} (1 - p _ {i j}) \\log (1 - \\hat {p} _ {i j}) \\} / N _ {p o s} \\tag {13} \\\\ + \\left\\| \\mathbf {M} _ {r} \\right\\| _ {1} + \\left\\| \\mathbf {M} _ {r} \\right\\| _ {2}, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 308, + 276, + 784, + 340 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where the first part is the focal loss based on binary cross-entropy, and the second part is the L1 and L2 regularization of $\\mathbf{M}_r$ , $\\alpha_k(k = 1,2,3)$ is hyperparameters that adjusts the weights of positive and negative samples, $\\gamma$ is a hyperparameter for focal loss, and $N_{pos}$ is the number of positive samples in each mini-batch.", + "bbox": [ + 214, + 349, + 782, + 409 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Relation Prediction Loss In this paper, the relation decoder also considers relation prediction as a set prediction based on a query. Therefore, during the training phase, we also use Hungarian Matching to assign labels for relation prediction. Due to the imbalanced relation classes in SGG, we use Seesaw loss [29] to dynamically adjust the gradients of samples of different classes. The relation loss is as follows:", + "bbox": [ + 214, + 430, + 782, + 513 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {r} = - \\sum_ {i = 1} ^ {C _ {r}} y _ {i} \\log \\left(\\hat {\\sigma} _ {i}\\right), \\quad \\hat {\\sigma} _ {i} = \\frac {e ^ {\\hat {r} _ {i}}}{\\sum_ {i \\neq j} ^ {C _ {r}} \\mathcal {S} _ {i j} e ^ {\\hat {r} _ {j}} + e ^ {\\hat {r} _ {i}}}, \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 346, + 523, + 782, + 559 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $C_r$ is the number of relation classes, $y_i \\in \\{0,1\\}$ , $1 \\leqslant i \\leqslant C_r$ is the one-hot ground truth label, $S_{ij}$ is a tunable balancing factor between different classes, detail can be found in [29].", + "bbox": [ + 214, + 566, + 782, + 613 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In summary, the total loss of our method is", + "bbox": [ + 238, + 614, + 553, + 628 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {e} + \\lambda_ {1} \\mathcal {L} _ {M g} + \\lambda_ {2} \\mathcal {L} _ {r}, \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 411, + 643, + 782, + 656 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Where $\\lambda_{1}$ and $\\lambda_{2}$ are hyperparameters used to respectively adjust multi-granularity learning loss $\\mathcal{L}_{Mg}$ and relation prediction loss $\\mathcal{L}_r$ .", + "bbox": [ + 214, + 667, + 797, + 700 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 214, + 722, + 374, + 739 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Experimental Settings", + "text_level": 1, + "bbox": [ + 214, + 753, + 444, + 768 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Datasets: We evaluate our proposed method on the following datasets: (1) Visual Genome (VG150) [13] is the most widely used dataset in SGG, consisting of the most frequent 150 object classes and 50 predicate classes. (2) Open Images V6 (OIv6) [15] is a large-scale dataset proposed by Google. We follow", + "bbox": [ + 212, + 779, + 782, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Mg-RMPN", + "bbox": [ + 653, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "the data processing and evaluation protocols in [17, 40]. OIv6 consists of 601 object classes and 30 predicate classes. (2) Generalized Question Answering (GQA200) [9] is another vision and language benchmark with more than 3.8M relation annotations. We follow the data processing in [7], which consists of Top-200 object classes and Top-100 predicate classes.", + "bbox": [ + 212, + 146, + 782, + 219 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Task & Evaluation Metrics: In this work, we focus on the Scene Graph Detection task, which detects all objects in an image and predicts their bounding boxes, labels, and relations. The initial evaluation metric Recall@K (R@K) was found to be dominated by head classes. Therefore, the mean Recall@K (mR@K) across all relations classes is proposed to evaluate unbiased SGG. However, focusing solely on mR@K and neglecting R@K can result in a tail bias. Therefore, we adopts the harmonic mean F@K of R@K and mR@K as an overall metric. For OIv6, the weighted evaluation metrics (wmAP rel, wnAP phr, scorewtd) are used for a more class-balanced evaluation.", + "bbox": [ + 212, + 220, + 784, + 354 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Implementation Details: We utilized the pre-trained deformable DETR [42] as the object detector. We set the loss hyperparameters $\\lambda_{1}$ and $\\lambda_{2}$ to 0.5 and 3 respectively. For the multi-granularity relationship matrices loss, we set $\\alpha_{k}(k = 1,2,3)$ to 0.75, 0.9, and 0.99, respectively, to control the weight of positive samples, with the focal loss parameter $\\gamma$ set to 2. The number of entities predicted by the detector $N_{e}$ is 100, the number of most probable subject-object pairs selected $N_{r}$ is 100, and the size of embedding dimensions $d$ is 256. Our model is implemented on 8 NVIDIA 3090 GPUs with learning rate $\\times 10^{-4}$ and batch size 16 for 24 epochs. More implementation details are shown in the Supplementary Material.", + "bbox": [ + 212, + 357, + 785, + 507 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 Comparison with State of the Arts", + "text_level": 1, + "bbox": [ + 214, + 529, + 549, + 544 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "VG150: Table 1 shows the comparison of our method on VG150, with methods divided into two-stage (above) and end-to-end (below). Due to the imbalance issue in SGG, the two-stage methods (Motifs, VCTree, GPS-Net, RelDN, PE-Net) perform better in R@K but show very poor performance in mR@K. Thus, many rebalancing techniques (VCTree+TDE, BGNN, SHA+GCL) have been developed. Although they improve mR@K, they inevitably impair R@K. For example, SHA+GCL, despite achieving the best mR@K, its recall performance significantly deteriorates, with R@50/100 being only 14.9/18.2.", + "bbox": [ + 212, + 551, + 784, + 672 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Although the end-to-end methods are inferior to the two-stage methods on R@K, it demonstrates better potential on mR@K due to its advantage in sparse relation prediction, which does not dilute the scarce tail-class samples. Furthermore, compared to the rebalancing techniques of the two-stage methods, the end-to-end approach does not overly harm R@K and can achieve superior overall performance F@K, such as SGTR and Pair-Net.", + "bbox": [ + 212, + 674, + 784, + 762 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Our Mg-RMPN leads comprehensively in end-to-end methods. Compared with two-stage methods, Mg-RMPN achieves competitive performance in R@K and significantly excels in mR@K, achieving the best overall performance F@K. For example, Mg-RMPN achieved 19.3/22.8 on F@50/100, surpassing the end-to-end and two-stage optimal methods Pair-Net and PE-Net, respectively.", + "bbox": [ + 212, + 763, + 784, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "L. Wang et al.", + "bbox": [ + 271, + 114, + 367, + 128 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/172f1254352a9b345e79d3965ca305b7e774d3ad9d882cc44a9341ab493c762d.jpg", + "table_caption": [ + "Table 1: Performance comparison at scene graph detection task on VG150. * denotes the results from [16]." + ], + "table_footnote": [], + "table_body": "
MethodR@20 R@50 R@100mR@20 mR@50 mR@100F@20 F@50 F@100
Motifs [36]25.532.837.25.06.87.98.411.313.0
VCTree [27]24.531.936.25.47.48.78.812.014.0
VCTree+TDE [26]14.019.423.26.99.311.19.212.615.0
GPS-Net [21]22.328.933.26.98.79.810.513.415.1
BGNN [17]-31.035.8-10.712.6-15.918.6
RelDN [39]-31.435.9-6.07.3-10.112.1
SHA+GCL [7]-14.918.214.217.920.9-16.319.5
PE-Net [40]-30.735.2-12.414.5-17.720.5
FCSGG [22]16.121.325.12.73.64.24.66.27.2
RelTR [6]21.227.5-6.810.8-10.315.5-
AS-Net* [3]-18.721.1-6.17.2-9.210.7
HOTR* [12]-23.527.7-9.412.0-13.416.7
SGTR [16]-24.628.4-12.015.2-16.119.8
Pair-Net [30]18.824.929.38.912.415.412.116.620.2
Mg-RMPN(DETR)21.027.131.39.913.516.213.518.021.3
Mg-RMPN(Ours)22.529.133.510.314.417.314.119.322.8
", + "bbox": [ + 215, + 184, + 785, + 444 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We also compared with Pair-Net that is also based on relationship matrix learning. The experimental results show that our method is comprehensively superior to it, which demonstrates the effectiveness of our proposed Mg-RMPN. Furthermore, we also present the results of Mg-RMPN based on the DETR object detector to eliminate the influence of the detector. The results show that our Mg-RMPN(DETR) still achieves the best overall performance among both one-stage and two-stage methods.", + "bbox": [ + 212, + 497, + 785, + 603 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Open Image V6: To validate the generalizability of Mg-RMPN, we performed experiments on Open Images V6 in Table 2 and compared with two-stage (above) and end-to-end (below) methods. Compared to classic SGG benchmarks, our method achieved the optimal results among all methods with 45.5, 77.8, 57.4 respectively in mR@50, R@50, F@50. Compared to open image benchmarks, we obtained competitive 35.5 and 36.4 in $\\mathrm{wmAP}_{\\mathrm{rel}}$ and $\\mathrm{wmAP}_{\\mathrm{phr}}$ , and achieved the best result 43.6 in the overall metric score $_{\\mathrm{wtd}}$ .", + "bbox": [ + 212, + 614, + 785, + 722 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "GQA200: We also conducted experiments on the more challenging GQA200, as shown in Table 3. Compared to two-stage methods (VTransE, Motifs, VCTree, SHA+GCL), our method's overall performance F@K is only slightly inferior to SHA+GCL on F@100, which sacrificed a significant amount of R@K to enhance mR@K by using rebalancing techniques. Compared to the end-to-end method Pair-Net, which is also based on relationship matrices, our method is comprehensively superior.", + "bbox": [ + 212, + 733, + 785, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Mg-RMPN", + "bbox": [ + 653, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/a20cd7efe416397520b77551a1932cec69569fac6ff7905d824035208aff488a.jpg", + "table_caption": [ + "Table 2: Performance comparison on OIV6. * denotes the results from [16,17]. †denote results reproduced with the authors' code." + ], + "table_footnote": [], + "table_body": "
MethodmR@50R@50F@50wmAPrelwmAPphrscorewtd
Motifs* [36]32.771.644.929.931.638.9
VCTree* [27]33.974.146.534.233.140.2
RelDN* [39]34.073.146.432.233.440.8
G-RCNN* [34]34.074.546.733.234.241.8
GPS-Net* [21]35.374.848.032.934.041.7
BGNN* [17]40.575.052.633.534.242.1
RelTR [6]-71.7-37.237.543.0
AS-Ne* [3]35.255.343.025.927.532.4
HOTR* [12]40.152.745.519.421.526.7
SGTR* [16]42.659.949.837.038.742.3
Pair-Net† [30]44.577.456.531.832.440.3
Mg-RMPN(Ours)45.577.857.435.536.443.6
", + "bbox": [ + 263, + 184, + 738, + 388 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/515f092d453578df0b2379233a8aeb17c4e98a0ad2882616d6c3abc25a3e34b3.jpg", + "table_caption": [ + "Table 3: Performance comparison at scene graph detection task on GQA200. * denotes the results from [7]. †denote results reproduced with the authors' code." + ], + "table_footnote": [], + "table_body": "
MethodR@50R@100mR@50mR@100F@50F@100
VTransE* [38]27.230.75.86.69.610.9
Motifs* [36]28.933.16.47.710.512.5
VCTree* [27]28.331.96.57.410.612.0
SHA+GCL* [7]14.817.917.820.116.218.9
Pair-Net† [30]20.223.410.612.613.916.4
Mg-RMPN (Ours)23.225.712.814.516.518.5
", + "bbox": [ + 274, + 440, + 728, + 563 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.3 Visualization of Multi-Granularity Relationship Matrices", + "text_level": 1, + "bbox": [ + 215, + 589, + 732, + 604 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Figure 3 presents a visualization of the multi-granularity relationship matrices learned by our Mg-RMPN, where (d), (e), (f) are the sparse relationship matrices learned after being filtered by GPM-1, GPM-2, GPM-3, respectively, and (a), (b), (c) are the corresponding ground truth relationship matrices defined in Section 3.6. The visualization results show that the dense negative samples in the relationship matrix are filtered layer by layer according to different granularities. The consistency of the generated multi-granularity relationship matrices with their ground truth indicates that the three GPMs of Mg-RMPN indeed focus on filtering negative samples of different granularities.", + "bbox": [ + 212, + 613, + 787, + 748 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Figures 3 (h) and (g), respectively, present a zoomed-in view of the final relationship matrix and its ground truth. We can observe that the predicted relationship matrix successfully captures the true relationships, for example: \"man wearing shirt\", \"man holding racket\", \"man wearing short\" and \"pole on fence\". Additionally, the relationship matrix also predicts some relationships that do not in the ground truth, as illustrated by the dashed lines in the rows", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "L. Wang et al.", + "bbox": [ + 271, + 114, + 367, + 128 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/88eef34e090e017fc0bd9b278b104f472bf5f0d55181d294db4b8087b487bd84.jpg", + "image_caption": [ + "(a) Ground truth $\\mathbf{M}_{F1}^{gt}$" + ], + "image_footnote": [], + "bbox": [ + 232, + 143, + 362, + 252 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/7f4dac97d7411486991dc6e8f4577243682de9501f1db16f2f66920843bb7d49.jpg", + "image_caption": [ + "(b) Ground truth $\\mathbf{M}_{\\pmb{F}_2}^{gt}$" + ], + "image_footnote": [], + "bbox": [ + 364, + 156, + 486, + 252 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/b1efc3e7320fa5abbc194866758cb6355c002a968debd24867c1edf86df96e62.jpg", + "image_caption": [ + "(c) Ground truth $\\mathbf{M}_{r3}^{gt}$" + ], + "image_footnote": [], + "bbox": [ + 488, + 156, + 614, + 252 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/cbf961e879467d824496405e8dc3935189dc0a31fb70385f987abee430bab51f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 616, + 143, + 769, + 268 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/388cd83ade6d94c4e7f913984517fe93d16f09004ede628cce35e0a493a3057f.jpg", + "image_caption": [ + "(d) Multi-granularity $\\mathbf{M}_{r1}$" + ], + "image_footnote": [], + "bbox": [ + 232, + 277, + 359, + 372 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/5e822dc4ef2812098e1887bd1083d937525a4850d0f18fa7e3c2e4550fd2293d.jpg", + "image_caption": [ + "(e) Multi-granularity $\\mathbf{M}_{r2}$" + ], + "image_footnote": [], + "bbox": [ + 361, + 277, + 486, + 373 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/24a68a31da06af6c687428d0263c295d987826ae1df5042e28a687bf274fd2ba.jpg", + "image_caption": [ + "(f) Multi-granularity $\\mathbf{M}_{r3}$" + ], + "image_footnote": [], + "bbox": [ + 488, + 277, + 616, + 372 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/074f5772091af3359ebfaf9bb09e99ad75a17ec8c2ac8cdba57953f4a205916f.jpg", + "image_caption": [ + "Fig. 3: Visualization of Multi-Granularity Relationship Matrices" + ], + "image_footnote": [], + "bbox": [ + 616, + 270, + 769, + 393 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/b69a63d3072eb708aace50b2939ff9dfd45abb474c363642fe1232553cdc0908.jpg", + "table_caption": [ + "Table 4: Ablation studies of Mg-RMPN components on VG150." + ], + "table_footnote": [], + "table_body": "
GPM-1GPM-2GPM-3R@20/50/100mR@20/50/100F@20/50/100
✓✘✘6.5/10.7/15.22.9/4.8/6.64.0/6.6/9.2
✘✓✘12.5/17.5/21.65.2/7.4/9.37.3/10.4/13.0
✓✓✘12.8/18.1/22.35.8/8.5/10.78.0/11.6/14.5
✘✘✓21.2/27.6/32.19.2/13.1/15.812.8/17.8/21.2
✓✘✓21.7/28.2/32.69.8/13.8/16.513.5/18.5/21.9
✘✓✓21.8/28.2/32.59.7/13.7/16.713.4/18.4/22.1
✓✓✓22.5/29.1/33.510.3/14.4/17.314.1/19.3/22.8
", + "bbox": [ + 246, + 462, + 754, + 590 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "and columns, which represent the relationships with the corresponding entity \"man\" as the subject and object, respectively. Due to the abundant presence of \"man\" in SGG, the relationship matrix tends to favor predicting relations that include \"man\", which aligns with the actual language induction bias. Therefore, the results of visualization indicate that our Mg-RMPN can accurately predict the sparse relationship matrix.", + "bbox": [ + 212, + 619, + 787, + 710 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.4 Ablation Studies", + "text_level": 1, + "bbox": [ + 214, + 736, + 401, + 750 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Component Analysis of Mg-RMPN. To assess the effectiveness of each GPM in Mg-RMPN, we performed ablation experiments in Table 4. In Mg-RMPN, each GPM focuses on negative sample filtering of different granularities, and the three GPMs collaborate to complete the learning of the sparse relationship matrix. GPM-1 and GPM-2 are used to help GPM-3 filter negative", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Mg-RMPN", + "bbox": [ + 653, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/9e52e5561fcccb77c2284235e15063cfc4097ca68ef1be3395a96ccec106d27b.jpg", + "table_caption": [ + "Table 5: Ablation studies of different input information for Mg-RMPN on VG150." + ], + "table_footnote": [], + "table_body": "
Visual Semantic PositionR@20/50/100mR@20/50/100F@20/50/100
✓XX21.2/27.6/32.19.8/13.6/16.413.4/18.2/21.7
✓X✓22.2/28.7/33.110.1/14.1/16.913.9/18.9/22.4
✓✓X21.2/27.6/32.29.8/13.8/16.813.4/18.4/22.1
✓✓✓22.5/29.1/33.510.3/14.4/17.314.1/19.3/22.8
", + "bbox": [ + 243, + 171, + 759, + 258 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "samples. Using GPM-1 and GPM-2 individually cannot make full use of the ground truth, hence their results are very poor. The results obtained using only GPM-3 are also not good enough. Further increasing the assistance of GPMs for negative sample filtering can significantly improve the performance of the model. Mg-RMPN fully utilized three GPMs to achieve optimal results, confirming that learning negative samples based on different granularities can effectively improve the model's learning ability for sparse relationship matrices.", + "bbox": [ + 212, + 290, + 782, + 396 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Different input information for Mg-RMPN This paper bridges entity detection and relation detection through a relationship matrix, making full use of the semantic and positional information of entities and solving the problem that end-to-end SGG relies solely on visual information to predict relations. Table 5 presents ablation studies of Mg-RMPN based on different input information. Compared to models that only utilize visual information, separately incorporating the semantic and positional information of entities can enhance the performance of the model. Furthermore, the improvement in introducing semantic and positional information simultaneously is significantly greater than in introducing a single modality of information, indicating that semantic and positional information are two complementary information. Hence, the semantic and positional information incorporated in our approach plays a crucial role in predicting relations. More hyperparameter analysis can be found in Supplementary Material.", + "bbox": [ + 212, + 398, + 784, + 609 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 645, + 357, + 660 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, our proposed end-to-end method bridges entity detection and relation detection through a sparse relationship matrix, which not only eliminates the need for post-processing of relation matching but also leverages the semantics and positional information of predicted entities to enhance relation prediction. To predict the sparse relationship matrix, we propose a multi-granularity sparse relationship matrix prediction network, which utilizes three gated pooling modules focusing on filtering negative samples at different granularities. Finally, a set of sparse, most likely subject-object pairs can be constructed and used for relation decoding. The experimental results demonstrate that our method achieves an optimal overall performance in both the end-to-end and two-stage methods.", + "bbox": [ + 212, + 688, + 785, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "L. Wang et al.", + "bbox": [ + 271, + 114, + 367, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work was supported by the National Natural Science Foundation of China (U21A20485, 61976170, 62088102) and the National Key R&D Program of China (2023YFB4704900).", + "bbox": [ + 215, + 175, + 784, + 220 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 243, + 321, + 258 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: European conference on computer vision. pp. 213-229. Springer (2020)", + "2. Chen, C., Zhan, Y., Yu, B., Liu, L., Luo, Y., Du, B.: Resistance training using prior bias: Toward unbiased scene graph generation. Proceedings of the AAAI Conference on Artificial Intelligence 36(1), 212-220 (Jun 2022)", + "3. Chen, M., Liao, Y., Liu, S., Chen, Z., Wang, F., Qian, C.: Reformulating hoi detection as adaptive set prediction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9004-9013 (2021)", + "4. Chen, S., Jin, Q., Wang, P., Wu, Q.: Say as you wish: Fine-grained control of image caption generation with abstract scene graphs. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9962-9971 (2020)", + "5. Chen, T., Yu, W., Chen, R., Lin, L.: Knowledge-embedded routing network for scene graph generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6163-6171 (2019)", + "6. Cong, Y., Yang, M.Y., Rosenhahn, B.: Reltr: Relation transformer for scene graph generation. IEEE Transactions on Pattern Analysis and Machine Intelligence (2023)", + "7. Dong, X., Gan, T., Song, X., Wu, J., Cheng, Y., Nie, L.: Stacked hybrid-attention and group collaborative learning for unbiased scene graph generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19427-19436 (2022)", + "8. Gao, L., Lyu, X., Guo, Y., Hu, Y., Li, Y.F., Xu, L., Shen, H.T., Song, J.: Informative scene graph generation via debiasing. arXiv preprint arXiv:2308.05286 (2023)", + "9. Hudson, D.A., Manning, C.D.: Gqa: A new dataset for real-world visual reasoning and compositional question answering. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 6700-6709 (2019)", + "10. Johnson, J., Gupta, A., Fei-Fei, L.: Image generation from scene graphs. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1219-1228 (2018)", + "11. Johnson, J., Krishna, R., Stark, M., Li, L.J., Shamma, D., Bernstein, M., Fei-Fei, L.: Image retrieval using scene graphs. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3668-3678 (2015)", + "12. Kim, B., Lee, J., Kang, J., Kim, E.S., Kim, H.J.: Hotr: End-to-end human-object interaction detection with transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 74-83 (2021)", + "13. Krishna, R., Zhu, Y., Groth, O., Johnson, J., Hata, K., Kravitz, J., Chen, S., Kalantidis, Y., Li, L.J., Shamma, D.A., et al.: Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision 123(1), 32-73 (2017)" + ], + "bbox": [ + 218, + 273, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Mg-RMPN", + "bbox": [ + 653, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "14. Kuhn, H.W.: The hungarian method for the assignment problem. Naval research logistics quarterly 2(1-2), 83-97 (1955)", + "15. Kuznetsova, A., Rom, H., Alldrin, N., Uijlings, J., Krasin, I., Pont-Tuset, J., Kamali, S., Popov, S., Malloci, M., Kolesnikov, A., et al.: The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. International Journal of Computer Vision 128(7), 1956-1981 (2020)", + "16. Li, R., Zhang, S., He, X.: Sgtr: End-to-end scene graph generation with transformer. In: proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 19486-19496 (2022)", + "17. Li, R., Zhang, S., Wan, B., He, X.: Bipartite graph network with adaptive message passing for unbiased scene graph generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11109-11119 (2021)", + "18. Li, W., Zhang, H., Bai, Q., Zhao, G., Jiang, N., Yuan, X.: Ppdl: Predicate probability distribution based loss for unbiased scene graph generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19447-19456 (2022)", + "19. Li, Y., Ma, T., Bai, Y., Duan, N., Wei, S., Wang, X.: Pastegan: A semi-parametric method to generate image from scene graph. Advances in Neural Information Processing Systems 32 (2019)", + "20. Lin, T.Y., Goyal, P., Girshick, R., He, K., Dollar, P.: Focal loss for dense object detection. In: Proceedings of the IEEE international conference on computer vision. pp. 2980-2988 (2017)", + "21. Lin, X., Ding, C., Zeng, J., Tao, D.: Gps-net: Graph property sensing network for scene graph generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3746-3753 (2020)", + "22. Liu, H., Yan, N., Mortazavi, M., Bhanu, B.: Fully convolutional scene graph generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11546-11556 (2021)", + "23. Lyu, X., Gao, L., Xie, J., Zeng, P., Tian, Y., Shao, J., Shen, H.T.: Generalized unbiased scene graph generation. arXiv preprint arXiv:2308.04802 (2023)", + "24. Ren, S., He, K., Girshick, R., Sun, J.: Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems 28 (2015)", + "25. Schroeder, B., Tripathi, S.: Structured query-based image retrieval using scene graphs. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. pp. 178-179 (2020)", + "26. Tang, K., Niu, Y., Huang, J., Shi, J., Zhang, H.: Unbiased scene graph generation from biased training. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 3716-3725 (2020)", + "27. Tang, K., Zhang, H., Wu, B., Luo, W., Liu, W.: Learning to compose dynamic tree structures for visual contexts. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 6619-6628 (2019)", + "28. Teney, D., Liu, L., van Den Hengel, A.: Graph-structured representations for visual question answering. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1-9 (2017)", + "29. Wang, J., Zhang, W., Zang, Y., Cao, Y., Pang, J., Gong, T., Chen, K., Liu, Z., Loy, C.C., Lin, D.: Seesaw loss for long-tailed instance segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9695-9704 (2021)", + "30. Wang, J., Wen, Z., Li, X., Guo, Z., Yang, J., Liu, Z.: Pair then relation: Pair-net for panoptic scene graph generation. arXiv preprint arXiv:2307.08699 (2023)" + ], + "bbox": [ + 217, + 147, + 784, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "L. Wang et al.", + "bbox": [ + 271, + 114, + 367, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "31. Wang, L., Yuan, Z., Chen, B.: Learning to generate an unbiased scene graph by using attribute-guided predicate features. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 2581-2589 (2023)", + "32. Xu, D., Zhu, Y., Choy, C.B., Fei-Fei, L.: Scene graph generation by iterative message passing. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5410-5419 (2017)", + "33. Yang, G., Zhang, J., Zhang, Y., Wu, B., Yang, Y.: Probabilistic modeling of semantic ambiguity for scene graph generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12527-12536 (2021)", + "34. Yang, J., Lu, J., Lee, S., Batra, D., Parikh, D.: Graph r-cnn for scene graph generation. In: Proceedings of the European conference on computer vision (ECCV). pp. 670-685 (2018)", + "35. Yu, J., Chai, Y., Wang, Y., Hu, Y., Wu, Q.: Cogtree: Cognition tree loss for unbiased scene graph generation. In: Zhou, Z.H. (ed.) Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence, IJCAI-21. pp. 1274-1280. International Joint Conferences on Artificial Intelligence Organization (8 2021)", + "36. Zellers, R., Yatskar, M., Thomson, S., Choi, Y.: Neural motifs: Scene graph parsing with global context. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5831-5840 (2018)", + "37. Zhang, A., Yao, Y., Chen, Q., Ji, W., et al, L.: Fine-grained scene graph generation with data transfer (2022)", + "38. Zhang, H., Kyaw, Z., Chang, S.F., Chua, T.S.: Visual translation embedding network for visual relation detection. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5532-5540 (2017)", + "39. Zhang, J., Shih, K.J., Elgammal, A., Tao, A., Catanzaro, B.: Graphical contrastive losses for scene graph parsing. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11535-11543 (2019)", + "40. Zheng, C., Lyu, X., Gao, L., Dai, B., Song, J.: Prototype-based embedding network for scene graph generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22783-22792 (2023)", + "41. Zhong, Y., Wang, L., Chen, J., Yu, D., Li, Y.: Comprehensive image captioning via scene graph decomposition. In: European Conference on Computer Vision. pp. 211-229. Springer (2020)", + "42. Zhu, X., Su, W., Lu, L., Li, B., Wang, X., Dai, J.: Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159 (2020)" + ], + "bbox": [ + 212, + 146, + 787, + 645 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Mg-RMPN", + "bbox": [ + 653, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/Multi-Granularity Sparse Relationship Matrix Prediction Network for End-to-End Scene Graph Generation/6b85535a-3ad0-4a54-882a-f4d4244fbf81_model.json b/2024/Multi-Granularity Sparse Relationship Matrix Prediction Network for End-to-End Scene Graph Generation/6b85535a-3ad0-4a54-882a-f4d4244fbf81_model.json new file mode 100644 index 0000000000000000000000000000000000000000..a3e39181d37b7cb61b3679dfc2d84fb493b16841 --- /dev/null +++ b/2024/Multi-Granularity Sparse Relationship Matrix Prediction Network for End-to-End Scene Graph Generation/6b85535a-3ad0-4a54-882a-f4d4244fbf81_model.json @@ -0,0 +1,2445 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.223, + 0.141, + 0.782, + 0.207 + ], + "angle": 0, + "content": "Multi-Granularity Sparse Relationship Matrix Prediction Network for End-to-End Scene Graph Generation" + }, + { + "type": "text", + "bbox": [ + 0.33, + 0.235, + 0.673, + 0.252 + ], + "angle": 0, + "content": "Lei Wang, Zejian Yuan, and Badong Chen" + }, + { + "type": "text", + "bbox": [ + 0.236, + 0.263, + 0.766, + 0.306 + ], + "angle": 0, + "content": "Institute of Artificial Intelligence and Robotics, Xi'an Jiaotong University, Xi'an 710049, China \nleiwangmail@stu.xjtu.edu.cn, {yuan.ze.jian, chenbd}@mail.xjtu.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.341, + 0.744, + 0.593 + ], + "angle": 0, + "content": "Abstract. Current end-to-end Scene Graph Generation (SGG) relies solely on visual representations to separately detect sparse relations and entities in an image. This leads to the issue where the predictions of entities do not contribute to the prediction of relations, necessitating post-processing to assign corresponding subjects and objects to the predicted relations. In this paper, we introduce a sparse relationship matrix that bridges entity detection and relation detection. Our approach not only eliminates the need for relation matching, but also leverages the semantics and positional information of predicted entities to enhance relation prediction. Specifically, a multi-granularity sparse relationship matrix prediction network is proposed, which utilizes three gated pooling modules focusing on filtering negative samples at different granularities, thereby obtaining a sparse relationship matrix containing entity pairs most likely to form relations. Finally, a set of sparse, most probable subject-object pairs can be constructed and used for relation decoding. Experimental results on multiple datasets demonstrate that our method achieves a new state-of-the-art overall performance. Our code is available at https://github.com/wanglei0618/Mg-RMPN." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.605, + 0.741, + 0.633 + ], + "angle": 0, + "content": "Keywords: Scene Graph Generation \\(\\cdot\\) End-to-End \\(\\cdot\\) Sparse Relationship Matrix \\(\\cdot\\) Multi-Granularity" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.658, + 0.377, + 0.674 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Scene Graph Generation (SGG) is a fundamental visual comprehension task that captures semantic information by detecting relation triplets \\(<\\) subject entity, predicate, object entity \\(>\\) in an image. This structured representation can facilitate many downstream tasks, such as image captioning [4, 41], visual question answering [9, 28], image retrieval [11, 25] and image generation [10, 19]. The end-to-end SGG methods [6, 16, 30] can directly generate sparse relations from an image based on a fixed number of relation queries, avoiding dense prediction that contains a lot of background relations in two-stage SGG. This leads to faster inference speed, and because fewer background relations are predicted, it exhibits superior predictive performance for rare samples of tail classes." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.37, + 0.129 + ], + "angle": 0, + "content": "L. Wang et al." + }, + { + "type": "image", + "bbox": [ + 0.273, + 0.145, + 0.425, + 0.311 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.334, + 0.318, + 0.353, + 0.33 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.434, + 0.144, + 0.596, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.503, + 0.318, + 0.523, + 0.33 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.604, + 0.155, + 0.73, + 0.316 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.657, + 0.318, + 0.674, + 0.33 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.341, + 0.788, + 0.397 + ], + "angle": 0, + "content": "Fig. 1: (a) End-to-end SGG pipeline. (b) Our proposed end-to-end pipeline based on multi-granularity sparse relationship matrix prediction network. (c) The average number of entities and relations in the VG150, OIV6, and GQA200 datasets demonstrates the sparsity of relations in an image." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.429, + 0.788, + 0.55 + ], + "angle": 0, + "content": "Current end-to-end approaches [6,16,22] still lag behind two-stage methods in performance, which is attributed to their reliance solely on visual representations to separately detect entities and relations, as shown in Figure 1 (a). This leads to two drawbacks: (1) It requires post-processing to match the corresponding subjects and objects for the relations, and the model's performance is constrained by the capabilities of matching. (2) Due to the inability to determine the subject and object in advance, entity predictions cannot help with relation prediction, and relation prediction can only rely on visual features." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.552, + 0.788, + 0.672 + ], + "angle": 0, + "content": "In this paper, we introduce a sparse relationship matrix that bridges entity detection and relation detection. Our approach not only eliminates the need for relation matching, but also leverages the semantics and positional information of predicted entities to enhance relation prediction. Figure 1 (b) shows our pipeline based on the \"pair then relation\" [30] framework, which first constructs a relationship matrix based on entity predictions to represent the possibility of forming a relation between two entities and then selects the sparsest, most likely subject-object pairs for relation prediction." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.673, + 0.788, + 0.779 + ], + "angle": 0, + "content": "Compared to all possible pairs formed by the entity proposals, the subject-object pairs with relations are extremely sparse. Figure 1 (c) illustrates the average entities and relations in the datasets. Taking VG150 [13] as an example, when generating \\( N_{e} = 100 \\) entities, there are a total of \\( (100(100 - 1)) = 9900 \\) possible combinations. But the average relations per image is only 5.8, which means that more than 99. \\( 9\\% \\) pairs are negative samples. Therefore, directly learning this extremely sparse relationship matrix is very challenging." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In this work, we propose a Multi-granularity sparse Relationship Matrix Prediction Network (Mg-RMPN) to partition the negative samples into different granularities, achieving layer-by-layer filtering of dense negative samples. Specifically, Mg-RMPN employs three Gated Pooling Modules (GPM) with identical" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.655, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Mg-RMPN" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.787, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.253 + ], + "angle": 0, + "content": "structures but independent parameters, each of which focuses on filtering negative samples at different granularities. The GPMs are used to generate relation matrices at different granularities and construct the corresponding ground truth matrices for supervised learning. The experimental results from various datasets reveal that our approach not only accurately identifies numerous head-class samples, but also excels in predicting sparse tail-class samples, achieving state-of-the-art overall performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.253, + 0.788, + 0.344 + ], + "angle": 0, + "content": "The contributions are summarized as follows: (1) An end-to-end SGG framework based on the relationship matrix is proposed that bridges entity detection and relation detection. (2) An Mg-RMPN is proposed to achieve sparse relationship matrix prediction based on multi-granularity negative sample learning. (3) Experimental results have confirmed that our method achieves state-of-the-art overall performance in both end-to-end and two-stage methods." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.362, + 0.398, + 0.379 + ], + "angle": 0, + "content": "2 Related Works" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.389, + 0.468, + 0.404 + ], + "angle": 0, + "content": "2.1 Scene Graph Generation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.41, + 0.789, + 0.622 + ], + "angle": 0, + "content": "Early work [5, 27, 32, 36] in SGG is based on the two-stage pipeline, first employing a Faster-RCNN [24] object detector to generate entity proposals, and the class of each entity is predicted. Then, entities are paired to form relations, and all possible pairs are classified. Its drawback lies that dense relation pairs include numerous background relations, which dilute the sparse tail-class samples, thus aggravating the imbalance issue. Therefore, some recent efforts [8, 23] have focused on addressing the imbalanced predictions, developing a series of rebalancing strategies such as logit adjustment [2, 26], loss reweighting [18, 33, 35], and data resampling [17, 31, 37]. Unlike these methods, we introduce a relationship matrix to generate sparse subject-object pairs for relation prediction. The sparse relations generated in this way have a higher signal-to-noise ratio, which can alleviate the imbalance problem. The results of the experiments validate that our method maintains outstanding performance for head classes and also demonstrates high performance for tail classes." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.639, + 0.57, + 0.654 + ], + "angle": 0, + "content": "2.2 End-to-End Scene Graph Generation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.789, + 0.841 + ], + "angle": 0, + "content": "End-to-end SGG, inspired by DETR [1], generates sparse relations directly from an image based on queries. For instance, [6] introduced the Relation Transformer to directly generate a set of relations from visual features. However, the generated relations did not match the corresponding subjects and objects, hence [16] proposed a graph assembling module to address the relation matching. [30] further introduced a \"pair then relation\" framework that predetermines subject-object pairs, circumventing the relation matching. However, these approaches suffer from the issue of relying solely on visual representations to detect relations and entities separately. Therefore, this paper introduces Mg-RMPN to predict a sparse relationship matrix to link entity detection and relation detection, not only eliminating the need for relation matching but also utilizing the predicted entity semantics and positional information to assist in relation prediction." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.37, + 0.129 + ], + "angle": 0, + "content": "L. Wang et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.145, + 0.331, + 0.161 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.178, + 0.422, + 0.192 + ], + "angle": 0, + "content": "3.1 Problem Definition" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.203, + 0.788, + 0.292 + ], + "angle": 0, + "content": "SGG aims to generate a scene graph \\(\\mathcal{G} = \\{\\mathcal{E},\\mathcal{R}\\}\\) from an input image \\(\\mathcal{I}\\). The scene graph consists of a set of \\(n\\) entities \\(\\mathcal{E} = \\{e_i\\}_{i=1}^n\\) and a set of \\(m\\) relations \\(\\mathcal{R} = \\{r_k\\}_{k=1}^m\\) between entities. The set of entities \\(\\mathcal{E}\\) can be further decomposed into a set of bounding boxes \\(\\mathcal{B} = \\{\\mathbf{b}_i\\}_{i=1}^n\\) and a set of class labels \\(\\mathcal{C} = \\{c_i\\}_{i=1}^n\\). The generation of a scene graph \\(\\mathcal{G}\\) can be formulated as the joint probability distribution:" + }, + { + "type": "equation", + "bbox": [ + 0.379, + 0.295, + 0.785, + 0.31 + ], + "angle": 0, + "content": "\\[\n\\Pr (\\mathcal {G} | \\mathcal {I}) = \\Pr (\\mathcal {B}, \\mathcal {C} | \\mathcal {I}) \\Pr (\\mathcal {R} | \\mathcal {I}, \\mathcal {B}, \\mathcal {C}), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.317, + 0.788, + 0.395 + ], + "angle": 0, + "content": "where \\(\\operatorname{Pr}(\\mathcal{B},\\mathcal{C}|\\mathcal{I})\\) represents the entity representation obtained from the object detector, including the class labels and bounding boxes. \\(\\operatorname{Pr}(\\mathcal{R}|\\mathcal{I},\\mathcal{B},\\mathcal{C})\\) represents the relation prediction based on pairs of entities by the relation decoder. The process from \\(\\operatorname{Pr}(\\mathcal{B},\\mathcal{C}|\\mathcal{I})\\) to \\(\\operatorname{Pr}(\\mathcal{R}|\\mathcal{I},\\mathcal{B},\\mathcal{C})\\) requires our Mg-RMPN to predict the pairs of entities most likely to form a relation from all possible pairs of entities." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.416, + 0.434, + 0.43 + ], + "angle": 0, + "content": "3.2 Overall Architecture" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.441, + 0.788, + 0.579 + ], + "angle": 0, + "content": "As shown in Figure 2 (a), our method comprises three modules: (1) entity detection, (2) relationship matrix prediction, and (3) relation prediction. The entity detection is responsible for generating a set of entities based on the object detector, including their visual representations, class labels, and bounding boxes. The relationship matrix prediction is responsible for producing a sparse relationship matrix that represents the relevance between two entities, and then obtaining the indices of the subject and object entities that are most likely to form a relation from all possible combinations. Finally, the relation of these sparse subject-object pairs is predicted through the relation decoder." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.6, + 0.398, + 0.616 + ], + "angle": 0, + "content": "3.3 Object Detector" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.625, + 0.788, + 0.763 + ], + "angle": 0, + "content": "This paper employs a deformable DETR [42] as the object detector, which contains a transformer encoder-decoder architecture with a set of entity queries. Here, the entity queries interact with encoder features through multi-scale deformable attention. Given an image \\(\\mathcal{I}\\), the object detector produces \\(N_{e}\\) entity representations \\(\\mathbf{Q}_e = \\{\\mathbf{q}_i^e\\}_{i=1}^{N_e} \\in \\mathbb{R}^{N_e \\times d}\\) from a set of learnable entity queries, where \\(d\\) is the embedding dimensions. Based on \\(\\mathbf{Q}_e\\), the object detector then follows with a classification head and a regression head to predict the entity's class predictions \\(\\mathbf{C} = \\{\\hat{\\mathbf{c}}_i\\}_{i=1}^n \\in \\mathbb{R}^{N_e \\times C_e}\\) and bounding boxes \\(\\mathbf{B} = \\{\\hat{\\mathbf{b}}_i\\}_{i=1}^n \\in \\mathbb{R}^{N_e \\times 4}\\), respectively, where \\(C_e\\) is the number of entity classes." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.784, + 0.753, + 0.8 + ], + "angle": 0, + "content": "3.4 Multi-Granularity Relationship Matrix Prediction Network" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Given \\(N_{e}\\) entities, Multi-granularity sparse Relationship Matrix Prediction Network (Mg-RMPN) generates a \\(N_{e} \\times N_{e}\\) adjacency relationship matrix \\(\\mathbf{M}_r\\). Each" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.655, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Mg-RMPN" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.231, + 0.144, + 0.77, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.237, + 0.256, + 0.734, + 0.425 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.435, + 0.785, + 0.464 + ], + "angle": 0, + "content": "Fig. 2: (a) Overall pipeline of our end-to-end SGG method. (b) Multi-granularity sparse Relationship Matrix Prediction Network (Mg-RMPN). GPM is Gated Pooling Module." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.492, + 0.784, + 0.537 + ], + "angle": 0, + "content": "node represents the probability of forming a relationship between two entities. Due to the sparse entities and relations in an image, \\(\\mathbf{M}_r\\) contains a large number of pairs without relations, that is, negative samples." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.538, + 0.785, + 0.628 + ], + "angle": 0, + "content": "Therefore, Mg-RMPN defined three \\(\\mathbf{M}_r\\) based on multi-granularity negative samples, where the three \\(\\mathbf{M}_r\\) are learned through three structurally identical but parameter-independent Gated Pooling Modules (GPM), as shown in Figure 2 (b). With the outputs of the GPMs, Mg-RMPN is able to generate the final sparse relationship matrix, enabling the identification of most probable subject-object pairs to predict relations." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.649, + 0.784, + 0.743 + ], + "angle": 0, + "content": "Gated Pooling Module Unlike [30], which only uses visual features, our GPM determines the correlation of two entities based on their visual representation, linguistic prior and location information. Given a set of visual representations \\(\\mathbf{Q}_e\\) of the entities in an image, GPM first uses two fully connected networks \\(\\mathcal{F}_{sub}\\) and \\(\\mathcal{F}_{obj}\\) to project it to the visual representations of the subject and object, \\(\\mathbf{Q}_{sub}\\) and \\(\\mathbf{Q}_{obj}\\), respectively, as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.273, + 0.751, + 0.785, + 0.772 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Q} _ {s u b} = \\mathcal {F} _ {s u b} \\left(\\mathbf {Q} _ {e}\\right) = \\left\\{\\mathbf {q} _ {i} ^ {s u b} \\right\\} _ {i = 1} ^ {N _ {e}}, \\quad \\mathbf {Q} _ {o b j} = \\mathcal {F} _ {o b j} \\left(\\mathbf {Q} _ {e}\\right) = \\left\\{\\mathbf {q} _ {j} ^ {o b j} \\right\\} _ {j = 1} ^ {N _ {e}}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.784, + 0.811 + ], + "angle": 0, + "content": "Subsequently, the visual similarity between the subject \\(i\\) and the object \\(j\\) can be calculated as \\(v_{ij}^{sim} = \\mathbf{q}_i^{sub} \\cdot (\\mathbf{q}_j^{obj})^{\\mathrm{T}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.811, + 0.785, + 0.842 + ], + "angle": 0, + "content": "Given a set of class prediction distributions \\(\\mathbf{C}\\) for entities, GPM uses the prior linguistic function \\(\\mathcal{F}_{prior}\\) from [26] to predict the correlation between two" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.37, + 0.129 + ], + "angle": 0, + "content": "L. Wang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.784, + 0.207 + ], + "angle": 0, + "content": "entities, \\( s_{ij}^{prior} = \\mathcal{F}_{prior}(\\hat{\\mathbf{c}}_i,\\hat{\\mathbf{c}}_j) \\). \\( \\mathcal{F}_{prior} \\) aims to capture the linguistic inductive bias that exists between two entities to enhance the prediction of the relational matrices. Similarly to [26], \\( \\mathcal{F}_{prior} \\) obtains a reasonable initialization based on the distribution of relations generated by entity pairs in the training dataset." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.208, + 0.785, + 0.35 + ], + "angle": 0, + "content": "Based on the positions \\(\\mathbf{B}\\) of the entities, GPM uses the function \\(\\mathcal{F}_{\\text{overlap}}\\) to obtain the relative position information between two entities \\(\\mathbf{b}_{ij}^{\\text{overlap}} = \\mathcal{F}_{\\text{overlap}}(\\hat{\\mathbf{b}}_i, \\hat{\\mathbf{b}}_j) \\in \\mathbb{R}^8\\), where 8 elements indicate difference in \\(x\\) and \\(y\\) coordinates, width ratio, height ratio, IOU, center distance, area ratio, and aspect ratio. Additionally, we can also obtain the union bounding box \\(\\hat{\\mathbf{b}}_{ij}^u \\in \\mathbb{R}^9\\) for the entity pairs, where 9 elements indicate the union bounding box coordinates \\((x_1, y_1, x_2, y_2)\\), center \\((\\frac{x_1 + x_2}{2}, \\frac{y_1 + y_2}{2})\\), size \\((x_2 - x_1, y_2 - y_1)\\) and area \\((x_2 - x_1)(y_2 - y_1)\\), respectively. Then, we can obtain the relevance representation of the subject \\(i\\) and object \\(j\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.389, + 0.361, + 0.785, + 0.381 + ], + "angle": 0, + "content": "\\[\n\\mathbf {e} _ {i j} = \\left[ v _ {i j} ^ {\\text {s i m}}, s _ {i j} ^ {\\text {p r i o r}}, \\mathbf {b} _ {i j} ^ {\\text {o v e r l a p}}, \\hat {\\mathbf {b}} _ {i j} ^ {u} \\right], \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.39, + 0.784, + 0.421 + ], + "angle": 0, + "content": "where \\([\\cdot, \\cdot]\\) denotes the concatenation operation. Subsequently, the feature representation of the entity pair can be expressed as" + }, + { + "type": "equation", + "bbox": [ + 0.323, + 0.432, + 0.785, + 0.453 + ], + "angle": 0, + "content": "\\[\n\\mathbf {f} _ {i j} = [ \\mathbf {q} _ {i} ^ {s u b}, \\mathbf {q} _ {j} ^ {o b j}, f _ {s} (s _ {i j} ^ {p r i o r}), p o s (\\mathbf {b} _ {i j} ^ {o v e r l a p}), p o s (\\hat {\\mathbf {b}} _ {i j} ^ {u}) ], \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.461, + 0.784, + 0.519 + ], + "angle": 0, + "content": "Where \\( f_{s} \\) is a fully-connected layer for language prior encoding, and \\( pos \\) is a fully-connected layer for positional encoding. Then, GPM uses a gated pooling layer \\( f_{gate} \\) to obtain the probability of relation between subject \\( i \\) and object \\( j \\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.379, + 0.52, + 0.785, + 0.541 + ], + "angle": 0, + "content": "\\[\n\\hat {p} _ {i j} = \\operatorname {S i g m o i d} \\left(\\sum f _ {\\text {g a t e}} \\left(\\mathbf {f} _ {i j}\\right) \\odot \\mathbf {e} _ {i j}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.547, + 0.784, + 0.606 + ], + "angle": 0, + "content": "where \\(\\odot\\) represents element-wise multiplication. Finally, we can obtain a \\(N_{e} \\times N_{e}\\) relationship matrix \\(\\mathbf{M}_{r} = \\mathrm{GPM}(\\mathbf{Q}_{e},\\mathbf{C},\\mathbf{B})\\). Note that we set the diagonal elements of \\(\\mathbf{M}_{r}\\) to zero, removing the cases where an entity serves as both subject and object simultaneously." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.784, + 0.704 + ], + "angle": 0, + "content": "Multi-Granularity Sparse Relationship Matrix Prediction Based on the output of three GPMs, we can construct three relationship matrices based on the granularity of negative samples and build the corresponding ground truth matrices (see Section 3.6) for supervised learning, enabling each GPM to focus on identifying negative samples of specific granularity." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.785, + 0.84 + ], + "angle": 0, + "content": "Since the \\( N_{e} \\) entity proposals are typically much greater than the number of entities contained in an image, resulting in a large number of background entities among all possible pairs of entities. Since background entities do not form relationships, we divide the negative samples in the relationship matrix into three levels of granularity: (1) high-confidence negative samples composed of background entities (background-background), (2) medium-confidence negative samples constructed from background and foreground entities (entity-background or background-entity), and (3) low-confidence negative samples composed of foreground entities (entity-entity)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.655, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Mg-RMPN" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.784, + 0.175 + ], + "angle": 0, + "content": "Then we can define three relationship matrices based on the output of the GPMs as follows." + }, + { + "type": "equation", + "bbox": [ + 0.395, + 0.189, + 0.562, + 0.202 + ], + "angle": 0, + "content": "\\[\n\\mathbf {M} _ {r 1} = \\operatorname {G P M - 1} \\left(\\mathbf {Q} _ {e}, \\mathbf {C}, \\mathbf {B}\\right)\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.395, + 0.207, + 0.784, + 0.22 + ], + "angle": 0, + "content": "\\[\n\\mathbf {M} _ {r 2} = \\mathbf {M} _ {r 1} + \\operatorname {G P M - 2} \\left(\\mathbf {Q} _ {e}, \\mathbf {C}, \\mathbf {B}\\right) \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.395, + 0.224, + 0.609, + 0.238 + ], + "angle": 0, + "content": "\\[\n\\mathbf {M} _ {r 3} = \\mathbf {M} _ {r 2} + \\operatorname {G P M - 3} (\\mathbf {Q} _ {e}, \\mathbf {C}, \\mathbf {B})\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.254, + 0.784, + 0.359 + ], + "angle": 0, + "content": "In Eq.6, each GPM can focus on learning the filtering of negative samples at different granularities, as follows: (1) GPM-1 focuses on learning from pairs between background and filters out a large number of high-confidence negative samples. (2) GPM-2 focuses on learning from pairs between background and entities and filters out medium-confidence negative samples. (3) GPM-3 focuses on learning from no relation pairs between entities and only needs to filter out entity pairs that do not form relationships from the remaining entity pairs." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.362, + 0.784, + 0.421 + ], + "angle": 0, + "content": "Based on the final sparse relationship matrix \\(\\mathbf{M}_{r3}\\), we can select the top-\\(N_{r}\\) subject-object pairs most likely to form relations and obtain their indices in the entity visual representations \\(\\mathbf{Q}_e\\), the entity's class predictions \\(\\mathbf{C}\\), and the bounding boxes \\(\\mathbf{B}\\)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.451, + 0.406, + 0.465 + ], + "angle": 0, + "content": "3.5 Relation Decoder" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.484, + 0.784, + 0.529 + ], + "angle": 0, + "content": "With the \\(\\mathbf{Q}_e\\), \\(\\mathbf{C}\\) and \\(\\mathbf{B}\\) of subjects and objects, we can obtain the subject-object pair representation \\(\\mathbf{Q}_{pair}\\), which integrates visual, semantic, and positional encodings as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.269, + 0.544, + 0.784, + 0.563 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Q} _ {p a i r} = \\left[ \\mathbf {Q} _ {e} ^ {s u b}, e m b (\\mathbf {C} ^ {s u b}), p o s (\\mathbf {B} ^ {s u b}), \\mathbf {Q} _ {e} ^ {o b j}, e m b (\\mathbf {C} ^ {o b j}), p o s (\\mathbf {B} ^ {o b j}) \\right]. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.576, + 0.784, + 0.592 + ], + "angle": 0, + "content": "where \\(emb\\) is a pre-trained Glove language model to acquire the word embedding." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.594, + 0.785, + 0.702 + ], + "angle": 0, + "content": "In this paper, we adopt the relation decoder in [30], which consists of transformer decoders in the style of DETR, to predict relations. In the relation decoder, we initialize a relation query \\(\\mathbf{Q}_r \\in \\mathbb{R}^{N_r \\times d}\\) as the query input, and the subject-object pair \\(\\mathbf{Q}_{pair}\\) is projected as the key and value of cross-attention. Subsequently, the relation representation after the Relation Decoder (RD) can be expressed as \\(\\tilde{\\mathbf{Q}}_r = \\mathrm{RD}(\\mathbf{Q}_r, \\mathbf{Q}_{pair})\\). Then, the relation prediction for the entity pairs can be expressed as \\(\\hat{\\mathbf{r}} = W_{cls} \\tilde{\\mathbf{Q}}_r\\), where \\(W_{cls}\\) is the liner relation classifier." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.732, + 0.334, + 0.747 + ], + "angle": 0, + "content": "3.6 Training" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.784, + 0.84 + ], + "angle": 0, + "content": "Our end-to-end SGG is divided into three subtasks: (1) an entity detection task based on the object detector, (2) a sparse relationship matrix prediction task based on the Mg-RMPN, and (3) a relation prediction task based on the relation decoder. During training, each subtask generates the corresponding supervisory information and losses, as follows." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.37, + 0.129 + ], + "angle": 0, + "content": "L. Wang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.254 + ], + "angle": 0, + "content": "Entity Detection Loss We uses the end-to-end deformable DETR [42] as object detector, which uses the set prediction loss proposed in DETR [1] by assigning the ground truth entities to the predictions. A cost function is applied to compute the matching cost between a prediction and a ground-truth entity. With the cost matrix, the entity prediction-ground truth assignment is computed with the Hungarian Matching [14]. Given a set of \\( N_{e} \\) entity proposals \\( \\{e_i\\}_{i=1}^{N_e} \\) from object detector, the set prediction loss for can be presented as:" + }, + { + "type": "equation", + "bbox": [ + 0.399, + 0.265, + 0.786, + 0.283 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {e} = \\Sigma_ {i = 1} ^ {N _ {e}} \\left[ \\mathcal {L} _ {c l s} ^ {e} + \\mathbf {1} _ {e _ {i} \\neq \\phi} \\mathcal {L} _ {b o x} ^ {e} \\right], \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.295, + 0.79, + 0.341 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{cls}^{e}\\) denotes the cross-entropy loss for label classification and \\(e_i \\neq \\phi\\) means that \\(<\\text{background}>\\) is not assigned to the \\(i\\)th entity prediction. \\(\\mathcal{L}_{box}^{e}\\) consists of \\(L_1\\) loss and generalized IoU loss for the box regression." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.365, + 0.789, + 0.412 + ], + "angle": 0, + "content": "Multi-granularity Relationship Matrices Loss To predict the final sparse relationship matrix, we need to define the supervisory information and loss functions during the training phase." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.429, + 0.789, + 0.503 + ], + "angle": 0, + "content": "Multi-granularity Supervision Matrices In the training phase, we assign labels to the entity proposals \\(\\{e_i\\}_{i=1}^{N_e}\\) by Hungarian matching and obtain the ground truth for each node in \\(\\mathbf{M}_r\\) regarding the subject, object, and whether they form a relation. In response to the sparsity of \\(\\mathbf{M}_r\\), we propose Mg-RMPN to achieve hierarchical identification of negative samples at different granularities." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.504, + 0.789, + 0.549 + ], + "angle": 0, + "content": "Consequently, based on the granularity of the negative samples that each GPM focuses on in Section 3.4, the corresponding ground truth can be obtained. (1) \\(\\mathbf{M}_{r1}^{gt}\\) filters out all pairs composed of background entities as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.241, + 0.56, + 0.786, + 0.595 + ], + "angle": 0, + "content": "\\[\n\\mathbf {M} _ {r 1} ^ {g t} = \\left\\{ \\right.p _ {i j} \\mid p _ {i j} = \\left\\{\\begin{array}{l l}0&\\text {i f} (e _ {i} = \\phi \\text {a n d} e _ {j} = \\phi) \\text {o r} i = j\\\\1&\\text {o t h e r w i s e}\\end{array}, 1 \\leq i \\leq N _ {e}, 1 \\leq j \\leq N _ {e} \\right\\}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.605, + 0.789, + 0.651 + ], + "angle": 0, + "content": "where \\( e_i = \\phi \\) (\\( e_j = \\phi \\)) means that \\( <\\text{background}> \\) is assigned to the \\( i \\)th (\\( j \\)th) entity prediction, \\( i = j \\) means removing self-connected entity pairs. (2) \\( \\mathbf{M}_{r2}^{gt} \\) further filters out all pairs composed of backgrounds and entities as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.243, + 0.662, + 0.786, + 0.697 + ], + "angle": 0, + "content": "\\[\n\\mathbf {M} _ {r 2} ^ {g t} = \\left\\{ \\right.p _ {i j} \\mid p _ {i j} = \\left\\{\\begin{array}{l l}0&\\text {i f} (e _ {i} = \\phi \\text {o r} e _ {j} = \\phi) \\text {o r} i = j\\\\1&\\text {o t h e r w i s e}\\end{array}, 1 \\leq i \\leq N _ {e}, 1 \\leq j \\leq N _ {e} \\right\\}. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.708, + 0.789, + 0.739 + ], + "angle": 0, + "content": "(3) \\(\\mathbf{M}_{r3}^{gt}\\) filters out all non-relationship pairs, and \\(\\mathbf{M}_{r3}^{gt}\\) is also the final ground truth relationship matrix of the scene graph, expressed as follows" + }, + { + "type": "equation", + "bbox": [ + 0.317, + 0.751, + 0.786, + 0.784 + ], + "angle": 0, + "content": "\\[\n\\mathbf {M} _ {r 3} ^ {g t} = \\left\\{p _ {i j} \\mid p _ {i j} = \\left\\{ \\begin{array}{l l} 0 & \\text {o t h e r w i s e} \\\\ 1 & \\text {i f} r _ {i j} = 1 \\end{array} , 1 \\leq i \\leq N _ {e}, 1 \\leq j \\leq N _ {e} \\right\\}, \\right. \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.795, + 0.789, + 0.845 + ], + "angle": 0, + "content": "where \\(r_{ij} = 1\\) implies that there is a relation between subject \\(i\\) and object \\(j\\). Here, the relationships of the negative sample subsets in \\(\\mathbf{M}_{r1}^{gt}\\), \\(\\mathbf{M}_{r2}^{gt}\\) and \\(\\mathbf{M}_{r3}^{gt}\\) are \\(\\mathbf{M}_{r1|p_{ij} = 0}^{gt} \\subseteq \\mathbf{M}_{r2|p_{ij} = 0}^{gt} \\subseteq \\mathbf{M}_{r3|p_{ij} = 0}^{gt}\\)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.655, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Mg-RMPN" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.147, + 0.784, + 0.218 + ], + "angle": 0, + "content": "Multi-granularity Relationship Matrices Loss Considering that the large number of negative samples in \\(\\mathbf{M}_r\\) have different confidence, we modify the binary cross-entropy function using focal loss [20], which applies weighted discrimination on well-classified samples, forcing the model to focus on wrongly classified samples. Then, the multi-granularity learning loss is" + }, + { + "type": "equation", + "bbox": [ + 0.293, + 0.227, + 0.784, + 0.243 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {M g} = \\mathcal {L} _ {R M} ^ {1} \\left(\\mathbf {M} _ {r 1}, \\mathbf {M} _ {r 1} ^ {g t}\\right) + \\mathcal {L} _ {R M} ^ {2} \\left(\\mathbf {M} _ {r 2}, \\mathbf {M} _ {r 2} ^ {g t}\\right) + \\mathcal {L} _ {R M} ^ {3} \\left(\\mathbf {M} _ {r 3}, \\mathbf {M} _ {r 3} ^ {g t}\\right), \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.22, + 0.251, + 0.697, + 0.268 + ], + "angle": 0, + "content": "where the relationship matrices loss \\(\\mathcal{L}_{RM}^{k}(k = 1,2,3)\\) is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.309, + 0.277, + 0.785, + 0.341 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {R M} ^ {k} \\left(\\mathbf {M} _ {r}, \\mathbf {M} _ {r} ^ {g t}\\right) = - \\sum_ {i} \\sum_ {j} \\left\\{\\alpha_ {k} \\left(1 - \\hat {p} _ {i j}\\right) ^ {\\gamma} p _ {i j} \\log (\\hat {p} _ {i j}) \\right. \\\\ + (1 - \\alpha_ {k}) \\hat {p} _ {i j} ^ {\\gamma} (1 - p _ {i j}) \\log (1 - \\hat {p} _ {i j}) \\} / N _ {p o s} \\tag {13} \\\\ + \\left\\| \\mathbf {M} _ {r} \\right\\| _ {1} + \\left\\| \\mathbf {M} _ {r} \\right\\| _ {2}, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.35, + 0.784, + 0.41 + ], + "angle": 0, + "content": "where the first part is the focal loss based on binary cross-entropy, and the second part is the L1 and L2 regularization of \\(\\mathbf{M}_r\\), \\(\\alpha_k(k = 1,2,3)\\) is hyperparameters that adjusts the weights of positive and negative samples, \\(\\gamma\\) is a hyperparameter for focal loss, and \\(N_{pos}\\) is the number of positive samples in each mini-batch." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.431, + 0.784, + 0.514 + ], + "angle": 0, + "content": "Relation Prediction Loss In this paper, the relation decoder also considers relation prediction as a set prediction based on a query. Therefore, during the training phase, we also use Hungarian Matching to assign labels for relation prediction. Due to the imbalanced relation classes in SGG, we use Seesaw loss [29] to dynamically adjust the gradients of samples of different classes. The relation loss is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.347, + 0.524, + 0.784, + 0.56 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {r} = - \\sum_ {i = 1} ^ {C _ {r}} y _ {i} \\log \\left(\\hat {\\sigma} _ {i}\\right), \\quad \\hat {\\sigma} _ {i} = \\frac {e ^ {\\hat {r} _ {i}}}{\\sum_ {i \\neq j} ^ {C _ {r}} \\mathcal {S} _ {i j} e ^ {\\hat {r} _ {j}} + e ^ {\\hat {r} _ {i}}}, \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.568, + 0.784, + 0.614 + ], + "angle": 0, + "content": "where \\( C_r \\) is the number of relation classes, \\( y_i \\in \\{0,1\\} \\), \\( 1 \\leqslant i \\leqslant C_r \\) is the one-hot ground truth label, \\( S_{ij} \\) is a tunable balancing factor between different classes, detail can be found in [29]." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.615, + 0.555, + 0.629 + ], + "angle": 0, + "content": "In summary, the total loss of our method is" + }, + { + "type": "equation", + "bbox": [ + 0.413, + 0.644, + 0.784, + 0.657 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {e} + \\lambda_ {1} \\mathcal {L} _ {M g} + \\lambda_ {2} \\mathcal {L} _ {r}, \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.669, + 0.798, + 0.701 + ], + "angle": 0, + "content": "Where \\(\\lambda_{1}\\) and \\(\\lambda_{2}\\) are hyperparameters used to respectively adjust multi-granularity learning loss \\(\\mathcal{L}_{Mg}\\) and relation prediction loss \\(\\mathcal{L}_r\\)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.723, + 0.375, + 0.74 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.755, + 0.446, + 0.77 + ], + "angle": 0, + "content": "4.1 Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.784, + 0.841 + ], + "angle": 0, + "content": "Datasets: We evaluate our proposed method on the following datasets: (1) Visual Genome (VG150) [13] is the most widely used dataset in SGG, consisting of the most frequent 150 object classes and 50 predicate classes. (2) Open Images V6 (OIv6) [15] is a large-scale dataset proposed by Google. We follow" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.129 + ], + "angle": 0, + "content": "L. Wang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.22 + ], + "angle": 0, + "content": "the data processing and evaluation protocols in [17, 40]. OIv6 consists of 601 object classes and 30 predicate classes. (2) Generalized Question Answering (GQA200) [9] is another vision and language benchmark with more than 3.8M relation annotations. We follow the data processing in [7], which consists of Top-200 object classes and Top-100 predicate classes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.222, + 0.785, + 0.356 + ], + "angle": 0, + "content": "Task & Evaluation Metrics: In this work, we focus on the Scene Graph Detection task, which detects all objects in an image and predicts their bounding boxes, labels, and relations. The initial evaluation metric Recall@K (R@K) was found to be dominated by head classes. Therefore, the mean Recall@K (mR@K) across all relations classes is proposed to evaluate unbiased SGG. However, focusing solely on mR@K and neglecting R@K can result in a tail bias. Therefore, we adopts the harmonic mean F@K of R@K and mR@K as an overall metric. For OIv6, the weighted evaluation metrics (wmAP rel, wnAP phr, scorewtd) are used for a more class-balanced evaluation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.358, + 0.787, + 0.508 + ], + "angle": 0, + "content": "Implementation Details: We utilized the pre-trained deformable DETR [42] as the object detector. We set the loss hyperparameters \\(\\lambda_{1}\\) and \\(\\lambda_{2}\\) to 0.5 and 3 respectively. For the multi-granularity relationship matrices loss, we set \\(\\alpha_{k}(k = 1,2,3)\\) to 0.75, 0.9, and 0.99, respectively, to control the weight of positive samples, with the focal loss parameter \\(\\gamma\\) set to 2. The number of entities predicted by the detector \\(N_{e}\\) is 100, the number of most probable subject-object pairs selected \\(N_{r}\\) is 100, and the size of embedding dimensions \\(d\\) is 256. Our model is implemented on 8 NVIDIA 3090 GPUs with learning rate \\(\\times 10^{-4}\\) and batch size 16 for 24 epochs. More implementation details are shown in the Supplementary Material." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.53, + 0.55, + 0.545 + ], + "angle": 0, + "content": "4.2 Comparison with State of the Arts" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.553, + 0.785, + 0.674 + ], + "angle": 0, + "content": "VG150: Table 1 shows the comparison of our method on VG150, with methods divided into two-stage (above) and end-to-end (below). Due to the imbalance issue in SGG, the two-stage methods (Motifs, VCTree, GPS-Net, RelDN, PE-Net) perform better in R@K but show very poor performance in mR@K. Thus, many rebalancing techniques (VCTree+TDE, BGNN, SHA+GCL) have been developed. Although they improve mR@K, they inevitably impair R@K. For example, SHA+GCL, despite achieving the best mR@K, its recall performance significantly deteriorates, with R@50/100 being only 14.9/18.2." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.675, + 0.785, + 0.763 + ], + "angle": 0, + "content": "Although the end-to-end methods are inferior to the two-stage methods on R@K, it demonstrates better potential on mR@K due to its advantage in sparse relation prediction, which does not dilute the scarce tail-class samples. Furthermore, compared to the rebalancing techniques of the two-stage methods, the end-to-end approach does not overly harm R@K and can achieve superior overall performance F@K, such as SGTR and Pair-Net." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.785, + 0.84 + ], + "angle": 0, + "content": "Our Mg-RMPN leads comprehensively in end-to-end methods. Compared with two-stage methods, Mg-RMPN achieves competitive performance in R@K and significantly excels in mR@K, achieving the best overall performance F@K. For example, Mg-RMPN achieved 19.3/22.8 on F@50/100, surpassing the end-to-end and two-stage optimal methods Pair-Net and PE-Net, respectively." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.655, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Mg-RMPN" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.145, + 0.787, + 0.174 + ], + "angle": 0, + "content": "Table 1: Performance comparison at scene graph detection task on VG150. * denotes the results from [16]." + }, + { + "type": "table", + "bbox": [ + 0.216, + 0.185, + 0.787, + 0.445 + ], + "angle": 0, + "content": "
MethodR@20 R@50 R@100mR@20 mR@50 mR@100F@20 F@50 F@100
Motifs [36]25.532.837.25.06.87.98.411.313.0
VCTree [27]24.531.936.25.47.48.78.812.014.0
VCTree+TDE [26]14.019.423.26.99.311.19.212.615.0
GPS-Net [21]22.328.933.26.98.79.810.513.415.1
BGNN [17]-31.035.8-10.712.6-15.918.6
RelDN [39]-31.435.9-6.07.3-10.112.1
SHA+GCL [7]-14.918.214.217.920.9-16.319.5
PE-Net [40]-30.735.2-12.414.5-17.720.5
FCSGG [22]16.121.325.12.73.64.24.66.27.2
RelTR [6]21.227.5-6.810.8-10.315.5-
AS-Net* [3]-18.721.1-6.17.2-9.210.7
HOTR* [12]-23.527.7-9.412.0-13.416.7
SGTR [16]-24.628.4-12.015.2-16.119.8
Pair-Net [30]18.824.929.38.912.415.412.116.620.2
Mg-RMPN(DETR)21.027.131.39.913.516.213.518.021.3
Mg-RMPN(Ours)22.529.133.510.314.417.314.119.322.8
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.498, + 0.787, + 0.604 + ], + "angle": 0, + "content": "We also compared with Pair-Net that is also based on relationship matrix learning. The experimental results show that our method is comprehensively superior to it, which demonstrates the effectiveness of our proposed Mg-RMPN. Furthermore, we also present the results of Mg-RMPN based on the DETR object detector to eliminate the influence of the detector. The results show that our Mg-RMPN(DETR) still achieves the best overall performance among both one-stage and two-stage methods." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.616, + 0.787, + 0.723 + ], + "angle": 0, + "content": "Open Image V6: To validate the generalizability of Mg-RMPN, we performed experiments on Open Images V6 in Table 2 and compared with two-stage (above) and end-to-end (below) methods. Compared to classic SGG benchmarks, our method achieved the optimal results among all methods with 45.5, 77.8, 57.4 respectively in mR@50, R@50, F@50. Compared to open image benchmarks, we obtained competitive 35.5 and 36.4 in \\(\\mathrm{wmAP}_{\\mathrm{rel}}\\) and \\(\\mathrm{wmAP}_{\\mathrm{phr}}\\), and achieved the best result 43.6 in the overall metric score\\(_{\\mathrm{wtd}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.787, + 0.84 + ], + "angle": 0, + "content": "GQA200: We also conducted experiments on the more challenging GQA200, as shown in Table 3. Compared to two-stage methods (VTransE, Motifs, VCTree, SHA+GCL), our method's overall performance F@K is only slightly inferior to SHA+GCL on F@100, which sacrificed a significant amount of R@K to enhance mR@K by using rebalancing techniques. Compared to the end-to-end method Pair-Net, which is also based on relationship matrices, our method is comprehensively superior." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.129 + ], + "angle": 0, + "content": "L. Wang et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.173 + ], + "angle": 0, + "content": "Table 2: Performance comparison on OIV6. * denotes the results from [16,17]. †denote results reproduced with the authors' code." + }, + { + "type": "table", + "bbox": [ + 0.264, + 0.185, + 0.74, + 0.389 + ], + "angle": 0, + "content": "
MethodmR@50R@50F@50wmAPrelwmAPphrscorewtd
Motifs* [36]32.771.644.929.931.638.9
VCTree* [27]33.974.146.534.233.140.2
RelDN* [39]34.073.146.432.233.440.8
G-RCNN* [34]34.074.546.733.234.241.8
GPS-Net* [21]35.374.848.032.934.041.7
BGNN* [17]40.575.052.633.534.242.1
RelTR [6]-71.7-37.237.543.0
AS-Ne* [3]35.255.343.025.927.532.4
HOTR* [12]40.152.745.519.421.526.7
SGTR* [16]42.659.949.837.038.742.3
Pair-Net† [30]44.577.456.531.832.440.3
Mg-RMPN(Ours)45.577.857.435.536.443.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.402, + 0.788, + 0.43 + ], + "angle": 0, + "content": "Table 3: Performance comparison at scene graph detection task on GQA200. * denotes the results from [7]. †denote results reproduced with the authors' code." + }, + { + "type": "table", + "bbox": [ + 0.275, + 0.441, + 0.73, + 0.564 + ], + "angle": 0, + "content": "
MethodR@50R@100mR@50mR@100F@50F@100
VTransE* [38]27.230.75.86.69.610.9
Motifs* [36]28.933.16.47.710.512.5
VCTree* [27]28.331.96.57.410.612.0
SHA+GCL* [7]14.817.917.820.116.218.9
Pair-Net† [30]20.223.410.612.613.916.4
Mg-RMPN (Ours)23.225.712.814.516.518.5
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.59, + 0.733, + 0.605 + ], + "angle": 0, + "content": "4.3 Visualization of Multi-Granularity Relationship Matrices" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.788, + 0.749 + ], + "angle": 0, + "content": "Figure 3 presents a visualization of the multi-granularity relationship matrices learned by our Mg-RMPN, where (d), (e), (f) are the sparse relationship matrices learned after being filtered by GPM-1, GPM-2, GPM-3, respectively, and (a), (b), (c) are the corresponding ground truth relationship matrices defined in Section 3.6. The visualization results show that the dense negative samples in the relationship matrix are filtered layer by layer according to different granularities. The consistency of the generated multi-granularity relationship matrices with their ground truth indicates that the three GPMs of Mg-RMPN indeed focus on filtering negative samples of different granularities." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Figures 3 (h) and (g), respectively, present a zoomed-in view of the final relationship matrix and its ground truth. We can observe that the predicted relationship matrix successfully captures the true relationships, for example: \"man wearing shirt\", \"man holding racket\", \"man wearing short\" and \"pole on fence\". Additionally, the relationship matrix also predicts some relationships that do not in the ground truth, as illustrated by the dashed lines in the rows" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.655, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Mg-RMPN" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.233, + 0.144, + 0.363, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.254, + 0.258, + 0.348, + 0.269 + ], + "angle": 0, + "content": "(a) Ground truth \\(\\mathbf{M}_{F1}^{gt}\\)" + }, + { + "type": "image", + "bbox": [ + 0.365, + 0.157, + 0.488, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.377, + 0.258, + 0.47, + 0.269 + ], + "angle": 0, + "content": "(b) Ground truth \\(\\mathbf{M}_{\\pmb{F}_2}^{gt}\\)" + }, + { + "type": "image", + "bbox": [ + 0.49, + 0.157, + 0.616, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.507, + 0.258, + 0.597, + 0.269 + ], + "angle": 0, + "content": "(c) Ground truth \\(\\mathbf{M}_{r3}^{gt}\\)" + }, + { + "type": "image", + "bbox": [ + 0.617, + 0.144, + 0.77, + 0.269 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.233, + 0.279, + 0.361, + 0.373 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.245, + 0.376, + 0.355, + 0.387 + ], + "angle": 0, + "content": "(d) Multi-granularity \\(\\mathbf{M}_{r1}\\)" + }, + { + "type": "image", + "bbox": [ + 0.362, + 0.279, + 0.488, + 0.374 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.371, + 0.376, + 0.481, + 0.387 + ], + "angle": 0, + "content": "(e) Multi-granularity \\(\\mathbf{M}_{r2}\\)" + }, + { + "type": "image", + "bbox": [ + 0.49, + 0.279, + 0.617, + 0.373 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.376, + 0.608, + 0.387 + ], + "angle": 0, + "content": "(f) Multi-granularity \\(\\mathbf{M}_{r3}\\)" + }, + { + "type": "image", + "bbox": [ + 0.617, + 0.271, + 0.771, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.285, + 0.406, + 0.716, + 0.421 + ], + "angle": 0, + "content": "Fig. 3: Visualization of Multi-Granularity Relationship Matrices" + }, + { + "type": "table_caption", + "bbox": [ + 0.285, + 0.437, + 0.716, + 0.451 + ], + "angle": 0, + "content": "Table 4: Ablation studies of Mg-RMPN components on VG150." + }, + { + "type": "table", + "bbox": [ + 0.248, + 0.463, + 0.756, + 0.591 + ], + "angle": 0, + "content": "
GPM-1GPM-2GPM-3R@20/50/100mR@20/50/100F@20/50/100
✓✘✘6.5/10.7/15.22.9/4.8/6.64.0/6.6/9.2
✘✓✘12.5/17.5/21.65.2/7.4/9.37.3/10.4/13.0
✓✓✘12.8/18.1/22.35.8/8.5/10.78.0/11.6/14.5
✘✘✓21.2/27.6/32.19.2/13.1/15.812.8/17.8/21.2
✓✘✓21.7/28.2/32.69.8/13.8/16.513.5/18.5/21.9
✘✓✓21.8/28.2/32.59.7/13.7/16.713.4/18.4/22.1
✓✓✓22.5/29.1/33.510.3/14.4/17.314.1/19.3/22.8
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.621, + 0.788, + 0.712 + ], + "angle": 0, + "content": "and columns, which represent the relationships with the corresponding entity \"man\" as the subject and object, respectively. Due to the abundant presence of \"man\" in SGG, the relationship matrix tends to favor predicting relations that include \"man\", which aligns with the actual language induction bias. Therefore, the results of visualization indicate that our Mg-RMPN can accurately predict the sparse relationship matrix." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.737, + 0.402, + 0.751 + ], + "angle": 0, + "content": "4.4 Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Component Analysis of Mg-RMPN. To assess the effectiveness of each GPM in Mg-RMPN, we performed ablation experiments in Table 4. In Mg-RMPN, each GPM focuses on negative sample filtering of different granularities, and the three GPMs collaborate to complete the learning of the sparse relationship matrix. GPM-1 and GPM-2 are used to help GPM-3 filter negative" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.129 + ], + "angle": 0, + "content": "L. Wang et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.225, + 0.144, + 0.777, + 0.159 + ], + "angle": 0, + "content": "Table 5: Ablation studies of different input information for Mg-RMPN on VG150." + }, + { + "type": "table", + "bbox": [ + 0.244, + 0.172, + 0.761, + 0.259 + ], + "angle": 0, + "content": "
Visual Semantic PositionR@20/50/100mR@20/50/100F@20/50/100
✓XX21.2/27.6/32.19.8/13.6/16.413.4/18.2/21.7
✓X✓22.2/28.7/33.110.1/14.1/16.913.9/18.9/22.4
✓✓X21.2/27.6/32.29.8/13.8/16.813.4/18.4/22.1
✓✓✓22.5/29.1/33.510.3/14.4/17.314.1/19.3/22.8
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.291, + 0.784, + 0.397 + ], + "angle": 0, + "content": "samples. Using GPM-1 and GPM-2 individually cannot make full use of the ground truth, hence their results are very poor. The results obtained using only GPM-3 are also not good enough. Further increasing the assistance of GPMs for negative sample filtering can significantly improve the performance of the model. Mg-RMPN fully utilized three GPMs to achieve optimal results, confirming that learning negative samples based on different granularities can effectively improve the model's learning ability for sparse relationship matrices." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.399, + 0.785, + 0.61 + ], + "angle": 0, + "content": "Different input information for Mg-RMPN This paper bridges entity detection and relation detection through a relationship matrix, making full use of the semantic and positional information of entities and solving the problem that end-to-end SGG relies solely on visual information to predict relations. Table 5 presents ablation studies of Mg-RMPN based on different input information. Compared to models that only utilize visual information, separately incorporating the semantic and positional information of entities can enhance the performance of the model. Furthermore, the improvement in introducing semantic and positional information simultaneously is significantly greater than in introducing a single modality of information, indicating that semantic and positional information are two complementary information. Hence, the semantic and positional information incorporated in our approach plays a crucial role in predicting relations. More hyperparameter analysis can be found in Supplementary Material." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.646, + 0.359, + 0.661 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.787, + 0.841 + ], + "angle": 0, + "content": "In this paper, our proposed end-to-end method bridges entity detection and relation detection through a sparse relationship matrix, which not only eliminates the need for post-processing of relation matching but also leverages the semantics and positional information of predicted entities to enhance relation prediction. To predict the sparse relationship matrix, we propose a multi-granularity sparse relationship matrix prediction network, which utilizes three gated pooling modules focusing on filtering negative samples at different granularities. Finally, a set of sparse, most likely subject-object pairs can be constructed and used for relation decoding. The experimental results demonstrate that our method achieves an optimal overall performance in both the end-to-end and two-stage methods." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.655, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Mg-RMPN" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.176, + 0.785, + 0.222 + ], + "angle": 0, + "content": "This work was supported by the National Natural Science Foundation of China (U21A20485, 61976170, 62088102) and the National Key R&D Program of China (2023YFB4704900)." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.244, + 0.323, + 0.26 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.274, + 0.785, + 0.315 + ], + "angle": 0, + "content": "1. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: European conference on computer vision. pp. 213-229. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.317, + 0.785, + 0.357 + ], + "angle": 0, + "content": "2. Chen, C., Zhan, Y., Yu, B., Liu, L., Luo, Y., Du, B.: Resistance training using prior bias: Toward unbiased scene graph generation. Proceedings of the AAAI Conference on Artificial Intelligence 36(1), 212-220 (Jun 2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.359, + 0.785, + 0.399 + ], + "angle": 0, + "content": "3. Chen, M., Liao, Y., Liu, S., Chen, Z., Wang, F., Qian, C.: Reformulating hoi detection as adaptive set prediction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9004-9013 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.401, + 0.785, + 0.44 + ], + "angle": 0, + "content": "4. Chen, S., Jin, Q., Wang, P., Wu, Q.: Say as you wish: Fine-grained control of image caption generation with abstract scene graphs. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9962-9971 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.442, + 0.785, + 0.481 + ], + "angle": 0, + "content": "5. Chen, T., Yu, W., Chen, R., Lin, L.: Knowledge-embedded routing network for scene graph generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6163-6171 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.483, + 0.785, + 0.522 + ], + "angle": 0, + "content": "6. Cong, Y., Yang, M.Y., Rosenhahn, B.: Reltr: Relation transformer for scene graph generation. IEEE Transactions on Pattern Analysis and Machine Intelligence (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.524, + 0.785, + 0.579 + ], + "angle": 0, + "content": "7. Dong, X., Gan, T., Song, X., Wu, J., Cheng, Y., Nie, L.: Stacked hybrid-attention and group collaborative learning for unbiased scene graph generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19427-19436 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.58, + 0.785, + 0.619 + ], + "angle": 0, + "content": "8. Gao, L., Lyu, X., Guo, Y., Hu, Y., Li, Y.F., Xu, L., Shen, H.T., Song, J.: Informative scene graph generation via debiasing. arXiv preprint arXiv:2308.05286 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.621, + 0.785, + 0.66 + ], + "angle": 0, + "content": "9. Hudson, D.A., Manning, C.D.: Gqa: A new dataset for real-world visual reasoning and compositional question answering. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 6700-6709 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.661, + 0.785, + 0.702 + ], + "angle": 0, + "content": "10. Johnson, J., Gupta, A., Fei-Fei, L.: Image generation from scene graphs. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1219-1228 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.703, + 0.785, + 0.743 + ], + "angle": 0, + "content": "11. Johnson, J., Krishna, R., Stark, M., Li, L.J., Shamma, D., Bernstein, M., Fei-Fei, L.: Image retrieval using scene graphs. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3668-3678 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.744, + 0.785, + 0.785 + ], + "angle": 0, + "content": "12. Kim, B., Lee, J., Kang, J., Kim, E.S., Kim, H.J.: Hotr: End-to-end human-object interaction detection with transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 74-83 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.786, + 0.785, + 0.84 + ], + "angle": 0, + "content": "13. Krishna, R., Zhu, Y., Groth, O., Johnson, J., Hata, K., Kravitz, J., Chen, S., Kalantidis, Y., Li, L.J., Shamma, D.A., et al.: Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision 123(1), 32-73 (2017)" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.274, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.129 + ], + "angle": 0, + "content": "L. Wang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.175 + ], + "angle": 0, + "content": "14. Kuhn, H.W.: The hungarian method for the assignment problem. Naval research logistics quarterly 2(1-2), 83-97 (1955)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.176, + 0.785, + 0.23 + ], + "angle": 0, + "content": "15. Kuznetsova, A., Rom, H., Alldrin, N., Uijlings, J., Krasin, I., Pont-Tuset, J., Kamali, S., Popov, S., Malloci, M., Kolesnikov, A., et al.: The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. International Journal of Computer Vision 128(7), 1956-1981 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.23, + 0.785, + 0.271 + ], + "angle": 0, + "content": "16. Li, R., Zhang, S., He, X.: Sgtr: End-to-end scene graph generation with transformer. In: proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 19486-19496 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.271, + 0.785, + 0.312 + ], + "angle": 0, + "content": "17. Li, R., Zhang, S., Wan, B., He, X.: Bipartite graph network with adaptive message passing for unbiased scene graph generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11109-11119 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.312, + 0.785, + 0.366 + ], + "angle": 0, + "content": "18. Li, W., Zhang, H., Bai, Q., Zhao, G., Jiang, N., Yuan, X.: Ppdl: Predicate probability distribution based loss for unbiased scene graph generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19447-19456 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.366, + 0.785, + 0.407 + ], + "angle": 0, + "content": "19. Li, Y., Ma, T., Bai, Y., Duan, N., Wei, S., Wang, X.: Pastegan: A semi-parametric method to generate image from scene graph. Advances in Neural Information Processing Systems 32 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.407, + 0.785, + 0.448 + ], + "angle": 0, + "content": "20. Lin, T.Y., Goyal, P., Girshick, R., He, K., Dollar, P.: Focal loss for dense object detection. In: Proceedings of the IEEE international conference on computer vision. pp. 2980-2988 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.448, + 0.785, + 0.489 + ], + "angle": 0, + "content": "21. Lin, X., Ding, C., Zeng, J., Tao, D.: Gps-net: Graph property sensing network for scene graph generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3746-3753 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.488, + 0.785, + 0.529 + ], + "angle": 0, + "content": "22. Liu, H., Yan, N., Mortazavi, M., Bhanu, B.: Fully convolutional scene graph generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11546-11556 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.529, + 0.785, + 0.556 + ], + "angle": 0, + "content": "23. Lyu, X., Gao, L., Xie, J., Zeng, P., Tian, Y., Shao, J., Shen, H.T.: Generalized unbiased scene graph generation. arXiv preprint arXiv:2308.04802 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.556, + 0.785, + 0.597 + ], + "angle": 0, + "content": "24. Ren, S., He, K., Girshick, R., Sun, J.: Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems 28 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.597, + 0.785, + 0.638 + ], + "angle": 0, + "content": "25. Schroeder, B., Tripathi, S.: Structured query-based image retrieval using scene graphs. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. pp. 178-179 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.638, + 0.785, + 0.678 + ], + "angle": 0, + "content": "26. Tang, K., Niu, Y., Huang, J., Shi, J., Zhang, H.: Unbiased scene graph generation from biased training. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 3716-3725 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.678, + 0.785, + 0.718 + ], + "angle": 0, + "content": "27. Tang, K., Zhang, H., Wu, B., Luo, W., Liu, W.: Learning to compose dynamic tree structures for visual contexts. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 6619-6628 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.718, + 0.785, + 0.759 + ], + "angle": 0, + "content": "28. Teney, D., Liu, L., van Den Hengel, A.: Graph-structured representations for visual question answering. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1-9 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.759, + 0.785, + 0.813 + ], + "angle": 0, + "content": "29. Wang, J., Zhang, W., Zang, Y., Cao, Y., Pang, J., Gong, T., Chen, K., Liu, Z., Loy, C.C., Lin, D.: Seesaw loss for long-tailed instance segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9695-9704 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.813, + 0.785, + 0.841 + ], + "angle": 0, + "content": "30. Wang, J., Wen, Z., Li, X., Guo, Z., Yang, J., Liu, Z.: Pair then relation: Pair-net for panoptic scene graph generation. arXiv preprint arXiv:2307.08699 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.655, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Mg-RMPN" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "31. Wang, L., Yuan, Z., Chen, B.: Learning to generate an unbiased scene graph by using attribute-guided predicate features. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 2581-2589 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.787, + 0.232 + ], + "angle": 0, + "content": "32. Xu, D., Zhu, Y., Choy, C.B., Fei-Fei, L.: Scene graph generation by iterative message passing. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5410-5419 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.787, + 0.273 + ], + "angle": 0, + "content": "33. Yang, G., Zhang, J., Zhang, Y., Wu, B., Yang, Y.: Probabilistic modeling of semantic ambiguity for scene graph generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12527-12536 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.787, + 0.314 + ], + "angle": 0, + "content": "34. Yang, J., Lu, J., Lee, S., Batra, D., Parikh, D.: Graph r-cnn for scene graph generation. In: Proceedings of the European conference on computer vision (ECCV). pp. 670-685 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.787, + 0.37 + ], + "angle": 0, + "content": "35. Yu, J., Chai, Y., Wang, Y., Hu, Y., Wu, Q.: Cogtree: Cognition tree loss for unbiased scene graph generation. In: Zhou, Z.H. (ed.) Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence, IJCAI-21. pp. 1274-1280. International Joint Conferences on Artificial Intelligence Organization (8 2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.371, + 0.787, + 0.412 + ], + "angle": 0, + "content": "36. Zellers, R., Yatskar, M., Thomson, S., Choi, Y.: Neural motifs: Scene graph parsing with global context. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5831-5840 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.412, + 0.787, + 0.438 + ], + "angle": 0, + "content": "37. Zhang, A., Yao, Y., Chen, Q., Ji, W., et al, L.: Fine-grained scene graph generation with data transfer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.787, + 0.481 + ], + "angle": 0, + "content": "38. Zhang, H., Kyaw, Z., Chang, S.F., Chua, T.S.: Visual translation embedding network for visual relation detection. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5532-5540 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.481, + 0.787, + 0.522 + ], + "angle": 0, + "content": "39. Zhang, J., Shih, K.J., Elgammal, A., Tao, A., Catanzaro, B.: Graphical contrastive losses for scene graph parsing. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11535-11543 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.522, + 0.787, + 0.564 + ], + "angle": 0, + "content": "40. Zheng, C., Lyu, X., Gao, L., Dai, B., Song, J.: Prototype-based embedding network for scene graph generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22783-22792 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.565, + 0.787, + 0.605 + ], + "angle": 0, + "content": "41. Zhong, Y., Wang, L., Chen, J., Yu, D., Li, Y.: Comprehensive image captioning via scene graph decomposition. In: European Conference on Computer Vision. pp. 211-229. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.606, + 0.787, + 0.646 + ], + "angle": 0, + "content": "42. Zhu, X., Su, W., Lu, L., Li, B., Wang, X., Dai, J.: Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159 (2020)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.646 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Multi-Granularity Sparse Relationship Matrix Prediction Network for End-to-End Scene Graph Generation/6b85535a-3ad0-4a54-882a-f4d4244fbf81_origin.pdf b/2024/Multi-Granularity Sparse Relationship Matrix Prediction Network for End-to-End Scene Graph Generation/6b85535a-3ad0-4a54-882a-f4d4244fbf81_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..07b5bfd528dd67c7e5e2567caa2e7b2f3074e852 --- /dev/null +++ b/2024/Multi-Granularity Sparse Relationship Matrix Prediction Network for End-to-End Scene Graph Generation/6b85535a-3ad0-4a54-882a-f4d4244fbf81_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e21c3fc84e6f9a1acd5c865379acdb3a74a51917eb3d5fd4b4b1e39b05c54b84 +size 891522 diff --git a/2024/Multi-Granularity Sparse Relationship Matrix Prediction Network for End-to-End Scene Graph Generation/full.md b/2024/Multi-Granularity Sparse Relationship Matrix Prediction Network for End-to-End Scene Graph Generation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..886a6530aae8385115019d1b0f5c639064911514 --- /dev/null +++ b/2024/Multi-Granularity Sparse Relationship Matrix Prediction Network for End-to-End Scene Graph Generation/full.md @@ -0,0 +1,351 @@ +# Multi-Granularity Sparse Relationship Matrix Prediction Network for End-to-End Scene Graph Generation + +Lei Wang, Zejian Yuan, and Badong Chen + +Institute of Artificial Intelligence and Robotics, Xi'an Jiaotong University, Xi'an 710049, China +leiwangmail@stu.xjtu.edu.cn, {yuan.ze.jian, chenbd}@mail.xjtu.edu.cn + +Abstract. Current end-to-end Scene Graph Generation (SGG) relies solely on visual representations to separately detect sparse relations and entities in an image. This leads to the issue where the predictions of entities do not contribute to the prediction of relations, necessitating post-processing to assign corresponding subjects and objects to the predicted relations. In this paper, we introduce a sparse relationship matrix that bridges entity detection and relation detection. Our approach not only eliminates the need for relation matching, but also leverages the semantics and positional information of predicted entities to enhance relation prediction. Specifically, a multi-granularity sparse relationship matrix prediction network is proposed, which utilizes three gated pooling modules focusing on filtering negative samples at different granularities, thereby obtaining a sparse relationship matrix containing entity pairs most likely to form relations. Finally, a set of sparse, most probable subject-object pairs can be constructed and used for relation decoding. Experimental results on multiple datasets demonstrate that our method achieves a new state-of-the-art overall performance. Our code is available at https://github.com/wanglei0618/Mg-RMPN. + +Keywords: Scene Graph Generation $\cdot$ End-to-End $\cdot$ Sparse Relationship Matrix $\cdot$ Multi-Granularity + +# 1 Introduction + +Scene Graph Generation (SGG) is a fundamental visual comprehension task that captures semantic information by detecting relation triplets $<$ subject entity, predicate, object entity $>$ in an image. This structured representation can facilitate many downstream tasks, such as image captioning [4, 41], visual question answering [9, 28], image retrieval [11, 25] and image generation [10, 19]. The end-to-end SGG methods [6, 16, 30] can directly generate sparse relations from an image based on a fixed number of relation queries, avoiding dense prediction that contains a lot of background relations in two-stage SGG. This leads to faster inference speed, and because fewer background relations are predicted, it exhibits superior predictive performance for rare samples of tail classes. + +![](images/ffba968c47daf7af70f546ad00059648ad59b6a4167a7122278183114fb04901.jpg) +(a) + +![](images/25b231acb2938b3714ad51294734b301f2375ed1f1dc452e95e18b3c647361ed.jpg) +(b) + +![](images/3ee8640a5287cfecdb9ad07fc0368e89580760db82751dbadde379c03aa80f58.jpg) +(c) +Fig. 1: (a) End-to-end SGG pipeline. (b) Our proposed end-to-end pipeline based on multi-granularity sparse relationship matrix prediction network. (c) The average number of entities and relations in the VG150, OIV6, and GQA200 datasets demonstrates the sparsity of relations in an image. + +Current end-to-end approaches [6,16,22] still lag behind two-stage methods in performance, which is attributed to their reliance solely on visual representations to separately detect entities and relations, as shown in Figure 1 (a). This leads to two drawbacks: (1) It requires post-processing to match the corresponding subjects and objects for the relations, and the model's performance is constrained by the capabilities of matching. (2) Due to the inability to determine the subject and object in advance, entity predictions cannot help with relation prediction, and relation prediction can only rely on visual features. + +In this paper, we introduce a sparse relationship matrix that bridges entity detection and relation detection. Our approach not only eliminates the need for relation matching, but also leverages the semantics and positional information of predicted entities to enhance relation prediction. Figure 1 (b) shows our pipeline based on the "pair then relation" [30] framework, which first constructs a relationship matrix based on entity predictions to represent the possibility of forming a relation between two entities and then selects the sparsest, most likely subject-object pairs for relation prediction. + +Compared to all possible pairs formed by the entity proposals, the subject-object pairs with relations are extremely sparse. Figure 1 (c) illustrates the average entities and relations in the datasets. Taking VG150 [13] as an example, when generating $N_{e} = 100$ entities, there are a total of $(100(100 - 1)) = 9900$ possible combinations. But the average relations per image is only 5.8, which means that more than 99. $9\%$ pairs are negative samples. Therefore, directly learning this extremely sparse relationship matrix is very challenging. + +In this work, we propose a Multi-granularity sparse Relationship Matrix Prediction Network (Mg-RMPN) to partition the negative samples into different granularities, achieving layer-by-layer filtering of dense negative samples. Specifically, Mg-RMPN employs three Gated Pooling Modules (GPM) with identical + +structures but independent parameters, each of which focuses on filtering negative samples at different granularities. The GPMs are used to generate relation matrices at different granularities and construct the corresponding ground truth matrices for supervised learning. The experimental results from various datasets reveal that our approach not only accurately identifies numerous head-class samples, but also excels in predicting sparse tail-class samples, achieving state-of-the-art overall performance. + +The contributions are summarized as follows: (1) An end-to-end SGG framework based on the relationship matrix is proposed that bridges entity detection and relation detection. (2) An Mg-RMPN is proposed to achieve sparse relationship matrix prediction based on multi-granularity negative sample learning. (3) Experimental results have confirmed that our method achieves state-of-the-art overall performance in both end-to-end and two-stage methods. + +# 2 Related Works + +# 2.1 Scene Graph Generation + +Early work [5, 27, 32, 36] in SGG is based on the two-stage pipeline, first employing a Faster-RCNN [24] object detector to generate entity proposals, and the class of each entity is predicted. Then, entities are paired to form relations, and all possible pairs are classified. Its drawback lies that dense relation pairs include numerous background relations, which dilute the sparse tail-class samples, thus aggravating the imbalance issue. Therefore, some recent efforts [8, 23] have focused on addressing the imbalanced predictions, developing a series of rebalancing strategies such as logit adjustment [2, 26], loss reweighting [18, 33, 35], and data resampling [17, 31, 37]. Unlike these methods, we introduce a relationship matrix to generate sparse subject-object pairs for relation prediction. The sparse relations generated in this way have a higher signal-to-noise ratio, which can alleviate the imbalance problem. The results of the experiments validate that our method maintains outstanding performance for head classes and also demonstrates high performance for tail classes. + +# 2.2 End-to-End Scene Graph Generation + +End-to-end SGG, inspired by DETR [1], generates sparse relations directly from an image based on queries. For instance, [6] introduced the Relation Transformer to directly generate a set of relations from visual features. However, the generated relations did not match the corresponding subjects and objects, hence [16] proposed a graph assembling module to address the relation matching. [30] further introduced a "pair then relation" framework that predetermines subject-object pairs, circumventing the relation matching. However, these approaches suffer from the issue of relying solely on visual representations to detect relations and entities separately. Therefore, this paper introduces Mg-RMPN to predict a sparse relationship matrix to link entity detection and relation detection, not only eliminating the need for relation matching but also utilizing the predicted entity semantics and positional information to assist in relation prediction. + +# 3 Method + +# 3.1 Problem Definition + +SGG aims to generate a scene graph $\mathcal{G} = \{\mathcal{E},\mathcal{R}\}$ from an input image $\mathcal{I}$ . The scene graph consists of a set of $n$ entities $\mathcal{E} = \{e_i\}_{i=1}^n$ and a set of $m$ relations $\mathcal{R} = \{r_k\}_{k=1}^m$ between entities. The set of entities $\mathcal{E}$ can be further decomposed into a set of bounding boxes $\mathcal{B} = \{\mathbf{b}_i\}_{i=1}^n$ and a set of class labels $\mathcal{C} = \{c_i\}_{i=1}^n$ . The generation of a scene graph $\mathcal{G}$ can be formulated as the joint probability distribution: + +$$ +\Pr (\mathcal {G} | \mathcal {I}) = \Pr (\mathcal {B}, \mathcal {C} | \mathcal {I}) \Pr (\mathcal {R} | \mathcal {I}, \mathcal {B}, \mathcal {C}), \tag {1} +$$ + +where $\operatorname{Pr}(\mathcal{B},\mathcal{C}|\mathcal{I})$ represents the entity representation obtained from the object detector, including the class labels and bounding boxes. $\operatorname{Pr}(\mathcal{R}|\mathcal{I},\mathcal{B},\mathcal{C})$ represents the relation prediction based on pairs of entities by the relation decoder. The process from $\operatorname{Pr}(\mathcal{B},\mathcal{C}|\mathcal{I})$ to $\operatorname{Pr}(\mathcal{R}|\mathcal{I},\mathcal{B},\mathcal{C})$ requires our Mg-RMPN to predict the pairs of entities most likely to form a relation from all possible pairs of entities. + +# 3.2 Overall Architecture + +As shown in Figure 2 (a), our method comprises three modules: (1) entity detection, (2) relationship matrix prediction, and (3) relation prediction. The entity detection is responsible for generating a set of entities based on the object detector, including their visual representations, class labels, and bounding boxes. The relationship matrix prediction is responsible for producing a sparse relationship matrix that represents the relevance between two entities, and then obtaining the indices of the subject and object entities that are most likely to form a relation from all possible combinations. Finally, the relation of these sparse subject-object pairs is predicted through the relation decoder. + +# 3.3 Object Detector + +This paper employs a deformable DETR [42] as the object detector, which contains a transformer encoder-decoder architecture with a set of entity queries. Here, the entity queries interact with encoder features through multi-scale deformable attention. Given an image $\mathcal{I}$ , the object detector produces $N_{e}$ entity representations $\mathbf{Q}_e = \{\mathbf{q}_i^e\}_{i=1}^{N_e} \in \mathbb{R}^{N_e \times d}$ from a set of learnable entity queries, where $d$ is the embedding dimensions. Based on $\mathbf{Q}_e$ , the object detector then follows with a classification head and a regression head to predict the entity's class predictions $\mathbf{C} = \{\hat{\mathbf{c}}_i\}_{i=1}^n \in \mathbb{R}^{N_e \times C_e}$ and bounding boxes $\mathbf{B} = \{\hat{\mathbf{b}}_i\}_{i=1}^n \in \mathbb{R}^{N_e \times 4}$ , respectively, where $C_e$ is the number of entity classes. + +# 3.4 Multi-Granularity Relationship Matrix Prediction Network + +Given $N_{e}$ entities, Multi-granularity sparse Relationship Matrix Prediction Network (Mg-RMPN) generates a $N_{e} \times N_{e}$ adjacency relationship matrix $\mathbf{M}_r$ . Each + +![](images/bca1af44e3398e6d5ccd55d8f755be506a7c50a8e8f153e4d74f93c8d4a9011e.jpg) + +![](images/6b05dda4c16792948784c67ff2c8ed53787ef133e697bc5080416e3c097918fc.jpg) +Fig. 2: (a) Overall pipeline of our end-to-end SGG method. (b) Multi-granularity sparse Relationship Matrix Prediction Network (Mg-RMPN). GPM is Gated Pooling Module. + +node represents the probability of forming a relationship between two entities. Due to the sparse entities and relations in an image, $\mathbf{M}_r$ contains a large number of pairs without relations, that is, negative samples. + +Therefore, Mg-RMPN defined three $\mathbf{M}_r$ based on multi-granularity negative samples, where the three $\mathbf{M}_r$ are learned through three structurally identical but parameter-independent Gated Pooling Modules (GPM), as shown in Figure 2 (b). With the outputs of the GPMs, Mg-RMPN is able to generate the final sparse relationship matrix, enabling the identification of most probable subject-object pairs to predict relations. + +Gated Pooling Module Unlike [30], which only uses visual features, our GPM determines the correlation of two entities based on their visual representation, linguistic prior and location information. Given a set of visual representations $\mathbf{Q}_e$ of the entities in an image, GPM first uses two fully connected networks $\mathcal{F}_{sub}$ and $\mathcal{F}_{obj}$ to project it to the visual representations of the subject and object, $\mathbf{Q}_{sub}$ and $\mathbf{Q}_{obj}$ , respectively, as follows: + +$$ +\mathbf {Q} _ {s u b} = \mathcal {F} _ {s u b} \left(\mathbf {Q} _ {e}\right) = \left\{\mathbf {q} _ {i} ^ {s u b} \right\} _ {i = 1} ^ {N _ {e}}, \quad \mathbf {Q} _ {o b j} = \mathcal {F} _ {o b j} \left(\mathbf {Q} _ {e}\right) = \left\{\mathbf {q} _ {j} ^ {o b j} \right\} _ {j = 1} ^ {N _ {e}}. \tag {2} +$$ + +Subsequently, the visual similarity between the subject $i$ and the object $j$ can be calculated as $v_{ij}^{sim} = \mathbf{q}_i^{sub} \cdot (\mathbf{q}_j^{obj})^{\mathrm{T}}$ . + +Given a set of class prediction distributions $\mathbf{C}$ for entities, GPM uses the prior linguistic function $\mathcal{F}_{prior}$ from [26] to predict the correlation between two + +entities, $s_{ij}^{prior} = \mathcal{F}_{prior}(\hat{\mathbf{c}}_i,\hat{\mathbf{c}}_j)$ . $\mathcal{F}_{prior}$ aims to capture the linguistic inductive bias that exists between two entities to enhance the prediction of the relational matrices. Similarly to [26], $\mathcal{F}_{prior}$ obtains a reasonable initialization based on the distribution of relations generated by entity pairs in the training dataset. + +Based on the positions $\mathbf{B}$ of the entities, GPM uses the function $\mathcal{F}_{\text{overlap}}$ to obtain the relative position information between two entities $\mathbf{b}_{ij}^{\text{overlap}} = \mathcal{F}_{\text{overlap}}(\hat{\mathbf{b}}_i, \hat{\mathbf{b}}_j) \in \mathbb{R}^8$ , where 8 elements indicate difference in $x$ and $y$ coordinates, width ratio, height ratio, IOU, center distance, area ratio, and aspect ratio. Additionally, we can also obtain the union bounding box $\hat{\mathbf{b}}_{ij}^u \in \mathbb{R}^9$ for the entity pairs, where 9 elements indicate the union bounding box coordinates $(x_1, y_1, x_2, y_2)$ , center $(\frac{x_1 + x_2}{2}, \frac{y_1 + y_2}{2})$ , size $(x_2 - x_1, y_2 - y_1)$ and area $(x_2 - x_1)(y_2 - y_1)$ , respectively. Then, we can obtain the relevance representation of the subject $i$ and object $j$ as follows: + +$$ +\mathbf {e} _ {i j} = \left[ v _ {i j} ^ {\text {s i m}}, s _ {i j} ^ {\text {p r i o r}}, \mathbf {b} _ {i j} ^ {\text {o v e r l a p}}, \hat {\mathbf {b}} _ {i j} ^ {u} \right], \tag {3} +$$ + +where $[\cdot, \cdot]$ denotes the concatenation operation. Subsequently, the feature representation of the entity pair can be expressed as + +$$ +\mathbf {f} _ {i j} = [ \mathbf {q} _ {i} ^ {s u b}, \mathbf {q} _ {j} ^ {o b j}, f _ {s} (s _ {i j} ^ {p r i o r}), p o s (\mathbf {b} _ {i j} ^ {o v e r l a p}), p o s (\hat {\mathbf {b}} _ {i j} ^ {u}) ], \tag {4} +$$ + +Where $f_{s}$ is a fully-connected layer for language prior encoding, and $pos$ is a fully-connected layer for positional encoding. Then, GPM uses a gated pooling layer $f_{gate}$ to obtain the probability of relation between subject $i$ and object $j$ as follows: + +$$ +\hat {p} _ {i j} = \operatorname {S i g m o i d} \left(\sum f _ {\text {g a t e}} \left(\mathbf {f} _ {i j}\right) \odot \mathbf {e} _ {i j}\right), \tag {5} +$$ + +where $\odot$ represents element-wise multiplication. Finally, we can obtain a $N_{e} \times N_{e}$ relationship matrix $\mathbf{M}_{r} = \mathrm{GPM}(\mathbf{Q}_{e},\mathbf{C},\mathbf{B})$ . Note that we set the diagonal elements of $\mathbf{M}_{r}$ to zero, removing the cases where an entity serves as both subject and object simultaneously. + +Multi-Granularity Sparse Relationship Matrix Prediction Based on the output of three GPMs, we can construct three relationship matrices based on the granularity of negative samples and build the corresponding ground truth matrices (see Section 3.6) for supervised learning, enabling each GPM to focus on identifying negative samples of specific granularity. + +Since the $N_{e}$ entity proposals are typically much greater than the number of entities contained in an image, resulting in a large number of background entities among all possible pairs of entities. Since background entities do not form relationships, we divide the negative samples in the relationship matrix into three levels of granularity: (1) high-confidence negative samples composed of background entities (background-background), (2) medium-confidence negative samples constructed from background and foreground entities (entity-background or background-entity), and (3) low-confidence negative samples composed of foreground entities (entity-entity). + +Then we can define three relationship matrices based on the output of the GPMs as follows. + +$$ +\mathbf {M} _ {r 1} = \operatorname {G P M - 1} \left(\mathbf {Q} _ {e}, \mathbf {C}, \mathbf {B}\right) +$$ + +$$ +\mathbf {M} _ {r 2} = \mathbf {M} _ {r 1} + \operatorname {G P M - 2} \left(\mathbf {Q} _ {e}, \mathbf {C}, \mathbf {B}\right) \tag {6} +$$ + +$$ +\mathbf {M} _ {r 3} = \mathbf {M} _ {r 2} + \operatorname {G P M - 3} (\mathbf {Q} _ {e}, \mathbf {C}, \mathbf {B}) +$$ + +In Eq.6, each GPM can focus on learning the filtering of negative samples at different granularities, as follows: (1) GPM-1 focuses on learning from pairs between background and filters out a large number of high-confidence negative samples. (2) GPM-2 focuses on learning from pairs between background and entities and filters out medium-confidence negative samples. (3) GPM-3 focuses on learning from no relation pairs between entities and only needs to filter out entity pairs that do not form relationships from the remaining entity pairs. + +Based on the final sparse relationship matrix $\mathbf{M}_{r3}$ , we can select the top- $N_{r}$ subject-object pairs most likely to form relations and obtain their indices in the entity visual representations $\mathbf{Q}_e$ , the entity's class predictions $\mathbf{C}$ , and the bounding boxes $\mathbf{B}$ . + +# 3.5 Relation Decoder + +With the $\mathbf{Q}_e$ , $\mathbf{C}$ and $\mathbf{B}$ of subjects and objects, we can obtain the subject-object pair representation $\mathbf{Q}_{pair}$ , which integrates visual, semantic, and positional encodings as follows: + +$$ +\mathbf {Q} _ {p a i r} = \left[ \mathbf {Q} _ {e} ^ {s u b}, e m b (\mathbf {C} ^ {s u b}), p o s (\mathbf {B} ^ {s u b}), \mathbf {Q} _ {e} ^ {o b j}, e m b (\mathbf {C} ^ {o b j}), p o s (\mathbf {B} ^ {o b j}) \right]. \tag {7} +$$ + +where $emb$ is a pre-trained Glove language model to acquire the word embedding. + +In this paper, we adopt the relation decoder in [30], which consists of transformer decoders in the style of DETR, to predict relations. In the relation decoder, we initialize a relation query $\mathbf{Q}_r \in \mathbb{R}^{N_r \times d}$ as the query input, and the subject-object pair $\mathbf{Q}_{pair}$ is projected as the key and value of cross-attention. Subsequently, the relation representation after the Relation Decoder (RD) can be expressed as $\tilde{\mathbf{Q}}_r = \mathrm{RD}(\mathbf{Q}_r, \mathbf{Q}_{pair})$ . Then, the relation prediction for the entity pairs can be expressed as $\hat{\mathbf{r}} = W_{cls} \tilde{\mathbf{Q}}_r$ , where $W_{cls}$ is the liner relation classifier. + +# 3.6 Training + +Our end-to-end SGG is divided into three subtasks: (1) an entity detection task based on the object detector, (2) a sparse relationship matrix prediction task based on the Mg-RMPN, and (3) a relation prediction task based on the relation decoder. During training, each subtask generates the corresponding supervisory information and losses, as follows. + +Entity Detection Loss We uses the end-to-end deformable DETR [42] as object detector, which uses the set prediction loss proposed in DETR [1] by assigning the ground truth entities to the predictions. A cost function is applied to compute the matching cost between a prediction and a ground-truth entity. With the cost matrix, the entity prediction-ground truth assignment is computed with the Hungarian Matching [14]. Given a set of $N_{e}$ entity proposals $\{e_i\}_{i=1}^{N_e}$ from object detector, the set prediction loss for can be presented as: + +$$ +\mathcal {L} _ {e} = \Sigma_ {i = 1} ^ {N _ {e}} \left[ \mathcal {L} _ {c l s} ^ {e} + \mathbf {1} _ {e _ {i} \neq \phi} \mathcal {L} _ {b o x} ^ {e} \right], \tag {8} +$$ + +where $\mathcal{L}_{cls}^{e}$ denotes the cross-entropy loss for label classification and $e_i \neq \phi$ means that $<\text{background}>$ is not assigned to the $i$ th entity prediction. $\mathcal{L}_{box}^{e}$ consists of $L_1$ loss and generalized IoU loss for the box regression. + +Multi-granularity Relationship Matrices Loss To predict the final sparse relationship matrix, we need to define the supervisory information and loss functions during the training phase. + +Multi-granularity Supervision Matrices In the training phase, we assign labels to the entity proposals $\{e_i\}_{i=1}^{N_e}$ by Hungarian matching and obtain the ground truth for each node in $\mathbf{M}_r$ regarding the subject, object, and whether they form a relation. In response to the sparsity of $\mathbf{M}_r$ , we propose Mg-RMPN to achieve hierarchical identification of negative samples at different granularities. + +Consequently, based on the granularity of the negative samples that each GPM focuses on in Section 3.4, the corresponding ground truth can be obtained. (1) $\mathbf{M}_{r1}^{gt}$ filters out all pairs composed of background entities as follows: + +$$ +\mathbf {M} _ {r 1} ^ {g t} = \left\{ \right.p _ {i j} \mid p _ {i j} = \left\{\begin{array}{l l}0&\text {i f} (e _ {i} = \phi \text {a n d} e _ {j} = \phi) \text {o r} i = j\\1&\text {o t h e r w i s e}\end{array}, 1 \leq i \leq N _ {e}, 1 \leq j \leq N _ {e} \right\}, \tag {9} +$$ + +where $e_i = \phi$ ( $e_j = \phi$ ) means that $<\text{background}>$ is assigned to the $i$ th ( $j$ th) entity prediction, $i = j$ means removing self-connected entity pairs. (2) $\mathbf{M}_{r2}^{gt}$ further filters out all pairs composed of backgrounds and entities as follows: + +$$ +\mathbf {M} _ {r 2} ^ {g t} = \left\{ \right.p _ {i j} \mid p _ {i j} = \left\{\begin{array}{l l}0&\text {i f} (e _ {i} = \phi \text {o r} e _ {j} = \phi) \text {o r} i = j\\1&\text {o t h e r w i s e}\end{array}, 1 \leq i \leq N _ {e}, 1 \leq j \leq N _ {e} \right\}. \tag {10} +$$ + +(3) $\mathbf{M}_{r3}^{gt}$ filters out all non-relationship pairs, and $\mathbf{M}_{r3}^{gt}$ is also the final ground truth relationship matrix of the scene graph, expressed as follows + +$$ +\mathbf {M} _ {r 3} ^ {g t} = \left\{p _ {i j} \mid p _ {i j} = \left\{ \begin{array}{l l} 0 & \text {o t h e r w i s e} \\ 1 & \text {i f} r _ {i j} = 1 \end{array} , 1 \leq i \leq N _ {e}, 1 \leq j \leq N _ {e} \right\}, \right. \tag {11} +$$ + +where $r_{ij} = 1$ implies that there is a relation between subject $i$ and object $j$ . Here, the relationships of the negative sample subsets in $\mathbf{M}_{r1}^{gt}$ , $\mathbf{M}_{r2}^{gt}$ and $\mathbf{M}_{r3}^{gt}$ are $\mathbf{M}_{r1|p_{ij} = 0}^{gt} \subseteq \mathbf{M}_{r2|p_{ij} = 0}^{gt} \subseteq \mathbf{M}_{r3|p_{ij} = 0}^{gt}$ . + +Multi-granularity Relationship Matrices Loss Considering that the large number of negative samples in $\mathbf{M}_r$ have different confidence, we modify the binary cross-entropy function using focal loss [20], which applies weighted discrimination on well-classified samples, forcing the model to focus on wrongly classified samples. Then, the multi-granularity learning loss is + +$$ +\mathcal {L} _ {M g} = \mathcal {L} _ {R M} ^ {1} \left(\mathbf {M} _ {r 1}, \mathbf {M} _ {r 1} ^ {g t}\right) + \mathcal {L} _ {R M} ^ {2} \left(\mathbf {M} _ {r 2}, \mathbf {M} _ {r 2} ^ {g t}\right) + \mathcal {L} _ {R M} ^ {3} \left(\mathbf {M} _ {r 3}, \mathbf {M} _ {r 3} ^ {g t}\right), \tag {12} +$$ + +where the relationship matrices loss $\mathcal{L}_{RM}^{k}(k = 1,2,3)$ is defined as + +$$ +\begin{array}{l} \mathcal {L} _ {R M} ^ {k} \left(\mathbf {M} _ {r}, \mathbf {M} _ {r} ^ {g t}\right) = - \sum_ {i} \sum_ {j} \left\{\alpha_ {k} \left(1 - \hat {p} _ {i j}\right) ^ {\gamma} p _ {i j} \log (\hat {p} _ {i j}) \right. \\ + (1 - \alpha_ {k}) \hat {p} _ {i j} ^ {\gamma} (1 - p _ {i j}) \log (1 - \hat {p} _ {i j}) \} / N _ {p o s} \tag {13} \\ + \left\| \mathbf {M} _ {r} \right\| _ {1} + \left\| \mathbf {M} _ {r} \right\| _ {2}, \\ \end{array} +$$ + +where the first part is the focal loss based on binary cross-entropy, and the second part is the L1 and L2 regularization of $\mathbf{M}_r$ , $\alpha_k(k = 1,2,3)$ is hyperparameters that adjusts the weights of positive and negative samples, $\gamma$ is a hyperparameter for focal loss, and $N_{pos}$ is the number of positive samples in each mini-batch. + +Relation Prediction Loss In this paper, the relation decoder also considers relation prediction as a set prediction based on a query. Therefore, during the training phase, we also use Hungarian Matching to assign labels for relation prediction. Due to the imbalanced relation classes in SGG, we use Seesaw loss [29] to dynamically adjust the gradients of samples of different classes. The relation loss is as follows: + +$$ +\mathcal {L} _ {r} = - \sum_ {i = 1} ^ {C _ {r}} y _ {i} \log \left(\hat {\sigma} _ {i}\right), \quad \hat {\sigma} _ {i} = \frac {e ^ {\hat {r} _ {i}}}{\sum_ {i \neq j} ^ {C _ {r}} \mathcal {S} _ {i j} e ^ {\hat {r} _ {j}} + e ^ {\hat {r} _ {i}}}, \tag {14} +$$ + +where $C_r$ is the number of relation classes, $y_i \in \{0,1\}$ , $1 \leqslant i \leqslant C_r$ is the one-hot ground truth label, $S_{ij}$ is a tunable balancing factor between different classes, detail can be found in [29]. + +In summary, the total loss of our method is + +$$ +\mathcal {L} = \mathcal {L} _ {e} + \lambda_ {1} \mathcal {L} _ {M g} + \lambda_ {2} \mathcal {L} _ {r}, \tag {15} +$$ + +Where $\lambda_{1}$ and $\lambda_{2}$ are hyperparameters used to respectively adjust multi-granularity learning loss $\mathcal{L}_{Mg}$ and relation prediction loss $\mathcal{L}_r$ . + +# 4 Experiments + +# 4.1 Experimental Settings + +Datasets: We evaluate our proposed method on the following datasets: (1) Visual Genome (VG150) [13] is the most widely used dataset in SGG, consisting of the most frequent 150 object classes and 50 predicate classes. (2) Open Images V6 (OIv6) [15] is a large-scale dataset proposed by Google. We follow + +the data processing and evaluation protocols in [17, 40]. OIv6 consists of 601 object classes and 30 predicate classes. (2) Generalized Question Answering (GQA200) [9] is another vision and language benchmark with more than 3.8M relation annotations. We follow the data processing in [7], which consists of Top-200 object classes and Top-100 predicate classes. + +Task & Evaluation Metrics: In this work, we focus on the Scene Graph Detection task, which detects all objects in an image and predicts their bounding boxes, labels, and relations. The initial evaluation metric Recall@K (R@K) was found to be dominated by head classes. Therefore, the mean Recall@K (mR@K) across all relations classes is proposed to evaluate unbiased SGG. However, focusing solely on mR@K and neglecting R@K can result in a tail bias. Therefore, we adopts the harmonic mean F@K of R@K and mR@K as an overall metric. For OIv6, the weighted evaluation metrics (wmAP rel, wnAP phr, scorewtd) are used for a more class-balanced evaluation. + +Implementation Details: We utilized the pre-trained deformable DETR [42] as the object detector. We set the loss hyperparameters $\lambda_{1}$ and $\lambda_{2}$ to 0.5 and 3 respectively. For the multi-granularity relationship matrices loss, we set $\alpha_{k}(k = 1,2,3)$ to 0.75, 0.9, and 0.99, respectively, to control the weight of positive samples, with the focal loss parameter $\gamma$ set to 2. The number of entities predicted by the detector $N_{e}$ is 100, the number of most probable subject-object pairs selected $N_{r}$ is 100, and the size of embedding dimensions $d$ is 256. Our model is implemented on 8 NVIDIA 3090 GPUs with learning rate $\times 10^{-4}$ and batch size 16 for 24 epochs. More implementation details are shown in the Supplementary Material. + +# 4.2 Comparison with State of the Arts + +VG150: Table 1 shows the comparison of our method on VG150, with methods divided into two-stage (above) and end-to-end (below). Due to the imbalance issue in SGG, the two-stage methods (Motifs, VCTree, GPS-Net, RelDN, PE-Net) perform better in R@K but show very poor performance in mR@K. Thus, many rebalancing techniques (VCTree+TDE, BGNN, SHA+GCL) have been developed. Although they improve mR@K, they inevitably impair R@K. For example, SHA+GCL, despite achieving the best mR@K, its recall performance significantly deteriorates, with R@50/100 being only 14.9/18.2. + +Although the end-to-end methods are inferior to the two-stage methods on R@K, it demonstrates better potential on mR@K due to its advantage in sparse relation prediction, which does not dilute the scarce tail-class samples. Furthermore, compared to the rebalancing techniques of the two-stage methods, the end-to-end approach does not overly harm R@K and can achieve superior overall performance F@K, such as SGTR and Pair-Net. + +Our Mg-RMPN leads comprehensively in end-to-end methods. Compared with two-stage methods, Mg-RMPN achieves competitive performance in R@K and significantly excels in mR@K, achieving the best overall performance F@K. For example, Mg-RMPN achieved 19.3/22.8 on F@50/100, surpassing the end-to-end and two-stage optimal methods Pair-Net and PE-Net, respectively. + +Table 1: Performance comparison at scene graph detection task on VG150. * denotes the results from [16]. + +
MethodR@20 R@50 R@100mR@20 mR@50 mR@100F@20 F@50 F@100
Motifs [36]25.532.837.25.06.87.98.411.313.0
VCTree [27]24.531.936.25.47.48.78.812.014.0
VCTree+TDE [26]14.019.423.26.99.311.19.212.615.0
GPS-Net [21]22.328.933.26.98.79.810.513.415.1
BGNN [17]-31.035.8-10.712.6-15.918.6
RelDN [39]-31.435.9-6.07.3-10.112.1
SHA+GCL [7]-14.918.214.217.920.9-16.319.5
PE-Net [40]-30.735.2-12.414.5-17.720.5
FCSGG [22]16.121.325.12.73.64.24.66.27.2
RelTR [6]21.227.5-6.810.8-10.315.5-
AS-Net* [3]-18.721.1-6.17.2-9.210.7
HOTR* [12]-23.527.7-9.412.0-13.416.7
SGTR [16]-24.628.4-12.015.2-16.119.8
Pair-Net [30]18.824.929.38.912.415.412.116.620.2
Mg-RMPN(DETR)21.027.131.39.913.516.213.518.021.3
Mg-RMPN(Ours)22.529.133.510.314.417.314.119.322.8
+ +We also compared with Pair-Net that is also based on relationship matrix learning. The experimental results show that our method is comprehensively superior to it, which demonstrates the effectiveness of our proposed Mg-RMPN. Furthermore, we also present the results of Mg-RMPN based on the DETR object detector to eliminate the influence of the detector. The results show that our Mg-RMPN(DETR) still achieves the best overall performance among both one-stage and two-stage methods. + +Open Image V6: To validate the generalizability of Mg-RMPN, we performed experiments on Open Images V6 in Table 2 and compared with two-stage (above) and end-to-end (below) methods. Compared to classic SGG benchmarks, our method achieved the optimal results among all methods with 45.5, 77.8, 57.4 respectively in mR@50, R@50, F@50. Compared to open image benchmarks, we obtained competitive 35.5 and 36.4 in $\mathrm{wmAP}_{\mathrm{rel}}$ and $\mathrm{wmAP}_{\mathrm{phr}}$ , and achieved the best result 43.6 in the overall metric score $_{\mathrm{wtd}}$ . + +GQA200: We also conducted experiments on the more challenging GQA200, as shown in Table 3. Compared to two-stage methods (VTransE, Motifs, VCTree, SHA+GCL), our method's overall performance F@K is only slightly inferior to SHA+GCL on F@100, which sacrificed a significant amount of R@K to enhance mR@K by using rebalancing techniques. Compared to the end-to-end method Pair-Net, which is also based on relationship matrices, our method is comprehensively superior. + +Table 2: Performance comparison on OIV6. * denotes the results from [16,17]. †denote results reproduced with the authors' code. + +
MethodmR@50R@50F@50wmAPrelwmAPphrscorewtd
Motifs* [36]32.771.644.929.931.638.9
VCTree* [27]33.974.146.534.233.140.2
RelDN* [39]34.073.146.432.233.440.8
G-RCNN* [34]34.074.546.733.234.241.8
GPS-Net* [21]35.374.848.032.934.041.7
BGNN* [17]40.575.052.633.534.242.1
RelTR [6]-71.7-37.237.543.0
AS-Ne* [3]35.255.343.025.927.532.4
HOTR* [12]40.152.745.519.421.526.7
SGTR* [16]42.659.949.837.038.742.3
Pair-Net† [30]44.577.456.531.832.440.3
Mg-RMPN(Ours)45.577.857.435.536.443.6
+ +Table 3: Performance comparison at scene graph detection task on GQA200. * denotes the results from [7]. †denote results reproduced with the authors' code. + +
MethodR@50R@100mR@50mR@100F@50F@100
VTransE* [38]27.230.75.86.69.610.9
Motifs* [36]28.933.16.47.710.512.5
VCTree* [27]28.331.96.57.410.612.0
SHA+GCL* [7]14.817.917.820.116.218.9
Pair-Net† [30]20.223.410.612.613.916.4
Mg-RMPN (Ours)23.225.712.814.516.518.5
+ +# 4.3 Visualization of Multi-Granularity Relationship Matrices + +Figure 3 presents a visualization of the multi-granularity relationship matrices learned by our Mg-RMPN, where (d), (e), (f) are the sparse relationship matrices learned after being filtered by GPM-1, GPM-2, GPM-3, respectively, and (a), (b), (c) are the corresponding ground truth relationship matrices defined in Section 3.6. The visualization results show that the dense negative samples in the relationship matrix are filtered layer by layer according to different granularities. The consistency of the generated multi-granularity relationship matrices with their ground truth indicates that the three GPMs of Mg-RMPN indeed focus on filtering negative samples of different granularities. + +Figures 3 (h) and (g), respectively, present a zoomed-in view of the final relationship matrix and its ground truth. We can observe that the predicted relationship matrix successfully captures the true relationships, for example: "man wearing shirt", "man holding racket", "man wearing short" and "pole on fence". Additionally, the relationship matrix also predicts some relationships that do not in the ground truth, as illustrated by the dashed lines in the rows + +![](images/88eef34e090e017fc0bd9b278b104f472bf5f0d55181d294db4b8087b487bd84.jpg) +(a) Ground truth $\mathbf{M}_{F1}^{gt}$ + +![](images/7f4dac97d7411486991dc6e8f4577243682de9501f1db16f2f66920843bb7d49.jpg) +(b) Ground truth $\mathbf{M}_{\pmb{F}_2}^{gt}$ + +![](images/b1efc3e7320fa5abbc194866758cb6355c002a968debd24867c1edf86df96e62.jpg) +(c) Ground truth $\mathbf{M}_{r3}^{gt}$ + +![](images/cbf961e879467d824496405e8dc3935189dc0a31fb70385f987abee430bab51f.jpg) + +![](images/388cd83ade6d94c4e7f913984517fe93d16f09004ede628cce35e0a493a3057f.jpg) +(d) Multi-granularity $\mathbf{M}_{r1}$ + +![](images/5e822dc4ef2812098e1887bd1083d937525a4850d0f18fa7e3c2e4550fd2293d.jpg) +(e) Multi-granularity $\mathbf{M}_{r2}$ + +![](images/24a68a31da06af6c687428d0263c295d987826ae1df5042e28a687bf274fd2ba.jpg) +(f) Multi-granularity $\mathbf{M}_{r3}$ + +![](images/074f5772091af3359ebfaf9bb09e99ad75a17ec8c2ac8cdba57953f4a205916f.jpg) +Fig. 3: Visualization of Multi-Granularity Relationship Matrices + +Table 4: Ablation studies of Mg-RMPN components on VG150. + +
GPM-1GPM-2GPM-3R@20/50/100mR@20/50/100F@20/50/100
✓✘✘6.5/10.7/15.22.9/4.8/6.64.0/6.6/9.2
✘✓✘12.5/17.5/21.65.2/7.4/9.37.3/10.4/13.0
✓✓✘12.8/18.1/22.35.8/8.5/10.78.0/11.6/14.5
✘✘✓21.2/27.6/32.19.2/13.1/15.812.8/17.8/21.2
✓✘✓21.7/28.2/32.69.8/13.8/16.513.5/18.5/21.9
✘✓✓21.8/28.2/32.59.7/13.7/16.713.4/18.4/22.1
✓✓✓22.5/29.1/33.510.3/14.4/17.314.1/19.3/22.8
+ +and columns, which represent the relationships with the corresponding entity "man" as the subject and object, respectively. Due to the abundant presence of "man" in SGG, the relationship matrix tends to favor predicting relations that include "man", which aligns with the actual language induction bias. Therefore, the results of visualization indicate that our Mg-RMPN can accurately predict the sparse relationship matrix. + +# 4.4 Ablation Studies + +Component Analysis of Mg-RMPN. To assess the effectiveness of each GPM in Mg-RMPN, we performed ablation experiments in Table 4. In Mg-RMPN, each GPM focuses on negative sample filtering of different granularities, and the three GPMs collaborate to complete the learning of the sparse relationship matrix. GPM-1 and GPM-2 are used to help GPM-3 filter negative + +Table 5: Ablation studies of different input information for Mg-RMPN on VG150. + +
Visual Semantic PositionR@20/50/100mR@20/50/100F@20/50/100
✓XX21.2/27.6/32.19.8/13.6/16.413.4/18.2/21.7
✓X✓22.2/28.7/33.110.1/14.1/16.913.9/18.9/22.4
✓✓X21.2/27.6/32.29.8/13.8/16.813.4/18.4/22.1
✓✓✓22.5/29.1/33.510.3/14.4/17.314.1/19.3/22.8
+ +samples. Using GPM-1 and GPM-2 individually cannot make full use of the ground truth, hence their results are very poor. The results obtained using only GPM-3 are also not good enough. Further increasing the assistance of GPMs for negative sample filtering can significantly improve the performance of the model. Mg-RMPN fully utilized three GPMs to achieve optimal results, confirming that learning negative samples based on different granularities can effectively improve the model's learning ability for sparse relationship matrices. + +Different input information for Mg-RMPN This paper bridges entity detection and relation detection through a relationship matrix, making full use of the semantic and positional information of entities and solving the problem that end-to-end SGG relies solely on visual information to predict relations. Table 5 presents ablation studies of Mg-RMPN based on different input information. Compared to models that only utilize visual information, separately incorporating the semantic and positional information of entities can enhance the performance of the model. Furthermore, the improvement in introducing semantic and positional information simultaneously is significantly greater than in introducing a single modality of information, indicating that semantic and positional information are two complementary information. Hence, the semantic and positional information incorporated in our approach plays a crucial role in predicting relations. More hyperparameter analysis can be found in Supplementary Material. + +# 5 Conclusion + +In this paper, our proposed end-to-end method bridges entity detection and relation detection through a sparse relationship matrix, which not only eliminates the need for post-processing of relation matching but also leverages the semantics and positional information of predicted entities to enhance relation prediction. To predict the sparse relationship matrix, we propose a multi-granularity sparse relationship matrix prediction network, which utilizes three gated pooling modules focusing on filtering negative samples at different granularities. Finally, a set of sparse, most likely subject-object pairs can be constructed and used for relation decoding. The experimental results demonstrate that our method achieves an optimal overall performance in both the end-to-end and two-stage methods. + +# Acknowledgements + +This work was supported by the National Natural Science Foundation of China (U21A20485, 61976170, 62088102) and the National Key R&D Program of China (2023YFB4704900). + +# References + +1. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: European conference on computer vision. pp. 213-229. Springer (2020) +2. Chen, C., Zhan, Y., Yu, B., Liu, L., Luo, Y., Du, B.: Resistance training using prior bias: Toward unbiased scene graph generation. Proceedings of the AAAI Conference on Artificial Intelligence 36(1), 212-220 (Jun 2022) +3. Chen, M., Liao, Y., Liu, S., Chen, Z., Wang, F., Qian, C.: Reformulating hoi detection as adaptive set prediction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9004-9013 (2021) +4. Chen, S., Jin, Q., Wang, P., Wu, Q.: Say as you wish: Fine-grained control of image caption generation with abstract scene graphs. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9962-9971 (2020) +5. Chen, T., Yu, W., Chen, R., Lin, L.: Knowledge-embedded routing network for scene graph generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6163-6171 (2019) +6. Cong, Y., Yang, M.Y., Rosenhahn, B.: Reltr: Relation transformer for scene graph generation. IEEE Transactions on Pattern Analysis and Machine Intelligence (2023) +7. Dong, X., Gan, T., Song, X., Wu, J., Cheng, Y., Nie, L.: Stacked hybrid-attention and group collaborative learning for unbiased scene graph generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19427-19436 (2022) +8. Gao, L., Lyu, X., Guo, Y., Hu, Y., Li, Y.F., Xu, L., Shen, H.T., Song, J.: Informative scene graph generation via debiasing. arXiv preprint arXiv:2308.05286 (2023) +9. Hudson, D.A., Manning, C.D.: Gqa: A new dataset for real-world visual reasoning and compositional question answering. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 6700-6709 (2019) +10. Johnson, J., Gupta, A., Fei-Fei, L.: Image generation from scene graphs. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1219-1228 (2018) +11. Johnson, J., Krishna, R., Stark, M., Li, L.J., Shamma, D., Bernstein, M., Fei-Fei, L.: Image retrieval using scene graphs. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3668-3678 (2015) +12. Kim, B., Lee, J., Kang, J., Kim, E.S., Kim, H.J.: Hotr: End-to-end human-object interaction detection with transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 74-83 (2021) +13. Krishna, R., Zhu, Y., Groth, O., Johnson, J., Hata, K., Kravitz, J., Chen, S., Kalantidis, Y., Li, L.J., Shamma, D.A., et al.: Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision 123(1), 32-73 (2017) + +14. Kuhn, H.W.: The hungarian method for the assignment problem. Naval research logistics quarterly 2(1-2), 83-97 (1955) +15. Kuznetsova, A., Rom, H., Alldrin, N., Uijlings, J., Krasin, I., Pont-Tuset, J., Kamali, S., Popov, S., Malloci, M., Kolesnikov, A., et al.: The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. International Journal of Computer Vision 128(7), 1956-1981 (2020) +16. Li, R., Zhang, S., He, X.: Sgtr: End-to-end scene graph generation with transformer. In: proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 19486-19496 (2022) +17. Li, R., Zhang, S., Wan, B., He, X.: Bipartite graph network with adaptive message passing for unbiased scene graph generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11109-11119 (2021) +18. Li, W., Zhang, H., Bai, Q., Zhao, G., Jiang, N., Yuan, X.: Ppdl: Predicate probability distribution based loss for unbiased scene graph generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19447-19456 (2022) +19. Li, Y., Ma, T., Bai, Y., Duan, N., Wei, S., Wang, X.: Pastegan: A semi-parametric method to generate image from scene graph. Advances in Neural Information Processing Systems 32 (2019) +20. Lin, T.Y., Goyal, P., Girshick, R., He, K., Dollar, P.: Focal loss for dense object detection. In: Proceedings of the IEEE international conference on computer vision. pp. 2980-2988 (2017) +21. Lin, X., Ding, C., Zeng, J., Tao, D.: Gps-net: Graph property sensing network for scene graph generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3746-3753 (2020) +22. Liu, H., Yan, N., Mortazavi, M., Bhanu, B.: Fully convolutional scene graph generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11546-11556 (2021) +23. Lyu, X., Gao, L., Xie, J., Zeng, P., Tian, Y., Shao, J., Shen, H.T.: Generalized unbiased scene graph generation. arXiv preprint arXiv:2308.04802 (2023) +24. Ren, S., He, K., Girshick, R., Sun, J.: Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems 28 (2015) +25. Schroeder, B., Tripathi, S.: Structured query-based image retrieval using scene graphs. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. pp. 178-179 (2020) +26. Tang, K., Niu, Y., Huang, J., Shi, J., Zhang, H.: Unbiased scene graph generation from biased training. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 3716-3725 (2020) +27. Tang, K., Zhang, H., Wu, B., Luo, W., Liu, W.: Learning to compose dynamic tree structures for visual contexts. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 6619-6628 (2019) +28. Teney, D., Liu, L., van Den Hengel, A.: Graph-structured representations for visual question answering. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1-9 (2017) +29. Wang, J., Zhang, W., Zang, Y., Cao, Y., Pang, J., Gong, T., Chen, K., Liu, Z., Loy, C.C., Lin, D.: Seesaw loss for long-tailed instance segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9695-9704 (2021) +30. Wang, J., Wen, Z., Li, X., Guo, Z., Yang, J., Liu, Z.: Pair then relation: Pair-net for panoptic scene graph generation. arXiv preprint arXiv:2307.08699 (2023) + +31. Wang, L., Yuan, Z., Chen, B.: Learning to generate an unbiased scene graph by using attribute-guided predicate features. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 2581-2589 (2023) +32. Xu, D., Zhu, Y., Choy, C.B., Fei-Fei, L.: Scene graph generation by iterative message passing. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5410-5419 (2017) +33. Yang, G., Zhang, J., Zhang, Y., Wu, B., Yang, Y.: Probabilistic modeling of semantic ambiguity for scene graph generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12527-12536 (2021) +34. Yang, J., Lu, J., Lee, S., Batra, D., Parikh, D.: Graph r-cnn for scene graph generation. In: Proceedings of the European conference on computer vision (ECCV). pp. 670-685 (2018) +35. Yu, J., Chai, Y., Wang, Y., Hu, Y., Wu, Q.: Cogtree: Cognition tree loss for unbiased scene graph generation. In: Zhou, Z.H. (ed.) Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence, IJCAI-21. pp. 1274-1280. International Joint Conferences on Artificial Intelligence Organization (8 2021) +36. Zellers, R., Yatskar, M., Thomson, S., Choi, Y.: Neural motifs: Scene graph parsing with global context. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5831-5840 (2018) +37. Zhang, A., Yao, Y., Chen, Q., Ji, W., et al, L.: Fine-grained scene graph generation with data transfer (2022) +38. Zhang, H., Kyaw, Z., Chang, S.F., Chua, T.S.: Visual translation embedding network for visual relation detection. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5532-5540 (2017) +39. Zhang, J., Shih, K.J., Elgammal, A., Tao, A., Catanzaro, B.: Graphical contrastive losses for scene graph parsing. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11535-11543 (2019) +40. Zheng, C., Lyu, X., Gao, L., Dai, B., Song, J.: Prototype-based embedding network for scene graph generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22783-22792 (2023) +41. Zhong, Y., Wang, L., Chen, J., Yu, D., Li, Y.: Comprehensive image captioning via scene graph decomposition. In: European Conference on Computer Vision. pp. 211-229. Springer (2020) +42. Zhu, X., Su, W., Lu, L., Li, B., Wang, X., Dai, J.: Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159 (2020) \ No newline at end of file diff --git a/2024/Multi-Granularity Sparse Relationship Matrix Prediction Network for End-to-End Scene Graph Generation/images.zip b/2024/Multi-Granularity Sparse Relationship Matrix Prediction Network for End-to-End Scene Graph Generation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..817d1662790b6d638753681b5c50830d8b5d7b7f --- /dev/null +++ b/2024/Multi-Granularity Sparse Relationship Matrix Prediction Network for End-to-End Scene Graph Generation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9225b2fbac185eb27c1d707f8c4733fdc428f65bc970d9fd73d70b19d171b87d +size 573549 diff --git a/2024/Multi-Granularity Sparse Relationship Matrix Prediction Network for End-to-End Scene Graph Generation/layout.json b/2024/Multi-Granularity Sparse Relationship Matrix Prediction Network for End-to-End Scene Graph Generation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..10563667256c9ccf2eb1e3d5c7bcbc33eeb6a05e --- /dev/null +++ b/2024/Multi-Granularity Sparse Relationship Matrix Prediction Network for End-to-End Scene Graph Generation/layout.json @@ -0,0 +1,10460 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 136, + 111, + 478, + 163 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 111, + 478, + 163 + ], + "spans": [ + { + "bbox": [ + 136, + 111, + 478, + 163 + ], + "type": "text", + "content": "Multi-Granularity Sparse Relationship Matrix Prediction Network for End-to-End Scene Graph Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 201, + 186, + 411, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 186, + 411, + 199 + ], + "spans": [ + { + "bbox": [ + 201, + 186, + 411, + 199 + ], + "type": "text", + "content": "Lei Wang, Zejian Yuan, and Badong Chen" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 144, + 208, + 468, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 144, + 208, + 468, + 242 + ], + "spans": [ + { + "bbox": [ + 144, + 208, + 468, + 242 + ], + "type": "text", + "content": "Institute of Artificial Intelligence and Robotics, Xi'an Jiaotong University, Xi'an 710049, China \nleiwangmail@stu.xjtu.edu.cn, {yuan.ze.jian, chenbd}@mail.xjtu.edu.cn" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 159, + 270, + 455, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 270, + 455, + 469 + ], + "spans": [ + { + "bbox": [ + 159, + 270, + 455, + 469 + ], + "type": "text", + "content": "Abstract. Current end-to-end Scene Graph Generation (SGG) relies solely on visual representations to separately detect sparse relations and entities in an image. This leads to the issue where the predictions of entities do not contribute to the prediction of relations, necessitating post-processing to assign corresponding subjects and objects to the predicted relations. In this paper, we introduce a sparse relationship matrix that bridges entity detection and relation detection. Our approach not only eliminates the need for relation matching, but also leverages the semantics and positional information of predicted entities to enhance relation prediction. Specifically, a multi-granularity sparse relationship matrix prediction network is proposed, which utilizes three gated pooling modules focusing on filtering negative samples at different granularities, thereby obtaining a sparse relationship matrix containing entity pairs most likely to form relations. Finally, a set of sparse, most probable subject-object pairs can be constructed and used for relation decoding. Experimental results on multiple datasets demonstrate that our method achieves a new state-of-the-art overall performance. Our code is available at https://github.com/wanglei0618/Mg-RMPN." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 159, + 479, + 453, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 479, + 453, + 501 + ], + "spans": [ + { + "bbox": [ + 159, + 479, + 453, + 501 + ], + "type": "text", + "content": "Keywords: Scene Graph Generation " + }, + { + "bbox": [ + 159, + 479, + 453, + 501 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 479, + 453, + 501 + ], + "type": "text", + "content": " End-to-End " + }, + { + "bbox": [ + 159, + 479, + 453, + 501 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 479, + 453, + 501 + ], + "type": "text", + "content": " Sparse Relationship Matrix " + }, + { + "bbox": [ + 159, + 479, + 453, + 501 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 479, + 453, + 501 + ], + "type": "text", + "content": " Multi-Granularity" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 521, + 230, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 521, + 230, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 521, + 230, + 533 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "Scene Graph Generation (SGG) is a fundamental visual comprehension task that captures semantic information by detecting relation triplets " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "<" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": " subject entity, predicate, object entity " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": ">" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": " in an image. This structured representation can facilitate many downstream tasks, such as image captioning [4, 41], visual question answering [9, 28], image retrieval [11, 25] and image generation [10, 19]. The end-to-end SGG methods [6, 16, 30] can directly generate sparse relations from an image based on a fixed number of relation queries, avoiding dense prediction that contains a lot of background relations in two-stage SGG. This leads to faster inference speed, and because fewer background relations are predicted, it exhibits superior predictive performance for rare samples of tail classes." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 167, + 114, + 260, + 246 + ], + "blocks": [ + { + "bbox": [ + 167, + 114, + 260, + 246 + ], + "lines": [ + { + "bbox": [ + 167, + 114, + 260, + 246 + ], + "spans": [ + { + "bbox": [ + 167, + 114, + 260, + 246 + ], + "type": "image", + "image_path": "ffba968c47daf7af70f546ad00059648ad59b6a4167a7122278183114fb04901.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 204, + 251, + 216, + 261 + ], + "lines": [ + { + "bbox": [ + 204, + 251, + 216, + 261 + ], + "spans": [ + { + "bbox": [ + 204, + 251, + 216, + 261 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 265, + 114, + 364, + 247 + ], + "blocks": [ + { + "bbox": [ + 265, + 114, + 364, + 247 + ], + "lines": [ + { + "bbox": [ + 265, + 114, + 364, + 247 + ], + "spans": [ + { + "bbox": [ + 265, + 114, + 364, + 247 + ], + "type": "image", + "image_path": "25b231acb2938b3714ad51294734b301f2375ed1f1dc452e95e18b3c647361ed.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 307, + 251, + 320, + 261 + ], + "lines": [ + { + "bbox": [ + 307, + 251, + 320, + 261 + ], + "spans": [ + { + "bbox": [ + 307, + 251, + 320, + 261 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 369, + 122, + 446, + 250 + ], + "blocks": [ + { + "bbox": [ + 369, + 122, + 446, + 250 + ], + "lines": [ + { + "bbox": [ + 369, + 122, + 446, + 250 + ], + "spans": [ + { + "bbox": [ + 369, + 122, + 446, + 250 + ], + "type": "image", + "image_path": "3ee8640a5287cfecdb9ad07fc0368e89580760db82751dbadde379c03aa80f58.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 402, + 251, + 412, + 261 + ], + "lines": [ + { + "bbox": [ + 402, + 251, + 412, + 261 + ], + "spans": [ + { + "bbox": [ + 402, + 251, + 412, + 261 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 270, + 482, + 314 + ], + "lines": [ + { + "bbox": [ + 130, + 270, + 482, + 314 + ], + "spans": [ + { + "bbox": [ + 130, + 270, + 482, + 314 + ], + "type": "text", + "content": "Fig. 1: (a) End-to-end SGG pipeline. (b) Our proposed end-to-end pipeline based on multi-granularity sparse relationship matrix prediction network. (c) The average number of entities and relations in the VG150, OIV6, and GQA200 datasets demonstrates the sparsity of relations in an image." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 339, + 482, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 339, + 482, + 435 + ], + "spans": [ + { + "bbox": [ + 130, + 339, + 482, + 435 + ], + "type": "text", + "content": "Current end-to-end approaches [6,16,22] still lag behind two-stage methods in performance, which is attributed to their reliance solely on visual representations to separately detect entities and relations, as shown in Figure 1 (a). This leads to two drawbacks: (1) It requires post-processing to match the corresponding subjects and objects for the relations, and the model's performance is constrained by the capabilities of matching. (2) Due to the inability to determine the subject and object in advance, entity predictions cannot help with relation prediction, and relation prediction can only rely on visual features." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 437, + 482, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 482, + 532 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 482, + 532 + ], + "type": "text", + "content": "In this paper, we introduce a sparse relationship matrix that bridges entity detection and relation detection. Our approach not only eliminates the need for relation matching, but also leverages the semantics and positional information of predicted entities to enhance relation prediction. Figure 1 (b) shows our pipeline based on the \"pair then relation\" [30] framework, which first constructs a relationship matrix based on entity predictions to represent the possibility of forming a relation between two entities and then selects the sparsest, most likely subject-object pairs for relation prediction." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 533, + 482, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 616 + ], + "type": "text", + "content": "Compared to all possible pairs formed by the entity proposals, the subject-object pairs with relations are extremely sparse. Figure 1 (c) illustrates the average entities and relations in the datasets. Taking VG150 [13] as an example, when generating " + }, + { + "bbox": [ + 130, + 533, + 482, + 616 + ], + "type": "inline_equation", + "content": "N_{e} = 100" + }, + { + "bbox": [ + 130, + 533, + 482, + 616 + ], + "type": "text", + "content": " entities, there are a total of " + }, + { + "bbox": [ + 130, + 533, + 482, + 616 + ], + "type": "inline_equation", + "content": "(100(100 - 1)) = 9900" + }, + { + "bbox": [ + 130, + 533, + 482, + 616 + ], + "type": "text", + "content": " possible combinations. But the average relations per image is only 5.8, which means that more than 99. " + }, + { + "bbox": [ + 130, + 533, + 482, + 616 + ], + "type": "inline_equation", + "content": "9\\%" + }, + { + "bbox": [ + 130, + 533, + 482, + 616 + ], + "type": "text", + "content": " pairs are negative samples. Therefore, directly learning this extremely sparse relationship matrix is very challenging." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "In this work, we propose a Multi-granularity sparse Relationship Matrix Prediction Network (Mg-RMPN) to partition the negative samples into different granularities, achieving layer-by-layer filtering of dense negative samples. Specifically, Mg-RMPN employs three Gated Pooling Modules (GPM) with identical" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 226, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 226, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 226, + 102 + ], + "type": "text", + "content": "L. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": "structures but independent parameters, each of which focuses on filtering negative samples at different granularities. The GPMs are used to generate relation matrices at different granularities and construct the corresponding ground truth matrices for supervised learning. The experimental results from various datasets reveal that our approach not only accurately identifies numerous head-class samples, but also excels in predicting sparse tail-class samples, achieving state-of-the-art overall performance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 200, + 482, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 200, + 482, + 272 + ], + "spans": [ + { + "bbox": [ + 130, + 200, + 482, + 272 + ], + "type": "text", + "content": "The contributions are summarized as follows: (1) An end-to-end SGG framework based on the relationship matrix is proposed that bridges entity detection and relation detection. (2) An Mg-RMPN is proposed to achieve sparse relationship matrix prediction based on multi-granularity negative sample learning. (3) Experimental results have confirmed that our method achieves state-of-the-art overall performance in both end-to-end and two-stage methods." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 286, + 243, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 286, + 243, + 300 + ], + "spans": [ + { + "bbox": [ + 131, + 286, + 243, + 300 + ], + "type": "text", + "content": "2 Related Works" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 308, + 286, + 319 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 308, + 286, + 319 + ], + "spans": [ + { + "bbox": [ + 131, + 308, + 286, + 319 + ], + "type": "text", + "content": "2.1 Scene Graph Generation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 324, + 482, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 324, + 482, + 492 + ], + "spans": [ + { + "bbox": [ + 130, + 324, + 482, + 492 + ], + "type": "text", + "content": "Early work [5, 27, 32, 36] in SGG is based on the two-stage pipeline, first employing a Faster-RCNN [24] object detector to generate entity proposals, and the class of each entity is predicted. Then, entities are paired to form relations, and all possible pairs are classified. Its drawback lies that dense relation pairs include numerous background relations, which dilute the sparse tail-class samples, thus aggravating the imbalance issue. Therefore, some recent efforts [8, 23] have focused on addressing the imbalanced predictions, developing a series of rebalancing strategies such as logit adjustment [2, 26], loss reweighting [18, 33, 35], and data resampling [17, 31, 37]. Unlike these methods, we introduce a relationship matrix to generate sparse subject-object pairs for relation prediction. The sparse relations generated in this way have a higher signal-to-noise ratio, which can alleviate the imbalance problem. The results of the experiments validate that our method maintains outstanding performance for head classes and also demonstrates high performance for tail classes." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 506, + 348, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 506, + 348, + 517 + ], + "spans": [ + { + "bbox": [ + 131, + 506, + 348, + 517 + ], + "type": "text", + "content": "2.2 End-to-End Scene Graph Generation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": "End-to-end SGG, inspired by DETR [1], generates sparse relations directly from an image based on queries. For instance, [6] introduced the Relation Transformer to directly generate a set of relations from visual features. However, the generated relations did not match the corresponding subjects and objects, hence [16] proposed a graph assembling module to address the relation matching. [30] further introduced a \"pair then relation\" framework that predetermines subject-object pairs, circumventing the relation matching. However, these approaches suffer from the issue of relying solely on visual representations to detect relations and entities separately. Therefore, this paper introduces Mg-RMPN to predict a sparse relationship matrix to link entity detection and relation detection, not only eliminating the need for relation matching but also utilizing the predicted entity semantics and positional information to assist in relation prediction." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 400, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 400, + 91, + 447, + 102 + ], + "type": "text", + "content": "Mg-RMPN" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 202, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 202, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 202, + 127 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 258, + 152 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 258, + 152 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 258, + 152 + ], + "type": "text", + "content": "3.1 Problem Definition" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 160, + 482, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 160, + 482, + 231 + ], + "spans": [ + { + "bbox": [ + 130, + 160, + 482, + 231 + ], + "type": "text", + "content": "SGG aims to generate a scene graph " + }, + { + "bbox": [ + 130, + 160, + 482, + 231 + ], + "type": "inline_equation", + "content": "\\mathcal{G} = \\{\\mathcal{E},\\mathcal{R}\\}" + }, + { + "bbox": [ + 130, + 160, + 482, + 231 + ], + "type": "text", + "content": " from an input image " + }, + { + "bbox": [ + 130, + 160, + 482, + 231 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 130, + 160, + 482, + 231 + ], + "type": "text", + "content": ". The scene graph consists of a set of " + }, + { + "bbox": [ + 130, + 160, + 482, + 231 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 160, + 482, + 231 + ], + "type": "text", + "content": " entities " + }, + { + "bbox": [ + 130, + 160, + 482, + 231 + ], + "type": "inline_equation", + "content": "\\mathcal{E} = \\{e_i\\}_{i=1}^n" + }, + { + "bbox": [ + 130, + 160, + 482, + 231 + ], + "type": "text", + "content": " and a set of " + }, + { + "bbox": [ + 130, + 160, + 482, + 231 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 130, + 160, + 482, + 231 + ], + "type": "text", + "content": " relations " + }, + { + "bbox": [ + 130, + 160, + 482, + 231 + ], + "type": "inline_equation", + "content": "\\mathcal{R} = \\{r_k\\}_{k=1}^m" + }, + { + "bbox": [ + 130, + 160, + 482, + 231 + ], + "type": "text", + "content": " between entities. The set of entities " + }, + { + "bbox": [ + 130, + 160, + 482, + 231 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 130, + 160, + 482, + 231 + ], + "type": "text", + "content": " can be further decomposed into a set of bounding boxes " + }, + { + "bbox": [ + 130, + 160, + 482, + 231 + ], + "type": "inline_equation", + "content": "\\mathcal{B} = \\{\\mathbf{b}_i\\}_{i=1}^n" + }, + { + "bbox": [ + 130, + 160, + 482, + 231 + ], + "type": "text", + "content": " and a set of class labels " + }, + { + "bbox": [ + 130, + 160, + 482, + 231 + ], + "type": "inline_equation", + "content": "\\mathcal{C} = \\{c_i\\}_{i=1}^n" + }, + { + "bbox": [ + 130, + 160, + 482, + 231 + ], + "type": "text", + "content": ". The generation of a scene graph " + }, + { + "bbox": [ + 130, + 160, + 482, + 231 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 130, + 160, + 482, + 231 + ], + "type": "text", + "content": " can be formulated as the joint probability distribution:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 231, + 233, + 480, + 245 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 233, + 480, + 245 + ], + "spans": [ + { + "bbox": [ + 231, + 233, + 480, + 245 + ], + "type": "interline_equation", + "content": "\\Pr (\\mathcal {G} | \\mathcal {I}) = \\Pr (\\mathcal {B}, \\mathcal {C} | \\mathcal {I}) \\Pr (\\mathcal {R} | \\mathcal {I}, \\mathcal {B}, \\mathcal {C}), \\tag {1}", + "image_path": "7f2246ea0025991509f1fb8b80c7b58b53ecb4a2b25730fa3411220e8464b975.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 251, + 482, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 251, + 482, + 312 + ], + "spans": [ + { + "bbox": [ + 130, + 251, + 482, + 312 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 251, + 482, + 312 + ], + "type": "inline_equation", + "content": "\\operatorname{Pr}(\\mathcal{B},\\mathcal{C}|\\mathcal{I})" + }, + { + "bbox": [ + 130, + 251, + 482, + 312 + ], + "type": "text", + "content": " represents the entity representation obtained from the object detector, including the class labels and bounding boxes. " + }, + { + "bbox": [ + 130, + 251, + 482, + 312 + ], + "type": "inline_equation", + "content": "\\operatorname{Pr}(\\mathcal{R}|\\mathcal{I},\\mathcal{B},\\mathcal{C})" + }, + { + "bbox": [ + 130, + 251, + 482, + 312 + ], + "type": "text", + "content": " represents the relation prediction based on pairs of entities by the relation decoder. The process from " + }, + { + "bbox": [ + 130, + 251, + 482, + 312 + ], + "type": "inline_equation", + "content": "\\operatorname{Pr}(\\mathcal{B},\\mathcal{C}|\\mathcal{I})" + }, + { + "bbox": [ + 130, + 251, + 482, + 312 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 251, + 482, + 312 + ], + "type": "inline_equation", + "content": "\\operatorname{Pr}(\\mathcal{R}|\\mathcal{I},\\mathcal{B},\\mathcal{C})" + }, + { + "bbox": [ + 130, + 251, + 482, + 312 + ], + "type": "text", + "content": " requires our Mg-RMPN to predict the pairs of entities most likely to form a relation from all possible pairs of entities." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 329, + 265, + 340 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 329, + 265, + 340 + ], + "spans": [ + { + "bbox": [ + 132, + 329, + 265, + 340 + ], + "type": "text", + "content": "3.2 Overall Architecture" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 349, + 482, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 349, + 482, + 458 + ], + "spans": [ + { + "bbox": [ + 130, + 349, + 482, + 458 + ], + "type": "text", + "content": "As shown in Figure 2 (a), our method comprises three modules: (1) entity detection, (2) relationship matrix prediction, and (3) relation prediction. The entity detection is responsible for generating a set of entities based on the object detector, including their visual representations, class labels, and bounding boxes. The relationship matrix prediction is responsible for producing a sparse relationship matrix that represents the relevance between two entities, and then obtaining the indices of the subject and object entities that are most likely to form a relation from all possible combinations. Finally, the relation of these sparse subject-object pairs is predicted through the relation decoder." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 475, + 243, + 487 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 475, + 243, + 487 + ], + "spans": [ + { + "bbox": [ + 132, + 475, + 243, + 487 + ], + "type": "text", + "content": "3.3 Object Detector" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 495, + 482, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 495, + 482, + 604 + ], + "spans": [ + { + "bbox": [ + 130, + 495, + 482, + 604 + ], + "type": "text", + "content": "This paper employs a deformable DETR [42] as the object detector, which contains a transformer encoder-decoder architecture with a set of entity queries. Here, the entity queries interact with encoder features through multi-scale deformable attention. Given an image " + }, + { + "bbox": [ + 130, + 495, + 482, + 604 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 130, + 495, + 482, + 604 + ], + "type": "text", + "content": ", the object detector produces " + }, + { + "bbox": [ + 130, + 495, + 482, + 604 + ], + "type": "inline_equation", + "content": "N_{e}" + }, + { + "bbox": [ + 130, + 495, + 482, + 604 + ], + "type": "text", + "content": " entity representations " + }, + { + "bbox": [ + 130, + 495, + 482, + 604 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_e = \\{\\mathbf{q}_i^e\\}_{i=1}^{N_e} \\in \\mathbb{R}^{N_e \\times d}" + }, + { + "bbox": [ + 130, + 495, + 482, + 604 + ], + "type": "text", + "content": " from a set of learnable entity queries, where " + }, + { + "bbox": [ + 130, + 495, + 482, + 604 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 130, + 495, + 482, + 604 + ], + "type": "text", + "content": " is the embedding dimensions. Based on " + }, + { + "bbox": [ + 130, + 495, + 482, + 604 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_e" + }, + { + "bbox": [ + 130, + 495, + 482, + 604 + ], + "type": "text", + "content": ", the object detector then follows with a classification head and a regression head to predict the entity's class predictions " + }, + { + "bbox": [ + 130, + 495, + 482, + 604 + ], + "type": "inline_equation", + "content": "\\mathbf{C} = \\{\\hat{\\mathbf{c}}_i\\}_{i=1}^n \\in \\mathbb{R}^{N_e \\times C_e}" + }, + { + "bbox": [ + 130, + 495, + 482, + 604 + ], + "type": "text", + "content": " and bounding boxes " + }, + { + "bbox": [ + 130, + 495, + 482, + 604 + ], + "type": "inline_equation", + "content": "\\mathbf{B} = \\{\\hat{\\mathbf{b}}_i\\}_{i=1}^n \\in \\mathbb{R}^{N_e \\times 4}" + }, + { + "bbox": [ + 130, + 495, + 482, + 604 + ], + "type": "text", + "content": ", respectively, where " + }, + { + "bbox": [ + 130, + 495, + 482, + 604 + ], + "type": "inline_equation", + "content": "C_e" + }, + { + "bbox": [ + 130, + 495, + 482, + 604 + ], + "type": "text", + "content": " is the number of entity classes." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 620, + 460, + 633 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 620, + 460, + 633 + ], + "spans": [ + { + "bbox": [ + 132, + 620, + 460, + 633 + ], + "type": "text", + "content": "3.4 Multi-Granularity Relationship Matrix Prediction Network" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "Given " + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "N_{e}" + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": " entities, Multi-granularity sparse Relationship Matrix Prediction Network (Mg-RMPN) generates a " + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "N_{e} \\times N_{e}" + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": " adjacency relationship matrix " + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_r" + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": ". Each" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 226, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 226, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 226, + 102 + ], + "type": "text", + "content": "L. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 141, + 114, + 471, + 201 + ], + "blocks": [ + { + "bbox": [ + 141, + 114, + 471, + 201 + ], + "lines": [ + { + "bbox": [ + 141, + 114, + 471, + 201 + ], + "spans": [ + { + "bbox": [ + 141, + 114, + 471, + 201 + ], + "type": "image", + "image_path": "bca1af44e3398e6d5ccd55d8f755be506a7c50a8e8f153e4d74f93c8d4a9011e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 145, + 202, + 449, + 336 + ], + "blocks": [ + { + "bbox": [ + 145, + 202, + 449, + 336 + ], + "lines": [ + { + "bbox": [ + 145, + 202, + 449, + 336 + ], + "spans": [ + { + "bbox": [ + 145, + 202, + 449, + 336 + ], + "type": "image", + "image_path": "6b05dda4c16792948784c67ff2c8ed53787ef133e697bc5080416e3c097918fc.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 344, + 480, + 367 + ], + "lines": [ + { + "bbox": [ + 131, + 344, + 480, + 367 + ], + "spans": [ + { + "bbox": [ + 131, + 344, + 480, + 367 + ], + "type": "text", + "content": "Fig. 2: (a) Overall pipeline of our end-to-end SGG method. (b) Multi-granularity sparse Relationship Matrix Prediction Network (Mg-RMPN). GPM is Gated Pooling Module." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 389, + 479, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 389, + 479, + 425 + ], + "spans": [ + { + "bbox": [ + 130, + 389, + 479, + 425 + ], + "type": "text", + "content": "node represents the probability of forming a relationship between two entities. Due to the sparse entities and relations in an image, " + }, + { + "bbox": [ + 130, + 389, + 479, + 425 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_r" + }, + { + "bbox": [ + 130, + 389, + 479, + 425 + ], + "type": "text", + "content": " contains a large number of pairs without relations, that is, negative samples." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 426, + 480, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 480, + 497 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 480, + 497 + ], + "type": "text", + "content": "Therefore, Mg-RMPN defined three " + }, + { + "bbox": [ + 130, + 426, + 480, + 497 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_r" + }, + { + "bbox": [ + 130, + 426, + 480, + 497 + ], + "type": "text", + "content": " based on multi-granularity negative samples, where the three " + }, + { + "bbox": [ + 130, + 426, + 480, + 497 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_r" + }, + { + "bbox": [ + 130, + 426, + 480, + 497 + ], + "type": "text", + "content": " are learned through three structurally identical but parameter-independent Gated Pooling Modules (GPM), as shown in Figure 2 (b). With the outputs of the GPMs, Mg-RMPN is able to generate the final sparse relationship matrix, enabling the identification of most probable subject-object pairs to predict relations." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 514, + 479, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 514, + 479, + 588 + ], + "spans": [ + { + "bbox": [ + 131, + 514, + 479, + 588 + ], + "type": "text", + "content": "Gated Pooling Module Unlike [30], which only uses visual features, our GPM determines the correlation of two entities based on their visual representation, linguistic prior and location information. Given a set of visual representations " + }, + { + "bbox": [ + 131, + 514, + 479, + 588 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_e" + }, + { + "bbox": [ + 131, + 514, + 479, + 588 + ], + "type": "text", + "content": " of the entities in an image, GPM first uses two fully connected networks " + }, + { + "bbox": [ + 131, + 514, + 479, + 588 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{sub}" + }, + { + "bbox": [ + 131, + 514, + 479, + 588 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 514, + 479, + 588 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{obj}" + }, + { + "bbox": [ + 131, + 514, + 479, + 588 + ], + "type": "text", + "content": " to project it to the visual representations of the subject and object, " + }, + { + "bbox": [ + 131, + 514, + 479, + 588 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{sub}" + }, + { + "bbox": [ + 131, + 514, + 479, + 588 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 514, + 479, + 588 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{obj}" + }, + { + "bbox": [ + 131, + 514, + 479, + 588 + ], + "type": "text", + "content": ", respectively, as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 167, + 594, + 480, + 611 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 167, + 594, + 480, + 611 + ], + "spans": [ + { + "bbox": [ + 167, + 594, + 480, + 611 + ], + "type": "interline_equation", + "content": "\\mathbf {Q} _ {s u b} = \\mathcal {F} _ {s u b} \\left(\\mathbf {Q} _ {e}\\right) = \\left\\{\\mathbf {q} _ {i} ^ {s u b} \\right\\} _ {i = 1} ^ {N _ {e}}, \\quad \\mathbf {Q} _ {o b j} = \\mathcal {F} _ {o b j} \\left(\\mathbf {Q} _ {e}\\right) = \\left\\{\\mathbf {q} _ {j} ^ {o b j} \\right\\} _ {j = 1} ^ {N _ {e}}. \\tag {2}", + "image_path": "4104180ba4a5e58c3361fd59b7b2b075bf184e209acb105a071c18450355ed6a.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 617, + 479, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 479, + 642 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 479, + 642 + ], + "type": "text", + "content": "Subsequently, the visual similarity between the subject " + }, + { + "bbox": [ + 130, + 617, + 479, + 642 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 617, + 479, + 642 + ], + "type": "text", + "content": " and the object " + }, + { + "bbox": [ + 130, + 617, + 479, + 642 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 130, + 617, + 479, + 642 + ], + "type": "text", + "content": " can be calculated as " + }, + { + "bbox": [ + 130, + 617, + 479, + 642 + ], + "type": "inline_equation", + "content": "v_{ij}^{sim} = \\mathbf{q}_i^{sub} \\cdot (\\mathbf{q}_j^{obj})^{\\mathrm{T}}" + }, + { + "bbox": [ + 130, + 617, + 479, + 642 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 642, + 480, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 642, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 642, + 480, + 666 + ], + "type": "text", + "content": "Given a set of class prediction distributions " + }, + { + "bbox": [ + 130, + 642, + 480, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{C}" + }, + { + "bbox": [ + 130, + 642, + 480, + 666 + ], + "type": "text", + "content": " for entities, GPM uses the prior linguistic function " + }, + { + "bbox": [ + 130, + 642, + 480, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{prior}" + }, + { + "bbox": [ + 130, + 642, + 480, + 666 + ], + "type": "text", + "content": " from [26] to predict the correlation between two" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 400, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 400, + 91, + 447, + 102 + ], + "type": "text", + "content": "Mg-RMPN" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 479, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 479, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 479, + 163 + ], + "type": "text", + "content": "entities, " + }, + { + "bbox": [ + 130, + 115, + 479, + 163 + ], + "type": "inline_equation", + "content": "s_{ij}^{prior} = \\mathcal{F}_{prior}(\\hat{\\mathbf{c}}_i,\\hat{\\mathbf{c}}_j)" + }, + { + "bbox": [ + 130, + 115, + 479, + 163 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 130, + 115, + 479, + 163 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{prior}" + }, + { + "bbox": [ + 130, + 115, + 479, + 163 + ], + "type": "text", + "content": " aims to capture the linguistic inductive bias that exists between two entities to enhance the prediction of the relational matrices. Similarly to [26], " + }, + { + "bbox": [ + 130, + 115, + 479, + 163 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{prior}" + }, + { + "bbox": [ + 130, + 115, + 479, + 163 + ], + "type": "text", + "content": " obtains a reasonable initialization based on the distribution of relations generated by entity pairs in the training dataset." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 164, + 480, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 164, + 480, + 277 + ], + "spans": [ + { + "bbox": [ + 130, + 164, + 480, + 277 + ], + "type": "text", + "content": "Based on the positions " + }, + { + "bbox": [ + 130, + 164, + 480, + 277 + ], + "type": "inline_equation", + "content": "\\mathbf{B}" + }, + { + "bbox": [ + 130, + 164, + 480, + 277 + ], + "type": "text", + "content": " of the entities, GPM uses the function " + }, + { + "bbox": [ + 130, + 164, + 480, + 277 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\text{overlap}}" + }, + { + "bbox": [ + 130, + 164, + 480, + 277 + ], + "type": "text", + "content": " to obtain the relative position information between two entities " + }, + { + "bbox": [ + 130, + 164, + 480, + 277 + ], + "type": "inline_equation", + "content": "\\mathbf{b}_{ij}^{\\text{overlap}} = \\mathcal{F}_{\\text{overlap}}(\\hat{\\mathbf{b}}_i, \\hat{\\mathbf{b}}_j) \\in \\mathbb{R}^8" + }, + { + "bbox": [ + 130, + 164, + 480, + 277 + ], + "type": "text", + "content": ", where 8 elements indicate difference in " + }, + { + "bbox": [ + 130, + 164, + 480, + 277 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 164, + 480, + 277 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 164, + 480, + 277 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 130, + 164, + 480, + 277 + ], + "type": "text", + "content": " coordinates, width ratio, height ratio, IOU, center distance, area ratio, and aspect ratio. Additionally, we can also obtain the union bounding box " + }, + { + "bbox": [ + 130, + 164, + 480, + 277 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{b}}_{ij}^u \\in \\mathbb{R}^9" + }, + { + "bbox": [ + 130, + 164, + 480, + 277 + ], + "type": "text", + "content": " for the entity pairs, where 9 elements indicate the union bounding box coordinates " + }, + { + "bbox": [ + 130, + 164, + 480, + 277 + ], + "type": "inline_equation", + "content": "(x_1, y_1, x_2, y_2)" + }, + { + "bbox": [ + 130, + 164, + 480, + 277 + ], + "type": "text", + "content": ", center " + }, + { + "bbox": [ + 130, + 164, + 480, + 277 + ], + "type": "inline_equation", + "content": "(\\frac{x_1 + x_2}{2}, \\frac{y_1 + y_2}{2})" + }, + { + "bbox": [ + 130, + 164, + 480, + 277 + ], + "type": "text", + "content": ", size " + }, + { + "bbox": [ + 130, + 164, + 480, + 277 + ], + "type": "inline_equation", + "content": "(x_2 - x_1, y_2 - y_1)" + }, + { + "bbox": [ + 130, + 164, + 480, + 277 + ], + "type": "text", + "content": " and area " + }, + { + "bbox": [ + 130, + 164, + 480, + 277 + ], + "type": "inline_equation", + "content": "(x_2 - x_1)(y_2 - y_1)" + }, + { + "bbox": [ + 130, + 164, + 480, + 277 + ], + "type": "text", + "content": ", respectively. Then, we can obtain the relevance representation of the subject " + }, + { + "bbox": [ + 130, + 164, + 480, + 277 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 164, + 480, + 277 + ], + "type": "text", + "content": " and object " + }, + { + "bbox": [ + 130, + 164, + 480, + 277 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 130, + 164, + 480, + 277 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 238, + 285, + 480, + 301 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 238, + 285, + 480, + 301 + ], + "spans": [ + { + "bbox": [ + 238, + 285, + 480, + 301 + ], + "type": "interline_equation", + "content": "\\mathbf {e} _ {i j} = \\left[ v _ {i j} ^ {\\text {s i m}}, s _ {i j} ^ {\\text {p r i o r}}, \\mathbf {b} _ {i j} ^ {\\text {o v e r l a p}}, \\hat {\\mathbf {b}} _ {i j} ^ {u} \\right], \\tag {3}", + "image_path": "b1bffa9a56200a42a33ca03926532a7f11d020e76fee3b496544f78e440fe911.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 308, + 479, + 333 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 308, + 479, + 333 + ], + "spans": [ + { + "bbox": [ + 131, + 308, + 479, + 333 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 308, + 479, + 333 + ], + "type": "inline_equation", + "content": "[\\cdot, \\cdot]" + }, + { + "bbox": [ + 131, + 308, + 479, + 333 + ], + "type": "text", + "content": " denotes the concatenation operation. Subsequently, the feature representation of the entity pair can be expressed as" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 197, + 342, + 480, + 358 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 342, + 480, + 358 + ], + "spans": [ + { + "bbox": [ + 197, + 342, + 480, + 358 + ], + "type": "interline_equation", + "content": "\\mathbf {f} _ {i j} = [ \\mathbf {q} _ {i} ^ {s u b}, \\mathbf {q} _ {j} ^ {o b j}, f _ {s} (s _ {i j} ^ {p r i o r}), p o s (\\mathbf {b} _ {i j} ^ {o v e r l a p}), p o s (\\hat {\\mathbf {b}} _ {i j} ^ {u}) ], \\tag {4}", + "image_path": "f752676f87a79a0baec1cc368f39fa088a141e64a4c9976553eb0578d67f29a9.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 365, + 479, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 365, + 479, + 411 + ], + "spans": [ + { + "bbox": [ + 131, + 365, + 479, + 411 + ], + "type": "text", + "content": "Where " + }, + { + "bbox": [ + 131, + 365, + 479, + 411 + ], + "type": "inline_equation", + "content": "f_{s}" + }, + { + "bbox": [ + 131, + 365, + 479, + 411 + ], + "type": "text", + "content": " is a fully-connected layer for language prior encoding, and " + }, + { + "bbox": [ + 131, + 365, + 479, + 411 + ], + "type": "inline_equation", + "content": "pos" + }, + { + "bbox": [ + 131, + 365, + 479, + 411 + ], + "type": "text", + "content": " is a fully-connected layer for positional encoding. Then, GPM uses a gated pooling layer " + }, + { + "bbox": [ + 131, + 365, + 479, + 411 + ], + "type": "inline_equation", + "content": "f_{gate}" + }, + { + "bbox": [ + 131, + 365, + 479, + 411 + ], + "type": "text", + "content": " to obtain the probability of relation between subject " + }, + { + "bbox": [ + 131, + 365, + 479, + 411 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 131, + 365, + 479, + 411 + ], + "type": "text", + "content": " and object " + }, + { + "bbox": [ + 131, + 365, + 479, + 411 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 131, + 365, + 479, + 411 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 231, + 411, + 480, + 428 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 411, + 480, + 428 + ], + "spans": [ + { + "bbox": [ + 231, + 411, + 480, + 428 + ], + "type": "interline_equation", + "content": "\\hat {p} _ {i j} = \\operatorname {S i g m o i d} \\left(\\sum f _ {\\text {g a t e}} \\left(\\mathbf {f} _ {i j}\\right) \\odot \\mathbf {e} _ {i j}\\right), \\tag {5}", + "image_path": "3140f1295c20fd7c50411d54e0b4ebc604e01b81c738ba87ecd3039b040626d4.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 433, + 479, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 433, + 479, + 479 + ], + "spans": [ + { + "bbox": [ + 131, + 433, + 479, + 479 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 433, + 479, + 479 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 131, + 433, + 479, + 479 + ], + "type": "text", + "content": " represents element-wise multiplication. Finally, we can obtain a " + }, + { + "bbox": [ + 131, + 433, + 479, + 479 + ], + "type": "inline_equation", + "content": "N_{e} \\times N_{e}" + }, + { + "bbox": [ + 131, + 433, + 479, + 479 + ], + "type": "text", + "content": " relationship matrix " + }, + { + "bbox": [ + 131, + 433, + 479, + 479 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{r} = \\mathrm{GPM}(\\mathbf{Q}_{e},\\mathbf{C},\\mathbf{B})" + }, + { + "bbox": [ + 131, + 433, + 479, + 479 + ], + "type": "text", + "content": ". Note that we set the diagonal elements of " + }, + { + "bbox": [ + 131, + 433, + 479, + 479 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{r}" + }, + { + "bbox": [ + 131, + 433, + 479, + 479 + ], + "type": "text", + "content": " to zero, removing the cases where an entity serves as both subject and object simultaneously." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 498, + 479, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 479, + 557 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 479, + 557 + ], + "type": "text", + "content": "Multi-Granularity Sparse Relationship Matrix Prediction Based on the output of three GPMs, we can construct three relationship matrices based on the granularity of negative samples and build the corresponding ground truth matrices (see Section 3.6) for supervised learning, enabling each GPM to focus on identifying negative samples of specific granularity." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 558, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 480, + 665 + ], + "type": "text", + "content": "Since the " + }, + { + "bbox": [ + 130, + 558, + 480, + 665 + ], + "type": "inline_equation", + "content": "N_{e}" + }, + { + "bbox": [ + 130, + 558, + 480, + 665 + ], + "type": "text", + "content": " entity proposals are typically much greater than the number of entities contained in an image, resulting in a large number of background entities among all possible pairs of entities. Since background entities do not form relationships, we divide the negative samples in the relationship matrix into three levels of granularity: (1) high-confidence negative samples composed of background entities (background-background), (2) medium-confidence negative samples constructed from background and foreground entities (entity-background or background-entity), and (3) low-confidence negative samples composed of foreground entities (entity-entity)." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 226, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 226, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 226, + 102 + ], + "type": "text", + "content": "L. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 479, + 138 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 479, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 479, + 138 + ], + "type": "text", + "content": "Then we can define three relationship matrices based on the output of the GPMs as follows." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 241, + 149, + 343, + 159 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 241, + 149, + 343, + 159 + ], + "spans": [ + { + "bbox": [ + 241, + 149, + 343, + 159 + ], + "type": "interline_equation", + "content": "\\mathbf {M} _ {r 1} = \\operatorname {G P M - 1} \\left(\\mathbf {Q} _ {e}, \\mathbf {C}, \\mathbf {B}\\right)", + "image_path": "d375342536d598f13e732bbe0a9e0bf2ab2acc628ef2caf4289900d40bbd42ed.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 241, + 163, + 479, + 174 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 241, + 163, + 479, + 174 + ], + "spans": [ + { + "bbox": [ + 241, + 163, + 479, + 174 + ], + "type": "interline_equation", + "content": "\\mathbf {M} _ {r 2} = \\mathbf {M} _ {r 1} + \\operatorname {G P M - 2} \\left(\\mathbf {Q} _ {e}, \\mathbf {C}, \\mathbf {B}\\right) \\tag {6}", + "image_path": "4aa801228576ff96bc1fbfc4a594b9c30c9e021737566daba93541d5ace4668f.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 241, + 177, + 372, + 188 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 241, + 177, + 372, + 188 + ], + "spans": [ + { + "bbox": [ + 241, + 177, + 372, + 188 + ], + "type": "interline_equation", + "content": "\\mathbf {M} _ {r 3} = \\mathbf {M} _ {r 2} + \\operatorname {G P M - 3} (\\mathbf {Q} _ {e}, \\mathbf {C}, \\mathbf {B})", + "image_path": "4ee21fcda25d1b345dcab5b1a904b300ec440db068544e1b7795aca348eb1250.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 201, + 479, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 201, + 479, + 284 + ], + "spans": [ + { + "bbox": [ + 130, + 201, + 479, + 284 + ], + "type": "text", + "content": "In Eq.6, each GPM can focus on learning the filtering of negative samples at different granularities, as follows: (1) GPM-1 focuses on learning from pairs between background and filters out a large number of high-confidence negative samples. (2) GPM-2 focuses on learning from pairs between background and entities and filters out medium-confidence negative samples. (3) GPM-3 focuses on learning from no relation pairs between entities and only needs to filter out entity pairs that do not form relationships from the remaining entity pairs." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 286, + 479, + 333 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 286, + 479, + 333 + ], + "spans": [ + { + "bbox": [ + 131, + 286, + 479, + 333 + ], + "type": "text", + "content": "Based on the final sparse relationship matrix " + }, + { + "bbox": [ + 131, + 286, + 479, + 333 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{r3}" + }, + { + "bbox": [ + 131, + 286, + 479, + 333 + ], + "type": "text", + "content": ", we can select the top-" + }, + { + "bbox": [ + 131, + 286, + 479, + 333 + ], + "type": "inline_equation", + "content": "N_{r}" + }, + { + "bbox": [ + 131, + 286, + 479, + 333 + ], + "type": "text", + "content": " subject-object pairs most likely to form relations and obtain their indices in the entity visual representations " + }, + { + "bbox": [ + 131, + 286, + 479, + 333 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_e" + }, + { + "bbox": [ + 131, + 286, + 479, + 333 + ], + "type": "text", + "content": ", the entity's class predictions " + }, + { + "bbox": [ + 131, + 286, + 479, + 333 + ], + "type": "inline_equation", + "content": "\\mathbf{C}" + }, + { + "bbox": [ + 131, + 286, + 479, + 333 + ], + "type": "text", + "content": ", and the bounding boxes " + }, + { + "bbox": [ + 131, + 286, + 479, + 333 + ], + "type": "inline_equation", + "content": "\\mathbf{B}" + }, + { + "bbox": [ + 131, + 286, + 479, + 333 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 357, + 248, + 368 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 357, + 248, + 368 + ], + "spans": [ + { + "bbox": [ + 132, + 357, + 248, + 368 + ], + "type": "text", + "content": "3.5 Relation Decoder" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 383, + 479, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 383, + 479, + 418 + ], + "spans": [ + { + "bbox": [ + 130, + 383, + 479, + 418 + ], + "type": "text", + "content": "With the " + }, + { + "bbox": [ + 130, + 383, + 479, + 418 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_e" + }, + { + "bbox": [ + 130, + 383, + 479, + 418 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 383, + 479, + 418 + ], + "type": "inline_equation", + "content": "\\mathbf{C}" + }, + { + "bbox": [ + 130, + 383, + 479, + 418 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 383, + 479, + 418 + ], + "type": "inline_equation", + "content": "\\mathbf{B}" + }, + { + "bbox": [ + 130, + 383, + 479, + 418 + ], + "type": "text", + "content": " of subjects and objects, we can obtain the subject-object pair representation " + }, + { + "bbox": [ + 130, + 383, + 479, + 418 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{pair}" + }, + { + "bbox": [ + 130, + 383, + 479, + 418 + ], + "type": "text", + "content": ", which integrates visual, semantic, and positional encodings as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 164, + 430, + 479, + 445 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 164, + 430, + 479, + 445 + ], + "spans": [ + { + "bbox": [ + 164, + 430, + 479, + 445 + ], + "type": "interline_equation", + "content": "\\mathbf {Q} _ {p a i r} = \\left[ \\mathbf {Q} _ {e} ^ {s u b}, e m b (\\mathbf {C} ^ {s u b}), p o s (\\mathbf {B} ^ {s u b}), \\mathbf {Q} _ {e} ^ {o b j}, e m b (\\mathbf {C} ^ {o b j}), p o s (\\mathbf {B} ^ {o b j}) \\right]. \\tag {7}", + "image_path": "6727c77b83f68a852500bead3d4267ff0a86d1cea7bc052b8f38ccdd74c47403.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 456, + 479, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 456, + 479, + 468 + ], + "spans": [ + { + "bbox": [ + 131, + 456, + 479, + 468 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 456, + 479, + 468 + ], + "type": "inline_equation", + "content": "emb" + }, + { + "bbox": [ + 131, + 456, + 479, + 468 + ], + "type": "text", + "content": " is a pre-trained Glove language model to acquire the word embedding." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 470, + 480, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 470, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 130, + 470, + 480, + 555 + ], + "type": "text", + "content": "In this paper, we adopt the relation decoder in [30], which consists of transformer decoders in the style of DETR, to predict relations. In the relation decoder, we initialize a relation query " + }, + { + "bbox": [ + 130, + 470, + 480, + 555 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_r \\in \\mathbb{R}^{N_r \\times d}" + }, + { + "bbox": [ + 130, + 470, + 480, + 555 + ], + "type": "text", + "content": " as the query input, and the subject-object pair " + }, + { + "bbox": [ + 130, + 470, + 480, + 555 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{pair}" + }, + { + "bbox": [ + 130, + 470, + 480, + 555 + ], + "type": "text", + "content": " is projected as the key and value of cross-attention. Subsequently, the relation representation after the Relation Decoder (RD) can be expressed as " + }, + { + "bbox": [ + 130, + 470, + 480, + 555 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{Q}}_r = \\mathrm{RD}(\\mathbf{Q}_r, \\mathbf{Q}_{pair})" + }, + { + "bbox": [ + 130, + 470, + 480, + 555 + ], + "type": "text", + "content": ". Then, the relation prediction for the entity pairs can be expressed as " + }, + { + "bbox": [ + 130, + 470, + 480, + 555 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{r}} = W_{cls} \\tilde{\\mathbf{Q}}_r" + }, + { + "bbox": [ + 130, + 470, + 480, + 555 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 470, + 480, + 555 + ], + "type": "inline_equation", + "content": "W_{cls}" + }, + { + "bbox": [ + 130, + 470, + 480, + 555 + ], + "type": "text", + "content": " is the liner relation classifier." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 579, + 204, + 591 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 579, + 204, + 591 + ], + "spans": [ + { + "bbox": [ + 132, + 579, + 204, + 591 + ], + "type": "text", + "content": "3.6 Training" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 605, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 479, + 665 + ], + "type": "text", + "content": "Our end-to-end SGG is divided into three subtasks: (1) an entity detection task based on the object detector, (2) a sparse relationship matrix prediction task based on the Mg-RMPN, and (3) a relation prediction task based on the relation decoder. During training, each subtask generates the corresponding supervisory information and losses, as follows." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 400, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 400, + 91, + 447, + 102 + ], + "type": "text", + "content": "Mg-RMPN" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "content": "Entity Detection Loss We uses the end-to-end deformable DETR [42] as object detector, which uses the set prediction loss proposed in DETR [1] by assigning the ground truth entities to the predictions. A cost function is applied to compute the matching cost between a prediction and a ground-truth entity. With the cost matrix, the entity prediction-ground truth assignment is computed with the Hungarian Matching [14]. Given a set of " + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "inline_equation", + "content": "N_{e}" + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "content": " entity proposals " + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "inline_equation", + "content": "\\{e_i\\}_{i=1}^{N_e}" + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "content": " from object detector, the set prediction loss for can be presented as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 244, + 209, + 481, + 224 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 244, + 209, + 481, + 224 + ], + "spans": [ + { + "bbox": [ + 244, + 209, + 481, + 224 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {e} = \\Sigma_ {i = 1} ^ {N _ {e}} \\left[ \\mathcal {L} _ {c l s} ^ {e} + \\mathbf {1} _ {e _ {i} \\neq \\phi} \\mathcal {L} _ {b o x} ^ {e} \\right], \\tag {8}", + "image_path": "6384d514d6643cbe0608d63378a4a3b76c744efffad2d54ba348baa7cf784bc2.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 233, + 483, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 233, + 483, + 270 + ], + "spans": [ + { + "bbox": [ + 131, + 233, + 483, + 270 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 233, + 483, + 270 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{cls}^{e}" + }, + { + "bbox": [ + 131, + 233, + 483, + 270 + ], + "type": "text", + "content": " denotes the cross-entropy loss for label classification and " + }, + { + "bbox": [ + 131, + 233, + 483, + 270 + ], + "type": "inline_equation", + "content": "e_i \\neq \\phi" + }, + { + "bbox": [ + 131, + 233, + 483, + 270 + ], + "type": "text", + "content": " means that " + }, + { + "bbox": [ + 131, + 233, + 483, + 270 + ], + "type": "inline_equation", + "content": "<\\text{background}>" + }, + { + "bbox": [ + 131, + 233, + 483, + 270 + ], + "type": "text", + "content": " is not assigned to the " + }, + { + "bbox": [ + 131, + 233, + 483, + 270 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 131, + 233, + 483, + 270 + ], + "type": "text", + "content": "th entity prediction. " + }, + { + "bbox": [ + 131, + 233, + 483, + 270 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{box}^{e}" + }, + { + "bbox": [ + 131, + 233, + 483, + 270 + ], + "type": "text", + "content": " consists of " + }, + { + "bbox": [ + 131, + 233, + 483, + 270 + ], + "type": "inline_equation", + "content": "L_1" + }, + { + "bbox": [ + 131, + 233, + 483, + 270 + ], + "type": "text", + "content": " loss and generalized IoU loss for the box regression." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 289, + 482, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 289, + 482, + 326 + ], + "spans": [ + { + "bbox": [ + 131, + 289, + 482, + 326 + ], + "type": "text", + "content": "Multi-granularity Relationship Matrices Loss To predict the final sparse relationship matrix, we need to define the supervisory information and loss functions during the training phase." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 339, + 482, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 339, + 482, + 398 + ], + "spans": [ + { + "bbox": [ + 131, + 339, + 482, + 398 + ], + "type": "text", + "content": "Multi-granularity Supervision Matrices In the training phase, we assign labels to the entity proposals " + }, + { + "bbox": [ + 131, + 339, + 482, + 398 + ], + "type": "inline_equation", + "content": "\\{e_i\\}_{i=1}^{N_e}" + }, + { + "bbox": [ + 131, + 339, + 482, + 398 + ], + "type": "text", + "content": " by Hungarian matching and obtain the ground truth for each node in " + }, + { + "bbox": [ + 131, + 339, + 482, + 398 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_r" + }, + { + "bbox": [ + 131, + 339, + 482, + 398 + ], + "type": "text", + "content": " regarding the subject, object, and whether they form a relation. In response to the sparsity of " + }, + { + "bbox": [ + 131, + 339, + 482, + 398 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_r" + }, + { + "bbox": [ + 131, + 339, + 482, + 398 + ], + "type": "text", + "content": ", we propose Mg-RMPN to achieve hierarchical identification of negative samples at different granularities." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 399, + 482, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 399, + 482, + 434 + ], + "spans": [ + { + "bbox": [ + 131, + 399, + 482, + 434 + ], + "type": "text", + "content": "Consequently, based on the granularity of the negative samples that each GPM focuses on in Section 3.4, the corresponding ground truth can be obtained. (1) " + }, + { + "bbox": [ + 131, + 399, + 482, + 434 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{r1}^{gt}" + }, + { + "bbox": [ + 131, + 399, + 482, + 434 + ], + "type": "text", + "content": " filters out all pairs composed of background entities as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 147, + 443, + 481, + 471 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 443, + 481, + 471 + ], + "spans": [ + { + "bbox": [ + 147, + 443, + 481, + 471 + ], + "type": "interline_equation", + "content": "\\mathbf {M} _ {r 1} ^ {g t} = \\left\\{ \\right.p _ {i j} \\mid p _ {i j} = \\left\\{\\begin{array}{l l}0&\\text {i f} (e _ {i} = \\phi \\text {a n d} e _ {j} = \\phi) \\text {o r} i = j\\\\1&\\text {o t h e r w i s e}\\end{array}, 1 \\leq i \\leq N _ {e}, 1 \\leq j \\leq N _ {e} \\right\\}, \\tag {9}", + "image_path": "5b39b93b9d671486ff62cedd1984d3eefcb6756309135d938b095cc84487aeda.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 479, + 482, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 479, + 482, + 515 + ], + "spans": [ + { + "bbox": [ + 131, + 479, + 482, + 515 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 479, + 482, + 515 + ], + "type": "inline_equation", + "content": "e_i = \\phi" + }, + { + "bbox": [ + 131, + 479, + 482, + 515 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 131, + 479, + 482, + 515 + ], + "type": "inline_equation", + "content": "e_j = \\phi" + }, + { + "bbox": [ + 131, + 479, + 482, + 515 + ], + "type": "text", + "content": ") means that " + }, + { + "bbox": [ + 131, + 479, + 482, + 515 + ], + "type": "inline_equation", + "content": "<\\text{background}>" + }, + { + "bbox": [ + 131, + 479, + 482, + 515 + ], + "type": "text", + "content": " is assigned to the " + }, + { + "bbox": [ + 131, + 479, + 482, + 515 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 131, + 479, + 482, + 515 + ], + "type": "text", + "content": "th (" + }, + { + "bbox": [ + 131, + 479, + 482, + 515 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 131, + 479, + 482, + 515 + ], + "type": "text", + "content": "th) entity prediction, " + }, + { + "bbox": [ + 131, + 479, + 482, + 515 + ], + "type": "inline_equation", + "content": "i = j" + }, + { + "bbox": [ + 131, + 479, + 482, + 515 + ], + "type": "text", + "content": " means removing self-connected entity pairs. (2) " + }, + { + "bbox": [ + 131, + 479, + 482, + 515 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{r2}^{gt}" + }, + { + "bbox": [ + 131, + 479, + 482, + 515 + ], + "type": "text", + "content": " further filters out all pairs composed of backgrounds and entities as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 148, + 524, + 481, + 552 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 524, + 481, + 552 + ], + "spans": [ + { + "bbox": [ + 148, + 524, + 481, + 552 + ], + "type": "interline_equation", + "content": "\\mathbf {M} _ {r 2} ^ {g t} = \\left\\{ \\right.p _ {i j} \\mid p _ {i j} = \\left\\{\\begin{array}{l l}0&\\text {i f} (e _ {i} = \\phi \\text {o r} e _ {j} = \\phi) \\text {o r} i = j\\\\1&\\text {o t h e r w i s e}\\end{array}, 1 \\leq i \\leq N _ {e}, 1 \\leq j \\leq N _ {e} \\right\\}. \\tag {10}", + "image_path": "6b8e4946882107f6402b1ae1711c5a93a7f6d96177cd7d8324dcb704c0546a08.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 560, + 482, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 560, + 482, + 585 + ], + "spans": [ + { + "bbox": [ + 132, + 560, + 482, + 585 + ], + "type": "text", + "content": "(3) " + }, + { + "bbox": [ + 132, + 560, + 482, + 585 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{r3}^{gt}" + }, + { + "bbox": [ + 132, + 560, + 482, + 585 + ], + "type": "text", + "content": " filters out all non-relationship pairs, and " + }, + { + "bbox": [ + 132, + 560, + 482, + 585 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{r3}^{gt}" + }, + { + "bbox": [ + 132, + 560, + 482, + 585 + ], + "type": "text", + "content": " is also the final ground truth relationship matrix of the scene graph, expressed as follows" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 194, + 594, + 481, + 620 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 594, + 481, + 620 + ], + "spans": [ + { + "bbox": [ + 194, + 594, + 481, + 620 + ], + "type": "interline_equation", + "content": "\\mathbf {M} _ {r 3} ^ {g t} = \\left\\{p _ {i j} \\mid p _ {i j} = \\left\\{ \\begin{array}{l l} 0 & \\text {o t h e r w i s e} \\\\ 1 & \\text {i f} r _ {i j} = 1 \\end{array} , 1 \\leq i \\leq N _ {e}, 1 \\leq j \\leq N _ {e} \\right\\}, \\right. \\tag {11}", + "image_path": "e2791d47a007588ee89e1ffe2338a13f6e09b6d05d1bd81c8b9b3ac10049b4cc.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 629, + 482, + 669 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 629, + 482, + 669 + ], + "spans": [ + { + "bbox": [ + 131, + 629, + 482, + 669 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 629, + 482, + 669 + ], + "type": "inline_equation", + "content": "r_{ij} = 1" + }, + { + "bbox": [ + 131, + 629, + 482, + 669 + ], + "type": "text", + "content": " implies that there is a relation between subject " + }, + { + "bbox": [ + 131, + 629, + 482, + 669 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 131, + 629, + 482, + 669 + ], + "type": "text", + "content": " and object " + }, + { + "bbox": [ + 131, + 629, + 482, + 669 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 131, + 629, + 482, + 669 + ], + "type": "text", + "content": ". Here, the relationships of the negative sample subsets in " + }, + { + "bbox": [ + 131, + 629, + 482, + 669 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{r1}^{gt}" + }, + { + "bbox": [ + 131, + 629, + 482, + 669 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 629, + 482, + 669 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{r2}^{gt}" + }, + { + "bbox": [ + 131, + 629, + 482, + 669 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 629, + 482, + 669 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{r3}^{gt}" + }, + { + "bbox": [ + 131, + 629, + 482, + 669 + ], + "type": "text", + "content": " are " + }, + { + "bbox": [ + 131, + 629, + 482, + 669 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{r1|p_{ij} = 0}^{gt} \\subseteq \\mathbf{M}_{r2|p_{ij} = 0}^{gt} \\subseteq \\mathbf{M}_{r3|p_{ij} = 0}^{gt}" + }, + { + "bbox": [ + 131, + 629, + 482, + 669 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 226, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 226, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 226, + 102 + ], + "type": "text", + "content": "L. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 131, + 116, + 479, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 116, + 479, + 172 + ], + "spans": [ + { + "bbox": [ + 131, + 116, + 479, + 172 + ], + "type": "text", + "content": "Multi-granularity Relationship Matrices Loss Considering that the large number of negative samples in " + }, + { + "bbox": [ + 131, + 116, + 479, + 172 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_r" + }, + { + "bbox": [ + 131, + 116, + 479, + 172 + ], + "type": "text", + "content": " have different confidence, we modify the binary cross-entropy function using focal loss [20], which applies weighted discrimination on well-classified samples, forcing the model to focus on wrongly classified samples. Then, the multi-granularity learning loss is" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 179, + 179, + 479, + 192 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 179, + 479, + 192 + ], + "spans": [ + { + "bbox": [ + 179, + 179, + 479, + 192 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {M g} = \\mathcal {L} _ {R M} ^ {1} \\left(\\mathbf {M} _ {r 1}, \\mathbf {M} _ {r 1} ^ {g t}\\right) + \\mathcal {L} _ {R M} ^ {2} \\left(\\mathbf {M} _ {r 2}, \\mathbf {M} _ {r 2} ^ {g t}\\right) + \\mathcal {L} _ {R M} ^ {3} \\left(\\mathbf {M} _ {r 3}, \\mathbf {M} _ {r 3} ^ {g t}\\right), \\tag {12}", + "image_path": "1ae2a551e79d155b7bf89da2d928f3ad88132baafa860a57a88ccfa30c16eac3.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 134, + 198, + 426, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 198, + 426, + 212 + ], + "spans": [ + { + "bbox": [ + 134, + 198, + 426, + 212 + ], + "type": "text", + "content": "where the relationship matrices loss " + }, + { + "bbox": [ + 134, + 198, + 426, + 212 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{RM}^{k}(k = 1,2,3)" + }, + { + "bbox": [ + 134, + 198, + 426, + 212 + ], + "type": "text", + "content": " is defined as" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 189, + 219, + 480, + 270 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 189, + 219, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 189, + 219, + 480, + 270 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {R M} ^ {k} \\left(\\mathbf {M} _ {r}, \\mathbf {M} _ {r} ^ {g t}\\right) = - \\sum_ {i} \\sum_ {j} \\left\\{\\alpha_ {k} \\left(1 - \\hat {p} _ {i j}\\right) ^ {\\gamma} p _ {i j} \\log (\\hat {p} _ {i j}) \\right. \\\\ + (1 - \\alpha_ {k}) \\hat {p} _ {i j} ^ {\\gamma} (1 - p _ {i j}) \\log (1 - \\hat {p} _ {i j}) \\} / N _ {p o s} \\tag {13} \\\\ + \\left\\| \\mathbf {M} _ {r} \\right\\| _ {1} + \\left\\| \\mathbf {M} _ {r} \\right\\| _ {2}, \\\\ \\end{array}", + "image_path": "df1ce1b957f6bf6a1b82e909a8fd16e2243d65789bd08a2c42596a31ab341ce0.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 277, + 479, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 277, + 479, + 324 + ], + "spans": [ + { + "bbox": [ + 131, + 277, + 479, + 324 + ], + "type": "text", + "content": "where the first part is the focal loss based on binary cross-entropy, and the second part is the L1 and L2 regularization of " + }, + { + "bbox": [ + 131, + 277, + 479, + 324 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_r" + }, + { + "bbox": [ + 131, + 277, + 479, + 324 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 277, + 479, + 324 + ], + "type": "inline_equation", + "content": "\\alpha_k(k = 1,2,3)" + }, + { + "bbox": [ + 131, + 277, + 479, + 324 + ], + "type": "text", + "content": " is hyperparameters that adjusts the weights of positive and negative samples, " + }, + { + "bbox": [ + 131, + 277, + 479, + 324 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 131, + 277, + 479, + 324 + ], + "type": "text", + "content": " is a hyperparameter for focal loss, and " + }, + { + "bbox": [ + 131, + 277, + 479, + 324 + ], + "type": "inline_equation", + "content": "N_{pos}" + }, + { + "bbox": [ + 131, + 277, + 479, + 324 + ], + "type": "text", + "content": " is the number of positive samples in each mini-batch." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 341, + 479, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 341, + 479, + 407 + ], + "spans": [ + { + "bbox": [ + 131, + 341, + 479, + 407 + ], + "type": "text", + "content": "Relation Prediction Loss In this paper, the relation decoder also considers relation prediction as a set prediction based on a query. Therefore, during the training phase, we also use Hungarian Matching to assign labels for relation prediction. Due to the imbalanced relation classes in SGG, we use Seesaw loss [29] to dynamically adjust the gradients of samples of different classes. The relation loss is as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 212, + 415, + 479, + 443 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 212, + 415, + 479, + 443 + ], + "spans": [ + { + "bbox": [ + 212, + 415, + 479, + 443 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {r} = - \\sum_ {i = 1} ^ {C _ {r}} y _ {i} \\log \\left(\\hat {\\sigma} _ {i}\\right), \\quad \\hat {\\sigma} _ {i} = \\frac {e ^ {\\hat {r} _ {i}}}{\\sum_ {i \\neq j} ^ {C _ {r}} \\mathcal {S} _ {i j} e ^ {\\hat {r} _ {j}} + e ^ {\\hat {r} _ {i}}}, \\tag {14}", + "image_path": "daa18391ab23f91e7340a623966209bf27729bed4080e27f07e6af3395419508.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 449, + 479, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 449, + 479, + 486 + ], + "spans": [ + { + "bbox": [ + 131, + 449, + 479, + 486 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 449, + 479, + 486 + ], + "type": "inline_equation", + "content": "C_r" + }, + { + "bbox": [ + 131, + 449, + 479, + 486 + ], + "type": "text", + "content": " is the number of relation classes, " + }, + { + "bbox": [ + 131, + 449, + 479, + 486 + ], + "type": "inline_equation", + "content": "y_i \\in \\{0,1\\}" + }, + { + "bbox": [ + 131, + 449, + 479, + 486 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 449, + 479, + 486 + ], + "type": "inline_equation", + "content": "1 \\leqslant i \\leqslant C_r" + }, + { + "bbox": [ + 131, + 449, + 479, + 486 + ], + "type": "text", + "content": " is the one-hot ground truth label, " + }, + { + "bbox": [ + 131, + 449, + 479, + 486 + ], + "type": "inline_equation", + "content": "S_{ij}" + }, + { + "bbox": [ + 131, + 449, + 479, + 486 + ], + "type": "text", + "content": " is a tunable balancing factor between different classes, detail can be found in [29]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 146, + 487, + 339, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 487, + 339, + 498 + ], + "spans": [ + { + "bbox": [ + 146, + 487, + 339, + 498 + ], + "type": "text", + "content": "In summary, the total loss of our method is" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 252, + 510, + 479, + 520 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 252, + 510, + 479, + 520 + ], + "spans": [ + { + "bbox": [ + 252, + 510, + 479, + 520 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {e} + \\lambda_ {1} \\mathcal {L} _ {M g} + \\lambda_ {2} \\mathcal {L} _ {r}, \\tag {15}", + "image_path": "2c9a96f136647a352e4bbcb3f86a500f29d21b0a8867f656ef75cb00bd6226f7.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 529, + 488, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 529, + 488, + 555 + ], + "spans": [ + { + "bbox": [ + 131, + 529, + 488, + 555 + ], + "type": "text", + "content": "Where " + }, + { + "bbox": [ + 131, + 529, + 488, + 555 + ], + "type": "inline_equation", + "content": "\\lambda_{1}" + }, + { + "bbox": [ + 131, + 529, + 488, + 555 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 529, + 488, + 555 + ], + "type": "inline_equation", + "content": "\\lambda_{2}" + }, + { + "bbox": [ + 131, + 529, + 488, + 555 + ], + "type": "text", + "content": " are hyperparameters used to respectively adjust multi-granularity learning loss " + }, + { + "bbox": [ + 131, + 529, + 488, + 555 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{Mg}" + }, + { + "bbox": [ + 131, + 529, + 488, + 555 + ], + "type": "text", + "content": " and relation prediction loss " + }, + { + "bbox": [ + 131, + 529, + 488, + 555 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_r" + }, + { + "bbox": [ + 131, + 529, + 488, + 555 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 572, + 229, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 572, + 229, + 586 + ], + "spans": [ + { + "bbox": [ + 131, + 572, + 229, + 586 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 131, + 597, + 272, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 597, + 272, + 609 + ], + "spans": [ + { + "bbox": [ + 131, + 597, + 272, + 609 + ], + "type": "text", + "content": "4.1 Experimental Settings" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 617, + 479, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 479, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 479, + 666 + ], + "type": "text", + "content": "Datasets: We evaluate our proposed method on the following datasets: (1) Visual Genome (VG150) [13] is the most widely used dataset in SGG, consisting of the most frequent 150 object classes and 50 predicate classes. (2) Open Images V6 (OIv6) [15] is a large-scale dataset proposed by Google. We follow" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 400, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 400, + 91, + 447, + 102 + ], + "type": "text", + "content": "Mg-RMPN" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 174 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 174 + ], + "type": "text", + "content": "the data processing and evaluation protocols in [17, 40]. OIv6 consists of 601 object classes and 30 predicate classes. (2) Generalized Question Answering (GQA200) [9] is another vision and language benchmark with more than 3.8M relation annotations. We follow the data processing in [7], which consists of Top-200 object classes and Top-100 predicate classes." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 175, + 480, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 175, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 130, + 175, + 480, + 281 + ], + "type": "text", + "content": "Task & Evaluation Metrics: In this work, we focus on the Scene Graph Detection task, which detects all objects in an image and predicts their bounding boxes, labels, and relations. The initial evaluation metric Recall@K (R@K) was found to be dominated by head classes. Therefore, the mean Recall@K (mR@K) across all relations classes is proposed to evaluate unbiased SGG. However, focusing solely on mR@K and neglecting R@K can result in a tail bias. Therefore, we adopts the harmonic mean F@K of R@K and mR@K as an overall metric. For OIv6, the weighted evaluation metrics (wmAP rel, wnAP phr, scorewtd) are used for a more class-balanced evaluation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 283, + 481, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 283, + 481, + 402 + ], + "spans": [ + { + "bbox": [ + 130, + 283, + 481, + 402 + ], + "type": "text", + "content": "Implementation Details: We utilized the pre-trained deformable DETR [42] as the object detector. We set the loss hyperparameters " + }, + { + "bbox": [ + 130, + 283, + 481, + 402 + ], + "type": "inline_equation", + "content": "\\lambda_{1}" + }, + { + "bbox": [ + 130, + 283, + 481, + 402 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 283, + 481, + 402 + ], + "type": "inline_equation", + "content": "\\lambda_{2}" + }, + { + "bbox": [ + 130, + 283, + 481, + 402 + ], + "type": "text", + "content": " to 0.5 and 3 respectively. For the multi-granularity relationship matrices loss, we set " + }, + { + "bbox": [ + 130, + 283, + 481, + 402 + ], + "type": "inline_equation", + "content": "\\alpha_{k}(k = 1,2,3)" + }, + { + "bbox": [ + 130, + 283, + 481, + 402 + ], + "type": "text", + "content": " to 0.75, 0.9, and 0.99, respectively, to control the weight of positive samples, with the focal loss parameter " + }, + { + "bbox": [ + 130, + 283, + 481, + 402 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 130, + 283, + 481, + 402 + ], + "type": "text", + "content": " set to 2. The number of entities predicted by the detector " + }, + { + "bbox": [ + 130, + 283, + 481, + 402 + ], + "type": "inline_equation", + "content": "N_{e}" + }, + { + "bbox": [ + 130, + 283, + 481, + 402 + ], + "type": "text", + "content": " is 100, the number of most probable subject-object pairs selected " + }, + { + "bbox": [ + 130, + 283, + 481, + 402 + ], + "type": "inline_equation", + "content": "N_{r}" + }, + { + "bbox": [ + 130, + 283, + 481, + 402 + ], + "type": "text", + "content": " is 100, and the size of embedding dimensions " + }, + { + "bbox": [ + 130, + 283, + 481, + 402 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 130, + 283, + 481, + 402 + ], + "type": "text", + "content": " is 256. Our model is implemented on 8 NVIDIA 3090 GPUs with learning rate " + }, + { + "bbox": [ + 130, + 283, + 481, + 402 + ], + "type": "inline_equation", + "content": "\\times 10^{-4}" + }, + { + "bbox": [ + 130, + 283, + 481, + 402 + ], + "type": "text", + "content": " and batch size 16 for 24 epochs. More implementation details are shown in the Supplementary Material." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 419, + 336, + 431 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 419, + 336, + 431 + ], + "spans": [ + { + "bbox": [ + 131, + 419, + 336, + 431 + ], + "type": "text", + "content": "4.2 Comparison with State of the Arts" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 437, + 480, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 480, + 533 + ], + "type": "text", + "content": "VG150: Table 1 shows the comparison of our method on VG150, with methods divided into two-stage (above) and end-to-end (below). Due to the imbalance issue in SGG, the two-stage methods (Motifs, VCTree, GPS-Net, RelDN, PE-Net) perform better in R@K but show very poor performance in mR@K. Thus, many rebalancing techniques (VCTree+TDE, BGNN, SHA+GCL) have been developed. Although they improve mR@K, they inevitably impair R@K. For example, SHA+GCL, despite achieving the best mR@K, its recall performance significantly deteriorates, with R@50/100 being only 14.9/18.2." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 534, + 480, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 534, + 480, + 604 + ], + "spans": [ + { + "bbox": [ + 130, + 534, + 480, + 604 + ], + "type": "text", + "content": "Although the end-to-end methods are inferior to the two-stage methods on R@K, it demonstrates better potential on mR@K due to its advantage in sparse relation prediction, which does not dilute the scarce tail-class samples. Furthermore, compared to the rebalancing techniques of the two-stage methods, the end-to-end approach does not overly harm R@K and can achieve superior overall performance F@K, such as SGTR and Pair-Net." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 605, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 480, + 665 + ], + "type": "text", + "content": "Our Mg-RMPN leads comprehensively in end-to-end methods. Compared with two-stage methods, Mg-RMPN achieves competitive performance in R@K and significantly excels in mR@K, achieving the best overall performance F@K. For example, Mg-RMPN achieved 19.3/22.8 on F@50/100, surpassing the end-to-end and two-stage optimal methods Pair-Net and PE-Net, respectively." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "text", + "content": "L. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 146, + 481, + 352 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 481, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 481, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 481, + 137 + ], + "type": "text", + "content": "Table 1: Performance comparison at scene graph detection task on VG150. * denotes the results from [16]." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 146, + 481, + 352 + ], + "lines": [ + { + "bbox": [ + 132, + 146, + 481, + 352 + ], + "spans": [ + { + "bbox": [ + 132, + 146, + 481, + 352 + ], + "type": "table", + "html": "
MethodR@20 R@50 R@100mR@20 mR@50 mR@100F@20 F@50 F@100
Motifs [36]25.532.837.25.06.87.98.411.313.0
VCTree [27]24.531.936.25.47.48.78.812.014.0
VCTree+TDE [26]14.019.423.26.99.311.19.212.615.0
GPS-Net [21]22.328.933.26.98.79.810.513.415.1
BGNN [17]-31.035.8-10.712.6-15.918.6
RelDN [39]-31.435.9-6.07.3-10.112.1
SHA+GCL [7]-14.918.214.217.920.9-16.319.5
PE-Net [40]-30.735.2-12.414.5-17.720.5
FCSGG [22]16.121.325.12.73.64.24.66.27.2
RelTR [6]21.227.5-6.810.8-10.315.5-
AS-Net* [3]-18.721.1-6.17.2-9.210.7
HOTR* [12]-23.527.7-9.412.0-13.416.7
SGTR [16]-24.628.4-12.015.2-16.119.8
Pair-Net [30]18.824.929.38.912.415.412.116.620.2
Mg-RMPN(DETR)21.027.131.39.913.516.213.518.021.3
Mg-RMPN(Ours)22.529.133.510.314.417.314.119.322.8
", + "image_path": "172f1254352a9b345e79d3965ca305b7e774d3ad9d882cc44a9341ab493c762d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 394, + 481, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 394, + 481, + 478 + ], + "spans": [ + { + "bbox": [ + 130, + 394, + 481, + 478 + ], + "type": "text", + "content": "We also compared with Pair-Net that is also based on relationship matrix learning. The experimental results show that our method is comprehensively superior to it, which demonstrates the effectiveness of our proposed Mg-RMPN. Furthermore, we also present the results of Mg-RMPN based on the DETR object detector to eliminate the influence of the detector. The results show that our Mg-RMPN(DETR) still achieves the best overall performance among both one-stage and two-stage methods." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 487, + 481, + 572 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 487, + 481, + 572 + ], + "spans": [ + { + "bbox": [ + 130, + 487, + 481, + 572 + ], + "type": "text", + "content": "Open Image V6: To validate the generalizability of Mg-RMPN, we performed experiments on Open Images V6 in Table 2 and compared with two-stage (above) and end-to-end (below) methods. Compared to classic SGG benchmarks, our method achieved the optimal results among all methods with 45.5, 77.8, 57.4 respectively in mR@50, R@50, F@50. Compared to open image benchmarks, we obtained competitive 35.5 and 36.4 in " + }, + { + "bbox": [ + 130, + 487, + 481, + 572 + ], + "type": "inline_equation", + "content": "\\mathrm{wmAP}_{\\mathrm{rel}}" + }, + { + "bbox": [ + 130, + 487, + 481, + 572 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 487, + 481, + 572 + ], + "type": "inline_equation", + "content": "\\mathrm{wmAP}_{\\mathrm{phr}}" + }, + { + "bbox": [ + 130, + 487, + 481, + 572 + ], + "type": "text", + "content": ", and achieved the best result 43.6 in the overall metric score" + }, + { + "bbox": [ + 130, + 487, + 481, + 572 + ], + "type": "inline_equation", + "content": "_{\\mathrm{wtd}}" + }, + { + "bbox": [ + 130, + 487, + 481, + 572 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 581, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 481, + 665 + ], + "type": "text", + "content": "GQA200: We also conducted experiments on the more challenging GQA200, as shown in Table 3. Compared to two-stage methods (VTransE, Motifs, VCTree, SHA+GCL), our method's overall performance F@K is only slightly inferior to SHA+GCL on F@100, which sacrificed a significant amount of R@K to enhance mR@K by using rebalancing techniques. Compared to the end-to-end method Pair-Net, which is also based on relationship matrices, our method is comprehensively superior." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 400, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 400, + 91, + 447, + 102 + ], + "type": "text", + "content": "Mg-RMPN" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 161, + 146, + 452, + 308 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "type": "text", + "content": "Table 2: Performance comparison on OIV6. * denotes the results from [16,17]. †denote results reproduced with the authors' code." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 161, + 146, + 452, + 308 + ], + "lines": [ + { + "bbox": [ + 161, + 146, + 452, + 308 + ], + "spans": [ + { + "bbox": [ + 161, + 146, + 452, + 308 + ], + "type": "table", + "html": "
MethodmR@50R@50F@50wmAPrelwmAPphrscorewtd
Motifs* [36]32.771.644.929.931.638.9
VCTree* [27]33.974.146.534.233.140.2
RelDN* [39]34.073.146.432.233.440.8
G-RCNN* [34]34.074.546.733.234.241.8
GPS-Net* [21]35.374.848.032.934.041.7
BGNN* [17]40.575.052.633.534.242.1
RelTR [6]-71.7-37.237.543.0
AS-Ne* [3]35.255.343.025.927.532.4
HOTR* [12]40.152.745.519.421.526.7
SGTR* [16]42.659.949.837.038.742.3
Pair-Net† [30]44.577.456.531.832.440.3
Mg-RMPN(Ours)45.577.857.435.536.443.6
", + "image_path": "a20cd7efe416397520b77551a1932cec69569fac6ff7905d824035208aff488a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 168, + 349, + 446, + 446 + ], + "blocks": [ + { + "bbox": [ + 132, + 318, + 482, + 340 + ], + "lines": [ + { + "bbox": [ + 132, + 318, + 482, + 340 + ], + "spans": [ + { + "bbox": [ + 132, + 318, + 482, + 340 + ], + "type": "text", + "content": "Table 3: Performance comparison at scene graph detection task on GQA200. * denotes the results from [7]. †denote results reproduced with the authors' code." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 168, + 349, + 446, + 446 + ], + "lines": [ + { + "bbox": [ + 168, + 349, + 446, + 446 + ], + "spans": [ + { + "bbox": [ + 168, + 349, + 446, + 446 + ], + "type": "table", + "html": "
MethodR@50R@100mR@50mR@100F@50F@100
VTransE* [38]27.230.75.86.69.610.9
Motifs* [36]28.933.16.47.710.512.5
VCTree* [27]28.331.96.57.410.612.0
SHA+GCL* [7]14.817.917.820.116.218.9
Pair-Net† [30]20.223.410.612.613.916.4
Mg-RMPN (Ours)23.225.712.814.516.518.5
", + "image_path": "515f092d453578df0b2379233a8aeb17c4e98a0ad2882616d6c3abc25a3e34b3.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 467, + 448, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 467, + 448, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 467, + 448, + 479 + ], + "type": "text", + "content": "4.3 Visualization of Multi-Granularity Relationship Matrices" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 486, + 482, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 482, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 482, + 593 + ], + "type": "text", + "content": "Figure 3 presents a visualization of the multi-granularity relationship matrices learned by our Mg-RMPN, where (d), (e), (f) are the sparse relationship matrices learned after being filtered by GPM-1, GPM-2, GPM-3, respectively, and (a), (b), (c) are the corresponding ground truth relationship matrices defined in Section 3.6. The visualization results show that the dense negative samples in the relationship matrix are filtered layer by layer according to different granularities. The consistency of the generated multi-granularity relationship matrices with their ground truth indicates that the three GPMs of Mg-RMPN indeed focus on filtering negative samples of different granularities." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "Figures 3 (h) and (g), respectively, present a zoomed-in view of the final relationship matrix and its ground truth. We can observe that the predicted relationship matrix successfully captures the true relationships, for example: \"man wearing shirt\", \"man holding racket\", \"man wearing short\" and \"pole on fence\". Additionally, the relationship matrix also predicts some relationships that do not in the ground truth, as illustrated by the dashed lines in the rows" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "text", + "content": "L. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 142, + 114, + 222, + 200 + ], + "blocks": [ + { + "bbox": [ + 142, + 114, + 222, + 200 + ], + "lines": [ + { + "bbox": [ + 142, + 114, + 222, + 200 + ], + "spans": [ + { + "bbox": [ + 142, + 114, + 222, + 200 + ], + "type": "image", + "image_path": "88eef34e090e017fc0bd9b278b104f472bf5f0d55181d294db4b8087b487bd84.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 155, + 204, + 212, + 213 + ], + "lines": [ + { + "bbox": [ + 155, + 204, + 212, + 213 + ], + "spans": [ + { + "bbox": [ + 155, + 204, + 212, + 213 + ], + "type": "text", + "content": "(a) Ground truth " + }, + { + "bbox": [ + 155, + 204, + 212, + 213 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{F1}^{gt}" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 223, + 124, + 298, + 200 + ], + "blocks": [ + { + "bbox": [ + 223, + 124, + 298, + 200 + ], + "lines": [ + { + "bbox": [ + 223, + 124, + 298, + 200 + ], + "spans": [ + { + "bbox": [ + 223, + 124, + 298, + 200 + ], + "type": "image", + "image_path": "7f4dac97d7411486991dc6e8f4577243682de9501f1db16f2f66920843bb7d49.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 230, + 204, + 287, + 213 + ], + "lines": [ + { + "bbox": [ + 230, + 204, + 287, + 213 + ], + "spans": [ + { + "bbox": [ + 230, + 204, + 287, + 213 + ], + "type": "text", + "content": "(b) Ground truth " + }, + { + "bbox": [ + 230, + 204, + 287, + 213 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{\\pmb{F}_2}^{gt}" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 299, + 124, + 376, + 200 + ], + "blocks": [ + { + "bbox": [ + 299, + 124, + 376, + 200 + ], + "lines": [ + { + "bbox": [ + 299, + 124, + 376, + 200 + ], + "spans": [ + { + "bbox": [ + 299, + 124, + 376, + 200 + ], + "type": "image", + "image_path": "b1efc3e7320fa5abbc194866758cb6355c002a968debd24867c1edf86df96e62.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 310, + 204, + 365, + 213 + ], + "lines": [ + { + "bbox": [ + 310, + 204, + 365, + 213 + ], + "spans": [ + { + "bbox": [ + 310, + 204, + 365, + 213 + ], + "type": "text", + "content": "(c) Ground truth " + }, + { + "bbox": [ + 310, + 204, + 365, + 213 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{r3}^{gt}" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 377, + 114, + 471, + 213 + ], + "blocks": [ + { + "bbox": [ + 377, + 114, + 471, + 213 + ], + "lines": [ + { + "bbox": [ + 377, + 114, + 471, + 213 + ], + "spans": [ + { + "bbox": [ + 377, + 114, + 471, + 213 + ], + "type": "image", + "image_path": "cbf961e879467d824496405e8dc3935189dc0a31fb70385f987abee430bab51f.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 142, + 220, + 220, + 295 + ], + "blocks": [ + { + "bbox": [ + 142, + 220, + 220, + 295 + ], + "lines": [ + { + "bbox": [ + 142, + 220, + 220, + 295 + ], + "spans": [ + { + "bbox": [ + 142, + 220, + 220, + 295 + ], + "type": "image", + "image_path": "388cd83ade6d94c4e7f913984517fe93d16f09004ede628cce35e0a493a3057f.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 149, + 297, + 217, + 306 + ], + "lines": [ + { + "bbox": [ + 149, + 297, + 217, + 306 + ], + "spans": [ + { + "bbox": [ + 149, + 297, + 217, + 306 + ], + "type": "text", + "content": "(d) Multi-granularity " + }, + { + "bbox": [ + 149, + 297, + 217, + 306 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{r1}" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 221, + 220, + 298, + 296 + ], + "blocks": [ + { + "bbox": [ + 221, + 220, + 298, + 296 + ], + "lines": [ + { + "bbox": [ + 221, + 220, + 298, + 296 + ], + "spans": [ + { + "bbox": [ + 221, + 220, + 298, + 296 + ], + "type": "image", + "image_path": "5e822dc4ef2812098e1887bd1083d937525a4850d0f18fa7e3c2e4550fd2293d.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 227, + 297, + 294, + 306 + ], + "lines": [ + { + "bbox": [ + 227, + 297, + 294, + 306 + ], + "spans": [ + { + "bbox": [ + 227, + 297, + 294, + 306 + ], + "type": "text", + "content": "(e) Multi-granularity " + }, + { + "bbox": [ + 227, + 297, + 294, + 306 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{r2}" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 299, + 220, + 377, + 295 + ], + "blocks": [ + { + "bbox": [ + 299, + 220, + 377, + 295 + ], + "lines": [ + { + "bbox": [ + 299, + 220, + 377, + 295 + ], + "spans": [ + { + "bbox": [ + 299, + 220, + 377, + 295 + ], + "type": "image", + "image_path": "24a68a31da06af6c687428d0263c295d987826ae1df5042e28a687bf274fd2ba.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 297, + 372, + 306 + ], + "lines": [ + { + "bbox": [ + 305, + 297, + 372, + 306 + ], + "spans": [ + { + "bbox": [ + 305, + 297, + 372, + 306 + ], + "type": "text", + "content": "(f) Multi-granularity " + }, + { + "bbox": [ + 305, + 297, + 372, + 306 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{r3}" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 377, + 214, + 471, + 312 + ], + "blocks": [ + { + "bbox": [ + 377, + 214, + 471, + 312 + ], + "lines": [ + { + "bbox": [ + 377, + 214, + 471, + 312 + ], + "spans": [ + { + "bbox": [ + 377, + 214, + 471, + 312 + ], + "type": "image", + "image_path": "074f5772091af3359ebfaf9bb09e99ad75a17ec8c2ac8cdba57953f4a205916f.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 174, + 321, + 438, + 333 + ], + "lines": [ + { + "bbox": [ + 174, + 321, + 438, + 333 + ], + "spans": [ + { + "bbox": [ + 174, + 321, + 438, + 333 + ], + "type": "text", + "content": "Fig. 3: Visualization of Multi-Granularity Relationship Matrices" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "table", + "bbox": [ + 151, + 366, + 462, + 468 + ], + "blocks": [ + { + "bbox": [ + 174, + 346, + 438, + 357 + ], + "lines": [ + { + "bbox": [ + 174, + 346, + 438, + 357 + ], + "spans": [ + { + "bbox": [ + 174, + 346, + 438, + 357 + ], + "type": "text", + "content": "Table 4: Ablation studies of Mg-RMPN components on VG150." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 151, + 366, + 462, + 468 + ], + "lines": [ + { + "bbox": [ + 151, + 366, + 462, + 468 + ], + "spans": [ + { + "bbox": [ + 151, + 366, + 462, + 468 + ], + "type": "table", + "html": "
GPM-1GPM-2GPM-3R@20/50/100mR@20/50/100F@20/50/100
✓✘✘6.5/10.7/15.22.9/4.8/6.64.0/6.6/9.2
✘✓✘12.5/17.5/21.65.2/7.4/9.37.3/10.4/13.0
✓✓✘12.8/18.1/22.35.8/8.5/10.78.0/11.6/14.5
✘✘✓21.2/27.6/32.19.2/13.1/15.812.8/17.8/21.2
✓✘✓21.7/28.2/32.69.8/13.8/16.513.5/18.5/21.9
✘✓✓21.8/28.2/32.59.7/13.7/16.713.4/18.4/22.1
✓✓✓22.5/29.1/33.510.3/14.4/17.314.1/19.3/22.8
", + "image_path": "b69a63d3072eb708aace50b2939ff9dfd45abb474c363642fe1232553cdc0908.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "table_body" + } + ], + "index": 18 + }, + { + "bbox": [ + 130, + 491, + 482, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 491, + 482, + 563 + ], + "spans": [ + { + "bbox": [ + 130, + 491, + 482, + 563 + ], + "type": "text", + "content": "and columns, which represent the relationships with the corresponding entity \"man\" as the subject and object, respectively. Due to the abundant presence of \"man\" in SGG, the relationship matrix tends to favor predicting relations that include \"man\", which aligns with the actual language induction bias. Therefore, the results of visualization indicate that our Mg-RMPN can accurately predict the sparse relationship matrix." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 131, + 583, + 246, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 583, + 246, + 594 + ], + "spans": [ + { + "bbox": [ + 131, + 583, + 246, + 594 + ], + "type": "text", + "content": "4.4 Ablation Studies" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "Component Analysis of Mg-RMPN. To assess the effectiveness of each GPM in Mg-RMPN, we performed ablation experiments in Table 4. In Mg-RMPN, each GPM focuses on negative sample filtering of different granularities, and the three GPMs collaborate to complete the learning of the sparse relationship matrix. GPM-1 and GPM-2 are used to help GPM-3 filter negative" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 400, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 400, + 91, + 447, + 102 + ], + "type": "text", + "content": "Mg-RMPN" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 149, + 136, + 465, + 205 + ], + "blocks": [ + { + "bbox": [ + 137, + 114, + 475, + 125 + ], + "lines": [ + { + "bbox": [ + 137, + 114, + 475, + 125 + ], + "spans": [ + { + "bbox": [ + 137, + 114, + 475, + 125 + ], + "type": "text", + "content": "Table 5: Ablation studies of different input information for Mg-RMPN on VG150." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 149, + 136, + 465, + 205 + ], + "lines": [ + { + "bbox": [ + 149, + 136, + 465, + 205 + ], + "spans": [ + { + "bbox": [ + 149, + 136, + 465, + 205 + ], + "type": "table", + "html": "
Visual Semantic PositionR@20/50/100mR@20/50/100F@20/50/100
✓XX21.2/27.6/32.19.8/13.6/16.413.4/18.2/21.7
✓X✓22.2/28.7/33.110.1/14.1/16.913.9/18.9/22.4
✓✓X21.2/27.6/32.29.8/13.8/16.813.4/18.4/22.1
✓✓✓22.5/29.1/33.510.3/14.4/17.314.1/19.3/22.8
", + "image_path": "9e52e5561fcccb77c2284235e15063cfc4097ca68ef1be3395a96ccec106d27b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 230, + 479, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 230, + 479, + 314 + ], + "spans": [ + { + "bbox": [ + 130, + 230, + 479, + 314 + ], + "type": "text", + "content": "samples. Using GPM-1 and GPM-2 individually cannot make full use of the ground truth, hence their results are very poor. The results obtained using only GPM-3 are also not good enough. Further increasing the assistance of GPMs for negative sample filtering can significantly improve the performance of the model. Mg-RMPN fully utilized three GPMs to achieve optimal results, confirming that learning negative samples based on different granularities can effectively improve the model's learning ability for sparse relationship matrices." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 316, + 480, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 316, + 480, + 483 + ], + "spans": [ + { + "bbox": [ + 130, + 316, + 480, + 483 + ], + "type": "text", + "content": "Different input information for Mg-RMPN This paper bridges entity detection and relation detection through a relationship matrix, making full use of the semantic and positional information of entities and solving the problem that end-to-end SGG relies solely on visual information to predict relations. Table 5 presents ablation studies of Mg-RMPN based on different input information. Compared to models that only utilize visual information, separately incorporating the semantic and positional information of entities can enhance the performance of the model. Furthermore, the improvement in introducing semantic and positional information simultaneously is significantly greater than in introducing a single modality of information, indicating that semantic and positional information are two complementary information. Hence, the semantic and positional information incorporated in our approach plays a crucial role in predicting relations. More hyperparameter analysis can be found in Supplementary Material." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 511, + 219, + 523 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 511, + 219, + 523 + ], + "spans": [ + { + "bbox": [ + 132, + 511, + 219, + 523 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "type": "text", + "content": "In this paper, our proposed end-to-end method bridges entity detection and relation detection through a sparse relationship matrix, which not only eliminates the need for post-processing of relation matching but also leverages the semantics and positional information of predicted entities to enhance relation prediction. To predict the sparse relationship matrix, we propose a multi-granularity sparse relationship matrix prediction network, which utilizes three gated pooling modules focusing on filtering negative samples at different granularities. Finally, a set of sparse, most likely subject-object pairs can be constructed and used for relation decoding. The experimental results demonstrate that our method achieves an optimal overall performance in both the end-to-end and two-stage methods." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "text", + "content": "L. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 139, + 480, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 139, + 480, + 175 + ], + "spans": [ + { + "bbox": [ + 132, + 139, + 480, + 175 + ], + "type": "text", + "content": "This work was supported by the National Natural Science Foundation of China (U21A20485, 61976170, 62088102) and the National Key R&D Program of China (2023YFB4704900)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 193, + 197, + 205 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 193, + 197, + 205 + ], + "spans": [ + { + "bbox": [ + 133, + 193, + 197, + 205 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 217, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 138, + 217, + 480, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 217, + 480, + 249 + ], + "spans": [ + { + "bbox": [ + 138, + 217, + 480, + 249 + ], + "type": "text", + "content": "1. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: European conference on computer vision. pp. 213-229. Springer (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 251, + 480, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 251, + 480, + 282 + ], + "spans": [ + { + "bbox": [ + 138, + 251, + 480, + 282 + ], + "type": "text", + "content": "2. Chen, C., Zhan, Y., Yu, B., Liu, L., Luo, Y., Du, B.: Resistance training using prior bias: Toward unbiased scene graph generation. Proceedings of the AAAI Conference on Artificial Intelligence 36(1), 212-220 (Jun 2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 284, + 480, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 284, + 480, + 316 + ], + "spans": [ + { + "bbox": [ + 138, + 284, + 480, + 316 + ], + "type": "text", + "content": "3. Chen, M., Liao, Y., Liu, S., Chen, Z., Wang, F., Qian, C.: Reformulating hoi detection as adaptive set prediction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9004-9013 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 317, + 480, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 317, + 480, + 348 + ], + "spans": [ + { + "bbox": [ + 138, + 317, + 480, + 348 + ], + "type": "text", + "content": "4. Chen, S., Jin, Q., Wang, P., Wu, Q.: Say as you wish: Fine-grained control of image caption generation with abstract scene graphs. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9962-9971 (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 350, + 480, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 350, + 480, + 380 + ], + "spans": [ + { + "bbox": [ + 138, + 350, + 480, + 380 + ], + "type": "text", + "content": "5. Chen, T., Yu, W., Chen, R., Lin, L.: Knowledge-embedded routing network for scene graph generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6163-6171 (2019)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 382, + 480, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 382, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 138, + 382, + 480, + 413 + ], + "type": "text", + "content": "6. Cong, Y., Yang, M.Y., Rosenhahn, B.: Reltr: Relation transformer for scene graph generation. IEEE Transactions on Pattern Analysis and Machine Intelligence (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 415, + 480, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 415, + 480, + 458 + ], + "spans": [ + { + "bbox": [ + 138, + 415, + 480, + 458 + ], + "type": "text", + "content": "7. Dong, X., Gan, T., Song, X., Wu, J., Cheng, Y., Nie, L.: Stacked hybrid-attention and group collaborative learning for unbiased scene graph generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19427-19436 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 459, + 480, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 459, + 480, + 490 + ], + "spans": [ + { + "bbox": [ + 138, + 459, + 480, + 490 + ], + "type": "text", + "content": "8. Gao, L., Lyu, X., Guo, Y., Hu, Y., Li, Y.F., Xu, L., Shen, H.T., Song, J.: Informative scene graph generation via debiasing. arXiv preprint arXiv:2308.05286 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 491, + 480, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 491, + 480, + 522 + ], + "spans": [ + { + "bbox": [ + 138, + 491, + 480, + 522 + ], + "type": "text", + "content": "9. Hudson, D.A., Manning, C.D.: Gqa: A new dataset for real-world visual reasoning and compositional question answering. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 6700-6709 (2019)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 523, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 523, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 134, + 523, + 480, + 555 + ], + "type": "text", + "content": "10. Johnson, J., Gupta, A., Fei-Fei, L.: Image generation from scene graphs. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1219-1228 (2018)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 556, + 480, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 556, + 480, + 588 + ], + "spans": [ + { + "bbox": [ + 134, + 556, + 480, + 588 + ], + "type": "text", + "content": "11. Johnson, J., Krishna, R., Stark, M., Li, L.J., Shamma, D., Bernstein, M., Fei-Fei, L.: Image retrieval using scene graphs. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3668-3678 (2015)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 589, + 480, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 589, + 480, + 621 + ], + "spans": [ + { + "bbox": [ + 134, + 589, + 480, + 621 + ], + "type": "text", + "content": "12. Kim, B., Lee, J., Kang, J., Kim, E.S., Kim, H.J.: Hotr: End-to-end human-object interaction detection with transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 74-83 (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 134, + 622, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 622, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 622, + 480, + 665 + ], + "type": "text", + "content": "13. Krishna, R., Zhu, Y., Groth, O., Johnson, J., Hata, K., Kravitz, J., Chen, S., Kalantidis, Y., Li, L.J., Shamma, D.A., et al.: Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision 123(1), 32-73 (2017)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 400, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 400, + 91, + 447, + 102 + ], + "type": "text", + "content": "Mg-RMPN" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 666 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "text", + "content": "14. Kuhn, H.W.: The hungarian method for the assignment problem. Naval research logistics quarterly 2(1-2), 83-97 (1955)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 139, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 139, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 133, + 139, + 480, + 182 + ], + "type": "text", + "content": "15. Kuznetsova, A., Rom, H., Alldrin, N., Uijlings, J., Krasin, I., Pont-Tuset, J., Kamali, S., Popov, S., Malloci, M., Kolesnikov, A., et al.: The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. International Journal of Computer Vision 128(7), 1956-1981 (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 182, + 480, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 182, + 480, + 214 + ], + "spans": [ + { + "bbox": [ + 133, + 182, + 480, + 214 + ], + "type": "text", + "content": "16. Li, R., Zhang, S., He, X.: Sgtr: End-to-end scene graph generation with transformer. In: proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 19486-19496 (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 214, + 480, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 214, + 480, + 247 + ], + "spans": [ + { + "bbox": [ + 133, + 214, + 480, + 247 + ], + "type": "text", + "content": "17. Li, R., Zhang, S., Wan, B., He, X.: Bipartite graph network with adaptive message passing for unbiased scene graph generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11109-11119 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 247, + 480, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 247, + 480, + 289 + ], + "spans": [ + { + "bbox": [ + 133, + 247, + 480, + 289 + ], + "type": "text", + "content": "18. Li, W., Zhang, H., Bai, Q., Zhao, G., Jiang, N., Yuan, X.: Ppdl: Predicate probability distribution based loss for unbiased scene graph generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19447-19456 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 289, + 480, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 289, + 480, + 322 + ], + "spans": [ + { + "bbox": [ + 133, + 289, + 480, + 322 + ], + "type": "text", + "content": "19. Li, Y., Ma, T., Bai, Y., Duan, N., Wei, S., Wang, X.: Pastegan: A semi-parametric method to generate image from scene graph. Advances in Neural Information Processing Systems 32 (2019)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 322, + 480, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 322, + 480, + 354 + ], + "spans": [ + { + "bbox": [ + 133, + 322, + 480, + 354 + ], + "type": "text", + "content": "20. Lin, T.Y., Goyal, P., Girshick, R., He, K., Dollar, P.: Focal loss for dense object detection. In: Proceedings of the IEEE international conference on computer vision. pp. 2980-2988 (2017)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 133, + 354, + 480, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 354, + 480, + 387 + ], + "spans": [ + { + "bbox": [ + 133, + 354, + 480, + 387 + ], + "type": "text", + "content": "21. Lin, X., Ding, C., Zeng, J., Tao, D.: Gps-net: Graph property sensing network for scene graph generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3746-3753 (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 133, + 386, + 480, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 386, + 480, + 418 + ], + "spans": [ + { + "bbox": [ + 133, + 386, + 480, + 418 + ], + "type": "text", + "content": "22. Liu, H., Yan, N., Mortazavi, M., Bhanu, B.: Fully convolutional scene graph generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11546-11556 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 133, + 418, + 480, + 440 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 418, + 480, + 440 + ], + "spans": [ + { + "bbox": [ + 133, + 418, + 480, + 440 + ], + "type": "text", + "content": "23. Lyu, X., Gao, L., Xie, J., Zeng, P., Tian, Y., Shao, J., Shen, H.T.: Generalized unbiased scene graph generation. arXiv preprint arXiv:2308.04802 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 133, + 440, + 480, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 440, + 480, + 472 + ], + "spans": [ + { + "bbox": [ + 133, + 440, + 480, + 472 + ], + "type": "text", + "content": "24. Ren, S., He, K., Girshick, R., Sun, J.: Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems 28 (2015)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 133, + 472, + 480, + 505 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 472, + 480, + 505 + ], + "spans": [ + { + "bbox": [ + 133, + 472, + 480, + 505 + ], + "type": "text", + "content": "25. Schroeder, B., Tripathi, S.: Structured query-based image retrieval using scene graphs. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. pp. 178-179 (2020)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 133, + 505, + 480, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 505, + 480, + 536 + ], + "spans": [ + { + "bbox": [ + 133, + 505, + 480, + 536 + ], + "type": "text", + "content": "26. Tang, K., Niu, Y., Huang, J., Shi, J., Zhang, H.: Unbiased scene graph generation from biased training. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 3716-3725 (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 536, + 480, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 536, + 480, + 568 + ], + "spans": [ + { + "bbox": [ + 133, + 536, + 480, + 568 + ], + "type": "text", + "content": "27. Tang, K., Zhang, H., Wu, B., Luo, W., Liu, W.: Learning to compose dynamic tree structures for visual contexts. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 6619-6628 (2019)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 133, + 568, + 480, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 568, + 480, + 601 + ], + "spans": [ + { + "bbox": [ + 133, + 568, + 480, + 601 + ], + "type": "text", + "content": "28. Teney, D., Liu, L., van Den Hengel, A.: Graph-structured representations for visual question answering. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1-9 (2017)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 133, + 601, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 601, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 133, + 601, + 480, + 643 + ], + "type": "text", + "content": "29. Wang, J., Zhang, W., Zang, Y., Cao, Y., Pang, J., Gong, T., Chen, K., Liu, Z., Loy, C.C., Lin, D.: Seesaw loss for long-tailed instance segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9695-9704 (2021)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 133, + 643, + 480, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 643, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 643, + 480, + 666 + ], + "type": "text", + "content": "30. Wang, J., Wen, Z., Li, X., Guo, Z., Yang, J., Liu, Z.: Pair then relation: Pair-net for panoptic scene graph generation. arXiv preprint arXiv:2307.08699 (2023)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "text", + "content": "L. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 511 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "31. Wang, L., Yuan, Z., Chen, B.: Learning to generate an unbiased scene graph by using attribute-guided predicate features. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 2581-2589 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 481, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 481, + 183 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 481, + 183 + ], + "type": "text", + "content": "32. Xu, D., Zhu, Y., Choy, C.B., Fei-Fei, L.: Scene graph generation by iterative message passing. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5410-5419 (2017)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 481, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 481, + 216 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 481, + 216 + ], + "type": "text", + "content": "33. Yang, G., Zhang, J., Zhang, Y., Wu, B., Yang, Y.: Probabilistic modeling of semantic ambiguity for scene graph generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12527-12536 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 216, + 481, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 481, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 481, + 248 + ], + "type": "text", + "content": "34. Yang, J., Lu, J., Lee, S., Batra, D., Parikh, D.: Graph r-cnn for scene graph generation. In: Proceedings of the European conference on computer vision (ECCV). pp. 670-685 (2018)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 249, + 481, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 481, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 481, + 293 + ], + "type": "text", + "content": "35. Yu, J., Chai, Y., Wang, Y., Hu, Y., Wu, Q.: Cogtree: Cognition tree loss for unbiased scene graph generation. In: Zhou, Z.H. (ed.) Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence, IJCAI-21. pp. 1274-1280. International Joint Conferences on Artificial Intelligence Organization (8 2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 293, + 481, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 481, + 326 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 481, + 326 + ], + "type": "text", + "content": "36. Zellers, R., Yatskar, M., Thomson, S., Choi, Y.: Neural motifs: Scene graph parsing with global context. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5831-5840 (2018)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 326, + 481, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 481, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 481, + 346 + ], + "type": "text", + "content": "37. Zhang, A., Yao, Y., Chen, Q., Ji, W., et al, L.: Fine-grained scene graph generation with data transfer (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 347, + 481, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 481, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 481, + 380 + ], + "type": "text", + "content": "38. Zhang, H., Kyaw, Z., Chang, S.F., Chua, T.S.: Visual translation embedding network for visual relation detection. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5532-5540 (2017)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 380, + 481, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 481, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 481, + 413 + ], + "type": "text", + "content": "39. Zhang, J., Shih, K.J., Elgammal, A., Tao, A., Catanzaro, B.: Graphical contrastive losses for scene graph parsing. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11535-11543 (2019)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 413, + 481, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 413, + 481, + 446 + ], + "spans": [ + { + "bbox": [ + 132, + 413, + 481, + 446 + ], + "type": "text", + "content": "40. Zheng, C., Lyu, X., Gao, L., Dai, B., Song, J.: Prototype-based embedding network for scene graph generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22783-22792 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 447, + 481, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 447, + 481, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 447, + 481, + 479 + ], + "type": "text", + "content": "41. Zhong, Y., Wang, L., Chen, J., Yu, D., Li, Y.: Comprehensive image captioning via scene graph decomposition. In: European Conference on Computer Vision. pp. 211-229. Springer (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 479, + 481, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 481, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 481, + 511 + ], + "type": "text", + "content": "42. Zhu, X., Su, W., Lu, L., Li, B., Wang, X., Dai, J.: Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159 (2020)" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 400, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 400, + 91, + 447, + 102 + ], + "type": "text", + "content": "Mg-RMPN" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file